Plugin subagents silently ignore `hooks`/`mcpServers`/`permissionMode`
frontmatter — these are honored only for user/project agents in
.claude/agents/ (code.claude.com/docs sub-agents). Setting them in a
plugin agent is dead config; `permissionMode` is MEDIUM because it
implies a restriction Claude Code does not apply (false security).
hooks/mcpServers are LOW.
Additive to PLH's agent-frontmatter loop (no new scanner, count stays
13). One humanizer pattern covers the three field titles. Hermetic
temp-fixture test (positive + negative). Suite 954 -> 957, byte-stable.
Part of v5.5.0 "steering-model I". Foundation (active-config-reader
enumeration) deferred to v5.6 with B — A/E are additive and don't
consume it.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
GO granted to implement v5.5.0 in a fresh session. STATE now opens
directly into the task with ordered, concrete first steps (Foundation
enumeration → A durability → E plugin-agent dead-config), the brief
pointer, and the impl-critical gotchas.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Verified 2026-06-20 against hooks.md + the 2.1.169 changelog: the
`post-session` hook in that changelog is a SELF-HOSTED-RUNNER
workspace-lifecycle hook (runs after the session, before the workspace
is deleted), NOT a settings.json hook event. It is absent from
hooks.md's 30-event list (all PascalCase), so config-audit was wrongly
treating a bogus `post-session` settings hook as valid — it now flags it.
Restructured the event test suite accordingly and added a negative test.
Resolved U1/U2 in the v5.5 plan (U1 refuted → removed; U2 `outputStyles`
plugin.json key confirmed → PLH unchanged). Suite stays 954, self-audit A/A.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Doc-grounded plan for the five new-functionality items greenlit after
checking the "seven steering mechanisms" framing against live CC docs
(2026-06-19): A durability/compaction findings, B load-pattern token
accounting, C output-style scanner, D mechanism-fit heuristic, E
agent-listing cost + plugin-agent dead-config.
Includes a full verification log (CONFIRMED/REFUTED/UNVERIFIED per claim
+ source), dependency graph, phased rollout (v5.5/v5.6/v5.7), testable
acceptance criteria, and open decisions. PLAN only — each feature awaits
its own GO. No code.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
HKV: add Setup, UserPromptExpansion, PostToolBatch to VALID_EVENTS,
verified live against code.claude.com/docs/en/hooks.md (2026-06-19). A
valid hook using one of these was wrongly flagged "will never fire" — a
user could delete a working hook. Made the "(N total)" hint dynamic so
it can't drift again. Flagged the unverified kebab 'post-session' in a
comment (an existing test depends on it; follow-up check needed).
RUL: reword the globs finding. Only `paths:` is documented; whether CC
ever read `globs` is unverified, so the old "deprecated/legacy" framing
overclaimed (Verifiseringsplikt). New wording steers to the documented
`paths:` field. Updated the coupled fix-engine title match and the
humanizer entry (which also carried the "field was renamed" overclaim).
Suite 950 -> 954 (badge bumped). self-audit A/A, scanner count 13. No
version bump — these land in the pending v5.4.1 patch.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
plugin-health flagged "CLAUDE.md missing <commands|agents|hooks> section" regardless of
whether the plugin actually had that component — e.g. graceful-handoff (no commands/ or
agents/ dir) got two spurious medium findings. Same over-report class as the model-field fix.
Now gated on component presence (pluginShipsComponent): a section is required only if the
plugin ships that component (commands/ or agents/ with .md, or hooks/hooks.json). Across the 5
stable plugins this drops 12 spurious findings to 3 legitimate ones (graceful-handoff hooks,
ai-psychosis commands+hooks). New fixture plugin-section-coverage proves both directions.
Found via the marketplace-wide review. Suite 950/950, self-audit A/A, scanner count 13.
tests badge 949 -> 950.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
plugin-health-scanner required `model`+`allowed-tools` on commands and `model`+`tools`
on agents, plus `name` on commands. Per primary docs these are OPTIONAL:
- Commands/skills (code.claude.com/docs slash-commands): "All fields are optional. Only
`description` is recommended." `name` defaults to the directory name.
- Subagents (code.claude.com/docs sub-agents): "Only `name` and `description` are required";
`model` defaults to `inherit`, `tools` inherits all.
REQUIRED_COMMAND_FRONTMATTER -> [description]; REQUIRED_AGENT_FRONTMATTER -> [name, description].
