Release version 0.1.0

feat(democracy): find the running stacks' worktrees live for the resource cleanup
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2026-10-30 18:29:49 +01:00 · 2026-10-30 17:41:10 +01:00 · 2026-10-30 17:37:47 +01:00 · 2026-10-30 17:35:27 +01:00 · 2026-10-30 16:54:37 +01:00 · 2026-10-30 16:54:37 +01:00
57 changed files with 3792 additions and 253 deletions
@@ -15,6 +15,7 @@ npm-debug.log*
 # Rendered skill install targets
 .agents/
 .claude/
 .mcp.json
 # Editor / OS
 .DS_Store
@@ -0,0 +1,36 @@
 # Changelog
 ## [0.1.0] - 2026-10-30
 Added:
 - Install: *make install* and *make project* install every locked skill
 - Install: enables caveman and ponytail; adds autotune and dream reminders
 - Commit: *commit-next* commits a bundle and stages the next one
 - Commit: *comments-clean* and *advocate* audit every commit before it lands
 - Commit: *confidence* and *break-risk* report measured fix and break numbers
 - Coordination: *democracy* elects the first agent to claim work as its owner
 - Coordination: *council.py* claims work; a hook blocks edits without a claim
 - Coordination: *funnel* keeps agent work in worktrees until it clears 90/5
 - Resources: *watch.py* blocks heavy jobs and cleans up before the disk fills
 - Cleanup: *cleanup* frees disk space after asking every peer agent
 - CI: *triage*, *observe* and *ci-artifacts* with *ci_failures.py* clustering
 - CI: *tune* finds speed-ups that never cut logic, coverage or correctness
 - Git: *merge-fix* with *find_fix.py* finds a fix on any branch and merges it
 - Reasoning: *active-listening* and *dialectic* pin requirements and causes
 - Review: *criticize* and *sceptic* give critique hardened by dialectic
 - Autonomy: *plan* hands an objective-oriented plan to *robot*
 - Quality: *test* runs Makefile targets; *--fix* loops until they are green
 - Quality: *refactor*, *no-defaults* and *scope-guard* keep changes lean
 - Skills: *autotune*, *autoskill*, *dream* and *essence* improve the skills
 - Release: *changelog* writes the pkgmgr release message since the last tag
 - Status: *status-matrix* lists everything in flight with its confidence
 - Help: *explain* and *help-skills* explain code and the skill set
 - Chat: *shortcuts* expands operator tokens such as *sobu* and *socon*
 - Security: *medusa* runs the MEDUSA scanner on repos and secret stores
 - Docs: the README skill map shows how the skills invoke each other
 Upgrade notes:
 - Install: restart the agent after *make install* to load skills and hooks
 - Resources: *watch.py* needs *~/.config/agents-council/config.json*
@@ -7,6 +7,135 @@ Every skill is pinned in `skills-lock.json` for reproducible installs.
 First-party skills live under `skills/` and are layered on top of the locked ones at install time.
 Compatible with Claude Code, Codex, Gemini CLI, Cursor, Copilot, Windsurf, Cline, and other agents.
 ## Skill map 🗺️
 How the first-party skills work together. A solid arrow means the skill invokes or relies on the other one; a dotted arrow means it fires automatically. Every skill that reports a result closes with the `confidence` block; those arrows are left out. The locked third-party skills (caveman, ponytail, archify) are not drawn.
 ```mermaid
 flowchart LR
  subgraph triggers [Triggers]
    commit_req([commit request])
    stage_req([staging or fix ready])
    edit_hook([edit in a git worktree])
    reminders([reminder hooks])
    code_task([code task])
  end
  subgraph commit [Commit]
    commit_next[commit-next]
    commit_no_verify[commit-no-verify]
    comments_clean[comments-clean]
    advocate[advocate]
    confidence[confidence]
    break_risk[break-risk]
    explain[explain]
  end
  subgraph coordination [Coordination]
    democracy[democracy]
    coordinate[coordinate]
    funnel[funnel]
    cleanup[cleanup]
  end
  subgraph ci [CI]
    observe[observe]
    triage[triage]
    ci_artifacts[ci-artifacts]
    tune[tune]
    merge_fix[merge-fix]
  end
  subgraph reasoning [Reasoning]
    active_listening[active-listening]
    dialectic[dialectic]
    criticize[criticize]
    sceptic[sceptic]
    auto_feature[auto-feature]
  end
  subgraph autonomy [Autonomy]
    plan[plan]
    robot[robot]
  end
  subgraph self_improvement [Self-improvement]
    dream[dream]
    autotune[autotune]
    autoskill[autoskill]
    essence[essence]
    shortcuts[shortcuts]
  end
  subgraph quality [Quality]
    test[test]
    refactor[refactor]
    no_defaults[no-defaults]
    scope_guard[scope-guard]
  end
  subgraph standalone [Standalone]
    changelog[changelog]
    help_skills[help-skills]
    status_matrix[status-matrix]
    medusa[medusa]
  end
  commit_req -.-> commit_next
  commit_req -.-> commit_no_verify
  stage_req -.-> confidence
  edit_hook -. council gate .-> democracy
  reminders -.-> dream
  reminders -.-> autotune
  code_task -.-> no_defaults
  code_task -.-> scope_guard
  commit_next --> merge_fix
  commit_next --> commit_no_verify
  commit_next --> explain
  commit_next --> dialectic
  commit_no_verify --> comments_clean
  commit_no_verify --> advocate
  advocate --> dialectic
  confidence --> break_risk
  confidence -. objective .-> autotune
  coordinate --> democracy
  cleanup --> democracy
  funnel --> democracy
  funnel --> dialectic
  democracy --> merge_fix
  observe --> triage
  observe --> ci_artifacts
  triage --> ci_artifacts
  triage --> democracy
  triage --> dialectic
  triage --> funnel
  triage --> tune
  tune --> democracy
  tune --> funnel
  tune --> merge_fix
  tune --> dialectic
  dialectic --> active_listening
  criticize --> dialectic
  criticize --> active_listening
  sceptic --> dialectic
  sceptic --> no_defaults
  auto_feature --> sceptic
  auto_feature --> active_listening
  plan --> robot
  plan --> active_listening
  robot --> dialectic
  dream --> dialectic
  dream --> autotune
  dream --> autoskill
  autoskill --> autotune
  autoskill --> shortcuts
  autotune --> essence
  test --> triage
  test --> dialectic
  test --> no_defaults
  test --> comments_clean
  refactor --> comments_clean
 ```
 ## Requirements 📦
 - Node.js (provides `npx skills`).
@@ -26,7 +155,7 @@ Copy the skills into a specific project (`<repo>/.agents/skills` and `<repo>/.cl
 make project TARGET=/path/to/repo
 ```
-Both commands also patch the target's `.claude/settings.json`: the caveman and ponytail plugins are enabled so their modes auto-activate on session start, and the `autotune` reminder hook is registered so the agent points you at `/autotune` at most once an hour. Existing settings are preserved.
+Both commands also patch the target's `.claude/settings.local.json` — the machine-local, untracked file, so a project's reviewed `.claude/settings.json` is never rewritten by an install: the caveman and ponytail plugins are enabled so their modes auto-activate on session start, and two reminder hooks are registered — `autotune` points you at `/autotune` at most once an hour, `dream` points you at `/dream` at most once every twelve hours per repository (stamped in that repository's git dir). A PreToolUse gate from the `democracy` skill blocks edits to files in a git worktree until the session holds a claim (`skills/democracy/scripts/council.py`), and its resource watch (`skills/democracy/scripts/watch.py`) measures disk, Docker and caches, cleans up low-risk data and blocks new heavy jobs before the disk runs out; it stays inactive until `~/.config/agents-council/config.json` exists. Existing settings are preserved.
 Restart your agent afterwards so it loads the new skills.
@@ -1,8 +1,8 @@
 #!/usr/bin/env node
 // Enable the caveman and ponytail plugins in a Claude Code settings.json so
-// their SessionStart hooks auto-activate both modes, and register the autotune
+// their SessionStart hooks auto-activate both modes, register the autotune
-// reminder hook. Merges non-destructively: existing marketplaces, plugins and
+// and dream reminder hooks, the democracy claim gate and the resource watch. Merges non-destructively: existing marketplaces,
-// hooks are preserved, unparseable files are left untouched.
+// plugins and hooks are preserved, unparseable files are left untouched.
 "use strict";
 const fs = require("fs");
@@ -19,6 +19,7 @@ const MARKETPLACES = {
  ponytail: { source: { source: "github", repo: "DietrichGebert/ponytail" } },
 };
 const PLUGINS = { "caveman@caveman": true, "ponytail@ponytail": true };
 const REMINDERS = ["autotune", "dream"];
 let data = {};
 if (fs.existsSync(settingsPath)) {
@@ -39,15 +40,41 @@ for (const [name, entry] of Object.entries(MARKETPLACES)) {
 }
 data.enabledPlugins = Object.assign(data.enabledPlugins || {}, PLUGINS);
 const reminder = path.join(path.dirname(settingsPath), "skills", "autotune", "hooks", "reminder.sh");
 data.hooks = data.hooks || {};
 data.hooks.UserPromptSubmit = data.hooks.UserPromptSubmit || [];
-const registered = JSON.stringify(data.hooks.UserPromptSubmit).includes("autotune/hooks/reminder.sh");
+const alreadyRegistered = JSON.stringify(data.hooks.UserPromptSubmit);
-if (!registered) {
+for (const skill of REMINDERS) {
  if (alreadyRegistered.includes(`${skill}/hooks/reminder.sh`)) continue;
  const reminder = path.join(path.dirname(settingsPath), "skills", skill, "hooks", "reminder.sh");
  data.hooks.UserPromptSubmit.push({
    hooks: [{ type: "command", command: `bash ${JSON.stringify(reminder)}` }],
  });
 }
 data.hooks.PreToolUse = data.hooks.PreToolUse || [];
 if (!JSON.stringify(data.hooks.PreToolUse).includes("democracy/scripts/council.py")) {
  const council = path.join(path.dirname(settingsPath), "skills", "democracy", "scripts", "council.py");
  data.hooks.PreToolUse.push({
    matcher: "Edit|Write|MultiEdit|NotebookEdit",
    hooks: [{ type: "command", command: `python3 ${JSON.stringify(council)} gate` }],
  });
 }
 const watch = path.join(path.dirname(settingsPath), "skills", "democracy", "scripts", "watch.py");
 const WATCH_HOOKS = [
  ["UserPromptSubmit", null, "watch"],
  ["PostToolUse", "Bash", "watch"],
  ["PreToolUse", "Bash", "block"],
 ];
 for (const [event, matcher, sub] of WATCH_HOOKS) {
  data.hooks[event] = data.hooks[event] || [];
  const registered = data.hooks[event].some((entry) =>
    (entry.hooks || []).some((hook) =>
      String(hook.command).includes("democracy/scripts/watch.py") && String(hook.command).endsWith(` ${sub}`)));
  if (registered) continue;
  const entry = { hooks: [{ type: "command", command: `python3 ${JSON.stringify(watch)} ${sub}` }] };
  data.hooks[event].push(matcher ? { matcher, ...entry } : entry);
 }
 fs.writeFileSync(settingsPath, JSON.stringify(data, null, 2) + "\n");
-console.log(`skills: enabled caveman + ponytail plugins and the autotune reminder in ${settingsPath}`);
+console.log(`skills: enabled caveman + ponytail plugins, the ${REMINDERS.join(" + ")} reminders, the democracy claim gate and resource watch in ${settingsPath}`);
@@ -23,10 +23,11 @@ if [[ ! -d "${TARGET}" ]]; then
 fi
 cd "${REPO_ROOT}"
 src="${REPO_ROOT}/.agents/skills"
 rm -rf "${src}"
 log "skills: restoring from skills-lock.json..."
 npx --yes skills experimental_install
 src="${REPO_ROOT}/.agents/skills"
 if [[ ! -d "${src}" ]]; then
 	warn "skills: ${src} not found after install."
 	exit 1
@@ -44,7 +45,7 @@ for dst in "${TARGET}/.agents/skills" "${TARGET}/.claude/skills"; do
 	cp -a "${src}/." "${dst}/"
 done
-settings="${TARGET}/.claude/settings.json"
+settings="${TARGET}/.claude/settings.local.json"
 if command -v node &>/dev/null; then
 	log "skills: auto-enabling caveman + ponytail plugins in ${settings}"
 	node "${REPO_ROOT}/scripts/enable-plugins.js" "${settings}"
@@ -1,83 +1,89 @@
 {
  "version": 1,
  "skills": {
    "archify": {
      "source": "tt-a1i/archify",
      "sourceType": "github",
      "skillPath": "archify/SKILL.md",
      "computedHash": "e9ef0923cda908886e667eb48e36b29e81d0a7a7b9fe2bc05f3779be271e5119"
    },
    "cavecrew": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/cavecrew/SKILL.md",
-      "computedHash": "c5527c994fbd4c22b36714e3b124a0f167a533d114ab164fb1d35e2123533917"
+      "computedHash": "2e122d819a812bef0e4958706f306e4346e78ecb5a4893997b99cbc19866f514"
    },
    "caveman": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman/SKILL.md",
-      "computedHash": "d18cdf73a5f5c496d681a43a6846336b110223bdffa6817b8992529c57f1a815"
+      "computedHash": "71b600c88c8d1d36f18c4875db23b53c5d6cb9ba3195e19451d6b1bb67db07c4"
    },
    "caveman-commit": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman-commit/SKILL.md",
-      "computedHash": "790a4eeace0be35c6691faf923518ba5bd50f1f1305d1101d09dd4971be94e00"
+      "computedHash": "b416cf134b0f33ddadeca7977aebdbb2d3db79a166d0fce2e1aabeec98132766"
    },
    "caveman-compress": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman-compress/SKILL.md",
-      "computedHash": "84517cd4cf7a49d8d2bc1baf00f61bc359306c6ad9389756b6937eff958bd374"
+      "computedHash": "9c8d35882a93e387a584cc4efaacb74433d379a2bd02e3e6f6ad4f7c30ff55ed"
    },
    "caveman-help": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman-help/SKILL.md",
-      "computedHash": "dd85267e76baad76995157e7b9f762dfa557cd58951ee92af0c283f48aa26537"
+      "computedHash": "c7a73a8c6afb6f871690666449209106aba847b4a9b0698686648960ece80012"
    },
    "caveman-review": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman-review/SKILL.md",
-      "computedHash": "fb7214a1c5793bae6ba8b1be4329e2e6f40dbec6dd911dfb335ad29f09c316a1"
+      "computedHash": "36a4007c5c0c6b90b5accfee1cd769a86e4963b651dece2ce5f17a223a0abb4b"
    },
    "caveman-stats": {
      "source": "JuliusBrussee/caveman",
      "sourceType": "github",
      "skillPath": "skills/caveman-stats/SKILL.md",
-      "computedHash": "47ce2de3d6cb39a75047b5c962e4eb3da15594e7397c94103e9a104d42626553"
+      "computedHash": "cc9fcd21f2a30cff5596b3cbb197fe0f4b1b30b58084a864350853c217ad301a"
    },
    "ponytail": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail/SKILL.md",
-      "computedHash": "210fa0831f649ce9f8a305f7b7c8cf2a0dd09ac4dd53067109941b96e7811d4e"
+      "computedHash": "061f887d9e3e735f85db3be3655a8936a00f24d5fb65ac4ab74a968a770fe342"
    },
    "ponytail-audit": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail-audit/SKILL.md",
-      "computedHash": "8b8d72e6ff70b65b6c9518d470fcf29dc696eea93d28f3efa10fdcc60cbd696d"
+      "computedHash": "1d718f893dabc01caad0dacef86444b5b4b3e1e59409527d66e8b5dd0cd85ae0"
    },
    "ponytail-debt": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail-debt/SKILL.md",
-      "computedHash": "0099b82174e3a4261828b4ec04d2c068b1b294fdab14f76701c2acd1f7c99539"
+      "computedHash": "f5d1d2ab891ff0a650a90b67235e487828e7d2a189de08a1ae073032de379407"
    },
    "ponytail-gain": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail-gain/SKILL.md",
-      "computedHash": "c45ecba7906960eea904218b623abf097ebff84ecbacca6d68da1d0687435f75"
+      "computedHash": "fca2297cf7b5fd26692f6bd94d398b522e19fa81ad6e2aaebf1933575f7a6803"
    },
    "ponytail-help": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail-help/SKILL.md",
-      "computedHash": "10e8382b1033c9a0fec0073de6d3de3846b4f5dca7d78b6df2d476a1f4cd43c6"
+      "computedHash": "4582025c3923072fb6c35b8cbd75adf4f7ce47df0eff43cc9a2a39808537fd61"
    },
    "ponytail-review": {
      "source": "DietrichGebert/ponytail",
      "sourceType": "github",
      "skillPath": "skills/ponytail-review/SKILL.md",
-      "computedHash": "4646594fcd75455533f7ff1425f4f1c5ba0fa65f88241bf262b1c4fc275f049a"
+      "computedHash": "800dc0c86402978e12359f701b9e0c81ba4607a0cd0027f4596d0f835a67eadc"
    }
  }
 }
@@ -1,11 +1,10 @@
 ---
 name: active-listening
 description: >
-  Gather every requirement before acting. Trigger at the start of any task
+  Ask focused questions until every requirement is pinned down, then reflect
-  whose scope, constraints, or success criteria are not fully pinned down, and
+  the task back. Trigger at the start of any task with unclear scope,
-  whenever another skill needs an operator-only fact — ask focused questions
+  constraints or success criteria, and when another skill needs an
-  until nothing needed is still open, then reflect the understanding back.
+  operator-only fact.
  Portable across projects.
 ---
 Do not act on a half-understood request. Listen actively until you could state
@@ -0,0 +1,88 @@
 ---
 name: advocate
 description: >
  Check a commit for legal soundness: licence compatibility of added code and
  dependencies, attribution, copied snippets, GDPR, trademarks. Run
  AUTOMATICALLY before every commit the operator requests; also on /advocate
  [ref|range], "rechtlich prüfen", "lizenz prüfen".
 ---
 The stance is the project's counsel: find what would make the change unlawful
 to distribute before it is distributed. This is not legal advice but a reading
 of the licence texts, with German and EU law where jurisdiction matters; every
 report says so in one line. Leaked secrets belong to `medusa`, engineering
 principles to `sceptic`.
 ## Target
 - Before a commit: the staging index (`git diff --cached`, plus every added or
  changed file in full where origin or licence is in question).
 - `/advocate <ref>` or `/advocate <a>..<b>`: that commit or range.
 - `/advocate` alone: the index if it is non-empty, else `HEAD`.
 ## Fast exit
 The target is in scope only if it adds or changes at least one of: a
 dependency manifest or lockfile; a vendored, generated, or third-party file; a
 LICENSE, NOTICE, or COPYING file or a licence header; code with a foreign
 origin (a source URL, a foreign copyright line, a block that reads unlike its
 surroundings); an external request (CDN, web font, analytics, embed, API);
 handling of personal data; a logo, icon, or third-party brand name. Nothing in
 scope: report `advocate: nothing licence-relevant in <target>` in one line and
 let the commit proceed.
 ## Project licence
 Read every declaration: LICENSE/COPYING, the manifest's licence field
 (`package.json`, `pyproject.toml`, `Cargo.toml`, `composer.json`, ...), SPDX
 headers, the README's licence section. Declarations that disagree are a
 finding. A project without any licence is a finding (all rights reserved),
 never a blocker of the commit at hand.
 ## The checks
 Only for a target in scope: read `checks.md` next to this file and apply its
 seven checks (inbound code, dependency licences, project licence fit,
 attribution, copied snippets, data protection, trademarks).
 ## Severity
 - **Blocker** - the commit would distribute what the project may not
  distribute: an incompatible inbound or dependency licence, or non-trivial
  code or assets without a licence.
 - **Finding** - fixable while keeping the change: missing attribution, NOTICE,
  or header; a data-protection gap; trademark use; a block of unclear origin.
 - **Note** - worth knowing, no action required.
 ## Blockers stop the commit
 Harden every blocker with the `dialectic` skill before reporting it. The
 antithesis tries: the code is trivial or the operator's own; the dependency is
 build- or test-only and never distributed; an exception applies (linking or
 Classpath exception, dual licensing, a licence grant in the issue or commit
 history). A blocker that survives stops the commit: present it with the
 options (remove, replace with a compatible alternative, relicense, obtain
 permission) and wait. The operator's explicit acceptance lets this one commit
 proceed and never carries over to the next. Findings and notes never stop a
 commit.
 ## Output
 One line with the target, the project licence, and "Keine Rechtsberatung" (in
 the operator's language). Then one table, blockers first:
 | Severity | Check | Location | What is wrong | Licence or law | Fix | Evidence |
 `Evidence` is `proven` (licence text, metadata, or file read; cite path:line
 or the registry URL) or `derived` (inferred). Close with the verdict (commit
 stops or proceeds) and the `confidence` skill's numbers.
 ## Rules
 - Never infer a licence from a package's name or reputation; read it. An
  unreachable registry makes the row `derived`, never clean.
 - Never invent a licence clause, a statute, or a court ruling; cite only what
  you read.
 - Read-only apart from stopping the commit: never edit licence files, add
  headers, or remove dependencies on your own. Propose the fix; the operator
  decides.
@@ -0,0 +1,37 @@
 # Advocate: the checks
 Apply each check to a target the `advocate` skill found in scope.
 1. **Inbound code** - every file or block of foreign origin carries a licence
   that the project licence can absorb in the direction of distribution. GPL
   code cannot enter a project distributed under MIT without relicensing the
   whole; Apache-2.0 cannot enter GPL-2.0-only; code without a licence is all
   rights reserved.
 2. **Dependency licences** - for each added or bumped dependency, read the
   licence of the resolved version (lockfile, registry metadata, or the
   package's own LICENSE), including the transitive dependencies the lockfile
   newly pulls. Watch copyleft reach (GPL, AGPL's network clause, LGPL static
   versus dynamic linking), source-available terms that are not open source
   (SSPL, BUSL, Commons Clause, "non-commercial"), and an upstream relicensing
   between the old and the new version.
 3. **Project licence fit** - the project's declarations still hold against
   what checks 1 and 2 found: a dependency that forces copyleft forces either
   the project licence or the dependency's removal.
 4. **Attribution** - retained copyright notices, the licence text shipped
   where required, an Apache-2.0 NOTICE propagated, modified-file notices
   (Apache-2.0 §4(b), GPL-3.0 §5(a)), SPDX headers where the project uses them,
   credit for CC BY material.
 5. **Copied snippets** - Stack Overflow content is CC BY-SA (attribution plus
   share-alike); gists, blog posts, other repositories, and verbatim
   reproductions of well-known code need a traceable origin and licence.
   Trivial code without originality (one-liners, usage dictated by an API) is
   not a finding.
 6. **Data protection (GDPR, TDDDG)** - new collection, storage, logging, or
   transfer of personal data; IP addresses sent to third parties through
   external fonts, CDNs, analytics, or embeds; storage on the user's device
   without consent (§ 25 TDDDG); transfers outside the EU. Name the data flow
   and the missing legal basis or consent mechanism.
 7. **Trademarks and branding** - third-party logos, icons, and product names
   used beyond plain reference: implied endorsement, use in the project's own
   name, modified logos. Icon sets carry their own licence, and a free licence
   on an icon does not license the trademark it depicts.
