A first-touch Edit/Write denial marks the file checked so the retry
passes. Sibling edits to the same file in the same parallel batch are
therefore judged against post-denial state and silently apply, leaving
the file in a state neither version intended.
Hooks see tool calls one at a time, so a batch-wide lock is not
possible. Instead make the partial application explicit: the Edit,
Write, MultiEdit, and condensed denials now name the file and warn
that other edits from the same batch may already have been applied,
and SKILL.md tells agents to send dependent edits sequentially and
re-read the file after a gated batch.
Expand proven-passive heredoc recognition beyond bare `cat` so documentation
writes via `tee`, `/bin/cat`, and `command cat` no longer trip the destructive
command detector on body text, while still failing closed for shells and pipes.
Two cases the word-level rewrite still got wrong. Short options that take an optional stuck value (-u[mode], -S[keyid]) end the cluster scan, so git commit -uno and -Sn are allowed while -nu stays blocked. Git accepts any unambiguous long-option prefix, so --no-veri and --no-verif on commit, push, merge and rebase are now blocked; --no-verbose stays allowed. Quoted data such as -m "--no-verify" is still treated as data. Independent exact-head review probed 34 commands in-process and against real git with no bypass and no false positive; hook test 35/35, eslint clean, CI 44/44 at the head.
pre-bash-commit-quality spawned Windows .cmd/.bat linters unquoted, so a spaced path failed, and passed --format compact, which ESLint 9 removed (#3075). Batch executables now run through cmd.exe with each argument carried in an env token and quoted, with quote, NUL, CR and LF rejected before spawn; non-batch Windows and POSIX paths keep direct argv spawn with shell false. ESLint uses its bundled default formatter, present on 8, 9 and 10. Regression tests cover the batch, non-batch and POSIX branches and the formatter change. Independent exact-head review passed with no P0/P1; CI 44/44 at the head.
Address #2921 and complete the segment-anchoring direction in #2979. Preserve explicit absolute exemptions while denying accidental matches in unrelated projects.
Forward-port #2925 for #2924 and verify ASCII and multibyte over-limit input suppression. Supersedes the overlapping direct-entrypoint fix in #2978.
Co-authored-by: jackie-cqz <2557911191@qq.com>
The context signal always rendered "N% of <window> window", including when
the window size was the assumed 200k default rather than a detected value.
On a 1M session whose transcript carries no [1m] marker, that produced
lines like:
[StrategicCompact] Context ~194k tokens (97% of 200k window)
while actual usage was ~19%. The user compacts on a false alarm, loses
context, and the resulting quality drop reads as a model regression.
The gap is structural: the context threshold defaults to 80% of the
window (160k on 200k), so the signal fires precisely in the 160k-200k
band where the size cannot be determined — above 200k the observed-tokens
fallback correctly infers 1M, and below 160k nothing fires.
Model id alone cannot close this. A tier may ship both a 200k and a 1M
variant under one id, so neither the known-family table nor a new entry
can distinguish them, and the transcript records no window field.
So stop asserting what isn't known: resolveContextWindow() now reports
whether the size was detected (env override, [1m] marker, known family,
or observed tokens > 200k) or assumed, and the hook omits the percentage
and window label when it was assumed. The token count, threshold, and
firing behaviour are unchanged.
resolveContextWindowTokens() keeps its existing signature and semantics.
Note: 3 pre-existing failures in tests/hooks/suggest-compact.test.js
reproduce identically on unmodified main and are untouched here.
Remove the redundant JSON-RPC initialize fallback from the consolidated MCP health-check batch. A routed 404 already proves the endpoint is reachable, and the real authenticated MCP call remains authoritative. Avoiding the fallback also prevents a stalled GET plus stalled POST from consuming twice the configured hook timeout.
The mcp-health-check preflight probes HTTP MCP servers with a bare GET.
Some Streamable HTTP servers route only POST /mcp and answer a bare GET
with 404 (Paper Desktop 0.5.3 is one). The probe scored that as down and
blocked every tool call for the server indefinitely, since the 30s
backoff just re-probes and re-fails.
A routed HTTP response of any status proves the endpoint is reachable,
which is all this preflight claims to check -- 400/401/403/405/406 are
already treated this way for the same reason. Add 404 to the set and let
the real MCP client validate the endpoint.
Adds a regression test that stands up a POST-only server (404 on GET,
200 on POST /mcp); it fails on the current code and passes with the fix.