This was over-reporting: every command without an explicit `model` got a spurious medium
finding (10 on the okr plugin alone). Found via the okr pilot review. Suite 949/949, self-audit
A/A, scanner count 13 (no new scanner).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Stop gitignoring STATE.md (remove from .gitignore) and commit it. Per the
updated ~/.claude/CLAUDE.md continuity rule, STATE.md is now tracked in every
repo and pushed to private Forgejo (never GitHub/public) — it survives fresh
clones, git clean, and branch switches instead of living only on disk.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Three additive findings extend existing PLH and SET scanners (scanner count
stays 13; --json/--raw byte-stable):
- CA-PLH-015 plugin-folder shadowing — a plugin.json component-path key in the
replaces-set (commands/agents/outputStyles) pointing at a custom path while
the default folder still exists; mirrors CC /doctor & claude plugin list.
- CA-PLH-016 skills:-array validation — each entry must resolve to a directory
inside the plugin root; flags non-string/escapes-root/not-found/not-a-directory;
mirrors claude plugin validate.
- CA-SET autoMode — structure (only environment/allow/soft_deny/hard_deny string
arrays; "$defaults" valid) = medium; dead-config (autoMode in shared
.claude/settings.json is not read by CC) = low.
Release mechanics: version 5.3.0 -> 5.4.0 (plugin.json); CHANGELOG [5.4.0];
README badge/TOC/What's-New/version-history; knowledge v5.4.0 scanner-backing
facts. Gates: suite 949/949, self-audit A/A + readmeCheck.passed (count 13),
SC-5 byte-equal, gitleaks clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
CA-SET autoMode shipped (3633571). All 3 Session B features built+pushed;
suite 936->949. Premise #3 confirmed against primary source (per-file-scope
gate passed). Next: Session C release (needs operator GO).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
settings-validator now validates the autoMode block (auto-mode classifier
config). Structure (medium): autoMode must be an object whose only keys are
environment/allow/soft_deny/hard_deny, each a string array ("$defaults" is a
valid entry); flags not-an-object, unknown-subkey, not-string-array. Dead-config
(low): Claude Code does not read autoMode from shared project settings
(.claude/settings.json), so an autoMode block committed there has no effect —
keyed on file.scope === 'project'.
Both premises primary-source-verified (code.claude.com/docs/en/auto-mode-config).
The plan's "test per-file scope first" gate passed: ConfigFile already carries
scope. SET is in the orchestrator; SC-5 re-checked, byte-equal (snapshot fixture
has no autoMode). Fixtures force-added (.claude/ is gitignored).
Tests +5 (944->949). Scanner count unchanged (13). --json/--raw byte-stable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
CA-PLH-016 shipped (76d5eda). Records that the plan's unverified "CC suggests
the parent directory" error text was dropped (not in primary docs).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
PLH now validates each entry of a plugin.json `skills` field (string|array):
every entry must resolve to an existing directory inside the plugin root.
One finding per bad entry (medium, plugin-hygiene), problem ∈ {non-string,
escapes-root, not-found, not-a-directory}. Mirrors `claude plugin validate`.
String|array normalized so a non-string top-level value is caught too.
Verifiseringsplikt: the plan's "CC suggests the parent directory" error text
is NOT in the primary docs — dropped. Only the four primary-source-verified
conditions are asserted (escape backed by the path-traversal rule).
Tests +3 (941->944). Scanner count unchanged (13). --json/--raw byte-stable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
CA-PLH-015 shipped (7abc5a1). Records the Verifiseringsplikt correction:
replaces-only field set (commands/agents/outputStyles); skills excluded
(adds-to-default), hooks/mcpServers/lspServers excluded (own merge rules).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
PLH now flags a plugin.json component-path key (commands/agents/outputStyles)
that replaces a default folder still present on disk — Claude Code stops
scanning that folder, so its contents are silently ignored (dead config).
Mirrors CC's /doctor & `claude plugin list` warning (v2.1.140+).
Field set pinned to the docs' "replaces" category only (Verifiseringsplikt,
code.claude.com/docs path-behavior-rules): skills is excluded (adds to the
default skills/ scan — both load) as are hooks/mcpServers/lspServers (own
merge rules); a custom path that addresses the default folder is not flagged.
Tests +5 (936->941). Scanner count unchanged (13). --json/--raw byte-stable.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Session A audit (read/plan only). Reconciled docs/cc-2.1.x-gap-matrix.md (the
v5.2.0 plan) against HEAD: the entire HIGH-priority false-positive cluster is
already CLOSED in v5.2.0 (settings keys + xhigh, 28 hook events, MCP POSIX/trust,
claude-md HIGH->MEDIUM reframe, DIS/CNF param-aware). The 8 unreleased commits add
incrementally; no active bug or false positive remains unshipped.
