perf(skills): cut skill tokens and load hooks from cached bytecode

Hooks start their script as __main__ from cached bytecode instead of recompiling it on every call, and the installer migrates registered old hook commands to that form. A disabled plugin is written as false instead of deleted, because a deleted key in settings.local.json lets a true in settings.json win; the auto-compact window drops to 300000. The coach hour prints one ready line chosen by the time of day instead of asking the agent to load the skill, kvb-cleanup moves its procedure into plan.md, and the token counter reads subagent transcripts too and counts with Counter and defaultdict instead of .get defaults.

🎯 Fix: 88 %
💥 Break: 8 %
💸 Tokens: –
⏱️ Velocity: –
🛡️ Security: 100 %
🏆 Score: 81
🚦 Gate: ⛔ fix 88 % < 90, break 8 % > 5

Created-with: https://s.veen.world/skills
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Kevin Veen-BirkenbachandClaude Opus 5.5 committed 2026-11-18 19:40:37 +00:00
1 parent d695bb2b58
commit 6c63c364bd
13 files changed
+250 -130

No files matched your search

+13 -7
View File
@@ -55,11 +55,12 @@ for (const [name, entry] of Object.entries(MARKETPLACES)) {
}
data.enabledPlugins = Object.assign(data.enabledPlugins || {}, PLUGINS);
for (const [plugin, marketplace] of Object.entries(RETIRED_PLUGINS)) {
delete data.enabledPlugins[plugin];
// Trip-wire: this file can be settings.local.json; a deleted key there lets a true in settings.json win, a false overrides it.
data.enabledPlugins[plugin] = false;
delete data.extraKnownMarketplaces[marketplace];
}
data.env = data.env || {};
if (!data.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW) data.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = "500000";
if (!data.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW) data.env.CLAUDE_CODE_AUTO_COMPACT_WINDOW = "300000";
const PYTHON = "/usr/bin/python3 -I";
const ABSOLUTE = { python3: PYTHON, bash: "/usr/bin/bash" };
require("./council-deny").withCouncilDeny(data);
@@ -129,15 +130,20 @@ const SCRIPT_HOOKS = [
["kvb-org-perform/scripts/log.py", "PostToolUse", "Bash", "persist", {}],
["kvb-agent-optimize/scripts/graph.py", "PreToolUse", null, "hook", { timeout: 30 }],
];
// Trip-wire: python3 recompiles the script it is started with on every call and caches only imported modules; loading it as __main__ by file path reuses its __pycache__.
const AS_MAIN = `'import importlib.util as u,sys;s=u.spec_from_file_location("__main__",sys.argv[1]);m=u.module_from_spec(s);sys.argv=sys.argv[1:];sys.modules["__main__"]=m;s.loader.exec_module(m)'`;
for (const [script, event, matcher, sub, extra] of SCRIPT_HOOKS) {
data.hooks[event] = data.hooks[event] || [];
const registered = data.hooks[event].some((entry) =>
(entry.hooks || []).some((hook) =>
String(hook.command).includes(script) && String(hook.command).endsWith(` ${sub}`)));
if (registered) continue;
const file = JSON.stringify(path.join(path.dirname(settingsPath), "skills", ...script.split("/")));
// Trip-wire: python3 on a missing file exits 2, which blocks every prompt or tool call; the test makes it exit 1.
const command = `[ -f ${file} ] && ${PYTHON} ${file} ${sub}`;
const command = `[ -f ${file} ] && ${PYTHON} -c ${AS_MAIN} ${file} ${sub}`;
const legacy = [`${PYTHON} ${file} ${sub}`, `[ -f ${file} ] && ${PYTHON} ${file} ${sub}`];
const registered = data.hooks[event].flatMap((entry) => entry.hooks || []).find((hook) =>
String(hook.command).includes(script) && String(hook.command).endsWith(` ${sub}`));
if (registered) {
if (legacy.includes(registered.command)) registered.command = command;
continue;
}
const entry = { hooks: [{ type: "command", command, ...extra }] };
data.hooks[event].push(matcher ? { matcher, ...entry } : entry);
}
+3 -2
View File
@@ -27,8 +27,9 @@ per day and never changes the exit code; pass its quotient or `?` to
`kvb-check-score --value`.
