The JWT chain scanner only accepted a header at the start of a segment or
after `-`, so `session_<JWT>` reached the audit and cost logs unredacted.
Treat `_` as a prefix boundary too and cover it in the tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Since expanded here-document segments are queued too, a long substitution puts every segment in the queue, and shift() copied it each time: 130,000 segments took about 20 s, past the hook timeout. An index keeps the walk linear (under 0.3 s).
Bash ends a decoded $'...' string at its first NUL, so `cat <<$'EOF\0X'`
ends at the line EOF, but the splitter looked for EOF\0X and swallowed
the commands after the body. A carriage return or a no-break space is
not a blank to Bash, so `echo foo\r# && npm run dev` runs the dev
server, but the splitter read the # as a comment. And spreading a long
substitution's segments into push() threw RangeError, which the hook's
catch turned into an allow. Truncate the decoded part at the NUL, use
Bash's blanks for word starts, and append the segments one at a time.
The tracker serialized any structured result with JSON.stringify before
redaction, so an encoding nested in it, such as a double-escaped JSON
token, reached tool-usage.jsonl whole, and Claude Code's Bash stdout went
in as text. Keep text only from the adapter's tool_output, as main did,
and from Claude Code's tool_response only the sizes of stdout and stderr
and its interrupted and image flags. Any other structure is left out.
The safety-classifier notice filter only covered the mechanical summary.
When generateSessionSummary runs (low context or message interval),
extractConversationText read the raw transcript and passed the notice to
the LLM as a user turn, so it could still reach saved context.
Move the check into one shared isClassifierNotice() in
scripts/lib/llm-summary.js and use it in both userMessageText() and
extractConversationText(). Add LLM-path regression tests, and split the
session-end classifier-notice tests into small functions (<50 lines).
Refs #3867
Assisted-by: Claude <noreply@anthropic.com>
The dev-guard half of 7632669d from #3735: segments that the splitter
reads from an expanded here-document body are scanned again, so a dev
command in a substitution nested inside one is still found.
(split from commit 7632669d28)
The lookbehind added to keep the JWT scan linear also skipped tokens right
after a `-`, so `--cookie "session=jwt-<JWT>"` reached the logs unredacted.
Scan each `[A-Za-z0-9_.-]` run once instead, checking every dot segment for
a header at its start or after a `-`, decoding at most once per segment so
long `ey-ey-...` runs stay linear.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
When a response is stopped by the safety classifier, the harness records
a notice ("Your response above was stopped by a safety classifier ...")
as a user entry in the transcript. session-end.js kept it as a user
message, so it was written under "### Tasks" and re-injected by
session-start.js on the next session.
userMessageText() now drops entries that start with that notice, next
to the existing HARNESS_NOISE filter. User messages that only mention
the phrase are kept, since the match is anchored to the start.
Fixes#3867
Assisted-by: Claude <noreply@anthropic.com>
- Replace the secret=/api_key= value regex with a single-pass, quote-aware
scanner so the whole Bash word is redacted when quoted and unquoted
fragments are joined (e.g. --secret='prefix-'suffix), while a value inside
an enclosing quoted string stops at that string's closing quote.
- Searching for the keyword itself (grep -n "secret =" file,
rg 'api_key:' dir) no longer redacts the rest of the command.
- Use a lookbehind instead of \b before the JWT header so long dot-free
ey-ey-... runs are scanned in linear time instead of quadratic.
- Add regression tests for joined fragments, keyword searches, and the
ey- performance case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Match JWT headers that start with ey or ew (whitespace after the opening
brace) and keep non-eyJ matches only when the header does not decode to
an object with an alg field. Parse secret= / api_key= values with escaped
double quotes, $'...' quoting and backslash escapes, and redact an
unterminated quoted value to the end of the command.
Refs #3846
sanitizeCommand only redacted keyword- or flag-shaped secrets, so a bare
Anthropic/OpenAI sk- key, a JWT, a PEM private key block, or a secret= /
api_key= / x-api-key: value was written verbatim to bash-commands.log and
cost-tracker.log. Add value-shape patterns for these and keep the key name
and separator for assignments so the log stays readable.