Operator GO 2026-06-19: "Release-only, ship-list tom" -> Session B SKIPPED. All
still-open gaps are M-effort enhancements -> defer to v5.4. v5.3.0 = docs +
knowledge + release of the 8 commits already on main.
Output: scope-decision block appended to docs/v5.3.0-release-plan.md (empty
ship-list + verbatim GO + full draft CHANGELOG bullets mapping all 8 commits +
What's New draft + 3 knowledge-backing entries for Session C); reconciliation
block prepended to docs/cc-2.1.x-gap-matrix.md. Version confirmed 5.3.0 (minor,
non-breaking). 9 commits 1576909..HEAD = 8 mapped + 1 excluded (the plan commit).
Three-session plan to release the 8 unreleased commits accumulated on main since
v5.2.0 as v5.3.0, with a fully updated README + CHANGELOG. Session A: gap-matrix
reconciliation + scope decision (operator GO on any new features). Session B:
conditional feature implementation (TDD). Session C: version sync, What's New,
CHANGELOG block, gates, tag, push. Each session has a testable Verifisering
section; key assumptions (non-breaking, docs-gate, number-only readme check)
flagged to test rather than trust. Mirrors docs/v5.2.0-release-plan.md.
The scanner table listed 12 rows while the badge, prose ("13 orchestrated
scanners"), and self-audit count all said 13 — the SKL skill-listing-budget
scanner (shipped in v5.2.0) was never added as a table row. self-audit
--check-readme didn't catch it: it checks the badge NUMBER against the
filesystem (13==13), not the prose table's completeness. Adds the SKL row
(CA-SKL-001 listing cap, CA-SKL-002 aggregate listing budget).
Commands are namespaced (/name:command), so a command name shared by two
differently-named plugins keeps both reachable — it is ambiguity, not a hard
conflict. The check now mirrors COL's plugin-vs-plugin skill finding: severity
LOW (was HIGH), category plugin-hygiene, COL-shaped details.namespaces, and a
group-first shape (one finding per command name listing every namespace, not
pairwise). It keys on the declared namespace (was folder basename) and fires
only across 2+ distinct namespaces — when plugins share a declared name, the
namespace-collision finding (medium) is the right signal, so this stays silent.
Removes the inaccurate "only one wins" humanizer entry. Adds fixtures
(duplicate-command-name; a shared command in duplicate-plugin-name's colliding
namespace) and 4 tests. Suite 932->936. self-audit A 97 / A 100, scanners 13.
Two plugins that declare the same `name` in plugin.json collapse into one
component namespace (/name:command, name:skill, agent "name"). Resolution
between two installed same-name plugins is undocumented, so one plugin's
commands/skills/agents are silently shadowed and unreachable. PLH now flags
this at medium severity, keying on the declared `name` (not folder basename,
via new declaredName on scanSinglePlugin) with a COL-shaped details.namespaces
payload. Name-less plugins are excluded from the collision map.
Search-first (code.claude.com/docs/en/plugins): plugin components are
namespaced by the declared name, so a plugin component can never shadow a
user/project one — only a same-name collision loses components. This refutes
the original "plugin vs user vs project shadowing" framing in the backlog.
Adds humanizer pattern, fixture (duplicate-plugin-name: 2 colliding + 2
name-less), and 3 tests. Suite 929->932. self-audit A 97 / A 100, scanners 13.
Extends the DIS scanner and its shared permission-rules lib with a third
documented Claude Code permission footgun. Verified verbatim against
code.claude.com/docs/en/permissions (fetched 2026-06-19).
CC's Tool(param:value) matching (2.1.178) is off-limits for a tool's own
canonicalizing field — CC ignores such a rule and emits a startup warning,
because e.g. Bash(command:rm *) is bypassable by a compound command. The
forbidden fields: command (Bash/PowerShell), file_path (Read/Edit/Write),
path (Grep/Glob), notebook_path (NotebookEdit), url (WebFetch).
- lib/permission-rules.mjs: new forbiddenParamRule(entry) returning
{ tool, key, hint } or null. Only the param:value form (colon present)
whose key equals the tool's forbidden field is flagged; Bash(npm:*),
WebFetch(domain:host), Agent(model:opus), and Bash(command) (no colon)
are left valid. FORBIDDEN_PARAMS map is the single source of truth.
- DIS: scans allow + deny + ask and splits severity by intent — deny/ask
hits are false security (medium: the block never applies), allow hits are
dead config (low: param:value matching is deny/ask-only). Two findings,
permissions-hygiene, CA-DIS-NNN.