- **Cost**: listing words delta x 1.33 x requests x 0.1 (cached) plus each
changed `SKILL.md`'s tokens delta x its invocations, per day of the last
7 days' transcripts (other runtimes: JSONL of that shape). A new skill
changed `SKILL.md`'s tokens delta x (its invocations + 0.1 x the later
turns it stays resident), per day of the last 7 days' session and
subagent transcripts (other runtimes: JSONL of that shape). A new skill
takes its category's median, named in the line.
- **Value**: tokens/day saved, derived by one `--evidence` line, e.g.
`"3829 SendMessage wakes/day x 10 % avoided x 43k"`; else `?:?`.
+71 -34
View File
@@ -4,9 +4,12 @@
Args:
--repo: skills repository root; its scripts/budget.py counts the listing.
--base: git ref the change starts from; the working tree is the after side.
--transcripts: Claude Code transcripts root, e.g. ~/.claude/projects; every */*.jsonl of the
last 7 days gives the requests (unique assistant message ids) and the invocations per skill
(unique Skill tool_use ids, a `<scope>:` prefix stripped). Read only; no text is printed.
--transcripts: Claude Code transcripts root, e.g. ~/.claude/projects; every session and
subagent transcript of the last 7 days (context.py of kvb-skill-silence lists them) gives
the requests (unique assistant message ids), the invocations per skill (unique Skill
tool_use ids, a `<scope>:` prefix stripped) and per skill its resident turns: every later
request of the same transcript, once per invocation, as Claude Code keeps an invoked
skill in context and re-injects it after a compaction. Read only; no text is printed.
--value: tokens per day the change saves; needs --evidence.
--evidence: one line that derives --value.
--json: print the figures as JSON instead of the lines.
@@ -23,10 +26,11 @@ Output:
Line 2: `💰 value 1:<n> (cost <c>/day, value <v>/day; <evidence>)` when the value is at least
the cost, `<n>:1` below it, `0:1` without a cost, `1:0` without a value; without --value
`💰 value ?:? (cost <c>/day; pass --value with evidence)`. Cost per day = listing words delta
* 1.33 tokens * requests per day * 0.1 (cached) + per changed SKILL.md its tokens delta * the
skill's invocations per day. A skill without a base SKILL.md takes the median invocations of
its category's base skills (the name up to its second dash; all base skills when it has
none), and the line names it.
* 1.33 tokens * requests per day * 0.1 (cached) + per changed SKILL.md its tokens delta *
(the skill's invocations per day + 0.1 * its resident turns per day). A skill without a base
SKILL.md takes the median invocations and the median resident turns of its category's base
skills (the name up to its second dash; all base skills when it has none), and the line
names them.
--json prints efficiency (null when new), the skills and listing totals, the tokens per skill
and the value figures.
value: `💰 measured value <v>/day (median daily context <b>→<a> tokens over <n> active days
@@ -49,6 +53,7 @@ import math
import statistics
import subprocess
import sys
from collections import Counter, defaultdict
from pathlib import Path, PurePosixPath
from types import ModuleType
@@ -192,55 +197,86 @@ def skill_uses(message: dict) -> list[tuple[str, str]]:
return found
def usage(root: Path, since: float) -> tuple[int, dict[str, int]]:
def entries(path: Path, since: float):
"""Args:
path: one transcript.
since: the epoch second the window starts.
Returns:
The assistant messages of the window, in transcript order.
"""
with path.open(encoding="utf-8", errors="ignore") as handle:
for line in handle:
if '"assistant"' not in line:
continue
try:
entry = json.loads(line)
stamp = datetime.datetime.fromisoformat(entry["timestamp"]).timestamp()
except (ValueError, KeyError, TypeError):
continue
message = entry.get("message")
if entry.get("type") == "assistant" and stamp >= since and isinstance(message, dict):
yield message
def usage(root: Path, since: float) -> tuple[int, dict[str, int], dict[str, int]]:
"""Args:
root: the transcripts root.
since: the epoch second the window starts.
Returns:
The requests, per skill its invocations, and per skill its resident turns.