@@ -0,0 +1,98 @@
 ---
 name: auto-feature
 description: >
  Derive missing work (sceptic pass for structural flaws, usability pass for
  missing features) and write requirement documents after single-select
  confirmation. Trigger on /auto-feature, "was fehlt noch", "welche features
  brauchen wir", or when the operator asks what to build next.
 ---
 Two blind spots produce different work: the code has defects of structure,
 the product has gaps of use. This skill runs both lenses, ranks what comes
 out, and lets the operator choose — it writes requirement documents, never
 implementations.
 ## Step 1: fix the scope
 Name the target in one line: repository, subsystem, role, or user-facing
 surface. Read the project's existing requirement documents first — their
 location, filename convention, and section structure are the template, and
 an already-written requirement is not a finding.
 ## Step 2: the sceptic pass
 Invoke the `sceptic` skill on the target. Its findings become requirement
 candidates by severity: a violated SPOT, SRP, or trust boundary is work the
 product needs whether or not a user ever notices. Carry each finding's
 `file:line` evidence forward — a requirement without it is a guess.
 ## Step 3: the usability pass
 Independently, look for what the product does not do yet. Anchor every gap
 in something observable in the repository, never in an invented persona:
 - **Entry** — can someone install, configure, and run it without reading the
  source? Missing `--help`, missing example config, an undocumented required
  variable.
 - **Recovery** — does a failure tell the user what to do next? Errors that
  print a stack trace instead of the remediation, no dry-run, no way to undo
  or uninstall, no idempotent re-run.
 - **Feedback** — does a long operation say anything while it works? Silent
  minutes, no progress, no exit summary.
 - **Empty and edge states** — first run with no data, zero results, an
  expired credential, a partially completed run.
 - **Discoverability** — a capability that exists but nothing surfaces:
  undocumented flags, features reachable only by reading tests.
 - **Accessibility and locale** — output that assumes colour, a terminal
  width, or one language.
 Each gap gets one line: `observed evidence -> who is blocked -> smallest
 change that unblocks them`. Drop anything you cannot anchor.
 ## Step 4: rank and ask
 Merge both lists, drop duplicates where a structural finding and a usability
 gap describe the same defect, and rank by blocked users first, blast radius
 second.
 Then apply `active-listening`: answer everything you can from the code, the
 docs, and the git history yourself, and put only the genuinely open decisions
 to the operator with `AskUserQuestion`, `multiSelect: false`, so every
 question renders as radio buttons. At most four questions per call, each
 option a real alternative — never one option plus filler:
 1. **Which candidates** become requirements this run, ranked, each option
   carrying its evidence in the description.
 2. **Which shape** the chosen one takes — for a feature typically
   `Minimal version that unblocks the user`, `Full version`, `Fix the
   underlying structure instead`, `Document the limitation and defer`.
 3. **Scope boundary** — what is explicitly out of scope, when the candidate
   could grow in more than one direction.
 4. **Placement** — which requirement file or directory receives it, when the
   project has more than one.
 Never assume an answer, never fold two decisions into one option, and never
 write a file before the answers are in.
 ## Step 5: write the requirements
 One document per feature — a requirement that needs "and" in its title is
 two requirements. Follow the project's existing template; when there is
 none, keep it to: problem with its evidence, the decided scope and its
 boundary, definition of done as checkable criteria, and the alternatives the
 operator rejected with the reason.
 Show the drafted document and ask for a final `Write it` / `Revise it` /
 `Discard` before touching disk.
 ## Rules
 - Requirements only: no implementation, no refactor, no staging as part of
  this skill.
 - Every requirement traces to evidence — a `file:line`, a failing path, a
  missing artifact. A requirement whose justification is "would be nice" is
  not written.
 - Record the rejected alternatives inside the document; the next reader
  needs to know what was already considered and dismissed.
 - Close with the `confidence` skill's numbers for the analysis: how sure the
  gaps are real, and what the passes did not cover.
@@ -1,12 +1,16 @@
 ---
 name: autoskill
 description: >
-  Detect operator repetition and turn it into a skill. Trigger when the
+  Turn operator repetition into a skill. Trigger when the operator gives the
-  operator gives an instruction, correction, or prompt pattern for roughly
+  same instruction, correction or prompt pattern about the third time in a
-  the third time in a conversation (or says they keep repeating themselves),
+  conversation, or says they keep repeating themselves, and the pattern is
-  and the pattern is general enough to reuse. Portable across projects.
+  reusable.
 ---
 Its objective is the `autotune` skill's **Objective** section: a skill is worth
 creating only when it cuts tokens or time and raises quality; judge every
 candidate against its three axes and rules.
 When you notice the operator repeating the same kind of instruction,
 correction, or prompt pattern (about three occurrences, exact wording may
 vary), do the following:
@@ -18,10 +22,12 @@ vary), do the following:
 3. On approval, write the skill:
   - a conversation shortcut or expansion: add it to the `shortcuts` skill
     table instead of creating a new skill.
-   - project-specific behavior: create it in the project's own skills
+   - anything else: persist it exactly as the `autotune` skill's
-     directory (e.g. `skills/<name>/SKILL.md` in the repository).
+     **Persistence** section specifies — locate the operator's skills
-   - portable behavior: create `skills/<name>/SKILL.md` in the operator's
+     repository, write `skills/<name>/SKILL.md` there (or in the current
-     skills repository.
+     project for project-specific behavior), then mirror the skill directory
     into `~/.claude/skills/` and `~/.agents/skills/`, plus the project's
     `.claude/skills/` and `.agents/skills/` when it is project-specific.
   Follow the local naming conventions and keep the skill a thin,
   single-purpose instruction; route to an authoritative doc when one
   exists instead of duplicating it.
@@ -1,18 +1,46 @@
 ---
 name: autotune
 description: >
-  Generate new agent skills and optimize existing ones so the agent works more
+  Generate new skills or optimize existing ones (fewer tokens, fewer errors,
-  efficiently: fewer input and output tokens, fewer errors, more focus, and
+  persistent learnings), each proposal confirmed by single-select question.
-  learnings that persist across sessions. Every proposal is confirmed with the
+  Trigger ONLY on an explicit request: /autotune, "autotune", "tune the
-  operator through single-select (radio button) questions. Trigger ONLY on an
+  skills"; never on your own initiative.
  explicit operator request (/autotune, "autotune", "tune the skills"); never
  run it on your own initiative. Portable across projects.
 ---
 Autotune turns observed friction into durable skills. It runs on demand only,
 it changes nothing without an explicit answer from the operator, and every
 skill it writes or rewrites lands in the operator's skills repository.
 ## Objective
 The shared objective of the self-improving skills (autotune, autoskill,
 confidence): every skill, rule, or report the agent produces about its own
 working is judged on three axes at once. A proposal that cannot name the axis
 it moves is not an optimization, it is a preference.
 - **Tokens** — fewer input and output tokens for the same result: read the range
  instead of the whole file, grep the saved log instead of re-running the
  command, route to an authoritative doc instead of duplicating it, report in
  lines instead of paragraphs.
 - **Time** — fewer round trips and less waiting: independent calls in parallel,
  one command that answers the question instead of three that circle it, no
  polling for work the harness will report on its own.
 - **Quality** — fewer errors and less rework: verification instead of
  assumption, and learnings that persist so the same mistake is not paid for
  twice.
 - A change must improve at least one axis and damage none. Output that got
  cheaper but lost correctness is a regression, not an optimization.
 - Quality is never the currency: validation, error handling, tests, and security
  steps are not what gets cut to save tokens or time.
 - Measure, do not guess: name the observed friction — the re-read, the retry,
  the correction the operator had to give twice — and what it cost. A guessed
  percentage is not a measurement.
 - The cheapest step is the one that does not run: drop work before optimizing
  it.
 - Persist it or pay again: an optimization that lives only in this session is
  re-derived in the next one, so it belongs in a skill or a memory entry.
 ## Trigger discipline
 - Run **only** when the operator asks: `/autotune`, "autotune", "tune the
@@ -65,10 +93,35 @@ for a final `Write it` / `Revise it` / `Discard` confirmation before touching
 disk. Rewrites of an existing skill additionally show a before/after diff of
 the changed lines in that confirmation.
 ## Persistence
 A skill that exists only under `.claude/skills` is lost on the next install, and
 one that exists only in the repository is not live until the next restart — so
 every skill is written to **both**, in this order:
 1. **Locate the operator's skills repository.** Take the first hit: the
   `SKILLS_REPO` environment variable; the current repository, when it holds
   `skills-lock.json` and a `skills/` directory; `pkgmgr path skills`; a git
   checkout named `skills` under the usual roots
   (`~/Repositories/*/*/skills`, `~/git`, `~/src`, `~/Projects`) carrying
   `skills-lock.json`. If none matches, ask the operator for the path with a
   single-select question — never skip the repository and write only the live
   copies.
 2. **Write the source of truth**: `skills/<name>/SKILL.md` in that repository
   (portable skills), or `skills/<name>/SKILL.md` in the current project
   (project-specific ones). This is the copy that gets committed.
 3. **Mirror it live**, so it works without a reinstall: copy the skill's whole
   directory to `~/.claude/skills/<name>/` and `~/.agents/skills/<name>/`, and
   for a project-specific skill additionally to that project's
   `.claude/skills/<name>/` and `.agents/skills/<name>/`. Copy the directory,
   not just `SKILL.md`, so hooks and references come along.
 Report all written paths. Deleting or renaming a skill follows the same list in
 reverse — remove the mirrors too, or the old name keeps firing.
 ## Step 3: write
- Portable skills go to `skills/<name>/SKILL.md` in the operator's skills
+- Write the skill to every location the **Persistence** section lists.
  repository; project-specific ones to that project's skills directory.
 - Frontmatter carries `name` and a `description` that states what the skill
  does **and** when to trigger it - the description is the only part loaded into
  every session, so it decides whether the skill ever fires.
@@ -76,7 +129,11 @@ the changed lines in that confirmation.
  behaviour. Route to an authoritative doc instead of duplicating it.
 - Optimizing an existing skill means the file gets shorter or sharper, not
  longer. Delete rules the agent already follows, merge duplicated ones, and
-  cut every sentence that does not change an action.
+  cut every sentence that does not change an action. Route that compaction
  through the `essence` skill instead of trimming by hand: it runs a health
  pass for dead pointers and rules duplicated across skills before cutting
  anything, ranks by what loads in every session, and shows the diff before it
  writes. It asks for its own scope — answer with the skill being optimized.
 - Report what changed in at most three lines, then refresh the reminder stamp
  so the next hint is an hour out: `date +%s > ~/.claude/.autotune-stamp`.
@@ -86,7 +143,6 @@ the changed lines in that confirmation.
  whole collection is not what autotune is for.
 - Never write a skill for one-off work, for secrets, or for behaviour an
  existing skill already covers - propose extending that skill instead.
- A skill that saves tokens but loses correctness is a regression; validation,
+- The **Objective** section decides what counts as an improvement: at least
-  error handling, and security steps are never the thing that gets trimmed.
+  one axis better, none damaged, and the claim named as observed friction rather
- Efficiency claims stay honest: name the observed friction the skill removes,
+  than a guessed percentage.
  never a guessed percentage.
@@ -0,0 +1,64 @@
 ---
 name: break-risk
 description: >
  Derive the break-risk percentage from measurement, not feel. Called by
  `confidence` on every report; also when the operator asks "what could this
  break?" or names a break-risk threshold.
 ---
 Break risk is the probability the change breaks something it did not intend to
 touch. It is not the inverse of fix confidence, and it is not a mood. Derive it
 by the steps below; a number produced without them is a guess wearing a
 percent sign.
 ## Fast path
 A change whose files nothing builds, deploys, or runs (documentation and
 comments only, or tests that ran green with it) has no consumers outside
 itself. Name the files (`git diff --cached --name-only`), report break risk
 under 1%, and skip the method. Any other file takes the full method.
 ## The method
 1. **Consumers**: enumerate every consumer of everything the change alters, in
   the committed state (`git grep <symbol> HEAD`) and of its output; say the
   count. A caveat never licenses a low number; an enumeration does.
 2. **Delta**: run the old and the new code on today's inputs and compare at
   the surface where the difference shows (rendered output, emitted config,
   returned set, wire bytes), not as the source reads.
 3. **Live or latent**: per finding, whether a shipped configuration reaches it.
   Only live findings move the number materially.
 4. **Matrix**: a verdict per deployment axis the project has, never averaged.
 5. **Failure mode**: fails closed and loud, or open and silent; the latter can
   outrank a likelier failure.
 6. **Evidence**: executed beats derived; a report, a summary or your own
   earlier message is not evidence; negative-control every check you cite,
   including one you did not write.
 `reference.md` next to this file explains each step with the trap it defeats.
 Read it when a step is new to you or its result is contested.
 ## Producing the number
 State it as a percentage next to fix confidence, never blended, and always with
 its residuals. Anchor roughly:
 - **Under 5%**: consumers enumerated, delta measured at the surface, every axis
  of the matrix covered by executed evidence, failure mode closed.
 - **5 to 15%**: measured, but one axis or one layer rests on reasoning — no live
  run, a mode not exercised, a runtime the render cannot prove.
 - **15 to 40%**: a real behaviour change whose consumers are enumerated but
  whose effect is unmeasured, or a latent finding with a plausible trigger.
 - **Above 40%**: a known live defect, an unenumerated consumer set, or a
  fail-open exposure without a verified control. Fix it rather than report it.
 Discount yourself after any error the operator caught in the same area during
 this session, and say that you are doing so. A region where you have already
 been wrong twice deserves a worse number than its paper evidence suggests.
 ## When the operator names a threshold
 A named threshold is a gate on the work, not a field in the report. Missing it
 is a work item: name the cheapest evidence that would close the gap, go get it,
 re-measure. Report a miss only when the evidence is genuinely out of reach, and
 name which piece.
@@ -0,0 +1,81 @@
 # Break risk: the method in detail
 Each step of the `break-risk` skill, with the trap it defeats.
 ## The consumer question, first and always
 Enumerate every consumer of every thing the change alters, before naming a
 number. A written caveat never licenses a low number; an enumeration does.
 - Search the **committed** state, not just the working tree. A symbol that looks
  unused in your branch may still be called from a file you have not opened
  (`git grep <symbol> HEAD`). Removing a function with zero external callers is
  a different risk class from removing one with nine.
 - Enumerate consumers of the **output**, not only of the code: who reads the
  variable, renders the template, parses the emitted block, imports the
  constant.
 - Say the count out loud. "Two consumers, both inside this change set" is
  evidence. "I don't think anything else uses it" is not.
 ## Measure the delta at the surface that matters
 Compare what the old code **actually produced** against what the new code
 produces, at the point where the difference becomes observable: rendered output,
 emitted config, returned set, wire bytes.
 The trap this defeats: reasoning about what the old code *looks like* it did. A
 sweep that appeared to forward sixteen ports forwarded **zero** in production,
 because the flag it gated on was unrendered template text. Had the delta been
 argued from the source rather than measured, the risk would have been sized
 against a behaviour that never existed.
 Extract the old implementation from version control and run it on today's
 inputs. If `DROPPED` is empty, nothing can regress; say so and move on.
 ## Classify every finding as live or latent
 For each way the change could break something, decide which it is and say so:
 - **Live**: a shipped configuration, inventory, variant or default reaches it.
 - **Latent**: reachable only under a configuration nobody ships today.
 Both belong in the report. Only live findings drive the number up materially. A
 latent finding with a named trigger is worth more to the operator than a vague
 worry, and it is honest about why the number is not zero.
 ## Walk the deployment matrix
 A change verified in one mode is unverified in the others. Enumerate the axes
 the project actually has, and state the verdict per cell rather than averaging
 them away. Typical axes: deployment mode, node role, topology, distro,
 filesystem, single versus multi host.
 Two numbers for two modes beats one blended number that is wrong in both.
 ## Weigh the failure mode, not only the probability
 A low-probability failure that **fails open** can outrank a likely one that
 fails closed. Ask what the world looks like when the change misbehaves:
 - Fails closed and loud: the service refuses to start, a gate goes red, a
  request errors. Cheap to detect, bounded blast radius.
 - Fails open and silent: a port stays reachable, a check passes vacuously, a
  credential keeps working, traffic is forwarded to an unintended place.
 Say which one applies. When a design choice is between an unverifiable control
 over an exposure and not creating the exposure, prefer the second and say why.
 ## Evidence rules
 - Executed beats derived. Run it, render it, read the output.
 - A subagent's report, a conversation summary, or your own earlier message is
  **not** evidence. Re-verify the claim yourself before it changes a number.
 - Negative control before claiming a check works: make the thing it should
  catch, watch it fail, restore. A check never observed failing is not known to
  work.
 - The same applies to a check you did not write. Citing an existing lint, test
  or gate as the reason a change is safe means reading its pattern, not its
  docstring, and watching it fail on the case you are trusting it to catch. A
  green gate whose rule never matched your case is silence, not a pass.
 - After a context compaction, rebuild attributions from the source; conclusions
  survive compaction, the evidence under them does not.
@@ -0,0 +1,135 @@
 ---
 name: changelog
 description: >
  Write the release message since the last release for `pkgmgr release`;
  prints it only, never edits CHANGELOG.md, tags or releases. Trigger on
  /changelog, "release message", "changelog entry", before cutting a release,
  or on pkgmgr's markdown-lint rejection.
 ---
 The message is one text that pkgmgr writes to four places: the CHANGELOG.md
 entry, the annotated release tag, `debian/changelog` and the RPM `%changelog`.
 Every rule below exists because one of those four would otherwise break or
 mangle it.
 ## 1. Find the range
 ```bash
 git fetch --tags
 git describe --tags --abbrev=0 --match 'v[0-9]*.[0-9]*.[0-9]*'
 git log --no-merges --format='%h %s%n%b' <tag>..HEAD
 ```
 - pkgmgr tags every release `vX.Y.Z`; the newest one reachable from HEAD is
  the base. No such tag: the range is the whole history, say so.
 - Drop the `Release version X.Y.Z` commits and any revert together with the
  commit it reverts inside the range.
 - A subject is a lead, not the change: read the diff (`git show --stat`, then
  the hunks) when the subject does not say what changed for the user.
 - When `gh` is available, check the previous tag's CI run
  (`gh run list --branch <tag> -L 1`). A failed run means that version was
  never published: say so under *Upgrade notes* and carry its upgrade notes
  forward to this release.
 ## 2. Sort into sections
 | Commit | Section |
 | --- | --- |
 | `feat` | Added, or Changed when it alters existing behaviour |
 | `fix`, `perf` | Fixed |
 | `!` or `BREAKING CHANGE:` | Breaking, plus one Upgrade notes line per required action |
 | `docs`, `style`, `refactor`, `test`, `ci`, `build`, `chore` | omitted unless a user sees the effect |
 | dependency bumps | one collapsed line, or omitted |
 When nothing user-visible remains (a test- or CI-only release), keep those
 commits under Fixed with an area prefix (`Tests:`, `CI:`) so the entry is not
 empty.
 Order: Breaking, Added, Changed, Fixed, Upgrade notes. Leave out empty
 sections.
 ## 3. Format
 ```text
 Fixed:
 - Backup: rsync -a stored no xattrs, ACLs or hardlinks; now -aHAX
 - Restore: *--empty* no longer aborts on a database with an extension
 Upgrade notes:
 - First backup re-copies every attribute-carrying volume once in full
 ```
 1. **Section labels are plain lines ending in `:`**, never `#` headings and
   never a bold-only line. pkgmgr creates the tag with `git tag -a -m`, whose
   default `strip` cleanup deletes every line starting with `#`. A bold-only
   line fails MD036.
 2. **No version heading, date or "Release X.Y.Z" line.** pkgmgr adds
   `## [X.Y.Z] - YYYY-MM-DD` itself.
 3. **One bullet, one line. Too long means shorter, never wrapped.** RPM turns
   every non-empty line into its own `-` entry and Debian indents each line,
   so a continuation line becomes a separate, broken entry.
 4. **Line length: the repository's MD013 limit**, which is markdownlint's
   default of 80 unless the repo's config changes it. Aim for 78 or less, so
   lines stay within 80 after Debian's two-space indent and RPM's two-character
   `-` prefix.
 5. **Lists use the marker CHANGELOG.md already uses**, no nesting, with a
   blank line before and after every list (MD004, MD032). MD004's default
   *consistent* style takes the marker of the first list in the file, and
   pkgmgr puts the new entry on top, so a marker that differs from the older
   entries turns every older list into a finding. Take the marker of the first
   list item in the existing CHANGELOG.md; use `-` only when there is no
   CHANGELOG.md yet, and the configured one when the repo's MD004 pins a style.
 6. **No backticks and no code blocks.** pkgmgr rewrites inline code to
   *italic* in all four targets, and a fenced block fails its fallback lint.
   Write *names*, *--flags* and *paths* in `*` italics yourself; use `*`, not
   `_`, for emphasis (MD049).
 7. **Bullet text is `Area: effect`**: the component the user knows (Backup,
   Restore, CLI, Tests), then what changed for them, not how the code
   changed. No commit SHAs, no em dashes: use `:` or `;`.
 8. **Plain hygiene**: no trailing whitespace (MD009), no double blank lines
   (MD012), no bare URLs: wrap them as `<https://...>` (MD034).
 ## 4. Lint it the way pkgmgr does
 pkgmgr wraps the message as `# Changelog`, blank line, `## [X.Y.Z] - date`,
 blank line, message, then runs `markdownlint-cli2` with the target repo's own
 config. Reproduce that exactly:
 1. Write that wrapped document to a scratch file outside the repository.
 2. From the repository root, so its config is picked up, run:
   ```bash
   markdownlint-cli2 --no-globs - < <scratch-file>
   ```
   *--no-globs* keeps the config's own globs from linting the whole tree too.
 3. A finding on line N points at message line N - 4. Shorten that line and
   lint again; print nothing until the result is 0 errors.
 4. The standalone check cannot see rules that compare the entry with the rest
   of the file (MD004 *consistent*, MD024 duplicate headings), but the
   repository's own markdown lint and pre-push gate do. Copy CHANGELOG.md to a
   scratch file, insert the wrapped heading and message below its
   `# Changelog` line, and lint that copy the same way. A finding in the older
   entries that the unmodified CHANGELOG.md does not produce comes from the
   new entry; fix the message, never the older entries.
 Without `markdownlint-cli2`, check by hand what pkgmgr's fallback checks plus
 the length limit: no line over the limit, no trailing whitespace, no
 backtick, no double blank line.
 When the operator pastes pkgmgr's rejection (`changelog entry:7:81 error
 MD013/line-length ...`), apply the same N - 4 mapping to the message they
 used, shorten only the lines it names, and lint again.
 ## 5. Output
 1. The message in one fenced `text` block, nothing else inside it.
 2. One line: `Bump: patch | minor | major` and why (breaking → major,
   `feat` → minor, otherwise patch).
 3. One line with the lint result (`markdownlint-cli2: 0 errors` or the manual
   check that ran).
 Then stop. Never edit CHANGELOG.md or a version file, never run
 `pkgmgr release`, never tag: the operator releases.
@@ -0,0 +1,78 @@
 ---
 name: ci-artifacts
 description: >
  Get the evidence out of a CI run: failed job logs (also mid-run) clustered
  by error line, and the artifacts that hold the cause. Trigger whenever a CI
  failure must be explained, and from triage.
 ---
 A job log tells you what broke. An artifact tells you why. Investigating a CI
 failure without downloading its artifacts is guessing with extra steps.