- 11 new tests (7 lib, 4 DIS) + 1 fixture forbidden-param-permissions
(force-added past .gitignore .claude/). Suite 918 -> 929. Snapshot
unchanged (SC-5 byte-equal), contamination grep clean, gitleaks clean.
README/CLAUDE document the broadened DIS mandate; test badge synced.
self-audit: PASS, configGrade A 97, pluginGrade A 100, scanners 13.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Add a char-based CML finding that mirrors Claude Code's own startup warning
("Large CLAUDE.md will impact performance (X chars > 40.0k)"). CC 2.1.169 scales
that threshold with the model's context window, so the finding anchors on the
conservative 200k window (we cannot observe the user's window; the anchor fires
earliest) and discloses the relaxed ~200,000-char figure at 1M context. MEDIUM
severity (token cost, not an adherence cliff — consistent with the v5.2.0 reframe).
Keyed on chars, not lines, so it is complementary to the existing 200/500-line
checks (which stay untouched): a file can be long by lines yet under budget (short
lines, e.g. large-cascade at 37k chars / 1024 lines), or short by lines yet over it.
Extract the shared 200k/1M context-window constants to scanners/lib/context-window.mjs
(single source of truth; skill-listing-budget.mjs now imports + re-exports them).
40.0k figure and context-window scaling verified against the CC changelog (2.1.169,
2026-06-08) and the live startup-warning text. +6 tests, new fixture large-claude-chars
(48,531 chars / 100 lines). Suite 918/918, self-audit PASS configGrade A 97.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Extends the DIS scanner and its shared permission-rules lib with two
documented Claude Code permission footguns. Verified verbatim against
code.claude.com/docs/en/permissions (fetched 2026-06-19).
- lib/permission-rules.mjs: new isIneffectiveAllowGlob(entry) — unanchored
tool-name globs in permissions.allow (`*`, `B*`, `mcp__*`) that CC silently
skips ("does not auto-approve anything"); valid only as a glob-free
`mcp__<server>__*`. Shared with CNF.
- lib/permission-rules.mjs: dominates() now treats the `Tool(*)` deny-all glob
as equivalent to a bare deny (covers a bare allow) — CC: "Bash(*) is
equivalent to Bash ... both forms remove the tool from Claude's context".
- DIS: new finding "Ineffective allow wildcard — Claude Code ignores this rule"
(low, permissions-hygiene, CA-DIS-NNN); the existing dead-allow finding now
also catches a bare allow killed by a Tool(*) deny.
- 9 new tests (5 lib, 4 DIS) + 2 fixtures (force-added past .gitignore .claude/).
Suite 903 -> 912. Snapshot unchanged, contamination grep clean. README/CLAUDE/
scanner-internals document the broadened DIS mandate; test badge synced.
self-audit: PASS, configGrade A 96, pluginGrade A 100, readme gate passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Chunk 2 of the disableBundledSkills GAP feature. Adds a conditional GAP check
that prescribes the `disableBundledSkills` lever — but only when the active
skill listing is measurably over budget (SKL's CA-SKL-002 overflow signal) and
the lever is un-pulled. It stays an opportunity, not noise.
Bundled skills (/code-review, /batch, /debug, /loop, /claude-api, …) live in the
CC binary, not on disk, so their exact cost is unmeasurable here — the finding
says so plainly, and frames the lever as zero-cost budget reclaim that leaves
the user's own skills untouched. CC 2.1.169+.
- Pure, exported bundledSkillsLeverFinding({leverPulled, aggregate}) → finding|null
(severity low, category token-efficiency, CA-GAP-NNN), wired into scan() via the
shared measureActiveSkillListing().
- Lever resolved via new isBundledSkillsDisabled(): env var + settings cascade
read directly, because discovery does NOT tag ~/.claude/settings.json (its
relPath lacks ".claude" when walked from the .claude root) — the dominant
user-scope location for this global preference would otherwise be missed.
- GAP scan() now reads HOME → existing GAP tests retrofitted to withHermeticHome
per the hermetic rule. Snapshots unchanged, contamination grep clean.
- 16 new tests (9 GAP, 7 lib). Suite 887 -> 903. README/CLAUDE.md document the
cross-scanner remediation; test counts synced. self-audit: PASS, configGrade
A 96, pluginGrade A 100, readme gate passed.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Chunk 1 of the disableBundledSkills GAP feature. Moves the per-description cap,
aggregate budget constants, calibration note, and the enumerate-and-measure step
out of skill-listing-scanner into scanners/lib/skill-listing-budget.mjs — so SKL
(diagnoses overflow) and the upcoming GAP check (prescribes disableBundledSkills)
consume one budget definition instead of two divergent copies.