"""
listed = load(CONTEXT, "context", "lists the transcripts").transcripts(root)
requests: set[str] = set()
invocations: dict[str, set[str]] = {}
for path in root.glob("*/*.jsonl"):
resident: Counter[str] = Counter()
for path in listed:
if path.stat().st_mtime < since:
continue
with path.open(encoding="utf-8", errors="ignore") as handle:
for line in handle:
if '"assistant"' not in line:
continue
try:
entry = json.loads(line)
stamp = datetime.datetime.fromisoformat(entry["timestamp"]).timestamp()
except (ValueError, KeyError, TypeError):
continue
message = entry.get("message")
if entry.get("type") != "assistant" or stamp < since or not isinstance(message, dict):
continue
if message.get("id"):
requests.add(str(message["id"]))
for skill, use in skill_uses(message):
invocations.setdefault(skill, set()).add(use)
return len(requests), {skill: len(uses) for skill, uses in invocations.items()}
live: Counter[str] = Counter()
for message in entries(path, since):
request = str(message.get("id") or "")
if request and request not in requests:
requests.add(request)
resident.update(live)
for skill, use in skill_uses(message):
seen = invocations.setdefault(skill, set())
if use not in seen:
seen.add(use)
live[skill] += 1
return len(requests), {skill: len(uses) for skill, uses in invocations.items()}, resident
def weigh(root: Path, listing_delta: int, body: dict[str, int], new: set[str], base_skills: list[str]) -> dict:
requests, invocations = usage(root, datetime.datetime.now(datetime.UTC).timestamp() - DAYS * 86400)
per_day = {skill: count / DAYS for skill, count in invocations.items()}
rates, assumed = {}, {}
requests, invocations, resident = usage(root, datetime.datetime.now(datetime.UTC).timestamp() - DAYS * 86400)
per_day = defaultdict(float, {skill: count / DAYS for skill, count in invocations.items()})
held = defaultdict(float, {skill: turns / DAYS for skill, turns in resident.items()})
rates, kept, assumed = {}, {}, {}
for skill in body:
if skill in new:
peers = [peer for peer in base_skills if category(peer) == category(skill)]
assumed[skill] = category(skill) if peers else ALL_SKILLS
pool = [per_day.get(peer, 0.0) for peer in peers or base_skills]
rates[skill] = statistics.median(pool) if pool else 0.0
pool = peers or base_skills
rates[skill] = statistics.median([per_day[peer] for peer in pool]) if pool else 0.0
kept[skill] = statistics.median([held[peer] for peer in pool]) if pool else 0.0
else:
rates[skill] = per_day.get(skill, 0.0)
rates[skill] = per_day[skill]
kept[skill] = held[skill]
listing = listing_delta * TOKENS_PER_WORD * (requests / DAYS) * CACHE_FACTOR
return {
"cost": listing + sum(delta * rates[skill] for skill, delta in body.items()),
"cost": listing + sum(delta * (rates[skill] + CACHE_FACTOR * kept[skill]) for skill, delta in body.items()),
"requests_per_day": requests / DAYS,
"invocations_per_day": rates,
"resident_turns_per_day": kept,
"median_for_new": assumed,
}
def value_line(weighed: dict, value: float | None, evidence: str | None) -> tuple[str, str]:
notes = [
f"{skill} new: {group} median {figure(weighed['invocations_per_day'][skill])} invocations/day"
f"{skill} new: {group} median {figure(weighed['invocations_per_day'][skill])} invocations/day, "
f"{figure(weighed['resident_turns_per_day'][skill])} resident turns/day"
for skill, group in sorted(weighed["median_for_new"].items())
]
cost = weighed["cost"]
@@ -331,6 +367,7 @@ def check(argv: list[str]) -> int:
"evidence": args.evidence,
"requests_per_day": round(weighed["requests_per_day"], 2),
"invocations_per_day": {s: round(r, 2) for s, r in weighed["invocations_per_day"].items()},
"resident_turns_per_day": {s: round(r, 2) for s, r in weighed["resident_turns_per_day"].items()},
"median_for_new": weighed["median_for_new"],
},
}))
+2 -63
View File
@@ -17,70 +17,9 @@ request and stop, or, ordered here by the operator, first
`${CLAUDE_SKILL_DIR}/../kvb-org-radio/scripts/radio.py post --to <chair>
"<plan>"`.
## Step 1: inventory
## Steps 1 to 4: inventory, owners, classes, proposal
Read-only, parallel, sized, only in the agent write scope (Claude Code:
`sandbox.filesystem.allowWrite`) and session scratch dirs, never another
user's files. Prefer project cleanup and teardown targets; obey raw-command
bans.
- **/tmp**: `pkgmgr gc --report --docker --root /tmp --root <write dir>`,
else `du -sh` per own entry; never system entries (`.*-unix`,
`systemd-private-*`, `ssh-*`, `tmux-*`, sockets).