 ## Getting the log
 For a GitHub Actions run, one call does it:
    ${CLAUDE_SKILL_DIR}/scripts/ci_failures.py <RUN_ID> [--repo <owner>/<repo>] [--artifacts]
 It prints how far HEAD moved since the run's commit, saves every finished
 failed job's log to `/tmp/ci-logs-<run>/` (colour stripped, even while the run
 is still in progress), and clusters the jobs by the normalised error line that
 ended them. Jobs still running are listed, not fetched: their API log can stop hours
 early. The clustering is a heuristic; check each boundary against the logs.
 Grep the saved files repeatedly. Re-fetching a log to find the line you just
 lost costs minutes; grepping the saved copy costs milliseconds.
 ## Getting the artifacts
    gh api /repos/<owner>/<repo>/actions/runs/<RUN_ID>/artifacts --paginate \
      --jq '.artifacts[] | "\(.name)\t\(.size_in_bytes)"'
    gh run download <RUN_ID> --repo <owner>/<repo> -n <NAME> -D <DIR>
 Choose by content, not by name length:
 - **Rescue / diagnostics bundles** carry the reverse-proxy access and error log,
  the per-service and per-container logs, the journal, and the event stream.
  This is where causes live. Prefer them.
 - **Test-report bundles** (Playwright and friends) are mostly video and traces
  and can run to hundreds of megabytes for one assertion string. Pull one only
  when you need the assertion text or the error-context file.
 - **Inventory / runtime dumps** settle "what was actually deployed" questions.
 Inside a rescue bundle, read in this order: the reverse-proxy log (it shows
 status code, upstream, and time-to-response for the exact failing request), the
 application container log, then the journal and event stream for restarts and
 OOM kills.
 An access-log line is a timing witness. A `499` means the *client* gave up; the
 server was still working. A `504` or an `upstream timed out ... while reading
 upstream` means your backend never answered. `ReqTime` versus `UpstreamTime`
 tells you whether the proxy or the upstream burned the seconds.
 ## Grep hygiene
 Two failure modes turn a search into a false conclusion:
 - **Clustered flags silently change the pattern.** Write them separately —
  `grep -r -a -i -l -F "needle" dir/`, not `-ariel`. A cluster where one letter
  takes an argument consumes your pattern and searches for something else, and
  the run still exits successfully.
 - **Substring matches lie.** A protocol token like `h2c` matches inside
  `veth2cb3f8c`; a short id matches inside a hash. Before concluding anything
  from a hit, print its context (`grep -o ".\{30\}needle.\{30\}"`) and look at
  it. Before concluding anything from a *miss*, prove the command was
  well-formed by searching for a string you know is present.
 Binary-ish logs need `-a`, or grep reports "binary file matches" and you see
 nothing.
 ## Rules
 - Download before theorising. Do it even when the job log looks conclusive: a
  log that explains the symptom rarely explains the cause.
 - Quote literally, with the file path and line number. Never paraphrase a log
  line you did not read, and never reconstruct one from memory.
 - A negative result is a finding only when the search was verified sound.
 - "The artifact does not settle this" is a legitimate answer. Say it instead of
  filling the gap with a plausible mechanism.
@@ -0,0 +1,149 @@
 #!/usr/bin/env python3
 """Fetch the failed jobs of a GitHub Actions run and cluster them by error line.
 Usage:
    ci_failures.py RUN_ID [--repo OWNER/REPO] [--out DIR] [--artifacts]
 Prints how far HEAD moved since the run's commit, saves each finished failed
 job's log (colour stripped) under DIR (default /tmp/ci-logs-RUN_ID), and
 groups the jobs by the normalised error line that ended them (the last one). Jobs still running are
 listed but not fetched: their API log can stop hours early. The clustering is
 a heuristic; confirm a cluster's boundary against the logs.
 """
 from __future__ import annotations
 import argparse
 import json
 import re
 import subprocess
 import sys
 from pathlib import Path
 FAILED = {"failure", "cancelled", "timed_out"}
 ANSI = re.compile(r"\x1b\[[0-9;]*[A-Za-z]")
 STAMP = re.compile(r"^\d{4}-\d\d-\d\dT[\d:.]+Z\s?")
 GENERIC = re.compile(r"Process completed with exit code|failed=0\b|FAILED - RETRYING")
 ANNOTATION = re.compile(r"##\[error\]")
 FATAL = re.compile(r"fatal: \[|FAILED!|Traceback|AssertionError|panic:")
 SIGNAL = re.compile(r"\b(ERROR|Error|FAILED|FAIL|fatal|Exception|failed)\b")
 TASK = re.compile(r"^TASK \[")
 REFUSAL = "terminal escape sequences"
 def run(*cmd: str) -> subprocess.CompletedProcess:
    return subprocess.run(list(cmd), capture_output=True, text=True, check=False)
 def clean(line: str) -> str:
    return STAMP.sub("", ANSI.sub("", line)).strip()
 def normalise(line: str) -> str:
    line = re.sub(r"\b[0-9a-f]{7,}\b", "<hex>", line)
    line = re.sub(r"\d+", "<n>", line)
    return line[:200]
 def output_lines(text: str) -> list[tuple[str, str]]:
    kept, echo, task = [], False, ""
    for raw in text.splitlines():
        line = clean(raw)
        if line.startswith("##[group]Run "):
            echo = True
        elif line.startswith("##[endgroup]"):
            echo = False
        elif TASK.match(line):
            task = line
        elif line and not echo and not GENERIC.search(line):
            kept.append((task, line))
    return kept
 def signature(text: str) -> str:
    lines = output_lines(text)
    for pattern in (ANNOTATION, FATAL, SIGNAL):
        for task, line in reversed(lines):
            if pattern.search(line):
                found = ANNOTATION.sub("", line).strip()
                return normalise(f"{task} {found}" if pattern is FATAL and task else found)
    return normalise(lines[-1][1]) if lines else "<empty log>"
 def cluster(jobs: list[tuple[str, str]]) -> dict[str, list[str]]:
    groups: dict[str, list[str]] = {}
    for job, text in jobs:
        groups.setdefault(signature(text), []).append(job)
    return dict(sorted(groups.items(), key=lambda item: -len(item[1])))
 def behind(sha: str) -> str:
    if run("git", "cat-file", "-e", f"{sha}^{{commit}}").returncode:
        return "behind=unknown (run commit not in this clone)"
    count = run("git", "rev-list", "--count", f"{sha}..HEAD").stdout.strip()
    moved = {p.split("/")[0] for p in run("git", "diff", "--name-only", f"{sha}..HEAD").stdout.split()}
    return f"behind={count} moved={','.join(sorted(moved)) or '-'}"
 def main() -> int:
    parser = argparse.ArgumentParser(prog="ci_failures")
    parser.add_argument("run_id")
    parser.add_argument("--repo")
    parser.add_argument("--out")
    parser.add_argument("--artifacts", action="store_true")
    args = parser.parse_args()
    repo = args.repo or run("gh", "repo", "view", "--json", "nameWithOwner", "-q", ".nameWithOwner").stdout.strip()
    if not repo:
        parser.error("no --repo and gh cannot resolve the current repository")
    meta = run("gh", "api", f"/repos/{repo}/actions/runs/{args.run_id}")
    # per_page=100 on a large matrix answers HTTP 502; 30 per page goes through.
    listing = run("gh", "api", "--paginate", f"/repos/{repo}/actions/runs/{args.run_id}/jobs?per_page=30",
                  "--jq", ".jobs[] | {id, name, status, conclusion}")
    if meta.returncode or listing.returncode:
        print((meta.stderr or listing.stderr).strip(), file=sys.stderr)
        return 1
    data = json.loads(meta.stdout)
    jobs = [json.loads(line) for line in listing.stdout.splitlines() if line]
    out = Path(args.out or f"/tmp/ci-logs-{args.run_id}")
    out.mkdir(parents=True, exist_ok=True)
    print(f"run={args.run_id} repo={repo} head={data['head_sha'][:10]} status={data['status']} "
          f"jobs={len(jobs)} {behind(data['head_sha'])}")
    running = [j["name"] for j in jobs if j["status"] != "completed"]
    if running:
        print(f"running ({len(running)}, not fetched): {', '.join(running)}")
    fetched, missing = [], []
    for job in jobs:
        if job["status"] != "completed" or job["conclusion"] not in FAILED:
            continue
        log = run("gh", "api", "--allow-escape-sequences", f"/repos/{repo}/actions/jobs/{job['id']}/logs")
        if log.returncode or REFUSAL in log.stdout[:200]:
            missing.append(f"{job['conclusion']}\t{job['name']}\t{log.stderr.strip()[:80]}")
            continue
        slug = re.sub(r"[^A-Za-z0-9._-]+", "_", job["name"])[:80]
        path = out / f"{job['id']}-{slug}.log"
        path.write_text("\n".join(ANSI.sub("", line) for line in log.stdout.splitlines()), encoding="utf-8")
        text = "##[error]cancelled" if job["conclusion"] == "cancelled" else log.stdout
        fetched.append((f"{job['conclusion']}\t{job['name']}\t{path}", text))
    if missing:
        (out / "no-log.txt").write_text("\n".join(missing) + "\n", encoding="utf-8")
        counts = {c: sum(m.startswith(c) for m in missing) for c in FAILED}
        summary = ", ".join(f"{n} {c}" for c, n in sorted(counts.items()) if n)
        print(f"no log for {len(missing)} jobs ({summary}; never started or expired): {out / 'no-log.txt'}")
    for number, (sig, jobs) in enumerate(cluster(fetched).items(), 1):
        print(f"cluster {number} ({len(jobs)} jobs): {sig}")
        for job in jobs[:3]:
            print(f"  {job}")
        if len(jobs) > 3:
            members = out / f"cluster-{number}.txt"
            members.write_text("\n".join(jobs) + "\n", encoding="utf-8")
            print(f"  ... +{len(jobs) - 3} more: {members}")
    if args.artifacts:
        listing = run("gh", "api", f"/repos/{repo}/actions/runs/{args.run_id}/artifacts", "--paginate",
                      "--jq", '.artifacts[] | "\\(.name)\\t\\(.size_in_bytes)"')
        print("artifacts:")
        print(listing.stdout.rstrip() or "  none")
    return 0
 if __name__ == "__main__":
    sys.exit(main())
@@ -0,0 +1,170 @@
 ---
 name: cleanup
 description: >
  Free disk space agents filled: ask peer sessions what they still need,
  inventory /tmp, Docker and agent dirs, and delete only what the operator
  picks. Trigger on /cleanup, "aufräumen", "bereinigen", "platte voll", "no
  space left on device".
 ---
 Nothing is deleted before the operator has picked it. Every candidate carries
 its owner, its size, and its risk.
 ## Step 0: the resource chair
 Shared resources belong to the resource chair of the `democracy` skill. If
 another session holds the chair, send it the cleanup request and stop, unless
 the operator ordered the cleanup in this session; then tell the chair before
 step 1, so its grants and queue stay consistent. The `democracy` resource
 watch already runs the configured low-risk cleanup at its stop threshold; this
 skill covers everything that needs the operator's pick.
 ## Step 1: ask the peers
 List every agent the host can message (in Claude Code: `ListAgents`). Only
 peers sharing this machine's filesystem and Docker daemon count: local
 sessions and your own subagents. Cloud sessions and remote sessions on other
 machines hold nothing here; skip them.
 Send each local peer one message (`SendMessage`) and go on with step 2 while
 they answer:
 ```text
 Cleanup planned by <own session name>. Reply to me with everything you still
 need, one item per line:
 - paths under /tmp, caches, worktrees
 - containers, compose or stack projects, volumes, images, networks
 - running jobs (deploys, tests, builds) and when they finish
 Reply "nothing" if you need none of it. Items you do not answer for stay.
 ```
 ## Step 2: inventory
 Read-only, in parallel, sizes included:
 - **Scope**: the directories agents may write to, taken from the host's
  sandbox or permission config (in Claude Code:
  `sandbox.filesystem.allowWrite` in `.claude/settings.json`) plus the session
  scratch dirs. Anything outside is out of scope, even when readable.
 - **/tmp**: `du -sh` per top-level entry the current user owns. System entries
  (`.X11-unix`, `.ICE-unix`, `systemd-private-*`, `ssh-*`, `tmux-*`, sockets)
  are never candidates.
 - **Docker**: `docker system df -v`; containers with state and compose project
  label, volumes with the containers using them
  (`docker ps -a --filter volume=<name>`), dangling and unused images, build
  cache, unused networks. For every stopped container (`exited`, `created`)
  record its volumes (`{{.Mounts}}`); for every volume no running container
  uses, record size, `CreatedAt` (`docker volume inspect`) and compose project
  label, anonymous ones included.
 - **Other agent dirs**: package and tool caches (pip, npm, pre-commit, act,
  gh, buildx), worktrees with `git status --short` and whether their branch is
  merged or pushed, Ansible and build temp dirs.
 - **Live use**: open handles (`fuser`, `lsof`) and running processes by name
  only (`ps -o pid,etime,comm`). Never print full command lines or
  environments: process arguments carry credentials. A sandboxed agent often
  sees only its own processes, so an empty handle list proves nothing there;
  then only entries untouched for more than 24 hours (`find -mtime +0`) are
  candidates, and newer ones stay as `keep (recent, possibly in use)`.
 - **Archive duties**: memory and project instructions that require copying
  something out before deletion (logs, reports, generated maps). Such an item
  is `archive first`, never a plain delete.
 Prefer the project's own cleanup and teardown targets (Makefile, scripts) over
 raw commands where they exist, and obey project rules that forbid raw
 commands.
 ## Step 3: attribute
 Give every candidate an owner: a peer's reply, a session id in the path, a
 worktree or compose project name, a container label, an open handle, or
 `unknown`. Then apply, first match wins:
 1. Claimed by a peer: `keep`.
 2. Attributable to a peer that has not answered: `keep (no answer from
   <peer>)`. Silence is not consent.
 3. Archive duty: `archive first`.
 4. Everything else is a candidate with a risk class:
 | Risk | Meaning | Typical items |
 | ---- | ------- | ------------- |
 | low | regenerated automatically, cheap | stopped containers without volumes, dangling images, unused networks, scratch dirs and logs of finished sessions |
 | medium | regenerated, but slow, rate-limited or network-bound | unused tagged images (re-pull, registry limits), build cache (cold rebuilds), package caches |
 | high | data loss or a broken run possible | volumes, stopped containers with volumes, running unclaimed containers or stacks, worktrees with uncommitted or unpushed work, unread reports, anything `unknown` |
 ## Step 4: propose
 One table, biggest first:
 | Area | Item | Size | Owner | Risk | Proposal | What breaks if wrong |
 Below it: reclaimable size per risk class, and every `keep` row grouped by
 reason, so the operator sees what was protected and why.
 Then ask with `AskUserQuestion` in two calls.
 **Call 1, the selection** (`multiSelect: true`, checklists, at most four
 questions): `/tmp`, `Docker`, `Other dirs`, and a separate `High risk`
 question that holds every high-risk candidate. Size, risk and worst case go
 into each option's description. An area with more than four options is
 grouped by owner or risk, or split into a further call. Only ticked items are
 selected; nothing is preselected.
 **Stopped containers and volumes** always get their own checklist
 (`multiSelect: true`, header `⚠ Volumes`, in a further call when call 1 is
 full). Every stopped container and every volume without a running container
 appears there unless a peer claimed it; none is dropped as "not proposed",
 `unknown` and anonymous ones included. The question text carries the warning:
 `Deleting a volume is irreversible; its data does not come back unless the
 description says how.` One option per stopped container together with its
 volumes (removing the container alone frees nothing of them), one per
 unattached volume, anonymous ones grouped. Each description gives size,
 `CreatedAt`, owner or compose project, and what is lost.
 **Call 2, the confirmation** (`multiSelect: false`, radio buttons), computed
 from the ticks: `Execute: <n> items, <size> (Recommended)`, `Revise the
 selection`, `Abort`. The execute description names the worst case of the
 ticked items. Revise repeats call 1; nothing is deleted before `Execute`.
 ## Step 5: execute
 - Re-check claims, open handles and running state right before each delete: a
  peer may have started something since the inventory.
 - Delete exactly the selected items, by name or path. No blanket prune
  (`docker system prune -a --volumes`, `rm -rf /tmp/*`) that sweeps beyond the
  selection. For more than a handful of paths, write the list to a file under
  `/tmp`, grep it against every claim once more, then delete with
  `xargs -d '\n' rm -rf -- < <list>`.
 - Locally built images free little on their own: their layers stay behind as
  build cache. After removing images, re-measure `docker system df`; the
  layers show up as reclaimable build cache. Offer
  `docker builder prune --filter until=24h` as a follow-up candidate, ask the
  peers that claimed build cache first, and run it outside their running
  builds.
 - A host permission layer (an auto-mode classifier, a deny rule) may refuse a
  delete. Never route around it with another command or a peer. Finish the
  rest, then put the refused item to the operator again as its own option: an
  explicit renewed selection may pass; if it is refused again, hand over the
  exact command with its list file for the operator to run.
 - Running stacks go down through the project's teardown target, never by
  removing their containers or volumes underneath them.
 - Archive-first items: copy out, verify the copy, then delete.
 - A delete failing on permissions (files owned by a container UID): use the
  project's container helper. Never sudo, never bypass the sandbox; without a
  helper, report the path.
 - Stop on the first unexpected error and report it.
 ## Step 6: report
 `df -h` and `docker system df` before and after, freed size per area, every
 skipped or failed item with its reason. Tell each peer that answered which of
 the removed items lay near its own.
 ## Rules
 - Questions, inventory and proposal change nothing. Deletion needs the
  operator's selection in this run; a selection from an earlier run does not
  carry over.
 - Never delete outside the agent write scope, another user's files, or system
  entries in `/tmp`.
 - Never invent a peer's answer or a size. Unmeasured is `unknown`, and
  `unknown` is high risk.
@@ -2,12 +2,9 @@
 name: comments-clean
 description: >
  Purge explanatory comments from staged and unstaged changes. Run
-  AUTOMATICALLY, without being asked, every time the operator requests a
+  AUTOMATICALLY before every commit or stage request (commit, commit
-  commit or stage in any wording (commit, commit no-verify, stage next,
+  no-verify, stage next, socon, sobu, socobu, "commite ..."); also on
-  socon, sobu, socobu, "commite ..."), BEFORE executing that request; also
+  /comments-clean or comment complaints.
  on /comments-clean or when the operator complains about comments.
  Explanatory comments are bad practice: they narrate code instead of
  letting it speak. Portable across projects.
 ---
 When the operator asks to commit or stage, run this audit first and only
@@ -23,9 +20,22 @@ added lines only) and delete every explanatory comment:
 2. Keep only: parameter/exit-code docs, nocheck/noqa/shellcheck directives,
   and project-sanctioned exception markers that name a concrete trip-wire.
   A relabeled narration is not an exception; delete it instead.
-3. Re-run the project's comment lint if one exists and re-stage the files
+3. Block-comment languages are not line-oriented. In Jinja (`{# … #}`), C-like
-   that were already staged.
+   (`/* … */`), HTML (`<!-- … -->`) and their kin, a closing delimiter alone on
-4. Penalty: if this audit finds anything you wrote yourself, admit it in
+   its own line looks exactly like a comment line and gets deleted with the
   rest — after which the opening delimiter swallows every live statement down
   to the next closer. Never strip these languages with a line filter. Delete
   the whole block by hand, then count the delimiters (`grep -c '{#'` against
   `grep -c '#}'`) and re-parse the file.
 4. Re-run the project's comment lint if one exists. A green lint does NOT prove
   the deletion was safe: a swallowed block is still syntactically valid, so
   both the lint and the language parser can pass over code that no longer
   executes. Diff against the pre-edit revision (`git show HEAD:<path>`) and
   confirm that only comment lines left.
 5. Do not stage in order to run a check (the `commit-next` skill's step 4
   says why). Re-stage after the audit exactly the files that were staged
   before it.
 6. Penalty: if this audit finds anything you wrote yourself, admit it in
   the report (file and count, no excuses), lower your confidence numbers
   for the affected bundle, and write the incident to persistent agent
   memory (extend the existing minimal-comments feedback entry with the
@@ -1,20 +1,36 @@
 ---
 name: commit-next
 description: >
-  Commit the staged bundle with --no-verify, then stage the next logical
+  Check every branch for the staged fix, commit the staged bundle with
-  bundle and brief the operator on it. Trigger on /commit-next or when the
+  --no-verify, then stage and brief the next bundle. Trigger on /commit-next
-  operator asks to commit and stage the next bundle in one step. Portable
+  or when the operator asks to commit and stage the next bundle in one step.
  across projects.
 ---
-1. Run the commit-no-verify skill on the currently staged changes
+1. `git fetch --all --prune --tags`, then hunt the staged fix across every
-   (comments-clean audit first, then `git commit --no-verify`, exact
+   ref with the merge-fix skill's `find_fix.py` (by `--grep`, by `-S`, by
   `--path`). Describe the staged bundle in one line first so a
   wrong search intent is visible. Only the search runs here — never
   cherry-pick or merge, and never touch the staging index or the tree.
 2. A similar fix on another branch is a stop: run the dialectic skill on the
   two candidates (staged version as thesis, foreign commit as antithesis)
   until one survives, then present sha, branch, both diffs, the verdict and
   the recommended move (keep staged / take theirs / combine). Wait for the
   operator's decision; do not commit.
 3. Nothing similar found, or the operator picked the staged version: run the
   commit-no-verify skill on the currently staged changes (comments-clean
   audit and advocate check first, then `git commit --no-verify`, exact
   staging index only).
-2. Group the remaining uncommitted changes into logical bundles (one
+4. Group the remaining uncommitted changes into logical bundles (one
   concern per bundle: same failure class, same mechanism, or same role)
-   and stage the next one; never mix concerns to save a commit.
+   and stage the next one; never mix concerns to save a commit. Staging
-3. For the newly staged bundle run the explain skill (what the changes do
+   happens here and nowhere else: never stage merely to run a lint or a
   test. Repository checks that enumerate through `git ls-files` already
   see tracked, modified files and read the worktree; only an untracked
   NEW file needs `git add` to become visible. Path-scope every `git add`
   to the bundle — `git add -A` sweeps up the operator's own edits and
   their local config.
 5. For the newly staged bundle run the explain skill (what the changes do
   and how the pieces interact, file:line-anchored) followed by the
   confidence skill (fix confidence, break risk, evidence, residuals).
-4. Close with what remains unstaged and wait for the operator's commit
+6. Close with what remains unstaged and wait for the operator's commit
   decision; never auto-commit the newly staged bundle.
@@ -1,14 +1,14 @@
 ---
 name: commit-no-verify
 description: >
-  Commit the currently staged changes with --no-verify (skip the pre-commit
+  Commit the staged changes with --no-verify. Trigger on /commit-no-verify or
-  hooks). Trigger on /commit-no-verify or when the operator asks for a
+  any request for a no-verify commit; invoking it IS the per-commit
-  no-verify commit in any wording. Invoking this skill IS the explicit
+  authorization --no-verify requires, never carried over to later commits.
  per-commit authorization that --no-verify requires; it never carries over
  to later commits. Portable across projects.
 ---
-1. Run the comments-clean audit over the staged and unstaged changes first.
+1. Run the comments-clean audit over the staged and unstaged changes first,
   then the advocate check over the staging index. An advocate blocker stops
   here.
 2. Commit exactly the staging index with `git commit --no-verify` and a
   Conventional-Commits message derived from the staged diff; never add
   unstaged files, never push.