- New lib: assessSkillListingBudget (pure aggregate math) + measureActiveSkillListing
(HOME-scoped enumerate-and-measure wrapper).
- SKL delegates measurement; all finding strings kept byte-identical. 18/18 SKL
tests pass unchanged → behavior-neutral refactor.
- 12 new lib unit tests pin the budget contract. Suite 875 -> 887.
- README badge + CLAUDE.md test counts synced (self-audit --check-readme: passed).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
New orchestrated scanner SKL (CA-SKL-001 1,536-char listing cap, CA-SKL-002
listing-budget sum) → 13 orchestrated scanners. Five validators refreshed for
the CC 2.1.114→181 settings/hook surface (xhigh effort, MessageDisplay +
post-session events). False positives eliminated in MCP and permissions
scanners. Hermetic HOME isolation across all CLI-spawning tests.
Version sync: plugin.json 5.1.0→5.2.0, README badges (version + tests-875+),
5 stale "12→13 scanners" prose fixes, What's New + version-history rewrite,
CHANGELOG [5.2.0] entry. 875/875 tests; self-audit configGrade A, pluginGrade A,
readmeCheck.passed:true.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Devil's advocate gap-verification (read-only Workflow, 9 skeptics) refuted the
blanket "all closed" claim by finding fix-cli.test.mjs was the one CLI-spawning
test still reading the real ~/.claude. fix-cli runs the SKL skill-listing
scanner (HOME-scoped) even with includeGlobal:false, so its manual findings
include CA-SKL-001 on a dev machine but not in clean CI.
This directly corrects 325182d, which listed fix-cli.test.mjs as "Proven safe,
left as-is (output byte-identical real vs empty HOME — fixable findings are
project-local HKV/RUL/SET, never SKL/COL)". That reasoning predated SKL being
wired into scan-orchestrator (7bb2547/66433fe) and was false: real HOME yields
manual=6 (incl. CA-SKL-001), hermetic manual=5 (5230 vs 4798 bytes).
- wrap all 5 fix-cli spawns in hermeticEnv() (matches the other 11 CLI tests)
- add a regression lock: a project-scoped run must surface no CA-SKL/CA-COL
- redirect the --apply backup check to HERMETIC_HOME — the test was also
writing backups into the real ~/.config-audit/backups on every run
- docs: stale "26 hook events" -> 28 (README:528, scanner-internals:73);
hook-validator.mjs comment April -> June 2026 (functional count already 28)
Re-audited all 12 CLI-spawning test files: 11 hermetic-helper, manifest custom
HOME-env, post-edit-verify safe-by-construction (early-exit only, never reaches
scanners). HOME-leak class now actually closed. Suite 875/875, no snapshot drift.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Brief for the next session (post-/clear): independently DISPROVE the
"CC 2.1.114-181 feilretting is complete" claim rather than re-assert it.
Designed to run as a Dynamic Workflow (Workflow tool): one read-only
Explore skeptic per scanner surface (settings/hooks/mcp/permissions-DIS/
claude-md/knowledge/token) tries to find a still-open or superficially-
fixed row, plus adversarial review of this week's CA-SKL-002 + HOME-leak
work, then synthesis into a punch-list or a verified attestation.
Includes the 3 claims to attack (P1 no active false positives, P2 fixes
correct-not-superficial, P3 new work regression-free), the surface-cluster
table with attack angles, a runnable script skeleton + output schemas, and
a read-only scope-fence (findings → report, no fixes without approval).
STATE.md (gitignored, polyrepo) updated on-disk to hand this off as the
next session's active task.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Follow-up to the posture-grade-stability fix in 66433fe. Audited every
test that spawns a CLI and found more of the same class: tests running
HOME-scoped scanners (SKL/COL) or the CLAUDE.md cascade against the
developer's real ~/.claude instead of an isolated HOME.
Fixed (env: hermeticEnv()):
- posture.test.mjs — runs full posture (SKL/COL/cascade); twin of
the posture-grade-stability leak, masked only
because its asserts are structural/relative
- drift-cli.test.mjs — ACTIVE bug: the CLI wrote baselines into the
real ~/.claude during the run (pollution); now
isolated, and afterEach cleanup wrapped in
withHermeticHome so it looks in the same HOME
- token-hotspots-cli.test.mjs — scan-orchestrator run executes SKL/COL on
real HOME; TOK reads the HOME cascade
- accurate-tokens.test.mjs — TOK reads the HOME cascade (kept the
ANTHROPIC_API_KEY deletion)
Proven safe, left as-is (no HOME-scoped scan affecting assertions, no
HOME writes): post-edit-verify.test.mjs (fast-path early-returns only),
fix-cli.test.mjs (output byte-identical real vs empty HOME — fixable
findings are project-local HKV/RUL/SET, never SKL/COL),
lint-default-output (caller already uses withHermeticHome).