- **Docker**, **worktrees**: read `docker.md` (its `⚠ Volumes` checklist
joins Step 4) and `worktrees.md` here when present.
- **Agent dirs**: pip, npm, pre-commit, act, gh, buildx caches; Ansible and
build temp dirs.
- **Council root** (`~/.local/state/agents-shared`): `archive/` only.
- **Register**: `council.py list --long`; `rasci.py artifacts`
(`kvb-org-artefacts`, `--ended` for candidates).
- **Queue**: `${CLAUDE_SKILL_DIR}/../kvb-org-queue-cleanup/scripts/cleanup.py
request`; go on meanwhile.
- **Live use**: `fuser` (exit 1 or 123: free), `lsof`, `ps -o
pid,etime,comm`; never full command lines or environments (credentials).
Sandboxed, an empty result proves nothing.
- **Archive duties**: memory or project rules requiring a copy before
deletion (logs, reports, generated maps).
## Step 2: ask the owners
Owner: register worker, path session id, worktree or compose project,
container label, open handle, else `unknown`. Once per local peer owning
unregistered candidates, `radio.py send call --to <peer>` the `ask.md` text;
go on meanwhile.
## Step 3: classify
First match wins: claimed by a peer, `keep`; in a register task not `done`,
`stopped` or `stale` (a branch merged into the default makes it
`stale`; an ended task's items are its worker's candidates), `keep (task <task>)`; a
running deploy, gate, test, swarm or build, `keep (running)`; a silent peer's,
`keep (no answer from <peer>)`, silence is not consent; sandboxed, touched
within 24 hours (`find -mmin -1440`; bfs fails on `-newermt`) and neither
owner-released nor operator-picked, `keep (recent, possibly in use)`; archive
duty, `archive first`, never a plain delete. Else a candidate:
- **low**, cheap to regenerate: stopped containers without volumes, dangling
images, unused networks, ended sessions' scratch and logs.
- **medium**, slow, rate-limited or networked to regenerate: unused tagged
images, build and package caches, clean merged worktrees.
- **high**, data loss or a broken run: volumes, stopped containers with
volumes, running unclaimed containers or stacks, dirty or unpushed
worktrees, 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 and every `keep` by
reason. Then `AskUserQuestion`:
1. `multiSelect: true`, up to four questions: `/tmp`, `Docker`, `Other
dirs`, `High risk` (every high-risk candidate); each option names size,
risk and worst case. Over four options, group by owner or risk or call
again; merge single items into one question: one-option questions fail.
Only ticks select, not an empty `Other` answer; nothing preselected.
2. `multiSelect: false`: `Execute: <n> items, <size> (Recommended)` naming
the ticked items' worst case, `Revise the selection` (repeats 1), `Abort`.
Read `plan.md` here and follow it to its `Execute` pick.
## Step 5: execute
+66
View File
@@ -0,0 +1,66 @@
# Plan: inventory, owners, classes, proposal
## Step 1: inventory
Read-only, parallel, sized, only in the agent write scope (Claude Code:
`sandbox.filesystem.allowWrite`) and session scratch dirs, never another
user's files. Prefer project cleanup and teardown targets; obey raw-command
bans.
- **/tmp**: `pkgmgr gc --report --docker --root /tmp --root <write dir>`,
else `du -sh` per own entry; never system entries (`.*-unix`,
`systemd-private-*`, `ssh-*`, `tmux-*`, sockets).
- **Docker**, **worktrees**: read `docker.md` (its `⚠ Volumes` checklist
joins Step 4) and `worktrees.md` here when present.
- **Agent dirs**: pip, npm, pre-commit, act, gh, buildx caches; Ansible and
build temp dirs.
- **Council root** (`~/.local/state/agents-shared`): `archive/` only.
- **Register**: `council.py list --long`; `rasci.py artifacts`
(`kvb-org-artefacts`, `--ended` for candidates).
- **Queue**: `${CLAUDE_SKILL_DIR}/../kvb-org-queue-cleanup/scripts/cleanup.py
request`; go on meanwhile.
- **Live use**: `fuser` (exit 1 or 123: free), `lsof`, `ps -o
pid,etime,comm`; never full command lines or environments (credentials).
Sandboxed, an empty result proves nothing.
- **Archive duties**: memory or project rules requiring a copy before
deletion (logs, reports, generated maps).