The preflight probe in mcp-health-check only ever sent a bare GET to the
server URL. Some Streamable HTTP MCP servers route POST exclusively and
answer any GET with 404 — api.telnyx.com/v2/mcp is one — so the probe
failed permanently against a perfectly healthy server.
404 is not in HEALTHY_HTTP_CODES, so every probe failed, the backoff
compounded to the 10-minute ceiling, and the hook blocked every tool call
for that server before it left the machine while `claude mcp list` still
reported it Connected.
Replay a failed GET as a real JSON-RPC initialize POST and accept that as
proof of life. Whitelisting 404 was the alternative, but it would mask
genuine outages on every other server.
Adds a regression test with a POST-only server that 404s all GETs and
validates the initialize body; it fails without this change.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
cost-tracker.js: RATE_TABLE priced all Opus models at the legacy $15/$75
tier and routed Fable/Mythos 5 to Sonnet rates, overstating Opus 5
sessions ~3x and understating Fable ~3.3x in costs.jsonl. Adds fable
($10/$50) and current opus ($5/$25) tiers, keeps Opus 4.0/4.1/3 on the
legacy tier, updates haiku to 4.5 pricing ($1/$5).
session-end.js: extractSessionSummary included local-command echoes
(<local-command-caveat>, <command-name>, <local-command-stdout>),
system reminders, tool_result carrier turns and isMeta entries in the
Tasks list, so SessionStart reloaded noise instead of user asks. Adds
a noise filter.
Both test suites pass (10/10 cost-tracker, 1/1 session-end).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Rebased onto origin/main (49128b576). Fixtures moved from
scripts/hooks/ to /tmp/ecc-pr2380-fixtures/ per reviewer feedback.
(Original commit b0e49036 was based on cc6724ee; main has since
refactored spawnShell to use a shellArgs variable and added
PowerShell .sh fallback paths. This rebase adapts the
const result = spawnSync(...) + __rawInput tagging pattern to
all three spawnSync call sites in spawnShell.)
* fix(skill-evolution): wire Skill PostToolUse tracker so skill-health shows real runs (#2463)
recordSkillExecution() had no production callers, so
~/.claude/state/skill-runs.jsonl was never written and
`scripts/skills-health.js --dashboard` always reported 0 runs.
Adds scripts/hooks/skill-run-tracker.js and registers it as an async
PostToolUse hook (matcher: Skill) in posttooluse-dispatcher.js, which is now
the single PostToolUse entrypoint on main.
Addresses the privacy and bounds review on #2555's sibling PR:
- No prompt text is persisted. task_description is synthesized as
"Skill invocation: <skill_id>"; tool_input.task_description/description/
prompt are never read.
- Every persisted string is bounded and charset-restricted. A skill id is an
identifier, so free text, newlines, or an over-long value are dropped rather
than truncated and written through.
- The JSONL sink is created 0600 and re-tightened on each append, repairing
files written before this bound existed.
- The sink is capped at MAX_RUN_RECORDS (5000), trimmed oldest-first, so the
append-only file can no longer grow without limit.
Tests cover the privacy guarantee (no prompt text reaches a record), the
identifier bounds, the file mode on POSIX, and the retention cap.
* fix(skill-evolution): re-register the tracker for PostToolUseFailure
The rebase onto current main dropped the hooks.json entry, which silently
resurrected the P1 from the earlier review round: deriveOutcome() still
branches on hook_event_name === 'PostToolUseFailure', but the PostToolUse
dispatcher does not fan that event out, so the branch was unreachable in
production. Hard Skill failures were dropped from telemetry entirely, which
inflates the dashboard success rate — the opposite of what #2463 asks for.
Restores the dedicated PostToolUseFailure entry (matcher Skill, id
post:skill:track, same run-with-flags wrapper and standard,strict gating as
the dispatcher registration). Verified end-to-end: a PostToolUseFailure
payload piped through run-with-flags now records outcome "failure".
Adds a regression test asserting the registration so a future rebase cannot
quietly drop it again.
---------
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
* fix(hooks): context-monitor noise — loop-detection false positives and per-call cost-warning spam
Two independent noise sources in the PostToolUse context monitor injected
agent-facing warnings on nearly every tool call:
1. LOOP WARNING false positives. hashToolCall() hashed only the first 160
chars of a Bash command, so distinct long commands sharing a prefix
(heredocs, long one-liners) collided and consecutive DIFFERENT calls
looked like a stuck loop. Additionally LOOP_THRESHOLD=3 against a
5-entry ring buffer fired on legitimate repetition (retries, polling).