Suite 875/875, no snapshot drift. No test regressed under isolation,
confirming none had a hidden real-HOME dependency.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Syklus 2 of Fase 4 Items 2+3. Flags when the sum of active skill
descriptions exceeds the listing budget (~2% of context, CC 2.1.32).
Design (operator-confirmed "fact-first, 200k anchor"):
- low severity (estimate) vs medium for the verified 1,536-char cap
- each description counted up to the 1,536 cap (what actually loads in
the listing) — avoids double-counting the tail CA-SKL-001 flags
- fires when sum > 2% x 200k = 4000 tok; evidence leads with the measured
sum + a calibration note that the budget scales 5x on 1M-context models
- aggregate emitted after the per-skill loop so the common case reads
001=cap, 002=aggregate (finding IDs are a sequential counter, not stable
semantic IDs — tests match on title, never NNN)
Also:
- tailored humanizer static entry for the aggregate title
- fix latent HOME leak in posture-grade-stability.test.mjs: it spawned
posture.mjs without hermeticEnv(), so a real ~/.claude leaked HOME-scoped
SKL/COL findings into the baseline grade (Token Efficiency A->B). Now
isolated like the 8 other CLI-spawning tests.
- docs sync: test count 868->875, scanner-internals, gap-matrix, plan status
Suite 875/875, no snapshot drift, self-audit clean.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Fase 4 Items 2+3 (CC 2.1.114→181 gap-review). New orchestrated scanner
`skill-listing-scanner.mjs` (prefix SKL) flags every active skill whose
description exceeds the verified 1,536-char listing cap (CC 2.1.105, changelog
L1502). Past the cap, Claude Code silently truncates the description the model
reads to route skill invocation — dropping the trigger phrases at the tail.
HOME-scoped over all user + plugin skills via enumerateSkills (COL is the model).
- CA-SKL-001 (medium): description > 1,536 chars. Remediation folds in Item
2(b) — recommends disableBundledSkills + skillOverrides + trimming
(designvalg A: no standalone GAP-check, which would fire for nearly everyone).
- Designvalg B: v1 ships the verified cap ONLY. The aggregate 2%-of-context
listing budget is deferred — it needs a context-window assumption that would
turn a verified fact into a guess (would carry a CALIBRATION_NOTE if added).
- Choice C: recognize the skillOverrides settings key (CC 2.1.129) in
KNOWN_KEYS. Left OUT of TYPE_CHECKS — the value is a per-skill object
(off/user-invocable-only/name-only), not a string; a 'string' check (as the
plan sketched) would create a NEW false positive. Verify-first deviation.
Registration: scan-orchestrator (13th scanner), humanizer (SKL → 'Wasted
tokens' + static/_default translations), scoring SCANNER_AREA_MAP (→ Token
Efficiency; no 11th area), README badge 12→13, CLAUDE.md (finding-id +
test-count), docs/scanner-internals.md, gap-matrix + plan status notes.
Snapshots reseeded hermetically (SEED_SNAPSHOT/UPDATE_SNAPSHOT): SKL entry with
0 findings in empty HOME, scanners_ok 11→12, claudeMdEstimatedTokens bump from
the CLAUDE.md edits flowing through the cascade. Contamination grep clean.
Suite 868/868 (856 baseline + 11 SKL + 1 skillOverrides). RED→GREEN logged
per cycle.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Plans the skill-listing token-management work (gap-matrix rows 167+169) for a fresh session. All version facts re-verified against the changelog cache before writing: disableBundledSkills v2.1.169; skill-listing budget = 2% of context (v2.1.32) + 1,536-char per-description cap (v2.1.105); skillOverrides v2.1.129.
Key findings from code investigation:
- Item 2(a) (disableBundledSkills unknown-key false-positive) is ALREADY fixed (Batch 1: KNOWN_KEYS + TYPE_CHECKS); matrix row 166 marked DONE.
- Item 2(b): a standalone binary GAP check would be noise; recommend folding the disableBundledSkills recommendation into Item 3 SKL scanner remediation (designvalg A).
- Item 3: new SKL scanner; lead with the verified 1,536-char truncation cap (high confidence); aggregate 2%-budget is an estimate needing a context-window assumption (designvalg B) — calibration-noted.