@@ -1,18 +1,21 @@
 ---
 name: confidence
 description: >
-  Report calibrated confidence unprompted. Trigger whenever you stage
+  Report calibrated fix confidence and break risk unprompted whenever you
-  changes, declare a fix or bundle ready, or summarize completed work the
+  stage changes, declare a fix or bundle ready, or summarize work the operator
-  operator may commit or deploy — the operator must never have to ask "how
+  may commit or deploy.
  sure are you?". Portable across projects.
 ---
 Every stage summary, fix report, or ready-to-commit bundle ends with a
-confidence block:
+confidence block.
 Invoke the `break-risk` skill every time, before writing the block; this skill
 reports the number it derives and never estimates one itself.
 1. **Fix confidence** (percent): probability the change does what it
-   claims. **Break risk** (percent): probability it breaks anything else.
+   claims. **Break risk** (percent): probability it breaks anything else,
-   Two separate numbers, never one blended figure.
+   derived via the `break-risk` skill. Two separate numbers, never one
   blended figure.
 2. **Evidence**: the strongest proof lines only — what was executed,
   reproduced, enumerated, or read from authoritative source. Distinguish
   proven (ran it, saw it) from derived (reasoned statically).
@@ -21,9 +24,17 @@ confidence block:
   Never a bare percentage without its residuals.
 Calibration rules:
 - Number the claim you verified, not the effort you spent; discount after
  any error the operator caught in the same area.
 - Distinguish "fixes the observed failure" from "makes the job green" when
  hidden layers may follow, and give both numbers.
 - Static analysis alone caps at the low nineties; only executed evidence
  (test, reproduction, live probe) justifies more.
 A threshold the operator named ("95% confidence, break risk ≤5%") is a gate on
 the work, handled as the `break-risk` skill's last section says, for both
 numbers.
 Objective: the `autotune` skill's **Objective** section. Two figures with their residuals replace a
 hedging paragraph and save the operator a verification round trip.
@@ -0,0 +1,11 @@
 ---
 name: coordinate
 description: >
  Coordinate a failure or fix with the agents on related branches through
  merges: claim, ask, report. Trigger on /coordinate and when a fix touches
  code other branches share.
 ---
 Run the `democracy` skill's **Lineage** and **Peer protocol** sections for the
 failure or fix at hand. They hold the whole procedure; this skill only keeps
 `/coordinate` working.
@@ -0,0 +1,93 @@
 ---
 name: criticize
 description: >
  Give constructive criticism of a problem, plan, design, decision, text or
  code from diverse aspects, each point hardened by dialectic. Trigger on
  /criticize, "kritisiere", "kritisch betrachten", "konstruktive Kritik", "was
  haeltst du davon", or requests for honest feedback.
 ---
 The stance is critical but constructive: the goal is a better result, not a
 verdict on its author. This skill judges any artifact from many angles; it is
 not limited to code (`sceptic` hunts violated engineering principles in code,
 `code-review` hunts bugs). Read-only: report the critique, change nothing
 unless the operator asks afterwards.
 ## Before you start: listen
 Invoke the `active-listening` skill. Pin down what is being criticized, what it
 is meant to achieve, for whom, and under which constraints. Derive what you can
 from the artifact and its surroundings first; ask only what only the operator
 knows. A critique against the wrong goal is noise.
 ## Steelman first
 Restate the artifact and its intent in its strongest form, in two or three
 sentences. Criticize that version, never a weaker one you made up. If you cannot
 steelman it, you have not understood it yet; go back to listening.
 ## The aspects
 Sweep every aspect that applies, and name the ones you skipped and why:
 1. **Goal fit** - does it solve the stated problem for the stated audience, or
   a neighbouring one?
 2. **Correctness** - are the facts, the logic, and the claimed cause-effect
   chains right? Which claim lacks evidence?
 3. **Completeness** - missing cases, missing stakeholders, missing failure
   paths, missing definition of done.
 4. **Consistency** - contradictions inside the artifact or against its
   neighbours; the same truth stated twice in two versions.
 5. **Simplicity** - mechanism, scope, or wording the problem never asked for;
   the smaller design that does the same.
 6. **Feasibility and cost** - effort, dependencies, order of steps, what blocks
   what, what it costs to run and to maintain.
 7. **Risk** - security at trust boundaries, data loss, reversibility, blast
   radius, what happens when it fails halfway.
 8. **People** - usability, accessibility, the reader's or user's actual path,
   who gains, who carries the burden.
 9. **Alternatives** - at least one materially different approach, compared
   honestly on the aspects above.
 10. **Ethics and compliance** - only where it touches personal data, licences,
    law, or fairness.
 ## Harden every point with dialectic
 Run each candidate critique through the `dialectic` skill before reporting it:
 - **Thesis**: the critique point with its evidence.
 - **Antithesis**: independent defenders argue that the artifact is right as it
  is, or that the constraint behind it is real; independent skeptics attack the
  suggested improvement for new problems it would cause.
 - **Synthesis**: keep the point, refine it, or drop it.
 Batch minor points into one dialectic round; run consequential points (high
 severity, expensive to act on, or challenging a premise) one by one.
 ## Output
 One line of scope, then the steelman. Then:
 - **Keep**: up to three strengths that must survive any change.
 - **Critique**: one ranked table, highest impact first:
 | Severity | Aspect | Observation | Why it matters | Suggestion | Evidence |
 `Evidence` is `proven` (read, run, reproduced, cited) or `derived` (reasoned).
 Close with at most three lines: how many candidate points dialectic dropped and
 why, the single change that helps most, and the `confidence` skill's numbers for
 the critique as a whole.
 ## Rules
 - Every critique carries a concrete suggestion. A complaint without one becomes
  an open question or is dropped.
 - Criticize the work, never the person. Direct wording, no sarcasm, no
  softening of a real problem, no inflation of a minor one.
 - A point that attacks a stated constraint or goal is labelled "challenges a
  premise" and kept apart from points that work within it.
 - No taste or formatting nitpicks unless the operator asks for them.
 - Answer in the operator's language.
 - The critique ends at the report. Do not fix, stage, or rewrite as part of
  this skill.
@@ -0,0 +1,182 @@
 ---
 name: democracy
 description: >
  Keep parallel agents from duplicating work: the first to claim it owns it;
  line representatives and a resource chair coordinate. Run AUTOMATICALLY
  before work another agent might also do (failure, cluster, feature, gate,
  deploy, cleanup), and on /democracy.
 ---
 Two agents doing the same work waste the second run and then collide in the
 merge. The remedy is an election nobody has to hold: whoever starts first owns
 the work. The structure is anarcho-syndicalist: councils (soviets) of
 delegates with a bound mandate, organised as a matrix, recallable by the
 operator at any time.
 ## Offices
 - **President of a piece of work**: the first agent that starts it. It does
  the whole piece. Every later agent hands its findings to the president and
  moves on to other work; it never builds a second version.
 - **Representative of an organisational line**: the first agent that works in
  the line. Lines come in as many dimensions as are useful: functional domain
  (security, GUI, identity, i18n), repository, branch line, host, cloud
  account. An agent belongs to one line per dimension.
 - **Organiser of a work cluster**: the first agent that starts a cluster
  cutting across lines (a failure cluster from `triage`, a cost cluster from
  `tune`, a feature touching several domains). It works the strand end to end
  and informs the representative of every line the strand touches before it
  changes that line's files.
 - **Chair**: the earliest-elected president still running. It chairs the
  council of all representatives and organisers and allocates the resources of
  the whole system everyone runs on. There is exactly one allocator at any
  time.
 Every office is a mandate, not a rank: it coordinates and allocates, it never
 takes over or directs the content of another agent's work. The operator
 outranks every office and appoints or recalls at will; an operator
 appointment outranks the election.
 ## Election
 The first to claim wins, and the claim is atomic, so two agents can never both
 be first. `${CLAUDE_SKILL_DIR}/scripts/council.py` keeps the claims in
 `/tmp/agents-council`; other runtimes call the same script from their skills
 directory with `--session <id>`.
 1. **Name the office**: `work/<slug>`, `cluster/<slug>`,
   `<dimension>/<line>` for a representative, or `chair`. The slug names the
   work concretely: the issue number, the failing line, the cost's log line.
 2. **Look first**: `council.py list`. A claim for the same work under another
   slug counts as first.
 3. **Claim**: `council.py claim <office> --work "<what, evidence, branch or
   worktree>"`. Exit 0: elected, or already yours. Exit 3: a live agent holds
   it, and its claim is printed. Exit 4: the holder's session is dead (see
   Succession).
 4. **Elected**: continue with the Peer protocol; its message is the
   announcement.
 5. **Not elected**: send the holder your findings and stop on that work.
 A Claude Code session cannot edit a file in a git worktree before it holds a
 claim: a PreToolUse hook (`council.py gate`) blocks the edit.
 Peers on other machines cannot see this host's claims. Between machines the
 earliest announcement by message time wins, and each host's claim directory
 records the result.
 ## Chair and resource allocation
 - The first agent elected on the system while `chair` is vacant also claims
  `chair`, in the same way.
 - An operator-appointed allocator holds the chair: its claim says
  `appointed: operator`, and no election replaces it while its session runs and
  the operator has not ended the appointment.
 - Before starting anything that consumes shared capacity (gates, deploys,
  iteration loops, swarm or cloud resources, cleanup of shared data), an agent
  sends the chair what, where, for how long and its expected disk need, waits
  for the grant, and reports the release. The chair keeps one queue, answers
  every request with its position, enforces the limits the operator set,
  checks every grant with `watch.py admit --need-gb <N>` (or `--kind <slot>`),
  and never cancels a running grant to free capacity.
 - Handover is explicit: the outgoing chair sends every peer the new holder,
  the queue and the running grants, and keeps allocating until the incoming
  chair confirms. Two allocators never run at once.
 ## Resource watch
 The chair keeps the system from running out instead of reacting once it has.
 `${CLAUDE_SKILL_DIR}/scripts/watch.py` does the measuring as hooks in every
 session, so no agent probes the disk by hand; it reads its thresholds, job
 patterns, cleanup commands and cache directories from
 `~/.config/agents-council/config.json` (keys in the script's docstring) and
 stays silent below the thresholds.
 - **Warn**: the chair and every session get the biggest consumers; the chair
  plans a cleanup and grants no large slot.
 - **Stop**: new heavy jobs are blocked in every session (teardown commands
  never), and the hook runs the configured low-risk cleanup by itself.
 - **Critical**: additionally a desktop alarm to the operator. A running stack
  is torn down only with the operator's OK.
 - A measured fall that would reach the critical threshold within the
  projection window raises the stage by one at once.
 - Caches together stay under their budget; the least recently used go first.
 `watch.py status` shows the last measurement. A `RESOURCE` line in a hook's
 context is addressed to the chair; any other session only adjusts its own
 plans to it.
 ## Succession
 - A piece of work or a cluster ends with its result, a proposal, or an
  abandoned attempt, reported to the peers who heard the claim. The holder
  then runs `council.py release <office>`, and the next agent that starts the
  work is elected afresh.
 - A representative or the chair that ends hands its office to the next-oldest
  live representative (by claim time) and announces it.
 - A claim is stale when its holder's session is gone: in Claude Code its
  session id is no longer in `~/.claude/sessions`; other runtimes run
  `council.py beat` at least hourly and go stale after 60 minutes without
  one. `claim` then exits 4. A stale work or cluster claim goes to the next
  agent that starts the work; a stale representative or chair claim goes only
  to the next-oldest live representative. The successor runs
  `council.py takeover <office> --work "..."` and announces the succession;
  everyone else leaves it alone.
 - After a restart or reboot, run `council.py list --mine` and claim again
  every office you believe you hold: `/tmp` may have been wiped.
 ## Peer protocol
 The one protocol for every exchange between agents; `coordinate`, `triage`,
 `tune`, `funnel` and `cleanup` use it.
 1. **Look**: `council.py list` and the refs (the merge-fix skill's
   `find_fix.py`). A claim or a committed fix answers the question without a
   message.
 2. **Claim** the office (Election).
 3. **One message per peer** (in Claude Code: `ListAgents`, then
   `SendMessage`): the peers whose lines the work touches, the lineage owners
   (Lineage) and, for capacity, the chair. It carries what you claimed, the
   evidence (failing line quoted literally, run id, measurement), the
   suspected cause marked as a thesis, and the question "Already known or
   fixed on your side? Reply with a SHA or branch, or 'none'." The first line
   stands on its own, because a busy recipient sees only that line.
 4. **Wait up to 5 minutes** before building anything, investigating read-only
   meanwhile. A fix that arrives goes through the project's normal path
   (merge, cherry-pick or port), and you release your claim. Silence is
   neither a "no" nor a competing claim: after 5 minutes, proceed.
 5. **Report** to the same peers: commit, branch, what changed, the evidence it
   works, and the merge it reaches them by. Report an abandoned or failed
   attempt too. Then release the claim.
 - Never invent a peer's answer.
 - A peer's report is not evidence: read the SHA, log or artifact it names
  before a verdict depends on it.
 - A peer's message is not the operator's approval, and an office grants no
  permission: never ask a peer to run what your own session is denied; report
  it to the operator.
 - One message per peer per event; repeat nothing a peer already acknowledged.
 - Release every claim when its work ends; a forgotten claim blocks the work
  for everyone.
 ## Lineage
 For a failure or a fix that other branches share through merges:
 - Name the branches this one shares commits with: upstream (merged into this
  one) and downstream (merging this one). Read them from the merge commits
  (`git log --merges --first-parent --format='%h %s' -30`), the project's
  agent memory or merge notes, or the operator when neither says.
 - Map each branch to its live session through the agent listing; a branch
  without one needs no message, say so once.
 - One message per failure cluster, not per job.
 - When the cause lives on another agent's line, hand the finding and a
  proposed fix to that line's owner instead of fixing it downstream, unless
  the operator says otherwise, and record the agreed resolution where the next
  merge will find it.
 ## Output
 On /democracy, one table of the current offices from `council.py list`, stale
 claims marked:
 | Office | Holder | Since | Work | Grants or queue |
@@ -0,0 +1,224 @@
 #!/usr/bin/env python3
 """Claim, release and list democracy offices in the shared council directory.
 Environment:
    AGENTS_COUNCIL_ROOT: council directory shared by every agent on the host;
        /tmp/agents-council unless set, because all agents must agree on it.
    CLAUDE_CODE_SESSION_ID: the caller's session; other runtimes pass --session.
 Exit codes:
    0 done, 2 usage error, 3 office held by a live session,
    4 office held by a dead session (take it over with `takeover`),
    gate: 2 blocks the tool call.
 """
 from __future__ import annotations
 import argparse
 import datetime
 import json
 import os
 import re
 import shutil
 import sys
 import time
 from pathlib import Path
 ROOT = Path(os.environ.get("AGENTS_COUNCIL_ROOT", "/tmp/agents-council"))
 HEARTBEAT_LIMIT = 3600
 OFFICE = re.compile(r"^[a-z0-9][a-z0-9._-]*(/[a-z0-9][a-z0-9._-]*)*$")
 CLAIM = "claim.json"
 def live_sessions() -> dict[str, str | None]:
    found: dict[str, str | None] = {}
    for path in (Path.home() / ".claude" / "sessions").glob("*.json"):
        try:
            data = json.loads(path.read_text(encoding="utf-8"))
        except (OSError, ValueError):
            continue
        if data.get("sessionId"):
            found[data["sessionId"]] = data.get("name")
    return found
 def caller(session: str | None) -> tuple[str, str]:
    if session:
        return session, "other"
    env = os.environ.get("CLAUDE_CODE_SESSION_ID")
    if env:
        return env, "claude-code"
    print("council: no session id; pass --session", file=sys.stderr)
    sys.exit(2)
 def office_dir(office: str) -> Path:
    if not OFFICE.match(office):
        print(f"council: invalid office {office!r}", file=sys.stderr)
        sys.exit(2)
    return ROOT / office
 def read_claim(path: Path) -> dict | None:
    try:
        return json.loads((path / CLAIM).read_text(encoding="utf-8"))
    except (FileNotFoundError, ValueError):
        return None
 def is_stale(claim: dict | None, live: dict[str, str | None]) -> bool:
    if claim is None:
        return False
    if claim["runtime"] == "claude-code":
        return claim["session"] not in live
    return time.time() - claim["beat"] > HEARTBEAT_LIMIT
 def claims() -> list[tuple[str, dict]]:
    found = []
    for path in sorted(ROOT.glob(f"**/{CLAIM}")):
        claim = read_claim(path.parent)
        if claim is not None:
            found.append((path.parent.relative_to(ROOT).as_posix(), claim))
    return found
 def write_claim(path: Path, office: str, sid: str, runtime: str, work: str) -> None:
    claim = {
        "office": office,
        "session": sid,
        "runtime": runtime,
        "name": live_sessions().get(sid),
        "since": datetime.datetime.now(datetime.timezone.utc).isoformat(timespec="seconds"),
        "beat": time.time(),
        "work": work,
    }
    (path / CLAIM).write_text(json.dumps(claim), encoding="utf-8")
 def cmd_claim(args: argparse.Namespace) -> int:
    sid, runtime = caller(args.session)
    path = office_dir(args.office)
    path.parent.mkdir(parents=True, exist_ok=True)
    try:
        path.mkdir()
    except FileExistsError:
        claim = read_claim(path)
        if claim is not None and claim["session"] == sid:
            write_claim(path, args.office, sid, runtime, args.work)
            print(f"held {args.office}")
            return 0
        print(json.dumps(claim) if claim else f"{args.office}: claim in progress")
        return 4 if is_stale(claim, live_sessions()) else 3
    write_claim(path, args.office, sid, runtime, args.work)
    print(f"elected {args.office}")
    return 0
 def cmd_takeover(args: argparse.Namespace) -> int:
    path = office_dir(args.office)
    claim = read_claim(path)
    if path.exists() and not is_stale(claim, live_sessions()):
        print(json.dumps(claim) if claim else f"{args.office}: claim in progress")
        return 3
    # Two successors racing between this check and the rename can both win;
    # democracy names exactly one successor per office to keep that window closed.
    if path.exists():
        tomb = path.with_name(f"{path.name}.stale-{os.getpid()}-{time.time_ns()}")
        try:
            path.rename(tomb)
        except FileNotFoundError:
            pass
        else:
            shutil.rmtree(tomb)
    return cmd_claim(args)
 def cmd_release(args: argparse.Namespace) -> int:
    sid, _ = caller(args.session)
    path = office_dir(args.office)
    claim = read_claim(path)
    if claim is None:
        print(f"{args.office}: not claimed")
        return 0
    if claim["session"] != sid and not is_stale(claim, live_sessions()):
        print(json.dumps(claim))
        return 3
    shutil.rmtree(path)
    print(f"released {args.office}")
    return 0
 def cmd_beat(args: argparse.Namespace) -> int:
    sid, _ = caller(args.session)
    for office, claim in claims():
        if claim["session"] == sid:
            claim["beat"] = time.time()
            (ROOT / office / CLAIM).write_text(json.dumps(claim), encoding="utf-8")
    return 0
 def cmd_list(args: argparse.Namespace) -> int:
    live = live_sessions()
    sid = None
    if args.mine:
        sid, _ = caller(args.session)
    for office, claim in claims():
        if sid is not None and claim["session"] != sid:
            continue
        state = "stale" if is_stale(claim, live) else "live"
        holder = claim["name"] or claim["session"]
        print(f"{office}\t{holder}\t{claim['since']}\t{state}\t{claim['work']}")
    return 0
 def in_git_worktree(target: Path) -> bool:
    return any((path / ".git").exists() for path in (target, *target.parents))
 def cmd_gate(_: argparse.Namespace) -> int:
    payload = json.load(sys.stdin)
    tool_input = payload.get("tool_input") or {}
    target = tool_input.get("file_path") or tool_input.get("notebook_path")
    if not target or not in_git_worktree(Path(target).resolve()):
        return 0
    if any(claim["session"] == payload.get("session_id") for _, claim in claims()):
        return 0
    script = Path(__file__).resolve()
    print(
        "democracy: this session holds no claim, so it may not edit files in a git "
        f"worktree. Claim the work first: python3 {script} claim work/<slug> "
        "--work '<one line>'. Exit 3 or 4 means another agent holds it; read the "
        "democracy skill.",
        file=sys.stderr,
    )
    return 2
 def main() -> int:
    parser = argparse.ArgumentParser(prog="council")
    sub = parser.add_subparsers(dest="cmd", required=True)
    for name, func in (("claim", cmd_claim), ("takeover", cmd_takeover)):
        cmd = sub.add_parser(name)
        cmd.add_argument("office")
        cmd.add_argument("--work", required=True)
        cmd.add_argument("--session")
        cmd.set_defaults(func=func)
    release = sub.add_parser("release")
    release.add_argument("office")
    release.add_argument("--session")
    release.set_defaults(func=cmd_release)
    beat = sub.add_parser("beat")
    beat.add_argument("--session")
    beat.set_defaults(func=cmd_beat)
    listing = sub.add_parser("list")
    listing.add_argument("--mine", action="store_true")
    listing.add_argument("--session")
    listing.set_defaults(func=cmd_list)
    sub.add_parser("gate").set_defaults(func=cmd_gate)
    args = parser.parse_args()
    return args.func(args)
 if __name__ == "__main__":
    sys.exit(main())
@@ -0,0 +1,359 @@
 #!/usr/bin/env python3
 """Watch disk, Docker and caches for the democracy resource chair.
 Usage:
    watch.py watch     hook (UserPromptSubmit, PostToolUse): measure at most once
                       per interval machine-wide, act on the stage, show alerts
    watch.py block     hook (PreToolUse on Bash): exit 2 for a heavy job while the
                       stage is stop or critical
    watch.py admit (--need-gb N | --kind KIND)
                       exit 0 when the grant leaves at least the stop threshold
                       free, 3 otherwise
    watch.py cleanup (--caches | --full)
                       trim caches to their budget; --full also runs the
                       configured low-risk cleanup commands
    watch.py status    print the last measurement
 Environment:
    AGENTS_COUNCIL_ROOT: shared state directory; /tmp/agents-council unless set,
        because every agent on the host must agree on it.
    AGENTS_COUNCIL_CONFIG: config file; ~/.config/agents-council/config.json
        unless set. Without a config file every command exits 0 and does nothing.
 Config keys, all required: disk, thresholds_gb {warn, stop, critical},
 projection_minutes, interval_seconds, cache_interval_seconds, heavy_commands,
 never_block (regexes), cleanup (shell commands), worktree_cleanup (shell
 commands with {worktree}), worktrees_command (shell command printing the
 running stacks' worktree paths, one per line), worktree_marker (path relative
 to a worktree that must exist), cache_dirs, cache_budget_percent,
 slot_need_gb {kind: GB}, notify (argv the alert text is appended to).