Fix: hash the full command (digest truncated, not the input — same
treatment the Edit/Write branch already got), and require all 5 of the
last 5 calls to be identical before warning.
2. COST NOTICE spam. run() deduped warnings on exact message text, but the
cost figure embedded in the text moves on nearly every call, so once a
session crossed $5 a 'new' COST NOTICE was injected per tool call for
the rest of the session. Context warnings had the same defect via the
remaining-% figure. Fix: dedupe on a stable per-tier key
(cost:notice/warning/critical, context:warning/critical, scope) so each
tier fires exactly once and re-fires only on genuine escalation. The
existing ECC_CONTEXT_MONITOR_COST_WARNINGS opt-out is unchanged.
Tests: loop threshold updated (5-of-5 fires, 4-of-5 does not), long
shared-prefix Bash hash regression, and a run()-level tier-dedupe test
(notice fires once, silent on cost tick, re-emits on escalation).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* refactor: keep context warning state immutable
---------
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
* feat(session-start): rank injected instincts by project/stack relevance
Instinct selection at SessionStart ranked purely by confidence, so a
high-confidence instinct about an unrelated stack could take an injection
slot from a lower-confidence instinct that is actually relevant to the
current project.
Rank by confidence + location/stack relevance instead: project-scoped
instincts, and instincts whose domain/trigger matches the detected stack
(languages/frameworks via detectProjectType, plus terraform/dbt markers),
get a small additive boost. The confidence>=threshold floor and the
injection cap are unchanged, and ranking degrades to confidence-only when
nothing matches or when ECC_INSTINCT_RELEVANCE_RANKING is set to off.
The ranking helpers live in scripts/lib/instinct-relevance.js with unit
coverage in tests/lib/, plus an end-to-end ordering test in tests/hooks/.
Completes part (b) of #2371; part (a) (configurable count + threshold)
shipped in #2413.
Fixes#2371
* refactor(session-start): drop redundant confidence tiebreaker in instinct sort
Greptile flagged that the secondary `right.confidence` comparison in
summarizeActiveInstincts' sort was dead code when relevance ranking is
disabled and, when enabled, was reached only on a floating-point tie of the
combined score — where it skipped the intended scope-label tiebreaker.
Remove it: the primary combined-score comparison already reduces to
confidence-only ordering when relevance is off, so behavior there is
unchanged; a genuine combined-score tie now falls through to the documented
scope-first, then id, order.
* test: isolate instinct relevance environment
---------
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
Feedback sent from the canvas only reached an agent through a live
/api/await long poll. When a turn ended with no await parked,
queueFeedback wrote the message to sessions.json and nothing ever
consumed it, so sending appeared to do nothing at all. The presence pill
made it worse: workingKeys had no expiry and the feedback handler never
broadcast presence, so it froze on "agent working" while nobody was
listening.
Delivery:
- Add the stop:plan-canvas-pending hook. It drains undelivered feedback
and blocks the Stop, handing the messages to the agent, so a canvas
message lands even when no await is running. Scoped to sessions under
cwd so parallel agents cannot swallow each other's feedback; set
ECC_PLAN_CANVAS_STOP_SCOPE=all to widen. Honors stop_hook_active and
fails open on every error path.
- run-with-flags.js did not await a hook's run(), so any async hook
silently degraded to pass-through. Fixed; plan-canvas-pending is the
only async hook today.
Presence and indicators:
- Presence is now ended/typing/thinking/listening/queued/waiting.
thinking and typing self-expire (90s/30s) and a 5s sweep pushes the
decay to an idle browser, so the pill can no longer stick.
- Broadcast presence when feedback is queued, and clear the activity
state when an agent reply lands.
- Add POST /api/session/:key/typing so agents can drive the indicator.
- Chat shows an animated dots bubble for thinking and typing, plus an
explicit note when a message is queued with nobody listening.
Respects prefers-reduced-motion.
- Send status reports what actually happened instead of always claiming
the agent will pick it up.
CLI and skill:
- Add `ecc-plan-canvas pending` and `typing <file> --state ...`.
- SKILL.md documents background await as the primary pattern and makes
replying in the canvas mandatory.
Tests: 6 new server cases covering queued presence, the typing endpoint,
state expiry and the sweep, plus a new hook suite covering delivery,
drain-once, stop_hook_active, cwd scoping and fail-open.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The Stop hook formats every JS/TS file edited during a response, grouped by the project root each file happens to sit in. That includes trees under .claude/plugins, which are third-party checkouts we only read.