- Reuse enumerateSkills() + parseFrontmatter; document the boundary vs TOK pattern F.
- Matrix row 170 keys (skillListingBudgetFraction/maxSkillDescriptionChars) do NOT exist (verified).
Plan doc carries exact file:line anchors, scanner-registration touchpoints, failing-test-first cycle specs, and 2 open design decisions to confirm before coding. STATE.md updated to resume here after /clear.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Fase 4 token-opt, Item 1 of gap-review NEXT STEP #2. The prompt-cache pattern corpus + TOK scanner were frozen at an "Opus 4.7" framing after CC shipped Opus 4.8 (default, 2.1.154) and Fable 5 (2.1.170). Model-era facts re-verified against the official changelog cache before editing.
The patterns are properties of prompt-caching, not of any model, so mechanic text is now model-neutral with a single "current default: Opus 4.8" anchor — preventing a re-freeze at the next model bump.
- rename knowledge/opus-4.7-patterns.md -> prompt-cache-patterns.md (git mv, history preserved); 6 reference sites updated
- TOK scanner: line-318 finding text (human-facing) made model-neutral; header + cache-prefix-scanner + CLI comments refreshed
- configuration-best-practices.md body + footnote 4.7 -> 4.8
- human-facing docs: commands/{tokens,help,manifest}.md, project CLAUDE.md, README, docs/scanner-internals.md
- gap-matrix row marked DONE; future Items 2/3 retargeted to new filename
Failing-test-first (Iron Law): +2 knowledge staleness guards (era-anchor + no-refreeze) +1 scanner assertion (no stale model anchor in finding text). Suite 853 -> 856 green; zero snapshot drift; self-audit A(97) PASS. CHANGELOG / v5 plan / ratified gap-plan keep historical opus-4.7-patterns refs (correct record of past state).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
`.mcp.json` has no per-server `trust` key — verified 2026-06-18 against
code.claude.com/docs/en/mcp + /settings. MCP server approval is
dialog/settings-based (enableAllProjectMcpServers / enabledMcpjsonServers /
disabledMcpjsonServers), never a JSON field. The scanner's "Missing trust
level" (CA-MCP-001, medium) and "Invalid trust level" (high) were false
positives flagging a field that does not exist.
- scanner: delete both trust checks + VALID_TRUST_LEVELS; drop `trust` from
VALID_SERVER_FIELDS so a stray `trust` is now flagged as an unknown field
- humanizer: remove the two trust-level entries
- knowledge (5 files): point to the real approval mechanism, not a trust field
- fixtures: scrub `trust` (incl. the invalid "local" in optimal-setup)
- tests: flip assertions (no trust-level finding; stray trust -> unknown
field) + add knowledge-staleness re-freeze guards
- snapshots: reseed (marketplace-medium .mcp.json -8 tokens, hermetic)
- gap-matrix: mark the trust verify-first item DONE
Suite: 853/853 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
The knowledge base was frozen at ~v2.1.111, describing an Opus-4.7 world
after CC shipped Opus 4.8 and Fable 5. Refreshed three agent-facing
knowledge files; every fact re-verified against the official changelog
(~/.claude/cache/changelog.md, CC 2.1.181) on 2026-06-18.
- feature-evolution.md: new Opus-4.8-era rows above the v2.1.111 freeze --
Opus 4.8 default + /effort xhigh (2.1.154), Fable 5 Mythos-class
(2.1.170), post-session hook (2.1.169), MessageDisplay (2.1.152),
/simplify -> /code-review (2.1.147), /config key=value (2.1.181).
- hook-events-reference.md: 26 -> 28 events (+MessageDisplay, +post-session);
documented Stop/SubagentStop additionalContext output field.
- claude-code-capabilities.md: 2026-06 model/effort lineup table;
bundled skills /simplify -> /code-review; documented /config key=value.
- cc-2.1.x-changelog-delta.md: marked SUPERSEDED by the gap-matrix.
Verification caught two version errors in STATE/matrix, corrected to the
changelog:
- Stop/SubagentStop additionalContext is 2.1.163, not 2.1.165 (2.1.165 was
"Bug fixes" only; matrix row 109 already said 2.1.163).
- settings `agent` field introduced 2.0.59; 2.1.157 = honored for
dispatched `claude agents` sessions, not the introduction.
New tests/knowledge/knowledge-staleness.test.mjs (8 tests) encodes the
verified facts as a re-freeze guard (RED before edits, GREEN after).
Full suite: 850/850 green (+8). self-audit PASS, A(100)/A(97).