Refs #3846
Address bot review on the Stop delivery path: the stop:evaluate-session
registration no longer carries async (an async Stop registration cannot
block), so delivery is synchronous, and the metadata fingerprint for the
changed entry is regenerated with the repo's own function; the delivery
marker is now claimed atomically with exclusive create and any write
failure fails closed instead of delivering without once-per-session
protection; the skill doc now marks the shell entry as counting-only,
requires the JavaScript evaluator through the plugin hooks for delivery,
and clarifies that Stop fires whenever Claude finishes responding; test
spawn helpers no longer coerce a failed spawn to exit 0.
Test plan:
- tests/hooks/evaluate-session.test.js: 25 passed, 0 failed (2 new:
marker write failure skips delivery, registration is synchronous);
pristine comparison: 22 passed, 3 failed (delivery, marker, registration)
- tests/hooks/hooks.test.js: 260 passed
- tests/hooks/stop-hooks-stdout.test.js: 48 passed
- tests/hooks/cursor-hook-dispatchers.test.js: 27 passed
- tests/integration/hooks.test.js: 27 passed
- scripts/ci/check-unicode-safety.js: passed
The v1 evaluator computed the learning signal but wrote it to stderr, which Claude Code never forwards to the model, so the learned-skills folder was never populated through this path.
Delivery now follows the maintainer-endorsed shape: an explicit stop_nudge_opt_in flag (default off, so existing installs are unchanged), the native Stop block contract (decision block plus reason) instead of plain stdout text, at most one delivery per session via a temp-dir marker, Stop-loop protection via stop_hook_active, and delivery only for sessions with a registered lease from SessionStart.
Tests: 7 new evaluator tests (opt-in off, no lease, one valid block then silence on repeat, stop_hook_active, short session, non-Stop event, missing session id). Evaluator suite 23/23 with the fix, failing on pristine; hooks suite 260/260; Stop stdout contract suite 48/48.
Limits stated plainly: end-to-end delivery in a live Claude Code session is not certified from this host (unit-level spawn plus parse only), and the fix ships inert by default until a user opts in.
config-protection.js and post-edit-typecheck.js attached process.stdin
listeners at module scope. tests/hooks/hook-input.test.js requires
config-protection.js, so that test never exits while stdin stays open.
Attach the listeners only when the script runs directly
(require.main === module), as stop-format-typecheck.js and quality-gate.js
already do.
post-edit-typecheck.js picked tsc diagnostics with line.includes(candidate),
so editing a root index.ts reported the errors of src/index.ts, and a
Windows relative path (src\index.ts) never matched tsc's forward-slash
output. Compare the file at the start of each diagnostic instead, the same
check diagnosticMatchesFile() does in stop-format-typecheck.js.
Refs #3821
On Windows, Claude Code kills an overrunning hook by killing the bash.exe
wrapper; the node.exe runner underneath survives. If another process still
holds the write end of the runner's stdin (a handle inherited by a process
spawned at the same moment), the runner never sees EOF and waits forever.
Orphaned runners were observed living 3-18 hours.
Add scripts/hooks/hook-watchdog.js and start it from the four hook entry
points (plugin-hook-bootstrap, run-with-flags, posttooluse-dispatcher,
lifecycle-hook-bootstrap). It exits the process when the parent PID is gone
(Windows) or the process was re-parented (POSIX), polled every 2s, or when a
hard deadline passes (330s, above the longest hooks.json timeout, or an
earlier inherited ECC_HOOK_DEADLINE_MS). Timers are unref'd so a finished
hook is never kept alive.
The regression test reproduces the leak with a detached runner whose stdin
write end is held by an unrelated process; it fails without the watchdog
(runner still alive after 8s) and passes with it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Restore the shipped payload-first contract from #3464/#3495 and the explicit source #3755 scope clarification. Two provider conversations must retain distinct IDs despite an inherited registered ECC parent ID. Normal and ambient producer24 tests, canonical resolver10 and independent differential review pass. Keep native importer guards; explicit provider/native correlation is queued separately.
Verify review discussion4419348312 against actual native producer/importer behavior. Prefer the independently sanitized ECC_SESSION_ID for the ECC-owned metrics feed, retain payload then legacy fallback for standalone hooks, and leave sibling hooks unchanged.
Actual Node producer plus read-only cached native binary reproduces zero imported metrics before the fix and exactly one tool/file/log after repeated sessions/status commands. Four producer regression controls and normal/adversarial inherited-env suites pass22/22. Focused coverage96.47percent lines and91.47percent branches; parent independent source review passed.
getopt_long accepts a unique prefix, so sudo --use root rm -rf build is
--user root and the resolver read root as the command. Expand unique
prefixes from sudo's long option table before the value/non-exec checks.