 """
 from __future__ import annotations
 import argparse
 import json
 import os
 import re
 import shlex
 import subprocess
 import sys
 import time
 from pathlib import Path
 ROOT = Path(os.environ.get("AGENTS_COUNCIL_ROOT", "/tmp/agents-council"))
 CONFIG = Path(
    os.environ.get(
        "AGENTS_COUNCIL_CONFIG",
        str(Path.home() / ".config" / "agents-council" / "config.json"),
    )
 )
 STAGES = ("ok", "warn", "stop", "critical")
 STATE = ROOT / "watch.json"
 ALERTS = ROOT / "alerts.jsonl"
 ALERT_TTL = 3600
 LOCK_TTL = 1800
 GB = 1e9
 def load_config() -> dict | None:
    if not CONFIG.is_file():
        return None
    return json.loads(CONFIG.read_text(encoding="utf-8"))
 def disk(cfg: dict) -> tuple[float, float]:
    st = os.statvfs(cfg["disk"])
    return st.f_bavail * st.f_frsize / GB, st.f_blocks * st.f_frsize / GB
 def read_json(path: Path) -> dict:
    try:
        return json.loads(path.read_text(encoding="utf-8"))
    except (OSError, ValueError):
        return {}
 def write_json(path: Path, data: dict) -> None:
    # A full disk refuses the write; the stage must still act, so it is dropped.
    try:
        path.parent.mkdir(parents=True, exist_ok=True)
        tmp = path.with_name(f"{path.name}.{os.getpid()}")
        tmp.write_text(json.dumps(data), encoding="utf-8")
        tmp.replace(path)
    except OSError:
        pass
 def stage_for(free: float, samples: list[list[float]], cfg: dict, now: float) -> str:
    limits = cfg["thresholds_gb"]
    level = 0
    for index, name in ((1, "warn"), (2, "stop"), (3, "critical")):
        if free < limits[name]:
            level = index
    window = cfg["projection_minutes"] * 60
    recent = [s for s in samples if now - s[0] <= window and now > s[0]]
    if recent and level < 3:
        then, before = recent[0]
        rate = (before - free) / (now - then)
        if rate > 0 and (free - limits["critical"]) / rate < window:
            level += 1
    return STAGES[level]
 def alert(stage: str, text: str) -> None:
    try:
        ROOT.mkdir(parents=True, exist_ok=True)
        with ALERTS.open("a", encoding="utf-8") as handle:
            handle.write(json.dumps({"id": time.time_ns(), "stage": stage, "text": text}) + "\n")
    except OSError:
        pass
 def unseen_alerts(session: str) -> list[str]:
    seen_path = ROOT / "seen" / session
    try:
        seen = int(seen_path.read_text(encoding="utf-8"))
    except (OSError, ValueError):
        seen = 0
    try:
        lines = ALERTS.read_text(encoding="utf-8").splitlines()
    except OSError:
        return []
    cutoff = time.time_ns() - ALERT_TTL * 10**9
    fresh = [json.loads(line) for line in lines if line]
    fresh = [a for a in fresh if a["id"] > max(seen, cutoff)]
    if fresh:
        write_text(seen_path, str(fresh[-1]["id"]))
    return [a["text"] for a in fresh]
 def write_text(path: Path, text: str) -> None:
    try:
        path.parent.mkdir(parents=True, exist_ok=True)
        path.write_text(text, encoding="utf-8")
    except OSError:
        pass
 def consumers() -> str:
    try:
        out = subprocess.run(
            ["docker", "system", "df", "--format", "{{.Type}}: {{.Size}} ({{.Reclaimable}} reclaimable)"],
            capture_output=True, text=True, timeout=30, check=False,
        ).stdout.strip()
    except (OSError, subprocess.TimeoutExpired):
        return "docker system df unavailable"
    return "; ".join(out.splitlines()) or "docker system df printed nothing"
 def notify(cfg: dict, text: str) -> None:
    try:
        subprocess.run([*cfg["notify"], text], capture_output=True, timeout=10, check=False)
    except (OSError, subprocess.TimeoutExpired):
        alert("critical", f"RESOURCE: desktop notification failed: {cfg['notify'][0]}")
 def spawn_cleanup(mode: str) -> None:
    subprocess.Popen(
        [sys.executable, str(Path(__file__).resolve()), "cleanup", mode],
        stdin=subprocess.DEVNULL, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
        start_new_session=True,
    )
 def act(cfg: dict, stage: str, previous: str, free: float) -> None:
    limits = cfg["thresholds_gb"]
    head = f"/ has {free:.1f} GB free (warn < {limits['warn']}, stop < {limits['stop']}, critical < {limits['critical']} GB)"
    if STAGES.index(stage) < STAGES.index(previous):
        alert(stage, f"RESOURCE {stage.upper()}: recovered, {head}.")
        return
    if stage == "warn":
        alert(stage, f"RESOURCE WARN: {head}. Biggest consumers: {consumers()}. Chair: plan a cleanup; peers: request no large slot.")
    elif stage in ("stop", "critical"):
        alert(stage, f"RESOURCE {stage.upper()}: {head}. New heavy jobs are blocked; low-risk cleanup is running. {consumers()}")
        spawn_cleanup("--full")
        if stage == "critical":
            notify(cfg, f"Agents council: {head}. Heavy jobs blocked; a running stack needs the operator's OK to go.")
 def cmd_watch(cfg: dict) -> int:
    payload = json.load(sys.stdin)
    now = time.time()
    state = read_json(STATE)
    if now - state.get("checked", 0) >= cfg["interval_seconds"]:
        free, _ = disk(cfg)
        samples = [s for s in state.get("samples", []) if now - s[0] <= cfg["projection_minutes"] * 60]
        stage = stage_for(free, samples, cfg, now)
        previous = state.get("stage", "ok")
        samples.append([now, free])
        state.update(checked=now, free=free, stage=stage, samples=samples)
        due = now - state.get("caches_checked", 0) >= cfg["cache_interval_seconds"]
        if due:
            state["caches_checked"] = now
        write_json(STATE, state)
        if stage != previous:
            act(cfg, stage, previous, free)
        elif due:
            spawn_cleanup("--caches")
    texts = unseen_alerts(payload["session_id"])
    if texts:
        print(json.dumps({"hookSpecificOutput": {
            "hookEventName": payload["hook_event_name"],
            "additionalContext": "\n".join(texts),
        }}))
    return 0
 def cmd_block(cfg: dict) -> int:
    payload = json.load(sys.stdin)
    command = (payload.get("tool_input") or {}).get("command", "")
    stage = read_json(STATE).get("stage", "ok")
    if stage not in ("stop", "critical"):
        return 0
    if any(re.search(p, command) for p in cfg["never_block"]):
        return 0
    if any(re.search(p, command) for p in cfg["heavy_commands"]):
        print(
            f"democracy: resource stage is {stage}; new heavy jobs are blocked until "
            "the disk recovers. Teardown commands still run. Ask the resource chair.",
            file=sys.stderr,
        )
        return 2
    return 0
 def cmd_admit(cfg: dict, args: argparse.Namespace) -> int:
    need = args.need_gb if args.need_gb is not None else cfg["slot_need_gb"][args.kind]
    free, _ = disk(cfg)
    left = free - need
    floor = cfg["thresholds_gb"]["stop"]
    verdict = "grant" if left >= floor else "deny"
    print(f"{verdict}: free {free:.1f} GB, need {need} GB, leaves {left:.1f} GB (floor {floor} GB)")
    return 0 if verdict == "grant" else 3
 def cache_files(dirs: list[str]) -> list[tuple[float, int, str]]:
    files = []
    for root in dirs:
        for base, _, names in os.walk(os.path.expanduser(root)):
            for name in names:
                path = os.path.join(base, name)
                try:
                    st = os.lstat(path)
                except OSError:
                    continue
                files.append((st.st_atime, st.st_size, path))
    return files
 def trim_caches(cfg: dict) -> float:
    _, total = disk(cfg)
    budget = total * GB * cfg["cache_budget_percent"] / 100
    files = sorted(cache_files(cfg["cache_dirs"]))
    used = sum(size for _, size, _ in files)
    freed = 0
    for _, size, path in files:
        if used - freed <= budget:
            break
        try:
            os.remove(path)
            freed += size
        except OSError:
            continue
    return freed / GB
 def running_worktrees(cfg: dict) -> list[str]:
    try:
        out = subprocess.run(
            cfg["worktrees_command"], shell=True, capture_output=True, text=True, timeout=60, check=False
        ).stdout
    except subprocess.TimeoutExpired:
        return []
    trees = sorted({line.strip() for line in out.splitlines() if line.strip()})
    return [t for t in trees if os.path.exists(os.path.join(t, cfg["worktree_marker"]))]
 def run_cleanup(cfg: dict) -> list[str]:
    commands = list(cfg["cleanup"])
    for tree in running_worktrees(cfg):
        commands += [t.format(worktree=shlex.quote(tree)) for t in cfg["worktree_cleanup"]]
    results = []
    for command in commands:
        try:
            code = subprocess.run(command, shell=True, capture_output=True, timeout=900, check=False).returncode
        except subprocess.TimeoutExpired:
            code = "timeout"
        results.append(f"{'ok' if code == 0 else f'failed ({code})'}: {command}")
    return results
 def cmd_cleanup(cfg: dict, args: argparse.Namespace) -> int:
    lock = ROOT / "cleanup.lock"
    ROOT.mkdir(parents=True, exist_ok=True)
    try:
        if time.time() - lock.stat().st_mtime > LOCK_TTL:
            lock.rmdir()
    except FileNotFoundError:
        pass
    try:
        lock.mkdir()
    except FileExistsError:
        return 0
    try:
        before, _ = disk(cfg)
        results = run_cleanup(cfg) if args.full else []
        cache_gb = trim_caches(cfg)
        after, _ = disk(cfg)
        failed = [r for r in results if not r.startswith("ok")]
        if args.full or cache_gb:
            alert(
                "cleanup",
                f"RESOURCE CLEANUP: freed {after - before:.1f} GB ({cache_gb:.1f} GB caches), "
                f"/ now {after:.1f} GB free; {len(results) - len(failed)} commands ok"
                + (f", failed: {' | '.join(failed)}" if failed else "."),
            )
    finally:
        lock.rmdir()
    return 0
 def cmd_status(cfg: dict) -> int:
    state = read_json(STATE)
    free, total = disk(cfg)
    print(f"stage={state.get('stage', 'unmeasured')} free={free:.1f}GB total={total:.1f}GB samples={len(state.get('samples', []))}")
    return 0
 def main() -> int:
    parser = argparse.ArgumentParser(prog="watch")
    sub = parser.add_subparsers(dest="cmd", required=True)
    sub.add_parser("watch")
    sub.add_parser("block")
    admit = sub.add_parser("admit")
    need = admit.add_mutually_exclusive_group(required=True)
    need.add_argument("--need-gb", type=float)
    need.add_argument("--kind")
    cleanup = sub.add_parser("cleanup")
    mode = cleanup.add_mutually_exclusive_group(required=True)
    mode.add_argument("--caches", action="store_true")
    mode.add_argument("--full", action="store_true")
    sub.add_parser("status")
    args = parser.parse_args()
    cfg = load_config()
    if cfg is None:
        return 0
    if args.cmd == "watch":
        return cmd_watch(cfg)
    if args.cmd == "block":
        return cmd_block(cfg)
    if args.cmd == "admit":
        return cmd_admit(cfg, args)
    if args.cmd == "cleanup":
        return cmd_cleanup(cfg, args)
    return cmd_status(cfg)
 if __name__ == "__main__":
    sys.exit(main())
@@ -1,11 +1,10 @@
 ---
 name: dialectic
 description: >
-  Reach a defensible conclusion through a dialectical process — thesis,
+  Reach a defensible conclusion by thesis, independent skeptic antithesis and
-  adversarial antithesis, synthesis — iterated until one thesis survives every
+  synthesis, iterated to ~99% confidence. Trigger for any consequential root
-  attack at ~99% confidence. Trigger for any consequential root-cause, design,
+  cause, design or decision where being wrong is expensive and the answer is
-  or decision where being wrong is expensive and the answer is not yet certain.
+  uncertain.
  Deep-inspects every accessible source. Portable across projects.
 ---
 Reach the truth by dialectic, not by first guess. Do not stop at a plausible
@@ -19,6 +18,10 @@ needs. A dialectic run on the wrong question wastes the whole loop. Return to
 `active-listening` mid-loop whenever an objection turns on a fact only the
 operator holds.
 When another skill or an autonomous run (robot, observe, a loop) invoked this
 skill, derive the requirements from the task and the evidence instead of
 asking; ask only when an operator-only fact blocks the conclusion.
 ## The loop
 1. **Thesis.** Form the strongest initial claim you can, grounded in evidence,
@@ -0,0 +1,135 @@
 ---
 name: dream
 description: >
  Consolidate the last 24 hours of work in this repository: what held, what
  broke, which hypotheses follow, into memory and improvement proposals.
  Trigger on /dream or when the operator asks to reflect on or learn from
  recent work.
 ---
 Sleep on the work: replay what happened in this repository, separate what held
 from what broke, and keep only the conclusions that survive an attack. A dream
 that flatters the dreamer is worthless — an unverified success counts as
 unknown, not as a win.
 ## Objective
 The dream exists to raise the confidence of every future run in this repository
 until the mistakes it found cannot happen again. Judge every finding, memory
 entry, and proposal by one question: does it make the next run measurably less
 likely to repeat this failure? What changes nothing about future behaviour is
 not worth recording.
 A class of error counts as closed only when something structural prevents it —
 a test, a lint rule, a CI check, a hook, a written rule in the agent
 instructions, or a skill. "Be more careful next time" closes nothing. Report the
 confidence the next run can now hold on each closed class, calibrated as the
 `confidence` skill requires, and name what is still open.
 ## Scope
 Everything is read from the repository the skill runs in (`git rev-parse
 --show-toplevel`); never mix in another project's history or memory. The default
 window is the last 24 hours — the operator may widen it (`/dream 7d`). State the
 repository and the window in one line before you start.
 Run only when the operator asks. The 12-hour reminder hook
 (`hooks/reminder.sh`, stamped per repository in its git dir) is a hint for the
 operator, not a trigger: print it and continue with their actual request.
 ## Step 1: recall
 Gather the window's actual record, not your recollection of it:
 - **History**: `git log --since` across all branches plus `--reflog`, with
  `--stat`; the reflog carries the resets, rebases, and amends the log hides.
 - **Working state**: `git status`, the staged and unstaged diff, stashes.
 - **CI**: the runs created in the window, their conclusions, and the failing
  logs of the red ones — the failing log line, not the job title.
 - **Reviews**: pull requests and issues touched in the window, and the review
  comments on them.
 - **Memory**: this project's memory directory and its `MEMORY.md` index — the
  findings already recorded, so the dream extends them instead of repeating them.
 - **Instructions**: `CLAUDE.md`, `AGENTS.md`, and the project's own skills, to
  see which rule a mistake actually violated.
 - **Sessions**: the readable session transcripts touched in the window, for the
  operator corrections that never reached a commit.
 ## Step 2: score
 Classify every action in the window, each with its evidence (`sha`, `file:line`,
 run id):
 - **Held** — shipped and still proven: green CI, a passing test that covers it,
  no revert, no follow-up fix.
 - **Broke** — reverted, amended, hot-fixed by a later commit, red CI, or
  corrected by the operator. A failure that ended green is still a failure; count
  the detour, not only the destination.
 - **Unknown** — no verification exists either way. Unknown is a finding in its
  own right, never a quiet pass.
 ## Step 3: hypothesize
 Turn the scored record into falsifiable hypotheses, each written as *cause →
 predicted consequence → what would disprove it*. Take a pattern (two or more
 occurrences) or a single expensive failure; drop the one-off noise. A hypothesis
 that nothing could disprove is not a hypothesis — cut it.
 ## Step 4: dialectic
 Put every surviving hypothesis through the `dialectic` skill: thesis,
 independent skeptics, synthesis, iterated to ~99%. What survives becomes a
 finding. What gets refuted is recorded as discarded **with the refutation**, so
 the next dream does not re-derive it.
 ## Step 5: write memory
 Persist the findings in this project's memory directory: one file per fact, one
 pointer line in `MEMORY.md`. Update the file that already covers a fact instead
 of adding a duplicate, and delete the memories this run disproved. Absolute
 dates, never "yesterday". Nothing that the repository already records — code
 structure, git history, `CLAUDE.md` — belongs in memory.
 ## Step 6: offer the improvements
 Two kinds come out of the findings:
 - **Code** — the change to this repository's source, tests, config, or CI that
  the finding calls for.
 - **Agent instructions** — the missing or wrong rule in `CLAUDE.md`,
  `AGENTS.md`, the project's own skills, or the contributor docs. Every mistake
  the agent repeated in the window is a candidate: quote the line that failed to
  prevent it, or name the rule that is absent, and propose the exact wording.
 Ask for each one with `AskUserQuestion` and `multiSelect: false` so it renders
 as radio buttons: the improvement, the evidence behind it, and the options
 `Apply now`, `Record as a task`, `Skip`. Ranked by evidence, one question per
 improvement. No repository file changes before an answer comes back.
 ## Step 7: autotune, then autoskill
 With the dialectic done and memory written, run `autotune` — hand it this run's
 findings as its evidence so it does not re-collect them — and then `autoskill`
 for the operator corrections the window revealed. Both keep their own
 confirmation steps; the dream never writes a skill on their behalf.
 Close the run by refreshing this repository's reminder stamp, so the next hint
 is twelve hours out: `date +%s > "$(git rev-parse --path-format=absolute --git-common-dir)/.dream-stamp"`.
 ## Rules
 - Evidence or unknown: a claim about what held needs a green signal, and its
  absence is reported as unknown rather than assumed.
 - Honest scoring: reverts, amends, and follow-up fixes are failures of the first
  attempt even when the final state is green.
 - Read-only until step 6: the dream inspects, and writes memory; repository code
  changes only on an explicit radio answer.
 - An empty window is a valid result — report "nothing in the last 24 hours in
  this repository" and stop. Never invent activity to have something to dream
  about.
 - Prefer the structural fix over the remembered one: a test, check, hook, or
  written rule that makes the error impossible beats a memory entry that only
  reminds you of it.
 - Report the whole ledger: held, broke, unknown, the surviving findings, the
  refuted hypotheses with why they fell, and the confidence each closed error
  class now carries into the next run.
@@ -0,0 +1,40 @@
 #!/usr/bin/env bash
 # UserPromptSubmit hook: emit the dream hint at most once per interval and per
 # repository. The stamp lives in the repository's git dir, so every repository
 # keeps its own cadence; outside a git repository the hook stays silent.
 # Stdout lands in the agent's context, which is why the payload tells the agent
 # to show the hint and keep going instead of running dream by itself.
 set -euo pipefail
 interval="${DREAM_INTERVAL_SECONDS:-43200}"
 if [[ -n "${DREAM_STAMP:-}" ]]; then
 	stamp="${DREAM_STAMP}"
 else
 	git_dir="$(git rev-parse --path-format=absolute --git-common-dir 2>/dev/null)" || exit 0
 	stamp="${git_dir}/.dream-stamp"
 fi
 now="$(date +%s)"
 last=0
 if [[ -f "${stamp}" ]]; then
 	read -r last <"${stamp}" || true
 	[[ "${last}" =~ ^[0-9]+$ ]] || last=0
 fi
 if ((now - last < interval)); then
 	exit 0
 fi
 mkdir -p "$(dirname "${stamp}")"
 printf '%s\n' "${now}" >"${stamp}"
 cat <<'HINT'
 DREAM HINT: over twelve hours since the last hint for this repository. Show the
 operator this notice verbatim as a short block, then continue with their
 request. Do NOT run dream, do not start replaying the history, do not ask
 questions about it.
    💤 /dream - consolidate the last 24h in this repository: what held, what
       broke, which hypotheses follow, into memory and improvement proposals.
       Runs only when you trigger it.
 HINT
@@ -0,0 +1,116 @@
 ---
 name: essence
 description: >
  Compact skills and memory to their essence after a health pass that finds
  dead weight. Trigger on /essence or when skills or MEMORY.md grow too long,
  duplicate, or eat too much context.
 ---
 Shorter, not thinner. Every fact that still earns its place survives; what goes
 is repetition, narration, and entries nothing loads any more.
 ## What actually costs context
 Compact by how often a thing is loaded, not by how large it is:
 1. **`MEMORY.md`** and every skill's **`description`** frontmatter load into
   *every* session. A wasted line here is paid on every single turn. Highest
   leverage by a wide margin.
 2. **Skill bodies** load when the skill fires. Worth tightening, paid only on use.
 3. **Memory files** load on recall. Lowest leverage — and the place where
   cutting evidence does the most damage.
 A 200-line skill body that fires twice a month costs less than four bloated
 `MEMORY.md` pointers. Rank the work accordingly and say so in the report.
 ## Step 1: ask for the scope
 Never assume the target. Ask once with `AskUserQuestion` and
 `multiSelect: true`, so the operator can tick several:
 - **Skills repository** — the portable `skills/*/SKILL.md` set.
 - **This project** — its `.claude/skills` and `.agents/skills`, plus the
  project's memory directory and `MEMORY.md`.
 - **Home** — `~/.claude` and `~/.agents` (skills, rules, commands, agents).
 Take nothing outside what was ticked.
 ## Step 2: doctor
 Before compacting anything, find what is already broken or dead. Report each
 finding with its path:
 - A `MEMORY.md` pointer whose file is gone, and a memory file no pointer names.
 - A dead `[[wikilink]]` — no memory of that name exists.
 - A memory asserting a `file:line`, function, or flag that no longer exists.
  Verify against the current tree; a stale citation is worse than no memory,
  because it reads as checked.
 - The same rule stated in two or more skills. One is authoritative, the rest
  are drift.
 - A first-party skill shadowing a locked/vendored one.
 - Frontmatter missing `name` or `description`, or a `description` that says
  what the skill does but never when it fires — such a skill costs context in
  every session and never triggers.
 Dead entries are deleted, not compacted. Fix the broken ones or report them;
 never compact around a stale citation, or the error survives in denser form.
 ## Step 3: compact
 Cut:
 - Anything the repository already records — code structure, git history, what
  `CLAUDE.md`/`AGENTS.md` already say.
 - A rule the agent follows anyway without being told.
 - The second and third statement of one rule; keep the authoritative one and
  point at it.
 - Narration, hedging, and worked examples beyond the one that carries the point.
 Keep, always:
 - The **why** and the evidence behind a memory — the date, the incident, the
  quoted correction. A rule stripped of its evidence degrades into an
  exhortation and stops changing behaviour.
 - Any fact with no other home. If it exists nowhere else, it is not redundancy.
 - Trigger wording in a `description`. Tighten it, never drop the "when".
 Dates stay absolute. Never introduce "recently" while shortening.
 ## Step 4: show the diff, then write
 Writing happens in place — an essence that lives beside the original saves
 nothing, because the original is still what loads.
 Preconditions, per target:
 - A git repository must have a clean status for the paths being rewritten, so
  every edit is revertable with git alone. Not clean: report and stop.
 - `~/.claude` and `~/.agents` are usually not under git. Copy each file to
  `<file>.pre-essence` first.
 Then present, per file: the diff, and the lines and characters saved. Ask for
 one confirmation covering the batch. Write only after it.
 ## Step 5: verify
 - Re-run the target's own validation if it has one (in a skills repository,
  its test suite: frontmatter, lock, shadowing).
 - Re-read each rewritten file and confirm every kept fact is still stated, not
  merely implied by something shorter.
 - Report the before/after loaded-context figure for the always-loaded surface
  (`MEMORY.md` plus all descriptions) separately from the rest. That number is
  the point of the run.
 ## Rules
 - Deleting is the goal; rewriting shorter is the fallback. Prefer removing an
  entry over polishing it.
 - Never merge two memories whose evidence differs. Two incidents supporting one
  rule are stronger than one; two rules sharing one incident is a fabrication.
 - A skill that has never fired is a candidate for deletion, not compaction —
  say so and let the operator decide.
 - Nothing is written without the Step 4 confirmation, and nothing outside the
  scope ticked in Step 1.