## Step 2: ask the owners
Owner: register worker, path session id, worktree or compose project,
container label, open handle, else `unknown`. Once per local peer owning
unregistered candidates, `radio.py send call --to <peer>` the `ask.md` text;
go on meanwhile.
## Step 3: classify
First match wins: claimed by a peer, `keep`; in a register task not `done`,
`stopped` or `stale` (a branch merged into the default makes it
`stale`; an ended task's items are its worker's candidates), `keep (task <task>)`; a
running deploy, gate, test, swarm or build, `keep (running)`; a silent peer's,
`keep (no answer from <peer>)`, silence is not consent; sandboxed, touched
within 24 hours (`find -mmin -1440`; bfs fails on `-newermt`) and neither
owner-released nor operator-picked, `keep (recent, possibly in use)`; archive
duty, `archive first`, never a plain delete. Else a candidate:
- **low**, cheap to regenerate: stopped containers without volumes, dangling
images, unused networks, ended sessions' scratch and logs.
- **medium**, slow, rate-limited or networked to regenerate: unused tagged
images, build and package caches, clean merged worktrees.
- **high**, data loss or a broken run: volumes, stopped containers with
volumes, running unclaimed containers or stacks, dirty or unpushed
worktrees, 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 and every `keep` by
reason. Then `AskUserQuestion`:
1. `multiSelect: true`, up to four questions: `/tmp`, `Docker`, `Other
dirs`, `High risk` (every high-risk candidate); each option names size,
risk and worst case. Over four options, group by owner or risk or call
again; merge single items into one question: one-option questions fail.
Only ticks select, not an empty `Other` answer; nothing preselected.
2. `multiSelect: false`: `Execute: <n> items, <size> (Recommended)` naming
the ticked items' worst case, `Revise the selection` (repeats 1), `Abort`.
+4 -4
View File
@@ -21,10 +21,10 @@ coach.py --profile <same> show [--json]
Areas: `psyche`, `body`, `sleep`, `food`, `tasks`, `delegate`. Every coaching
skill beyond this one is named `kvb-coach-<name>`.
`hooks/reminder.sh` triggers the short form at the first prompt of every full
hour, once across all agents (stamped in `~/.claude/.coach-hours/`): at most
three lines, the most urgent area first, then the agent continues with the
operator's request.
`hooks/reminder.sh` prints one fixed tip at the first prompt of every full
hour, once across all agents (stamped in `~/.claude/.coach-hours/`), chosen
by the hour: sleep late, food at meal times, else moving. The agent passes
it on in one line without loading this skill.
## Step 1: read the situation
+7 -5
View File
@@ -9,8 +9,10 @@ mkdir -p "${hours}"
mkdir "${hours}/${hour}" 2>/dev/null || exit 0
find "${hours}" -mindepth 1 -maxdepth 1 -type d ! -name "${hour}" -exec rmdir {} +
cat <<'HINT'
COACH HOUR: a new full hour began. Before answering the operator's request,
run the kvb-coach skill in its short form: at most three lines, the most
urgent area first, then continue with the request.
HINT
case "${hour##*T}" in
22 | 23 | 0[0-5]) tip="😴 Late: wind down, screens off soon, seven to nine hours of sleep." ;;
1[2-3] | 1[8-9]) tip="🥗 Meal time: a real meal with vegetables and protein, and a glass of water." ;;
*) tip="🚶 Stand up, stretch neck, hips and back for two minutes, and drink a glass of water." ;;
esac
printf 'COACH HOUR: open your answer with this one line for the operator, word for word, then continue with the request; load no skill for it:\n%s (More: /kvb-coach)\n' "${tip}"
+14 -2
View File
@@ -2,7 +2,8 @@
"""Report how much context agent sessions re-read per turn, from Claude Code's session transcripts.
Args:
--root: transcripts root, e.g. ~/.claude/projects (every */*.jsonl below).
--root: transcripts root, e.g. ~/.claude/projects (every session transcript
*/*.jsonl and every subagent transcript */*/subagents/*.jsonl below).
--days: how many days back, one row per day, newest last.
--threshold: a context size in tokens; per threshold the report adds the
share of turns above it and the share of context tokens above it.
@@ -34,10 +35,21 @@ def fail(message: str) -> None:
sys.exit(2)
def transcripts(root: Path) -> list[Path]:
"""Args:
root: the transcripts root, e.g. ~/.claude/projects.
Returns:
Every session transcript <project>/<session>.jsonl and every subagent
transcript <project>/<session>/subagents/<agent>.jsonl below it.