The >500-line check emitted HIGH severity with "Files over 500 lines
significantly reduce Claude's adherence to instructions." CC 2.1.169
scaled the "too long" warning by context window, and the plugin's own
configuration-best-practices.md:97 footnote already says raw line count
is a Sonnet-era heuristic superseded by cache-prefix stability — so the
absolute HIGH + universal adherence claim is now-wrong.
- >500 lines: HIGH -> MEDIUM, reworded to token-cost-every-turn +
smaller-context-model caveat + cache-prefix pointer; notes CC 2.1.169
scales the threshold by context window.
- >200 lines: stays MEDIUM, dropped the absolute "optimal adherence"
framing for the same token/context-window framing.
Aligns the scanner with anti-patterns.md:7 (CA-CML-001 = medium) and
configuration-best-practices.md:97. No snapshot impact (byte snapshots
use a 24-line fixture CLAUDE.md).
Full suite: 842/842 green (+5). self-audit PASS, A(100)/A(97).
The DIS scanner collapsed Tool(param) rules to the bare tool name, so
Agent(model:opus) deny + Agent(model:sonnet) allow (and the same for
WebFetch(domain:...)) were flagged as dead config — a false positive now
that CC 2.1.178 matches Tool(param:value) and 2.1.172 adds domain rules.
The conflict-detector shared the blind spot from the other side: a
wildcard deny like WebFetch(domain:*) did not cover a
WebFetch(domain:good.com) allow, so a genuine cross-scope conflict was
missed (false negative).
New shared scanners/lib/permission-rules.mjs:
- parseRule / paramMatches (glob)
- dominates(deny, allow) -> DIS dead-allow (deny fully covers allow)
- rulesIntersect(a, b) -> CNF cross-scope conflict (match sets intersect)
DIS now delegates to dominates; conflict-detector :156 delegates to
rulesIntersect. A bare deny still covers all params, so true positives
are preserved (Bash deny + Bash(npm:*) allow still flagged).
Re-seeded the marketplace-medium snapshots: the false-positive CA-DIS
finding (Read(src/**) allow + Read(./.env) deny) is correctly gone. This
changes snapshot CONTENT only — envelope schema is unchanged, so --json
and --raw stay byte-stable.
Full suite: 837/837 green (+25). self-audit PASS, A(100)/A(97).
The COL collision-scanner and the CLAUDE.md cascade resolve ~/.claude from
process.env.HOME (active-config-reader). Snapshot/byte CLIs were spawned with
the developer's real HOME, so they picked up installed plugins/skills and the
user CLAUDE.md — making the v5.0.0 + default-output snapshots machine- and
time-dependent. They were seeded 2026-05-01 with COL=1 (a real ~/.claude skill
collision) and drifted to COL=0 after the polyrepo split: 26 pre-existing
failures unrelated to Batch 1.
Fix (test-only, no production change):
- tests/helpers/hermetic-home.mjs — empty temp HOME, mirroring the pattern
collision.test.mjs already uses for the COL unit test.
- 7 harnesses spawn CLIs (or call lint()) under the hermetic HOME, so output
depends only on committed fixtures. Determinism verified across runs.
- Re-seeded all snapshots under hermetic HOME via SEED_SNAPSHOT/UPDATE_SNAPSHOT
(added a SEED guard to the frozen v5.0.0 byte tests). Snapshots now reflect
the fixture alone (COL=0, fixture-only activeConfig counts).
- Also re-seeded the unused env-aware snapshots (manifest/whats-active/
plugin-health), which had baked dozens of real ~/.claude skill/plugin names
into the committed repo — privacy cleanup.
Full suite: 812/812 green, stable across 3 runs.
Clears false positives on valid .mcp.json (gap matrix, Batch 1):
- ${CLAUDE_PROJECT_DIR} is auto-injected at runtime (CC 2.1.139) and never
needs an env block — now allowlisted.
- POSIX expansions like ${VAR%pattern} / ${VAR:-default} are resolved by
Claude Code (CC 2.1.142); the env-var regex now matches only bare
${IDENTIFIER}, so operator expressions are skipped.
Genuine bare unreferenced vars are still flagged (broken-project regression
intact). The MCP `trust` field is untouched — it is verify-first (point 4),
not part of Batch 1.
Tests: hermetic runtime temp-fixture; 23/23 MCP green, both directions covered.
Ref: docs/cc-2.1.x-gap-matrix.md
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ter3E2JSi1Khgmuf2kady8
Pre-trekexecute snapshot of in-progress CLAUDE.md/SKILL.md edits and
extracted docs/ files. Captured as one commit so /trekexecute claude-design
can run against a clean working tree.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