 - `caveman-compress` is not a rival: it rewrites the wording of one named file,
  this removes what should not be there at all. Run this one first — compressing
  a dead entry is pure waste.
@@ -1,10 +1,9 @@
 ---
 name: explain
 description: >
-  Explain code and its interactions. Trigger when the operator asks what a
+  Explain code and its interactions. Trigger when the operator asks what code
-  piece of code does, how a mechanism works, why components interact the way
+  does, how a mechanism works, why components interact, or invokes /explain on
-  they do, or invokes /explain on a file, function, role, or module.
+  a file, function, role or module.
  Portable across projects.
 ---
 Explain the named code so an engineer who has never seen it can follow it:
@@ -0,0 +1,113 @@
 ---
 name: funnel
 description: >
  Keep agent work in its own git worktree until it reaches 90% fix confidence
  and under 5% break risk; below that bar it becomes a proposal. Trigger when
  delegating work to agents, when agents work in parallel, or on /funnel.
 ---
 # Funnel
 The main line only receives changes that have earned it. Everything else stays
 in a sidetree until it does, or leaves as a proposal.
 ## The bar
 - **Fix confidence of at least 90 %** and **break risk under 5 %**, both
  reported by the `confidence` skill, the break risk derived through the
  `break-risk` skill.
 - The operator may set a different bar. Never lower it to make a change fit.
 ## The stages
 1. **Brief.** One concern per agent. Name the goal, the exact base commit, the
   scope that may change, the evidence expected back, a hard time cap (default
   60 minutes) and what the agent must not do: no pushes, no gates or deploys
   unless the resource chair of the `democracy` skill granted them, read-only
   probes on shared live systems.
 2. **Sidetree.** The agent works in its own worktree on its own branch,
   created from the exact base commit, so its edits never land under a running
   gate or deploy in the main line. Use the host's worktree isolation when it
   has one.
 3. **Evidence.** The agent runs targeted tests, adds a test for the new
   behaviour, and makes a negative control: the test fails without the change.
   Numbers come from the same source before and after.
 4. **Self-assessment.** The agent reports fix confidence and break risk with
   their residuals and commits on its branch.
 5. **Independent review.** Before anything is merged, run the `dialectic` skill
   with skeptics who did not write the change. A hole they find returns the
   change to stage 3.
 6. **Merge.** At or above the bar, merge the sidetree branch into the main line
   following the repository's commit rules, then re-verify on the main line
   with the project's gate.
 7. **Tear everything down.** Once the branch is merged, remove everything the
   sidetree brought into being, so the next round starts from a clean
   machine. See "After the merge" below.
 ## After the merge
 A merged sidetree leaves nothing behind. Work through the list in this order,
 and verify each step instead of assuming it:
 1. **Stop its jobs**: background runs, watchers, scheduled loops and agents
   that belong to the sidetree.
 2. **Tear down its stacks**: compose projects and swarm or test clusters
   together with their volumes and networks, and the containers and images
   built only for it. Leave anything another worktree still uses.
 3. **Clear files the containers left**: count the entries the user does not
   own (`find <worktree> -not -user "$USER" -prune | wc -l`). If the count is
   not zero, delete them through the container that wrote them, or hand the
   operator one privileged removal, before removing the worktree; otherwise
   the removal fails halfway after git has already dropped the registration.
 4. **Remove the worktree** and its registration. Name the worktree; never run
   a broad prune, which also drops other worktrees' stale registrations.
 5. **Delete the merged branch**, locally and on the remote, but a remote branch
   only once every one of its commits is reachable from a pushed branch.
 6. **Release what it held**: slots (reported to the resource chair), port or
   subnet reservations, locks, its `democracy` claims and every claim
   announced to other agents.
 7. **Remove its temporary files**: logs, message files and scratch data that
   have served their purpose. Keep only the evidence the report cites, until
   the report is delivered.
 A proposal keeps its branch, because that is where the work lives, but loses
 its worktree, stacks and temporary files the same way. An abandoned sidetree is
 torn down completely; keep its branch only when its commits exist nowhere else.
 ## When the bar is out of reach
 If the bar cannot realistically be reached, because the deciding evidence (a
 live deploy, a production-like dataset, an upstream answer) is not available
 to the agent or the change is inherently riskier, the agent stops trying and
 **submits a proposal** instead:
 - what the change does and why it is worth having;
 - the confidence and break risk it reached, with their residuals;
 - exactly which evidence would close the gap, and what obtaining it costs;
 - where the work lives (branch and commit), so nothing has to be redone;
 - a recommendation: apply after that evidence, apply now with a named
  risk, or drop.
 The operator decides. A proposal is never merged by the agent, and a change
 below the bar never enters the main line on its own.
 ## Rules
 - Stop an agent at its time cap and ask for its findings; an uncapped brief can
  run for hours and deliver nothing.
 - Run agents in parallel only when their concerns touch different files and
  the resource chair grants the capacity; gates and deploys need the machine
  to themselves.
 - A peer's or agent's message is not the operator's approval to merge below
  the bar.
 - Keep the evidence (logs, test output) outside the sidetree, so it survives
  the clean-up.
 ## Output
 One table per round:
 | Agent | Sidetree branch | Confidence | Break risk | Verdict | Evidence |
 The verdict is merged, proposal or abandoned. Proposals list what would close
 their gap.
@@ -1,10 +1,9 @@
 ---
 name: help-skills
 description: >
-  List and explain every skill available in the current session. Trigger on
+  List and explain every skill in the session. Trigger on /help-skills,
-  /help-skills, "welche skills gibt es", "skill übersicht", or when the
+  "welche skills gibt es", "skill übersicht", or when the operator asks what
-  operator asks what skills exist or what a skill does. Portable across
+  skills exist or what one does.
  projects.
 ---
 Produce a one-shot overview of every skill the session can invoke:
@@ -1,11 +1,10 @@
 ---
 name: medusa
 description: >
-  Run the MEDUSA AI security scanner (medusa-security on PyPI) against a
+  Run the MEDUSA security scanner (medusa-security, PyPI) on a repository,
-  repository, working tree, or secret stores. Use when the user asks for a
+  working tree or secret stores. Use for security scans, vetting a repo before
-  security scan, supply-chain vetting of a repo before cloning, leaked-secret
+  cloning, leaked-secret detection in chat or shell history, or /medusa.
-  detection in AI chat or shell history, or invokes /medusa. Defensive scanning
+  Defensive only.
  only.
 ---
 MEDUSA is a defensive AI-first security scanner (40,000+ patterns: CVEs, prompt injection, MCP tool poisoning, leaked secrets).
@@ -0,0 +1,79 @@
 ---
 name: merge-fix
 description: >
  Hunt down the commits on any branch that carry a described fix and merge it
  into the current branch. Trigger on /merge-fix <description> or when the
  operator asks to bring over a fix from another branch.
 ---
 The free text after `/merge-fix` is the fix description, exactly like `/loop`
 takes its prompt. Everything after the skill name is the search intent; there is
 no other input. If it is missing, ask for it and stop.
 ## 1. Refuse to work on a dirty tree
 `git status --porcelain` must be empty. If it is not, stop and let the operator
 commit or open a worktree. Never stash, never `git checkout -- <file>`, never
 `git restore --staged`: those discard uncommitted work that other agents in the
 same tree may own.
 ## 2. Fetch everything
 ```bash
 git fetch --all --prune --tags
 ```
 That is the "pull all branches" step. Do not check branches out one by one; all
 searching below runs against the fetched refs.
 ## 3. Turn the description into search terms
 Derive from the operator's text: 2-4 keywords for commit messages, plus any
 symbol, string, or path the fix must have touched. State the derived terms in
 one line so a wrong reading is visible before the search.
 ## 4. Search every ref
 Search by keyword, by literal and by path in one call; they find different
 things:
 ```bash
 ${CLAUDE_SKILL_DIR}/scripts/find_fix.py --grep '<keyword>' -S '<symbol or literal>' --path '<path or glob>'
 ```
 It drops every hit already in HEAD (`in_head` counts them; all in HEAD means
 the fix is already in, say so and stop) and prints the rest with their
 branches and files. Then read the actual change: `git show --stat <sha>` and
 the diff. A matching commit subject is a lead, not evidence, and `-S` also
 matches the commit that removed the literal; the diff decides.
 ## 5. Present the candidates
 One table: sha, branch(es), subject, files touched, one line on what the diff
 really does. Let the operator pick unless exactly one candidate survives step 4
 and its diff plainly is the described fix.
 ## 6. Bring it over
 - One or a few commits, rest of the branch unwanted: `git cherry-pick -x <sha>…`
  (`-x` records the origin).
 - The whole branch is wanted: `git merge --no-ff <branch>`.
 - Conflicts: resolve them by reading both sides. Never `-X ours`/`-X theirs`,
  never `git checkout --ours/--theirs` wholesale — that silently drops one
  side's fix. Unresolvable → `git cherry-pick --abort` / `git merge --abort`,
  report what conflicted, change nothing else.
 ## 7. Report
 Show `git log --oneline <base>..HEAD` and the resulting diffstat, then apply the
 confidence skill: what came over, what did not, what still has to be verified.
 Run the project's test or quality gate only when the operator asks for it.
 ## Rules
 - Never push and never force anything; the merge stays local until the operator
  ships it.
 - Never rewrite history to make a cherry-pick fit (no rebase of the target
  branch, no amend of foreign commits).
 - If the search finds nothing, say so plainly and list the terms tried. Do not
  reimplement the fix under this skill — that is a different task.
@@ -0,0 +1,67 @@
 #!/usr/bin/env python3
 """Find commits on any ref that match a fix, skipping those already in HEAD.
 Usage:
    find_fix.py [--grep KEYWORD]... [-S LITERAL]... [--path PATH]... [--fetch]
 Prints one line per candidate: sha, up to three branches, subject, files
 touched. -S also matches commits that removed the literal; read the diff.
 Run it inside the repository; it never checks anything out.
 """
 from __future__ import annotations
 import argparse
 import subprocess
 import sys
 LIMIT = 30
 def git(*args: str) -> subprocess.CompletedProcess:
    return subprocess.run(["git", *args], capture_output=True, text=True, check=False)
 def search(args: argparse.Namespace) -> list[str]:
    queries = [["--regexp-ignore-case", f"--grep={k}"] for k in args.grep]
    queries += [[f"-S{s}"] for s in args.S]
    queries += [["--", p] for p in args.path]
    found: list[str] = []
    for query in queries:
        for sha in git("log", "--all", "--format=%H", *query).stdout.split():
            if sha not in found:
                found.append(sha)
    return found
 def describe(sha: str) -> str:
    subject = git("log", "-1", "--format=%h %s", sha).stdout.strip()
    branches = git("branch", "-a", "--contains", sha, "--format=%(refname:short)").stdout.split()
    files = [f for f in git("show", "--name-only", "--format=", sha).stdout.split("\n") if f]
    shown = ",".join(branches[:3]) + (f",+{len(branches) - 3}" if len(branches) > 3 else "")
    return f"{subject}\t{shown or '-'}\t{len(files)} files: {','.join(files[:5])}"
 def main() -> int:
    parser = argparse.ArgumentParser(prog="find_fix")
    parser.add_argument("--grep", action="append", default=[])
    parser.add_argument("-S", action="append", default=[])
    parser.add_argument("--path", action="append", default=[])
    parser.add_argument("--fetch", action="store_true")
    args = parser.parse_args()
    if not (args.grep or args.S or args.path):
        parser.error("give at least one --grep, -S or --path")
    if args.fetch:
        git("fetch", "--all", "--prune", "--tags")
    hits = search(args)
    outside = [sha for sha in hits if git("merge-base", "--is-ancestor", sha, "HEAD").returncode]
    print(f"hits={len(hits)} in_head={len(hits) - len(outside)} candidates={len(outside)}")
    for sha in outside[:LIMIT]:
        print(describe(sha))
    if len(outside) > LIMIT:
        print(f"... {len(outside) - LIMIT} more, narrow the search")
    return 0
 if __name__ == "__main__":
    sys.exit(main())
@@ -1,15 +1,10 @@
 ---
 name: no-defaults
 description: >
-  Forbid default values for parameters, settings, and environment lookups in
+  Forbid default values for parameters, settings and environment lookups;
-  every language. Load and apply AUTOMATICALLY on every task that writes,
+  demand explicit values and fail loudly. Apply AUTOMATICALLY on every task
-  edits, generates, refactors, or reviews code, config, templates, schemas,
+  that writes, edits, generates, refactors or reviews code, config, templates,
-  or infrastructure — never wait to be asked. A default silently substitutes
+  schemas or infrastructure.
  a guess for a value the caller failed to supply, so a misconfiguration
  becomes wrong behaviour instead of a loud failure. Demand strict, explicit
  values and fail loudly when one is missing. `false`, `null` and `""` are the
  only free defaults; every other default needs a written justification.
  Portable across projects.
 ---
 A default value is a guess, written at the point where the caller's intent is
@@ -59,7 +54,7 @@ with `false`, `null`, or `""` is permitted, and with any other value is permitte
 only when justified in a comment.
 | Language | Never | Instead |
-|---|---|---|
+| --- | --- | --- |
 | Bash | `${VAR:-x}`, `${VAR:=x}`, `: "${VAR:=x}"` | `: "${VAR:?set VAR in <file>}"` |
 | Python | `def f(a=1)`, `d.get(k, x)`, `getattr(o, n, x)`, `os.environ.get(k, x)`, `argparse(default=)` | required args, `d[k]`, `os.environ[k]`, `add_argument(required=True)` |
 | Jinja / Ansible | `\| default(x)`, `lookup('env', k) \| default('')` | `\| mandatory`, `assert` on the var |
@@ -1,10 +1,9 @@
 ---
 name: observe
 description: >
-  Watch a named target (a CI run, a pipeline, a build, a log) on a fixed
+  Watch a named target (CI run, pipeline, build, log) on a cadence and triage
-  cadence and apply the triage skill whenever new failures appear. Trigger on
+  new failures as they appear. Trigger on /observe <target> or when the
-  /observe <target> or when the operator asks to keep monitoring something and
+  operator asks to keep monitoring and fix regressions.
  fix regressions as they surface. Portable across projects.
 ---
 Keep a target under continuous watch and act only when it degrades. The loop is
@@ -27,7 +26,8 @@ gave one): after each observation, schedule the next wake-up yourself and pass
 the same `/observe` prompt back so the next firing repeats the watch.
 1. **Poll the target.** Pull its current state (job list, check conclusions, the
-   log's tail). Enumerate the failures exactly as the `triage` skill's step 1
+   log's tail); for a GitHub Actions run, the ci-artifacts skill's
   `ci_failures.py <run>`. Enumerate the failures exactly as the `triage` skill's step 1
   does: failure, cancelled, timed-out; a downstream aggregate that fails only
   because an upstream did is not its own failure.
 2. **Diff against the baseline.** A failure is *new* only if it is not in the
@@ -1,10 +1,9 @@
 ---
 name: plan
 description: >
-  Turn a request into an objective-oriented plan and hand it to the robot skill
+  Turn a request into an objective-oriented plan and hand it to the robot
-  for autonomous execution. Trigger on /plan or when the operator asks for a
+  skill. Trigger on /plan or when the operator asks for a plan, breakdown or
-  plan, a breakdown, or a strategy for a task before any work starts. Portable
+  strategy before work starts.
  across projects.
 ---
 Produce a plan whose every item is an outcome with a check, then execute it
@@ -1,11 +1,9 @@
 ---
 name: refactor
 description: >
-  Refactor every file that is currently uncommitted in the git working tree -
+  Refactor every uncommitted file (staged, unstaged, untracked) up to the
-  staged, unstaged and untracked - up to the project's coding standards,
+  project's standards without changing behaviour. Trigger on /refactor or when
-  without changing behaviour. Trigger on /refactor or when the operator asks to
+  the operator asks to clean up, tidy or polish the current changes.
  clean up, tidy, or polish the current changes before committing. Portable
  across projects.
 ---
 The unit of work is the uncommitted change set, not the repository. Behaviour
@@ -1,10 +1,9 @@
 ---
 name: robot
 description: >
-  Work fully autonomously to a goal. Trigger when the operator wants hands-off,
+  Work fully autonomously to a goal: no check-ins, no stopping before the goal,
-  uninterrupted execution — the agent drives the task to completion without
+  only commands already cleared. Trigger when the operator wants hands-off,
-  check-ins, never stops before the goal is reached, and uses only commands it
+  uninterrupted execution.
  is already cleared to run. Portable across projects.
 ---
 Operate as an autonomous robot: given a goal, reach it without hand-holding.
@@ -0,0 +1,86 @@
 ---
 name: sceptic
 description: >
  Analyse a repository, path or diff with a hostile eye for violated
  principles: SPOT, DRY, KISS, SRP, security at trust boundaries. Trigger on
  /sceptic, "sei skeptisch", "kritische analyse", "review the architecture",
  or "what is wrong with the code".
 ---
 The stance is adversarial: the code is guilty until its structure proves
 otherwise. This skill hunts violated principles, not bugs (`code-review`),
 not bloat (`ponytail-audit`), not leaked secrets (`medusa`). Read-only —
 report the findings, change nothing unless the operator asks afterwards.
 ## Scope
 The operator names the target: a repository, a directory, a role, or a diff.
 Without a target, take the whole repository and rank by blast radius. State
 the scope in one line before reporting, and name explicitly what you did not
 read — an audit that silently covered half the tree reads as clean.
 ## The lenses
 Apply each independently; a finding belongs to exactly one:
 1. **SPOT** — one truth, one place. A value, a version, a hostname, a
   schema, a list of names that exists twice is a divergence waiting to
   happen. Name both locations and the drift they will cause.
 2. **DRY** — repeated *knowledge*, not repeated text. Two blocks that look
   alike but change for different reasons are correctly separate; merging
   them is the violation. Prove the shared reason to change, or drop the
   finding.
 3. **SRP** — one reason to change per unit. Look for the module that both
   parses and persists, the function whose name contains "and", the role
   that configures and deploys and monitors. The test: name the two
   stakeholders who would each demand a change here.
 4. **KISS** — mechanism the problem never asked for: an abstraction with one
   implementation, a config value that never varies, a cache with no
   measured pressure, a retry loop around a deterministic call.
 5. **Security** — only at trust boundaries: unvalidated input crossing into
   a query, a shell, a template, or a path; authorization checked in the
   caller instead of the resource; a secret in a default, a log, or an
   image layer; a permission wider than the operation needs. Defensive
   findings only.
 6. **Silent defaults** — a fallback that substitutes a guess for a value the
   caller failed to supply, turning a misconfiguration into wrong behaviour
   instead of a loud failure. See the `no-defaults` skill.
 ## Method
 - **Evidence before verdict.** Every finding carries `file:line` you actually
  read. Grep for the sibling occurrences before calling something a
  duplicate — one hit is not a pattern.
 - **Name the failure.** A finding must state what concretely goes wrong:
  which value drifts, which caller breaks, which input reaches which sink.
  A finding whose worst case is "feels wrong" gets deleted, not softened.
 - **Attack your own finding.** For anything consequential or expensive to
  fix, run it through the `dialectic` skill before reporting it — the
  strongest counter-argument is usually "these two are not the same
  knowledge" or "this boundary is already checked upstream".
 - **Respect intent.** A deliberate marker (`ponytail:`, a documented
  exception, a nocheck directive with a reason) is a known ceiling, not a
  finding. Report it only when the ceiling has since been exceeded.
 ## Output
 One ranked table, worst blast radius first:
 | Severity | Principle | Location | What is wrong | Smallest fix |
 Then three closing lines at most: what the scope excluded, which findings
 survived `dialectic` versus which are still derived-only, and the single
 finding to fix first. No per-finding paragraphs — the operator asks when a
 row needs unpacking.
 ## Rules
 - Never report formatting, naming taste, or missing comments; a linter owns
  those and the operator did not ask.
 - Distinguish proven (grepped, read, reproduced) from derived (reasoned
  statically) per finding, and close with the `confidence` skill's numbers
  for the audit as a whole.
 - No silent caps: if you ranked and cut, say how many findings were dropped
  and by what rule.
 - Skepticism ends at the report. Do not fix, stage, or refactor as part of
  this skill.
@@ -0,0 +1,34 @@
 ---
 name: scope-guard
 description: >
  Deliver exactly what was asked and nothing more. Trigger on every task that
  creates files, commands or commits, on unreviewed autonomous runs, on
  hand-over requests, and when the operator asks why something unrequested
  exists.
 ---
 The requested scope is the deliverable. A problem you noticed behind the
 request is not an invitation to solve it.
 Before creating anything the request did not name, answer one question: which
 sentence of the request asks for this? No sentence, no artifact.
 1. No permanent scaffolding for a one-off need. A migration you run once needs
   a throwaway script, not a CLI subcommand that ships forever. Do the thing,
   then delete the scaffolding.
 2. Path-scope every `git add`. Never `-A`, never `.`. The operator's tree
   carries their own edits and their local config, and sweeping those into
   your bundle makes them undo your work by hand.
 3. An adjacent defect gets reported in one line, not fixed — unless it blocks
   the request. Fixing it quietly puts an unreviewed change in someone else's
   commit.
 4. An instruction to hand something over is complete on its own. Print the
   artifact in the shape asked for: do not optimize, reorder or deduplicate it,
   do not wrap it in a command that consumes it, and do not append commentary,
   next steps, confidence blocks or offers. A narrowing qualifier ("just the
   line") re-shapes the earlier request; re-emit narrower without re-deriving
   the content. A problem with the artifact is worth one sentence before it,
   never a rewrite of it.
 When something genuinely seems missing, say so in one sentence and finish the
 requested scope anyway. Widening the scope is the operator's call.
@@ -1,10 +1,9 @@
 ---
 name: shortcuts
 description: >
-  Portable operator conversation shortcuts, reusable across every project.
+  Operator conversation shortcuts. Trigger when the message is exactly one of
-  Trigger when the user's message is exactly one of these tokens (optionally
+  these tokens, optionally with arguments: sobu, soco, socobu, socon, sona,
-  followed by arguments): sobu, soco, socobu, socon, sona, sote. Expand the
+  sote. Expand it and act on the expanded instruction.
  shortcut to the listed meaning and act on the expanded instruction.
 ---
 When an operator message is one of the shortcuts below, expand it to the meaning and act on it.
@@ -0,0 +1,61 @@
 ---
 name: status-matrix
 description: >
  Print one matrix of everything in flight: tasks, workflows, background jobs,
  with confidence and break risk per item. Trigger on /status-matrix,
  "status", "übersicht", "wo stehen wir", or what is done, left or running.
 ---
 One-shot overview, no side effects: this skill reads state and reports it. It
 never edits files, stages, commits, deploys, or starts work — the operator
 decides what to act on after reading the matrix.
 ## Collect the state
 Gather in parallel, skip any source the session does not have, and never invent
 a row for a source you could not read:
 1. **Tasks** — the session's todo list, the operator's stated goals this
   session, and any requirement or plan documents the work follows.
 2. **Working tree** — `git status --short` and `git log --oneline -10`:
   uncommitted files are open work, commits since the branch point are closed
   work. Untracked files count as open.
 3. **Running** — background bash jobs, spawned agents not yet reported,
   scheduled `/loop` wake-ups, and remote runs (CI, deploys). Poll their live
   state; a job you started is not "running" because you started it.
 4. **Remote** — open PRs and issues for the branch, and their check
   conclusions, when a forge CLI is available.