"""
return [*root.glob("*/*.jsonl"), *root.glob("*/*/subagents/*.jsonl")]
def scan(root: Path, since: float) -> tuple[dict[str, tuple[str, int]], dict[str, int]]:
turns: dict[str, tuple[str, int]] = {}
compactions: dict[str, int] = {}
for path in root.glob("*/*.jsonl"):
for path in transcripts(root):
if path.stat().st_mtime < since:
continue
with path.open(encoding="utf-8", errors="ignore") as handle:
+6
View File
@@ -89,6 +89,12 @@ class TestCoachHour(unittest.TestCase):
self.assertIn("kvb-coach", self._run("2026-11-03T15"))
self.assertEqual(sorted(p.name for p in self.hours.iterdir()), ["2026-11-03T15"])
def test_the_hour_picks_one_fixed_tip_and_loads_no_skill(self):
tips = {hour: self._run(f"2026-11-03T{hour}").splitlines() for hour in ("23", "04", "12", "19", "09", "16")}
firsts = {hour: lines[1].split(" ", 1)[0] for hour, lines in tips.items()}
self.assertEqual(firsts, {"23": "😴", "04": "😴", "12": "🥗", "19": "🥗", "09": "🚶", "16": "🚶"})
self.assertTrue(all(len(lines) == 2 and "load no skill" in lines[0] for lines in tips.values()))
def test_a_binary_planted_on_the_inherited_path_never_runs(self):
planted = Path(self.tmp.name) / "planted"
planted.mkdir()
+4 -1
View File
@@ -36,6 +36,9 @@ class TestContext(unittest.TestCase):
]
(self.root / "project").mkdir()
(self.root / "project" / "session.jsonl").write_text("\n".join(lines) + "\n", encoding="utf-8")
subagents = self.root / "project" / "session" / "subagents"
subagents.mkdir(parents=True)
(subagents / "agent-a1.jsonl").write_text(turn(self.today, "s1", 0) + "\n", encoding="utf-8")
def _run(self, *args: str) -> subprocess.CompletedProcess:
return subprocess.run(
@@ -52,7 +55,7 @@ class TestContext(unittest.TestCase):
)
self.assertEqual(
(today["day"], today["turns"], today["tokens"], today["compactions"], today["over"]["300000"]),
(self.today, 2, 450_100, 1, {"turns": 50, "tokens": 22}),
(self.today, 3, 450_200, 1, {"turns": 33, "tokens": 22}),
)
def test_usage_errors_exit_2(self):
+1 -1
View File
@@ -858,7 +858,7 @@ class TestGuard(unittest.TestCase):
self.assertEqual(hits, [])
def test_the_agent_root_is_named_where_it_is_inventoried_or_mounted(self):
self.assertIn("~/.local/state/agents-shared", (REPO_ROOT / "skills" / "kvb-cleanup" / "SKILL.md").read_text(encoding="utf-8"))
self.assertIn("~/.local/state/agents-shared", (REPO_ROOT / "skills" / "kvb-cleanup" / "plan.md").read_text(encoding="utf-8"))
self.assertIn("~/.local/state/agents-shared.", (CLIENT / "offices.py").read_text(encoding="utf-8"))
mounts = (REPO_ROOT / "docs" / "requirements" / "003-operator-workbench.md").read_text(encoding="utf-8").split("- **Mounts**:", 1)[1].split("\n\n", 1)[0]
self.assertIn("`~/.local/state/agents-shared`", mounts)
+37 -4
View File
@@ -3,6 +3,7 @@
from __future__ import annotations
import json
import re
import shutil
import subprocess
import tempfile
@@ -222,6 +223,35 @@ class TestSettingsRegistration(unittest.TestCase):
result = subprocess.run(["sh", "-c", gate["command"]], capture_output=True, check=False)
self.assertEqual(result.returncode, 1)
def test_a_hook_runs_its_script_as_main_from_cached_bytecode_and_keeps_its_exit_code(self):
(gate,) = self._hooks(self._patch(), "PreToolUse", "kvb-org-queue/scripts/queue.py")
script = Path(self.tmp.name) / "skills" / "kvb-org-queue" / "scripts" / "queue.py"
script.parent.mkdir(parents=True)
script.write_text("import sys\nprint(__name__, sys.argv[1:], sys.argv[0] == __file__)\nsys.exit(3)\n", encoding="utf-8")