--help joins the non-executing modes.
sudo accepts VAR=value between its options (parse_args.c is_envar), so
sudo FOO=1 -u root rm -rf resolved -u as the command. Keep reading options
past assignments. Add sudo -R/--chroot, -a/--auth-type, -c/--login-class
and doas -a as value flags, and stop a short-cluster scan at any unknown
letter, so sudo -R/live rm -rf no longer reads the l in /live as -l.
doas -C only checks config, so it is non-executing; the existing doas -C dd
case moves to the passive list.
sudo -l/--list, -v, -K, -V and -e/--edit (and doas -L) do not run the
command operand, so sudo -l rm -rf build was denied as destructive. The
privilege-wrapper resolver now reports no executable command for those
modes. Short clusters stop at the first value flag, so sudo -ul rm -rf
(user l) stays destructive.
isDestructiveRm/isDestructiveGit keyed on argv[0], so sudo rm -rf,
command rm -rf and env git reset --hard read as calls to the wrapper and
were allowed once the routine shell gate was satisfied. Run both checks on
the unwrapLeadWrappers argv too, the resolution the dd and sh -c paths
already use. Overlaps #2832, which also reworks sh -c parsing and has
conflicts with main; this is the minimal rm/Git part.
Fixes#3762. pre:bash:gateguard-fact-force was missing from
FAIL_CLOSED_ON_TRUNCATION_HOOKS, so a Bash payload over the stdin cap
exited 0 while the Edit and PowerShell GateGuard ids blocked. Share one
GateGuard id set for the fail-closed list and the off-switch branch, and
add runner-level tests for all three ids plus each off switch.
Fixes#3761. config-protection only read tool_input.file_path, so a MultiEdit
with the config path inside edits[] edited an existing ruff.toml/.eslintrc
without the block. Check every effective target and block if any is an
existing protected config.
Integrates #3697 on current main. Prefer the observer project ID, retain a sole legacy ID until the new observer folder exists, and release the ending session lease across same-root aliases. Replace the unsafe live merge hint with read-only project inspection, and validate alias IDs before path use. #3711 independently identified the computed-ID selection gap; its early switch without observer data and unchanged SessionEnd path were not adopted.
Port #3656 onto current main, preserving both the contributor fix and the unrelated current changelog entry. Includes the follow-up glob normalization from 2c11fd74.
- bash-hook-dispatcher: honor the reader's truncated flag. pre-Bash
blocks (exit 2) because safety checks need the complete request;
post-Bash skips its hooks and does not echo a partial payload.
- resolveStdinIdleTimeout: treat whitespace-only values as unset instead
of Number(' ') === 0 silently disabling the deadline.
- isCompleteJsonObject: do not accept a JSON array as a hook payload.
- tests: wait for the child's close event so stdio is drained before
asserting; cover both bash-hook-dispatcher modes.
- hooks/README.md: keep the example below the 5 s per-hook timeout.
Refs #3665
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
setTimeout fires after 1 ms for any delay above 2147483647 ms, so a large
ECC_HOOK_INPUT_IDLE_TIMEOUT_MS turned the deadline into ~1 ms and marked
every hook input incomplete. Clamp to the maximum with a diagnostic, and
surface the invalid-value diagnostic when readStdinRaw resolves the env.
Refs #3665
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
readStdinRaw() only settled on end, error or close. If the parent kept
stdin open without writing EOF, every hook that reads through it waited
until the harness killed it (an 8-10 s stall per call).
Add an idle deadline (default 3000 ms, ECC_HOOK_INPUT_IDLE_TIMEOUT_MS,
0 disables). It is reset on each chunk, unref()ed, and on expiry detaches
from and pauses the stream so a held-open pipe cannot keep the hook alive.
Input that already parses as a complete JSON object is accepted; anything
else is marked truncated so fail-closed safety hooks still block.
bash-hook-dispatcher.js now reads through the shared reader instead of its
own end/error-only loop.
Fixes#3665
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The bounded name parts (64 characters around the credential word) let a longer name such as API_KEY_ plus 65 characters go unredacted, and the field-bound bare value left `,beta` of CLIENT_SECRET=alpha,beta in the summary.
Find each name and separator first, starting a name only where a run of name characters starts, so every run is scanned once whatever its length, and test the whole name for the credential word. A bare value after a quoted JSON key still ends at the next field; any other runs to the next space.