Formatting them writes to code the user does not own. It also does real damage when a repo's committed code has drifted from its own formatter config: the rewrite is not a no-op but a wholesale reformat, so an unrelated bugfix ends up carrying hundreds of untouched lines. I hit this contributing to this repo — a 162-line fix arrived as a 478-line diff, most of it reformatted code the change never went near.
Skips both the user-level install root and a project-local one, mirroring the lookup in scripts/harness-audit.js. Paths are resolved before the prefix comparison, and a sibling such as .claude/plugins-backup does not match. The user own .claude config outside plugins is still formatted.
Adds 7 tests for the predicate, plus an end-to-end check that a clone file listed in the accumulator is left byte-identical. Suite 16 to 23.
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
isCommitNoVerifyShortFlag anchored on the first character, so it only recognised the flag when it led the cluster. Git clusters short options, which means git commit -an is -a plus the bypass flag and skips the hooks. -sn and -vn slip through the same way, while -na and -nm are caught — the difference is position, not intent.
Scanning now walks the cluster and stops at a value-taking option, since that option swallows the rest as its inline value. The n in -mn stays message text, and the existing -tn case keeps working.
Adds 4 tests: the three clustered forms that were escaping, plus -mn to pin the inline-value boundary. Verified the three fail against current main. Suite 25 to 29.
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
* fix(hooks,lib): fix hook detection and parsing edge cases
- auto-tmux-dev: dev\b -> dev(?![\w-]) so one-shot dev-build/dev-docs scripts
are not detached into tmux; align command shapes (yarn run dev, bun dev) with
pre-bash-dev-server-block.js DEV_PATTERN.
- pre-bash-commit-quality: skip obvious non-secret placeholders (env refs,
${...}, <...>, whitelisted tokens) in the api-key rule without suppressing
real high-entropy secrets; make -m message extraction quote- and
escaped-quote-aware so `-m "fix: \"x\""` / apostrophes are not truncated.
- pre-compact: annotate the CURRENT worktree's session (match **Worktree:** /
legacy **Project:**) instead of the newest *-session.tmp across all projects,
layered onto the LLM-summary flow from #2388; a present-but-blank Worktree
header is treated as non-legacy (no foreign project fallback).
- shell-substitution: stop double-appending a trailing backslash in an
unterminated backtick span.
- utils readStdinJson: on overflow, settle and resolve {} immediately (clear
timer + listeners) instead of waiting for end/timeout and parsing a partial
prefix; surface the overflow on stderr.
Regression tests added/extended (new tests/hooks/pre-compact.test.js).
Addresses review feedback on #2405. The earlier block-no-verify change was
dropped: its message-value skip on merge/cherry-pick/am/rebase would let
`git rebase -m --no-verify` bypass the hook (rebase's -m is the boolean
--merge), a false-negative worse than the contrived false-positive it fixed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(ci): align hook fixtures and drain oversized stdin
---------
Co-authored-by: djpjronline-netizen <276112803+djpjronline-netizen@users.noreply.github.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: haelyra <49814733+haelyra@users.noreply.github.com>
Claude Code writes one transcript JSONL line per content block, so a
single API response (one message.id) spans multiple assistant lines that
each repeat the same message.usage. sumUsageFromTranscript summed every
line, inflating token totals and estimated_cost_usd roughly 2.5-3x.
Verified on a real session: 704 assistant lines but only 286 unique
message.ids (2.46 lines/response on average); line-summing reported
$866.52 while the deduped total is $332.62. Usage payloads are identical
across lines of the same id (0/286 varied), so counting once per id is
equivalent to taking the last line per id.
Fix: collect usage into a Map keyed by message.id (last line wins) and
sum unique entries. Lines without a message.id (older transcript shapes)
keep the previous per-line behavior via a synthetic key, so existing
tests and old transcripts are unaffected.
Adds a regression test: a response split into 3 content-block lines with
the same message.id is counted exactly once.
Note: rows already written to ~/.claude/metrics/costs.jsonl by the old
code carry inflated token counts and estimates (except rows whose cost
came from the harness-cost cache, where cost is authoritative but token
counts are still inflated). Downstream consumers may want to annotate
history; this change intentionally does not rewrite the raw log.
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Preserve complete Stop-hook stdout through lifecycle wrappers, wait for queued output to flush before exiting, bound child output with a larger explicit buffer, and add end-to-end regressions for large, multibyte, dry-run, and failure cases.