runs = [subprocess.run(["sh", "-c", gate["command"]], capture_output=True, text=True, check=False) for _ in range(2)]
self.assertEqual([(run.returncode, run.stdout) for run in runs], [(3, "__main__ ['block'] True\n")] * 2)
self.assertEqual([path.name.split(".")[0] for path in (script.parent / "__pycache__").iterdir()], ["queue"])
def test_every_bash_gate_answers_under_the_cached_loader_as_it_does_started_directly(self):
data = self._patch()
(Path(self.tmp.name) / "skills").symlink_to(REPO_ROOT / "skills")
def answer(command: str, home: Path) -> tuple:
home.mkdir()
env = {"PATH": "/usr/bin:/bin", "HOME": str(home), "AGENTS_COUNCIL_ROOT": str(home / "council"), "XDG_STATE_HOME": str(home / "state")}
payload = json.dumps({"session_id": "s", "hook_event_name": "PreToolUse", "tool_name": "Bash", "tool_input": {"command": "ls"}, "cwd": "/"})
done = subprocess.run(["sh", "-c", command], input=payload, cwd="/", env=env, capture_output=True, text=True, timeout=60, check=False)
return done.returncode, done.stdout.replace(str(home), "~"), done.stderr.replace(str(home), "~")
bash = [entry for entry in data["hooks"]["PreToolUse"] if "matcher" not in entry or "Bash" in entry["matcher"].split("|")]
gates = [re.search(r' ("[^"]+\.py") (\w+)$', hook["command"]) for entry in bash for hook in entry["hooks"] if "importlib" in hook["command"]]
self.assertEqual(len(gates), 4)
for n, gate in enumerate(gates):
with self.subTest(script=gate[1]):
cached = answer(gate.string, Path(self.tmp.name) / f"cached-{n}")
direct = answer(f"/usr/bin/python3 -I {gate[1]} {gate[2]}", Path(self.tmp.name) / f"direct-{n}")
self.assertEqual(cached, direct)
def test_news_hook_moves_to_radio(self):
base = Path(self.tmp.name) / "skills"
self.settings.write_text(
@@ -243,14 +273,16 @@ class TestSettingsRegistration(unittest.TestCase):
encoding="utf-8",
)
data = self._patch()
self.assertNotIn("codebase-memory@codebase-memory", data["enabledPlugins"])
self.assertIs(data["enabledPlugins"]["codebase-memory@codebase-memory"], False)
self.assertNotIn("codebase-memory", data["extraKnownMarketplaces"])
self.assertTrue(data["enabledPlugins"]["caveman@caveman"])
self.settings.write_text("{}", encoding="utf-8")
self.assertIs(self._patch()["enabledPlugins"]["codebase-memory@codebase-memory"], False)
def test_compact_window_is_set_once_and_never_overwritten(self):
self.assertEqual(self._patch()["env"]["CLAUDE_CODE_AUTO_COMPACT_WINDOW"], "500000")
self.settings.write_text(json.dumps({"env": {"CLAUDE_CODE_AUTO_COMPACT_WINDOW": "300000"}}), encoding="utf-8")
self.assertEqual(self._patch()["env"]["CLAUDE_CODE_AUTO_COMPACT_WINDOW"], "300000")
self.settings.write_text(json.dumps({"env": {"CLAUDE_CODE_AUTO_COMPACT_WINDOW": "500000"}}), encoding="utf-8")
self.assertEqual(self._patch()["env"]["CLAUDE_CODE_AUTO_COMPACT_WINDOW"], "500000")
def test_hooks_run_the_system_python_isolated_and_old_entries_move_to_it(self):
old = f'[ -f "{self.tmp.name}/skills/kvb-org-queue/scripts/queue.py" ] && python3 "{self.tmp.name}/skills/kvb-org-queue/scripts/queue.py" block'
@@ -266,7 +298,8 @@ class TestSettingsRegistration(unittest.TestCase):
with self.subTest(command=command):
self.assertIn("/usr/bin/python3 -I ", command)
self.assertNotRegex(command, r"(^|&& )python3 ")
self.assertEqual(len(self._hooks(data, "PreToolUse", "kvb-org-queue/scripts/queue.py")), 1)
(gate,) = self._hooks(data, "PreToolUse", "kvb-org-queue/scripts/queue.py")