 ## The matrix
 One table, most-blocking first, one row per item:
 | Item | State | Conf. | Break risk | Evidence / next step |
 - **State**: `open`, `running`, `blocked`, `done`. `done` means verified —
  a passing run, a green check, a reproduction that now succeeds. Work that
  merely looks finished is `open`.
 - **Conf.** and **Break risk**: two separate percentages per the `confidence`
  skill — probability the item does what it claims, and probability it breaks
  something else. Only for rows that changed something; `—` for pure
  observation rows.
 - **Evidence / next step**: for `done`, the strongest proof line (what ran,
  what it returned). For everything else, the single next action.
 Then at most three lines below the table:
 - **Blocked on**: what needs the operator (a decision, a credential, an
  approval) — omit when nothing does.
 - **Residuals**: what keeps the lowest confidence number below 100.
 - **Running**: what will report back on its own, so the operator does not wait
  on it manually.
 ## Rules
 - Cover everything in scope, including work the operator did not ask about —
  a forgotten background job or an unpushed commit is exactly what the matrix
  exists to surface.
 - Report what the poll returned, never a predicted result for work still in
  flight; a running item has no confidence number yet.
 - Compress: one row per item, no per-item paragraphs. If a row needs
  explaining, the operator will ask.
@@ -1,95 +0,0 @@
 ---
 name: test-fix
 description: >
  Run the project's Makefile test targets in a loop - run, diagnose, fix, re-run -
  until every selected target is green or the loop is provably stuck. Trigger on
  /test-fix or when the operator asks to make the tests pass, fix the failing
  tests, or get the suite green. Portable across projects.
 ---
 Same target discovery and selection as the `test` skill - invoke it for step 1
 and step 2 rather than reimplementing the grep and the selection list. This
 skill owns what happens after the first red result: the fix loop.
 The contract is narrow: the code becomes correct, the tests stay honest.
 ## The loop
 One iteration is: run -> read the failure -> one root cause -> one fix ->
 re-run. Never batch several speculative fixes into one iteration; a green run
 after three simultaneous changes proves nothing about which one mattered.
 ### Run
 Run the failing target alone, not the whole selection - the fastest command
 that reproduces the failure. Widen back to the full selection only when that
 target is green.
 ### Diagnose
 Read the actual error output before touching a file. Apply the `triage` skill
 to reach a verified root cause; for a failure whose cause is genuinely unclear
 after one look, apply `dialectic` rather than guessing twice.
 Decide explicitly which side is wrong, and say so in one line before editing:
 - **Code is wrong**: the test states the intended behaviour. Fix the code.
 - **Test is wrong**: the test encodes an outdated or incorrect expectation.
  Fixing it is allowed *only* with a stated reason for why the old expectation
  was wrong - never because the code disagrees with it.
 - **Environment is wrong**: missing dependency, stale container, absent env var,
  unbuilt artefact. Fix the environment or report it; do not patch code around
  a broken environment.
 ### Fix
 Smallest change that addresses the root cause. Follow the project's existing
 idioms and the `no-defaults` and `comments-clean` rules.
 Forbidden, in every iteration, without explicit operator approval:
 - deleting, skipping, `xfail`-ing, or commenting out a failing test
 - loosening an assertion to whatever the code currently produces
 - catching or swallowing the exception the test was written to surface
 - adding retries, sleeps, or reruns to paper over a flaky failure
 - disabling a linter rule, a type check, or a test target from the selection
 Any of these is a report item, not a fix. A suite that is green because the
 failing test no longer runs is a regression disguised as success.
 ### Re-run
 Re-run the same target. Then, once it passes, re-run every target that was
 already green - a fix that breaks a neighbouring suite is not a fix. Only
 after the full selection is green is the loop done.
 ## Stopping
 Stop and report, without asking for permission to stop:
 - **Green**: full selection passes. Report and stop.
 - **No progress**: the same failure survives 3 iterations, or the failure count
  stops falling across 3 iterations. Report the root cause reached so far, what
  was tried, and why each attempt failed.
 - **Fix exceeds the mandate**: the real fix is an API change, a dependency bump,
  a schema migration, or a redesign. Name it, show the minimal diff it would
  need, and let the operator decide.
 - **Oscillation**: fixing A re-breaks B and vice versa. Report both with the
  conflict between them - that is a design problem, not a test problem.
 - **Flaky**: a target passes and fails without any change between runs. Report
  it as flaky with both outputs; never "fix" it by rerunning until green.
 Never loop silently. Emit one line per iteration: target, failure, hypothesis,
 change made.
 ## Report
 - Per iteration: what failed, the root cause, the fix, the result.
 - Final state of every selected target, pass or fail, with the verbatim output
  of anything still failing.
 - Everything deliberately not done: tests judged wrong but left alone,
  environment issues, out-of-mandate fixes.
 - The exact command reproducing the final state.
 State calibrated confidence in the fixes per the `confidence` skill. Do not
 commit unless the operator asked for a commit.
@@ -1,10 +1,9 @@
 ---
 name: test
 description: >
-  Discover the test targets of the project's Makefile, let the operator pick
+  Discover the Makefile's test targets, let the operator pick, run and report;
-  which ones to run, then run them and report. Trigger on /test or when the
+  fix mode loops until green. Trigger on /test, /test --fix, or when the
-  operator asks to run the tests, the test suite, or a specific test target
+  operator asks to run the tests, make them pass, or get the suite green.
  without naming the exact command. Portable across projects.
 ---
 The Makefile is the source of truth for how this project runs tests. Never
@@ -73,5 +72,59 @@ Per target: pass/fail plus the failing test names and the exact error output,
 quoted verbatim. Never paraphrase a failure. Then one line with the exact
 command to reproduce the failure alone.
-Do not fix what failed unless the operator asks - report first. If the operator
+Do not fix what failed unless the operator asks - report first. A request to
-does ask for a fix, apply the `triage` skill to it.
+make the tests pass, fix them, or get the suite green (`/test --fix`) switches
 to fix mode.
 ## Fix mode
 The contract is narrow: the code becomes correct, the tests stay honest.
 One iteration is: run -> read the failure -> one root cause -> one fix ->
 re-run. Never batch several speculative fixes into one iteration; a green run
 after three simultaneous changes proves nothing about which one mattered.
 - **Run** the failing target alone, the fastest command that reproduces the
  failure. Widen back to the full selection only when that target is green.
 - **Diagnose** from the actual error output before touching a file. Apply the
  `triage` skill to reach a verified root cause; for a cause still unclear
  after one look, apply `dialectic` rather than guessing twice. Say in one line
  which side is wrong:
  - **Code**: the test states the intended behaviour. Fix the code.
  - **Test**: it encodes an outdated or incorrect expectation. Fix it only with
    a stated reason why the old expectation was wrong, never because the code
    disagrees with it.
  - **Environment**: missing dependency, stale container, absent env var,
    unbuilt artefact. Fix the environment or report it; do not patch code
    around it.
 - **Fix** with the smallest change that addresses the root cause, in the
  project's idioms and under the `no-defaults` and `comments-clean` rules.
  Forbidden without explicit operator approval: deleting, skipping, `xfail`-ing
  or commenting out a failing test; loosening an assertion to what the code
  produces; swallowing the exception the test surfaces; retries, sleeps or
  reruns over a flaky failure; disabling a linter rule, a type check or a
  target. Any of these is a report item, not a fix: a suite that is green
  because the failing test no longer runs is a regression disguised as success.
 - **Re-run** the same target, then every target that was already green; only
  a full green selection ends the loop.
 Stop and report, without asking for permission to stop:
 - **Green**: the full selection passes.
 - **No progress**: the same failure survives 3 iterations, or the failure count
  stops falling across 3 iterations. Report the root cause reached, what was
  tried, and why each attempt failed.
 - **Beyond the mandate**: the real fix is an API change, a dependency bump, a
  schema migration, or a redesign. Name it with its minimal diff and let the
  operator decide.
 - **Oscillation**: fixing A re-breaks B and vice versa. Report both; that is a
  design problem, not a test problem.
 - **Flaky**: a target passes and fails without any change. Report both
  outputs; never rerun until green.
 Emit one line per iteration: target, failure, hypothesis, change. The final
 report gives per iteration what failed, the root cause, the fix and the
 result; the final state of every selected target with the verbatim output of
 anything still failing; everything deliberately not done; and the command
 reproducing the final state. State confidence per the `confidence` skill. Do
 not commit unless the operator asked for a commit.
@@ -1,10 +1,9 @@
 ---
 name: triage
 description: >
-  Triage a failing CI run, pipeline, or job set to a verified root cause and a
+  Triage a failing CI run, pipeline or job set to a verified root cause and a
-  real fix. Trigger when a build, test, or deploy pipeline has one or more failed
+  real fix, one cluster per worktree. Trigger when a build, test or deploy
-  jobs. Applies the dialectic skill to every failing job independently. Portable
+  pipeline has failed jobs.
  across projects.
 ---
 Drive a red pipeline to green by root cause, not by guesswork. Treat every
@@ -13,6 +12,23 @@ root cause is proven.
 ## Procedure
 For a GitHub Actions run, the ci-artifacts skill's `ci_failures.py <run>`
 does the mechanics of steps 0 to 2 in one call: distance to HEAD, the failed
 jobs' logs, and a first clustering. The judgement in those steps stays yours.
 0. **Check the run against HEAD before anything else.** A CI run is a snapshot
   of one commit, and the branch has moved since. Run
   `git log --oneline <run-sha>..HEAD | wc -l` as the first command of the
   triage. Zero means the run speaks for the current tree and you proceed.
   Non-zero means it does not, and you scope the job set down **before**
   enumerating, clustering, or spawning anything — every job you investigate
   against stale code is work thrown away, and a fix proposed there risks
   reverting someone else's. Two cheap discriminators do it without opening an
   artifact: `git diff --name-only <run-sha>..HEAD -- <source dir> | cut -d/ -f2 | sort -u`
   for which components already moved, and a grep of the saved job logs for the
   literal error signature of each cause already fixed in that range. What
   survives both is the job set the rest of this procedure applies to; say how
   many jobs you dropped and why.
 1. **Enumerate every failing job.** Pull the run's job list and isolate each job
   whose conclusion is failure, cancelled, or timed-out. A downstream aggregate
   job that fails only because an upstream job did is not a separate root cause —
@@ -21,7 +37,14 @@ root cause is proven.
   actual failing line, not by job name. Jobs whose messages differ in anything
   but the interpolated identifiers (host, app, port, path) belong in separate
   clusters.
-3. **Download artifacts per cluster.** For each cluster, fetch every artifact of
+3. **Coordinate every cluster.** Run the `democracy` skill's Lineage and Peer
   protocol for each cluster before fixing it: claim it, ask the lineage
   whether someone already fixes it, and take their fix or keep yours. Keep
   investigating while you wait for answers.
 4. **Download artifacts per cluster.** Use the `ci-artifacts` skill for the
   mechanics — it covers per-job logs while a run is still in progress, which
   bundle to pull, and the grep hygiene that keeps a false positive from
   becoming a root cause. For each cluster, fetch every artifact of
   one representative job: reports, rescue diagnostics, inventories, container
   logs. Pull a second member's artifacts only when the representative's
   evidence does not explain the whole cluster. The job log usually names an
@@ -29,21 +52,39 @@ root cause is proven.
   exception and the state dumps are inside the artifact. Do this even when the
   log looks conclusive, because a log that explains the symptom rarely
   explains the cause.
-4. **Dialectic per cluster.** For EACH cluster, invoke the `dialectic` skill:
+5. **One funnel per cluster.** Hand every cluster you claimed in step 3 to the
   `funnel` skill: one agent per cluster, working in its own worktree on its own
   branch created from the exact current HEAD, with that cluster as its only
   concern, the artifact paths of step 4 in its brief and a hard time cap.
   Steps 6 to 8 run inside that sidetree, so no cluster's edits land under
   another cluster's work or under a running gate or deploy. Run clusters in
   parallel only when their fixes touch different files and the machine has
   room. A split cluster (step 7) gets a funnel of its own.
 6. **Dialectic per cluster.** For EACH cluster, invoke the `dialectic` skill:
   form a thesis about the root cause from evidence (the job log, its artifacts,
   the code at the run's commit, git history), attack it with independent
-   skeptics, and iterate to a ~99% thesis. The clusters are independent, so run
+   skeptics, and iterate to a ~99% thesis.
-   their investigations in parallel where the tooling allows.
+7. **Distinguish shared vs hidden causes.** A cluster is a hypothesis, not a
 5. **Distinguish shared vs hidden causes.** A cluster is a hypothesis, not a
   proof: if the dialectic shows one cluster splitting into two causes, split it.
   When a job hides a second failure behind the first, keep going until the job
   is actually green, not just past the first error.
-6. **Fix at the root.** Apply the real fix in the repository for each proven root
+8. **Fix at the root, through the funnel.** Apply the real fix for each proven
-   cause. Never mask a failure — no retry-until-pass, no disabling the check, no
+   root cause in the cluster's sidetree, with a test and a negative control as
-   soft-skip. If a failure is genuinely external (upstream outage, flaky infra),
+   the `funnel` skill demands. The fix enters the main line only at the funnel's
-   confirm that with evidence and surface it honestly instead of fixing around
+   bar (at least 90 % fix confidence, under 5 % break risk) after its
-   it.
+   independent review; below the bar the cluster ends as a proposal for the
-7. **Follow the run.** While the run is still in progress, re-check it
+   operator, never as a merge. Never mask a failure: no retry-until-pass, no
   disabling the check, no soft-skip. If a failure is genuinely external
   (upstream outage, flaky infra), confirm that with evidence and surface it
   honestly instead of fixing around it. Report every fix, every proposal, and
   every abandoned attempt to the same agents through the `democracy` Peer protocol,
   then remove the sidetree.
 9. **Tune what ran out of time.** Only for a cluster whose jobs died on their
   time budget (timed out, or cancelled by the runner's limit), or when the
   operator asks: hand the job log that ran furthest to the `tune` skill. Its
   candidates run through funnels of their own, separate from the cluster's fix
   sidetree, and never delay or block the fix.
 10. **Follow the run.** While the run is still in progress, re-check it
   periodically and triage each newly failed job as it appears. The run is done
   only when it has finished and every failure has a verified fix.
@@ -51,7 +92,20 @@ root cause is proven.
 - Executed evidence beats reasoning: read the actual failing log line and the
  artifact, not the job title.
 - A derived source is not evidence. A conversation summary, your own earlier
  message, a cluster table you built ten minutes ago, or a subagent's report
  proves nothing on its own. Only the job log, the artifact, the code at the
  run's commit, and command output you just ran count. Quote them literally; if
  you cannot quote it, you do not know it.
 - Re-derive after a context compaction. A summary carried across a compaction
  boundary preserves your conclusions but not the evidence under them, and a
  cluster assignment that was wrong stays wrong while reading as settled fact.
  Before acting on any pre-compaction attribution, rebuild it from the job logs.
 - Classify each failure: a real defect (fix it), a flake or infra hiccup (prove
-  it is transient before dismissing), or already-fixed-by-a-later-commit (verify
+  it is transient before dismissing), or already-fixed-by-a-later-commit. For
-  the fix is present in the run's commit before claiming it).
+  that last class the comparison is `<run-sha>..HEAD`, never the run's commit
- Report a per-job verdict: root cause, the evidence that proves it, and the fix.
+  alone: the run cannot contain a fix that landed after it, so looking there
  can only ever confirm the failure. Name the commit that carries the fix.
 - Report a per-job verdict: root cause, the artifact path and line that prove
  it, the fix, the cluster's sidetree branch, and the funnel verdict (merged or
  proposal). A verdict that cites only the job log has not finished step 4.
@@ -0,0 +1,192 @@
 ---
 name: tune
 description: >
  Find and apply performance optimizations (CI, deploy, build, gate, hot
  paths) without reducing logic, coverage or correctness. Trigger on /tune,
  when a job dies on its time budget, when a pipeline or deploy grows slower,
  or asked where time goes.
 ---
 # Tune
 One rule outranks every optimization: **nothing that reduces logic, coverage or
 correctness counts as a speed-up.**
 ## Once per session: listen first
 The first time this skill runs in a session, invoke the `active-listening`
 skill before measuring anything, and ask three questions through the host's
 structured question mechanism. Offer each as a checklist (multi-select) with
 suggestions, and allow free text:
 1. **What to examine.** Suggestions: CI wall-clock, deploy duration, gate and
   test runtime, image pulls and build caches, per-item loops and process
   spawns, idempotency (work repeated on a second run), network round trips,
   resource use (CPU, memory, disk, I/O), and language or runtime choice for
   hot paths.
 2. **Which performance boundaries to keep.** Suggestions: time budgets per job,
   deploy or gate; resource caps (memory, CPU, disk headroom); a minimum saving
   per change (default 10 s per run); the measurement source that counts as
   the verdict.
 3. **Which limits apply to optimizations.** Suggestions: the scope that may be
   changed (files, roles, services); whether new dependencies, toolchains or
   languages may be introduced; whether infrastructure or CI configuration may
   change; the merge bar (default: the `funnel` skill's 90 % confidence and
   under 5 % break risk); a time cap per delegated agent (default 60 minutes).
 Keep the answers for the rest of the session and do not ask again. Work only
 inside them; a candidate outside them becomes a proposal.
 When another skill (triage, observe) or an autonomous run invoked tune, skip
 the questions: examine the handed-over log, apply the defaults above, and ask
 only when an operator-only fact blocks.
 ## Forbidden, whatever it saves
 - Dropping a test, a spec, a play, a drill or a verification pass.
 - Raising a timeout or budget so a slow thing fits instead of getting faster.
 - Swallowed failures, retry-until-pass, a soft skip, a narrowed assertion.
 - Trading error semantics, credential masking, concurrency safety or a lint's
  intent for speed. A per-item failure stays loud and names its item.
 A change that makes the clock smaller by testing less is reported to the
 operator as a coverage cut for them to decide, never applied as a tuning.
 ## Measure first
 Never tune what has not been measured, and read numbers from logs, not by eye.
 - **Task profilers** (Ansible `profile_tasks` and similar) print the duration
  of the *previous* task under the *next* task's header. Pair each header with
  the following timestamp line, or read the end-of-run recap.
 - **Build logs** (BuildKit): pair `#N [service x/y] RUN …` with `#N DONE Xs`
  by step number; `CACHED` marks a hit.
 - **Gates**: per-target wall-clock tables and failed-target lines. An exit code
  piped through `tee` is not a verdict.
 - **Code paths**: a profiler (py-spy, cProfile, perf, flame graphs) tells
  whether time goes to CPU, to I/O and waits, or to process start-up.
 - Stream every run to a log file and compare runs by their logs; never re-run
  to recover a lost number.
 - **Rank by fixed cost x count.** Measure the per-call fixed cost (one
  application console boot, one package-manager call, one container exec) and
  multiply by how often the run pays it. A one-second boot paid twenty-eight
  times is a candidate; a sixty-second download paid once is not.
 - **The baseline is the newest log of the current code.** An older log may
  still carry the step a previous optimization replaced, and a saving measured
  against it is fiction.
 ## Invalid comparisons
 Discard a before/after pair, not just its noise, when:
 - Another gate, deploy or heavy job shared the machine during either run.
 - The two runs differ in image or cache state. A cold build after a cache prune
  or a Dockerfile edit is expected: before calling a cold build a defect,
  compare the instruction text and the base image digest across both logs.
 - One run pays a one-time cost (first install, migration, purge, model
  download) the other does not.
 - A long wait hides retries or a failure; count retry lines before calling it
  slow.
 - A fresh-install pass is compared with an update pass. Analyse each pass of a
  multi-pass run separately.
 ## Before you optimize
 - **Search every branch** (the merge-fix skill's `find_fix.py -S <symbol> --path <path>`) for a change
  that already removes or reshapes the step. A fix on a sibling branch can make
  the whole candidate worth zero.
 - A step that is long by design (a backup and restore drill, an end-to-end
  suite) is not a candidate. Its cost buys its guarantee.
 - Apply the session's saving floor: a smaller saving is listed as "below
  floor" and dropped.
 - Pruning build caches to free disk costs a cold rebuild of every custom image
  on the next run. Name that cost before anyone prunes.
 ## Coordinate every cost cluster
 A cost cluster is one place the time goes: one step, one hot path, one
 repeated call, named by its log line and its number. Before investigating or
 implementing a cluster, run the `democracy` skill's Peer protocol for it with
 every agent session on this machine, not only the branch lineage.
 - **The first to identify a cluster works it.** "First" is the `democracy`
  claim `cluster/<slug>`, whose `--work` names the cluster's log line and
  number; between machines, the earliest announcement by message time. A
  claim without a measurement does not count.
 - **Someone was earlier:** send them your measurements and findings, then move
  on to the next cluster. Never build a second optimization of the same
  cluster.
 - **A claim ends** when its holder reports the result, a proposal or an
  abandoned attempt and releases it, or when the operator reassigns it.
 ## Levers that usually pay
 - **Batch per-item process spawns**: N calls that each boot an interpreter or
  an application console become one process running N commands (one app
  console boot, one module, one batched API request). Keep each item's failure
  named and the exit code honest.
 - **Skip no-op work** a previous step already did, such as a separate enable
  after an install that already enabled.
 - **Idempotency flip-flops**: any `changed` on a second consecutive run.
 - **Image pulls and builds**: pull only what the host lacks; keep build caches
  warm.
 - **Overlap waits with work** only where concurrent execution is proven safe.
  Schema migrations and lock-holding steps are usually not.
 ## Language and runtime
 Look for language optimizations where they make sense, for example replacing a
 Python hot path with Rust:
 - **Only for measured CPU-bound work.** When the profile shows the time in
  computation inside the process (parsing, hashing, rendering, tight loops),
  a compiled language can win by an order of magnitude. When the time goes to
  process start-up, network round trips or waiting, batching in the current
  language wins first, and a rewrite changes nothing.
 - **Prefer the narrowest cut**: rewrite the hot function as an extension
  module (for Python and Rust: PyO3 and maturin) or a small binary behind the
  same interface, before rewriting a whole tool.
 - **Count the total cost**: a toolchain in CI, builds per target architecture,
  shipping binaries to every host, the team's ability to maintain the code,
  and test parity with the old implementation.
 - **Prove it on a prototype**: the same input, the same machine state, a before
  and an after number, and the old test suite passing against the new code.
 - A language change is applied only when the session's limits allow new
  languages or toolchains; otherwise it is submitted as a proposal with the
  prototype's numbers.
 ## Delegating an optimization
 Hand every optimization to an agent through the `funnel` skill: one sidetree
 per candidate, branched from an exact commit, with a hard time cap and "measure
 first, report the time split before implementing". The agent probes live
 systems read-only and never starts gates or deploys. A broad brief without a
 cap can run for hours and deliver nothing; stop an agent at its cap and collect
 its findings.
 Before merging, run independent skeptics through the `dialectic` skill. The
 antithesis must try to show that the candidate removes coverage, hides a
 failure, or only moves the cost somewhere the measurement does not look.
 ## The loop
 1. Measure, and name the single biggest cost with its log line and number.
 2. Coordinate the cluster with the other agents; work it only when you
   identified it first, otherwise hand over your findings and take the next.
   Then form one candidate for it, after the branch search and the floor check.
 3. Harden it with the `dialectic` skill.
 4. Funnel it through a sidetree; merge what reaches the bar, submit the rest as
   proposals. Re-measure from the same source under the same conditions and
   report a before and an after number.