self.assertIn("spec_from_file_location", gate["command"])
def test_reminders_run_the_system_bash_and_old_entries_move_to_it(self):
old = f'bash "{self.tmp.name}/skills/kvb-dream/hooks/reminder.sh"'
+22 -7
View File
@@ -29,7 +29,7 @@ BUDGET = REPO_ROOT / "scripts" / "budget.py"
SKILL_A = "---\nname: kvb-a\ndescription: >\n one two three\n---\n"
SKILL_B = "---\nname: kvb-b\ndescription: >\n four five\n---\n"
NO_VALUE = "💰 value ?:? (cost 0/day; pass --value with evidence)"
NEW_B = "💰 value ?:? (cost 0/day; pass --value with evidence; kvb-b new: all skills median 0 invocations/day)"
NEW_B = "💰 value ?:? (cost 0/day; pass --value with evidence; kvb-b new: all skills median 0 invocations/day, 0 resident turns/day)"
CHEAP = "print('cheap')\n"
COSTLY = "data = list(range(2_000_000))\nprint(sum(data))\n"
LINE = re.compile(
@@ -149,6 +149,7 @@ class TestTokens(GitRepo):
"evidence": None,
"requests_per_day": 0,
"invocations_per_day": {"kvb-b": 0},
"resident_turns_per_day": {"kvb-b": 0},
"median_for_new": {"kvb-b": "all skills"},
},
},
@@ -194,13 +195,13 @@ class TestTokens(GitRepo):
result.stdout.splitlines(),
[
"💸 tokens 869 % (skills 13→113 tokens; listing 3→5 words)",
"💰 value ?:? (cost 203/day; pass --value with evidence)",
"💰 value ?:? (cost 333/day; pass --value with evidence)",
],
)
self.assertEqual(result.returncode, 3)
for value, line in (
("4053", "💰 value 1:20 (cost 203/day, value 4053/day; 4053 tokens/day saved)"),
("50", "💰 value 4.1:1 (cost 203/day, value 50/day; 4053 tokens/day saved)"),
("4053", "💰 value 1:12 (cost 333/day, value 4053/day; 4053 tokens/day saved)"),
("50", "💰 value 6.7:1 (cost 333/day, value 50/day; 4053 tokens/day saved)"),
):
with self.subTest(value=value):
result = self.tokens("--value", value, "--evidence", "4053 tokens/day saved")
@@ -209,16 +210,30 @@ class TestTokens(GitRepo):
self.assertEqual(
found,
{
"quotient": "1:20",
"cost": 202.66,
"quotient": "1:12",
"cost": 332.66,
"value": 4053,
"evidence": "e",
"requests_per_day": 10,
"invocations_per_day": {"kvb-a": 2},
"resident_turns_per_day": {"kvb-a": 13},
"median_for_new": {},
},
)
def test_subagent_transcripts_count_and_an_invoked_skill_stays_resident(self):
subagent = self.base / "projects" / "project" / "session" / "subagents" / "agent-a1.jsonl"
subagent.parent.mkdir(parents=True)
lines = [assistant("s0", self.recent, ("u0", "kvb-a")), *(assistant(f"s{n}", self.recent) for n in range(1, 14))]
subagent.write_text("\n".join(lines) + "\n", encoding="utf-8")
self.write("skills/kvb-a/SKILL.md", SKILL_A.ljust(452, "y"))
found = json.loads(self.tokens("--json").stdout)["value"]
self.assertEqual(
(found["requests_per_day"], found["invocations_per_day"], found["resident_turns_per_day"]),
(2, {"kvb-a": round(1 / 7, 2)}, {"kvb-a": round(13 / 7, 2)}),
)
self.assertEqual(found["cost"], round(100 * (1 / 7 + 0.1 * 13 / 7), 2))
def test_new_skill_takes_its_category_median(self):
self.write("skills/kvb-c-one/SKILL.md", skill_md("kvb-c-one", "one"))
self.write("skills/kvb-c-two/SKILL.md", skill_md("kvb-c-two", "two"))
@@ -231,7 +246,7 @@ class TestTokens(GitRepo):
result = self.tokens()
self.assertEqual(
result.stdout.splitlines()[1],
"💰 value ?:? (cost 202/day; pass --value with evidence; kvb-c-new new: kvb-c median 2 invocations/day)",
"💰 value ?:? (cost 1872/day; pass --value with evidence; kvb-c-new new: kvb-c median 2 invocations/day, 167 resident turns/day)",
)
self.assertEqual(result.returncode, 3)