 5. Repeat until the next candidate is worth less than the run that measures it.
 ## Output
 One table, biggest saving first:
 | Candidate | Before | After | Source | Verdict |
 Rejected candidates stay in the table with the reason they fell: below floor,
 coverage cut, outside the session's limits, invalid comparison, made moot by
 another branch, or cap reached without a measured result. Proposals carry what
 would close their gap.
@@ -37,6 +37,27 @@ class TestAutotuneSkill(unittest.TestCase):
        self.assertIn("multiSelect: false", text)
 class TestSkillPersistence(unittest.TestCase):
    """Both skill writers must persist to the skills repository AND the live dirs."""
    def test_autotune_names_every_target(self):
        text = SKILL.read_text(encoding="utf-8")
        for target in (
            "SKILLS_REPO",
            "skills/<name>/SKILL.md",
            "~/.claude/skills/",
            "~/.agents/skills/",
        ):
            self.assertIn(target, text)
    def test_autoskill_routes_to_the_same_persistence(self):
        text = (REPO_ROOT / "skills" / "autoskill" / "SKILL.md").read_text(
            encoding="utf-8"
        )
        for target in ("Persistence", "~/.claude/skills/", "~/.agents/skills/"):
            self.assertIn(target, text)
 class TestAutotuneHook(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
@@ -70,7 +91,9 @@ class TestSettingsRegistration(unittest.TestCase):
        self.settings = Path(self.tmp.name) / "settings.json"
    def _patch(self) -> dict:
-        subprocess.run(["node", str(PATCHER), str(self.settings)], check=True, capture_output=True)
+        subprocess.run(
            ["node", str(PATCHER), str(self.settings)], check=True, capture_output=True
        )
        return json.loads(self.settings.read_text(encoding="utf-8"))
    def test_hook_registered_once(self):
@@ -81,11 +104,19 @@ class TestSettingsRegistration(unittest.TestCase):
    def test_existing_hooks_preserved(self):
        self.settings.write_text(
-            json.dumps({"hooks": {"UserPromptSubmit": [{"hooks": [{"type": "command", "command": "true"}]}]}}),
+            json.dumps(
                {
                    "hooks": {
                        "UserPromptSubmit": [
                            {"hooks": [{"type": "command", "command": "true"}]}
                        ]
                    }
                }
            ),
            encoding="utf-8",
        )
        data = self._patch()
-        self.assertEqual(len(data["hooks"]["UserPromptSubmit"]), 2)
+        self.assertEqual(len(data["hooks"]["UserPromptSubmit"]), 4)
 if __name__ == "__main__":
@@ -0,0 +1,96 @@
 """Exercise ci_failures clustering and find_fix's ref search."""
 from __future__ import annotations
 import importlib.util
 import shutil
 import subprocess
 import sys
 import tempfile
 import unittest
 from pathlib import Path
 REPO_ROOT = Path(__file__).resolve().parents[1]
 CI_FAILURES = REPO_ROOT / "skills" / "ci-artifacts" / "scripts" / "ci_failures.py"
 FIND_FIX = REPO_ROOT / "skills" / "merge-fix" / "scripts" / "find_fix.py"
 def _load(path: Path):
    spec = importlib.util.spec_from_file_location(path.stem, path)
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module
 class TestCiFailures(unittest.TestCase):
    def setUp(self):
        self.ci = _load(CI_FAILURES)
    def test_jobs_differing_only_in_identifiers_share_a_cluster(self):
        one = "2026-10-30T10:00:00.1Z setup ok\n##[error]Timeout on host 10.0.0.12 after 300s\n"
        two = "\x1b[31m2026-10-30T11:00:00.1Z ##[error]Timeout on host 10.0.0.47 after 299s\x1b[0m\n"
        other = "PLAY RECAP failed=0\nfatal: unable to reach registry\n"
        groups = self.ci.cluster([("a", one), ("b", two), ("c", other)])
        self.assertEqual(list(groups.values()), [["a", "b"], ["c"]])
    def test_command_echo_and_retries_are_skipped_and_task_named(self):
        text = (
            '##[group]Run make deploy || echo "failed"\n'
            "make deploy\n"
            "##[endgroup]\n"
            "FAILED - RETRYING: wait (3 retries left)\n"
            "TASK [web-app-y : Probe] ****\n"
            "Traceback (most recent call last): recovered on retry\n"
            "TASK [web-app-x : Wait for health] ****\n"
            'fatal: [localhost]: FAILED! => {"msg": "timeout 120"}\n'
            "##[error]Process completed with exit code 2.\n"
        )
        self.assertEqual(
            self.ci.signature(text),
            'TASK [web-app-x : Wait for health] **** fatal: [localhost]: FAILED! => {"msg": "timeout <n>"}',
        )
    def test_generic_exit_code_line_is_not_the_signature(self):
        text = "Error: disk full\n##[error]Process completed with exit code 1.\n"
        self.assertEqual(self.ci.signature(text), "Error: disk full")
@unittest.skipUnless(shutil.which("git"), "git not installed")
 class TestFindFix(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
        self.addCleanup(self.tmp.cleanup)
        self.repo = Path(self.tmp.name)
        self.env = {
            "PATH": "/usr/bin:/bin",
            "HOME": str(self.repo),
            "GIT_AUTHOR_NAME": "t",
            "GIT_AUTHOR_EMAIL": "t@t",
            "GIT_COMMITTER_NAME": "t",
            "GIT_COMMITTER_EMAIL": "t@t",
        }
        self._git("init", "-q", "-b", "main")
        self._git("commit", "-q", "--allow-empty", "-m", "fix: probe in head")
        self._git("switch", "-q", "-c", "side")
        self._git("commit", "-q", "--allow-empty", "-m", "fix: probe on side")
        self._git("switch", "-q", "main")
    def _git(self, *args: str) -> None:
        subprocess.run(["git", *args], check=True, capture_output=True, env=self.env, cwd=self.repo)
    def test_only_commits_outside_head_are_candidates(self):
        result = subprocess.run(
            [sys.executable, str(FIND_FIX), "--grep", "probe"],
            capture_output=True,
            text=True,
            check=True,
            env=self.env,
            cwd=self.repo,
        )
        self.assertIn("hits=2 in_head=1 candidates=1", result.stdout)
        self.assertIn("fix: probe on side", result.stdout)
        self.assertNotIn("probe in head", result.stdout)
 if __name__ == "__main__":
    unittest.main()
@@ -0,0 +1,88 @@
 """Exercise the democracy council script: election, staleness, release, gate."""
 from __future__ import annotations
 import json
 import subprocess
 import sys
 import tempfile
 import unittest
 from pathlib import Path
 REPO_ROOT = Path(__file__).resolve().parents[1]
 COUNCIL = REPO_ROOT / "skills" / "democracy" / "scripts" / "council.py"
 class TestCouncil(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
        self.addCleanup(self.tmp.cleanup)
        self.home = Path(self.tmp.name) / "home"
        self.root = Path(self.tmp.name) / "council"
        self.sessions = self.home / ".claude" / "sessions"
        self.sessions.mkdir(parents=True)
        self._live("s1", "One")
        self._live("s2", "Two")
    def _live(self, sid: str, name: str) -> None:
        (self.sessions / f"{sid}.json").write_text(
            json.dumps({"sessionId": sid, "name": name}), encoding="utf-8"
        )
    def _run(self, sid: str, *args: str, stdin: str = "") -> subprocess.CompletedProcess:
        env = {
            "PATH": "/usr/bin:/bin",
            "HOME": str(self.home),
            "AGENTS_COUNCIL_ROOT": str(self.root),
            "CLAUDE_CODE_SESSION_ID": sid,
        }
        return subprocess.run(
            [sys.executable, str(COUNCIL), *args],
            capture_output=True,
            text=True,
            env=env,
            input=stdin,
            check=False,
        )
    def test_first_claim_wins_and_second_is_refused(self):
        self.assertEqual(self._run("s1", "claim", "work/x", "--work", "w").returncode, 0)
        second = self._run("s2", "claim", "work/x", "--work", "w")
        self.assertEqual(second.returncode, 3)
        self.assertIn('"session": "s1"', second.stdout)
    def test_dead_holder_is_stale_and_can_be_taken_over(self):
        self._run("s1", "claim", "work/x", "--work", "w")
        (self.sessions / "s1.json").unlink()
        self.assertEqual(self._run("s2", "claim", "work/x", "--work", "w").returncode, 4)
        self.assertEqual(self._run("s2", "takeover", "work/x", "--work", "w").returncode, 0)
        self.assertIn('"session": "s2"', (self.root / "work/x/claim.json").read_text())
    def test_takeover_of_a_live_holder_is_refused(self):
        self._run("s1", "claim", "work/x", "--work", "w")
        self.assertEqual(self._run("s2", "takeover", "work/x", "--work", "w").returncode, 3)
    def test_only_the_holder_releases(self):
        self._run("s1", "claim", "work/x", "--work", "w")
        self.assertEqual(self._run("s2", "release", "work/x").returncode, 3)
        self.assertEqual(self._run("s1", "release", "work/x").returncode, 0)
        self.assertFalse((self.root / "work/x").exists())
    def test_gate_blocks_unclaimed_edits_in_a_git_worktree_only(self):
        repo = Path(self.tmp.name) / "repo"
        (repo / ".git").mkdir(parents=True)
        inside = json.dumps({"session_id": "s1", "tool_input": {"file_path": str(repo / "f")}})
        outside = json.dumps(
            {"session_id": "s1", "tool_input": {"file_path": str(self.home / "f")}}
        )
        self.assertEqual(self._run("s1", "gate", stdin=inside).returncode, 2)
        self.assertEqual(self._run("s1", "gate", stdin=outside).returncode, 0)
        self._run("s1", "claim", "work/x", "--work", "w")
        self.assertEqual(self._run("s1", "gate", stdin=inside).returncode, 0)
    def test_invalid_office_is_a_usage_error(self):
        self.assertEqual(self._run("s1", "claim", "../x", "--work", "w").returncode, 2)
 if __name__ == "__main__":
    unittest.main()
@@ -0,0 +1,138 @@
 """Validate the dream reminder hook and its settings registration."""
 from __future__ import annotations
 import json
 import shutil
 import subprocess
 import tempfile
 import unittest
 from pathlib import Path
 REPO_ROOT = Path(__file__).resolve().parents[1]
 SKILL = REPO_ROOT / "skills" / "dream" / "SKILL.md"
 HOOK = REPO_ROOT / "skills" / "dream" / "hooks" / "reminder.sh"
 PATCHER = REPO_ROOT / "scripts" / "enable-plugins.js"
 def _run_hook(
    cwd: Path, stamp: Path | None = None, interval: str = "43200"
 ) -> subprocess.CompletedProcess:
    env = {
        "PATH": "/usr/bin:/bin",
        "HOME": str(cwd),
        "DREAM_INTERVAL_SECONDS": interval,
    }
    if stamp is not None:
        env["DREAM_STAMP"] = str(stamp)
    return subprocess.run(
        ["bash", str(HOOK)],
        capture_output=True,
        text=True,
        check=True,
        cwd=str(cwd),
        env=env,
    )
 class TestDreamSkill(unittest.TestCase):
    def test_skill_is_trigger_only(self):
        text = SKILL.read_text(encoding="utf-8")
        self.assertIn("name: dream", text)
        self.assertIn("multiSelect: false", text)
 class TestDreamHook(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
        self.addCleanup(self.tmp.cleanup)
        self.cwd = Path(self.tmp.name)
        self.stamp = self.cwd / ".claude" / ".dream-stamp"
    def test_first_run_hints_and_stamps(self):
        result = _run_hook(self.cwd, self.stamp)
        self.assertIn("/dream", result.stdout)
        self.assertTrue(self.stamp.is_file())
    def test_second_run_is_silent_within_interval(self):
        _run_hook(self.cwd, self.stamp)
        self.assertEqual(_run_hook(self.cwd, self.stamp).stdout, "")
    def test_hint_returns_after_the_interval(self):
        _run_hook(self.cwd, self.stamp)
        self.assertIn("/dream", _run_hook(self.cwd, self.stamp, interval="0").stdout)
    def test_corrupt_stamp_does_not_crash(self):
        self.stamp.parent.mkdir(parents=True, exist_ok=True)
        self.stamp.write_text("not-a-timestamp\n", encoding="utf-8")
        self.assertIn("/dream", _run_hook(self.cwd, self.stamp).stdout)
    def test_outside_a_repository_it_stays_silent(self):
        self.assertEqual(_run_hook(self.cwd).stdout, "")
    @unittest.skipUnless(shutil.which("git"), "git not installed")
    def test_stamp_is_per_repository(self):
        for name in ("one", "two"):
            repo = self.cwd / name
            repo.mkdir()
            subprocess.run(
                ["git", "init", "-q", str(repo)], check=True, capture_output=True
            )
            self.assertIn("/dream", _run_hook(repo).stdout)
            self.assertTrue((repo / ".git" / ".dream-stamp").is_file())
            self.assertEqual(_run_hook(repo).stdout, "")
    @unittest.skipUnless(shutil.which("git"), "git not installed")
    def test_linked_worktrees_share_the_stamp(self):
        repo = self.cwd / "main"
        tree = self.cwd / "tree"
        env = {
            "PATH": "/usr/bin:/bin",
            "HOME": str(self.cwd),
            "GIT_AUTHOR_NAME": "t",
            "GIT_AUTHOR_EMAIL": "t@t",
            "GIT_COMMITTER_NAME": "t",
            "GIT_COMMITTER_EMAIL": "t@t",
        }
        repo.mkdir()
        for args in (
            ["init", "-q"],
            ["commit", "-q", "--allow-empty", "-m", "init"],
            ["worktree", "add", "-q", str(tree)],
        ):
            subprocess.run(
                ["git", *args], check=True, capture_output=True, env=env, cwd=str(repo)
            )
        self.assertIn("/dream", _run_hook(repo).stdout)
        self.assertEqual(_run_hook(tree).stdout, "")
@unittest.skipUnless(shutil.which("node"), "node not installed")
 class TestSettingsRegistration(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
        self.addCleanup(self.tmp.cleanup)
        self.settings = Path(self.tmp.name) / "settings.json"
    def _patch(self) -> dict:
        subprocess.run(
            ["node", str(PATCHER), str(self.settings)], check=True, capture_output=True
        )
        return json.loads(self.settings.read_text(encoding="utf-8"))
    def test_hook_registered_once(self):
        self._patch()
        data = self._patch()
        entries = json.dumps(data["hooks"]["UserPromptSubmit"])
        self.assertEqual(entries.count("dream/hooks/reminder.sh"), 1)
        self.assertEqual(entries.count("autotune/hooks/reminder.sh"), 1)
        gate = json.dumps(data["hooks"]["PreToolUse"])
        self.assertEqual(gate.count("democracy/scripts/council.py"), 1)
        self.assertEqual(gate.count("democracy/scripts/watch.py"), 1)
        self.assertEqual(entries.count("democracy/scripts/watch.py"), 1)
        post = json.dumps(data["hooks"]["PostToolUse"])
        self.assertEqual(post.count("democracy/scripts/watch.py"), 1)
 if __name__ == "__main__":
    unittest.main()
@@ -12,6 +12,8 @@ LOCK_FILE = REPO_ROOT / "skills-lock.json"
 HASH_RE = re.compile(r"^[0-9a-f]{64}$")
 REQUIRED_FIELDS = ("source", "sourceType", "skillPath", "computedHash")
 DESCRIPTION_WORDS = 40
 LISTING_WORDS = 1300
 class TestSkillsLock(unittest.TestCase):
@@ -61,6 +63,16 @@ class TestSkillsLock(unittest.TestCase):
                    "description:", frontmatter, "frontmatter missing description"
                )
    def test_descriptions_fit_the_listing_budget(self):
        total = 0
        for skill_md in sorted((REPO_ROOT / "skills").glob("*/SKILL.md")):
            frontmatter = skill_md.read_text(encoding="utf-8").split("---", 2)[1]
            words = len(frontmatter.split("description:", 1)[1].split()) - 1
            total += words
            with self.subTest(skill=skill_md.parent.name):
                self.assertLessEqual(words, DESCRIPTION_WORDS)
        self.assertLessEqual(total, LISTING_WORDS)
    def test_first_party_skills_do_not_shadow_locked_skills(self):
        locked = set(self.lock["skills"].keys())
        for skill_dir in sorted((REPO_ROOT / "skills").glob("*/")):
@@ -0,0 +1,31 @@
 """The self-improving skills must share the objective held by autotune."""
 from __future__ import annotations
 import unittest
 from pathlib import Path
 REPO_ROOT = Path(__file__).resolve().parents[1]
 SKILLS = REPO_ROOT / "skills"
 ROUTERS = ("autoskill", "confidence")
 class TestOptimizeObjective(unittest.TestCase):
    def test_autotune_names_the_three_axes(self):
        text = (SKILLS / "autotune" / "SKILL.md").read_text(encoding="utf-8")
        self.assertIn("## Objective", text)
        for axis in ("**Tokens**", "**Time**", "**Quality**"):
            self.assertIn(axis, text)
    def test_routers_point_at_it(self):
        for skill in ROUTERS:
            text = (SKILLS / skill / "SKILL.md").read_text(encoding="utf-8")
            self.assertIn(
                "`autotune` skill's **Objective**",
                text,
                f"{skill} does not route to the autotune objective",
            )
 if __name__ == "__main__":
    unittest.main()
@@ -0,0 +1,151 @@
 """Exercise the democracy resource watch: stages, projection, block, admit, alerts, caches."""
 from __future__ import annotations
 import importlib.util
 import json
 import os
 import subprocess
 import sys
 import tempfile
 import time
 import unittest
 from pathlib import Path
 REPO_ROOT = Path(__file__).resolve().parents[1]
 WATCH = REPO_ROOT / "skills" / "democracy" / "scripts" / "watch.py"
 LIMITS = {"warn": 40, "stop": 15, "critical": 12}
 def _load():
    spec = importlib.util.spec_from_file_location("watch", WATCH)
    module = importlib.util.module_from_spec(spec)
    spec.loader.exec_module(module)
    return module
 class TestStage(unittest.TestCase):
    def setUp(self):
        self.watch = _load()
        self.cfg = {"thresholds_gb": LIMITS, "projection_minutes": 30}
    def test_thresholds(self):
        now = time.time()
        for free, stage in ((50, "ok"), (30, "warn"), (14, "stop"), (5, "critical")):
            self.assertEqual(self.watch.stage_for(free, [], self.cfg, now), stage)
    def test_fast_fall_escalates_one_stage(self):
        now = time.time()
        self.assertEqual(self.watch.stage_for(32, [[now - 600, 48]], self.cfg, now), "stop")
        self.assertEqual(self.watch.stage_for(32, [[now - 600, 32.1]], self.cfg, now), "warn")
 class TestWatchCommands(unittest.TestCase):
    def setUp(self):
        self.tmp = tempfile.TemporaryDirectory()
        self.addCleanup(self.tmp.cleanup)
        self.base = Path(self.tmp.name)
        self.root = self.base / "council"
        self.config = self.base / "config.json"
        self.cache = self.base / "cache"
        self.cache.mkdir()
        self.cfg = {
            "disk": str(self.base),
            "thresholds_gb": LIMITS,
            "projection_minutes": 30,
            "interval_seconds": 0,
            "cache_interval_seconds": 86400,
            "heavy_commands": ["^make compose-deploy", "^make swarm-"],
            "never_block": ["^make swarm-down"],
            "cleanup": [],
            "worktree_cleanup": [],
            "worktrees_command": "true",
            "worktree_marker": "Makefile",
            "cache_dirs": [str(self.cache)],
            "cache_budget_percent": 3,
            "slot_need_gb": {"swarm": 30},
            "notify": ["true"],
        }
        self._write_config()
    def _write_config(self):
        self.config.write_text(json.dumps(self.cfg), encoding="utf-8")
    def _run(self, *args: str, stdin: str = "") -> subprocess.CompletedProcess:
        env = {
            "PATH": str(self.base),
            "HOME": str(self.base),
            "AGENTS_COUNCIL_ROOT": str(self.root),
            "AGENTS_COUNCIL_CONFIG": str(self.config),
        }
        return subprocess.run(
            [sys.executable, str(WATCH), *args],
            capture_output=True, text=True, env=env, input=stdin, check=False,
        )
    def _stage(self, stage: str) -> None:
        self.root.mkdir(exist_ok=True)
        (self.root / "watch.json").write_text(json.dumps({"stage": stage}), encoding="utf-8")
    def _bash(self, command: str) -> str:
        return json.dumps({"tool_input": {"command": command}})
    def test_block_only_heavy_jobs_at_stop_and_never_teardown(self):
        self._stage("stop")
        self.assertEqual(self._run("block", stdin=self._bash("make compose-deploy apps=x")).returncode, 2)
        self.assertEqual(self._run("block", stdin=self._bash("make swarm-down")).returncode, 0)
        self.assertEqual(self._run("block", stdin=self._bash("ls")).returncode, 0)
        self._stage("warn")
        self.assertEqual(self._run("block", stdin=self._bash("make compose-deploy")).returncode, 0)
    def test_admit_keeps_the_stop_floor(self):
        self.cfg["thresholds_gb"] = {"warn": 0, "stop": 0, "critical": 0}
        self._write_config()
        self.assertEqual(self._run("admit", "--need-gb", "0.001").returncode, 0)
        self.assertEqual(self._run("admit", "--need-gb", "1e9").returncode, 3)
    def test_warn_alert_reaches_each_session_once(self):
        self.cfg["thresholds_gb"] = {"warn": 1e9, "stop": -1, "critical": -2}
        self._write_config()
        def hook(session: str) -> str:
            payload = {"session_id": session, "hook_event_name": "UserPromptSubmit"}
            return self._run("watch", stdin=json.dumps(payload)).stdout
        self.assertIn("RESOURCE WARN", hook("s1"))
        self.assertEqual(hook("s1"), "")
        self.assertIn("RESOURCE WARN", hook("s2"))
    def test_cache_trim_evicts_least_recently_used_first(self):
        old, new = self.cache / "old", self.cache / "new"
        old.write_bytes(b"x" * 4096)
        new.write_bytes(b"y" * 4096)
        os.utime(old, (1_000_000, 1_000_000))
        os.utime(new, (2_000_000, 2_000_000))
        st = os.statvfs(self.base)
        self.cfg["cache_budget_percent"] = (4096 + 100) / (st.f_blocks * st.f_frsize) * 100
        self._write_config()
        self.assertEqual(self._run("cleanup", "--caches").returncode, 0)
        self.assertFalse(old.exists())
        self.assertTrue(new.exists())
    def test_full_cleanup_reaches_only_marked_running_worktrees(self):
        marked, foreign = self.base / "marked", self.base / "foreign"
        for tree in (marked, foreign):
            tree.mkdir()
        (marked / "Makefile").write_text("", encoding="utf-8")
        self.cfg["worktrees_command"] = f"echo {marked}; echo {foreign}; echo {marked}"
        self.cfg["worktree_cleanup"] = ["echo x >> {worktree}/cleaned"]
        self._write_config()
        self.assertEqual(self._run("cleanup", "--full").returncode, 0)
        self.assertEqual((marked / "cleaned").read_text(encoding="utf-8"), "x\n")
        self.assertFalse((foreign / "cleaned").exists())
    def test_without_config_every_command_is_a_no_op(self):
        self.config.unlink()
        result = self._run("status")
        self.assertEqual((result.returncode, result.stdout), (0, ""))
 if __name__ == "__main__":
    unittest.main()