Claude Code Hooks

Claude Code Hooks

Claude Code fires hooks at tool-call boundaries. A hook is a shell command that receives a JSON event on stdin and, for blocking hooks, decides via its exit code whether the tool call proceeds. This is the only mechanism that structurally stops a behavior — a prose rule in CLAUDE.md is a suggestion the completion drive can override; a hook is a wall.

When a hook is the right tool (and when it isn't)

Write a hook when a rule keeps getting violated even though it's already written down. The tell: you added the prose rule, it read clearly, and the behavior recurred anyway — because at the moment of action, attention is 100% on "get the thing done" and the reminder loses. That recurrence is the signal to move the rule from prose (advisory) to a hook (enforced). Governance rule of thumb: Tier-0 irreversible action + only prose, no hook → it should be a hook. (Tier-0 here = an action whose damage cannot be undone from inside the session: destroying uncommitted work, pushing secrets to a remote, deleting files, publishing something outward. The test is reversibility, not severity.)

Do not reach for a hook when: the rule has never actually recurred (don't pre-build guards for hypothetical mistakes — cost with no proven benefit), or the "rule" is a judgment call with no mechanical signature (a hook can only match tokens/patterns; it can't judge whether a design is good).

If the symptom is “it keeps reviewing / waiting / retrying,” do not assume the answer is another hook. First complete the Loop Contract in rule 7 and read pitfall #36. The loop may be created entirely by an agent repeatedly applying a prose rule.

Hook types and what the exit code means

• PreToolUse is the workhorse for stopping a tool call — the four types in this table are the ones this file teaches, not the complete set of blockable events, and the official hooks reference (docs.claude.com / code.claude.com, not the references/ files in this bundle — those cover only the four types above) now lists many more blockable events, including UserPromptSubmit, PreCompact, TeammateIdle, and task and config events. If what you need to gate is not a tool call, look there before forcing it onto PreToolUse. matcher selects the tool (Bash, Agent, WebFetch, …) — and how it is matched depends on the characters you use: a matcher containing only letters, digits, _, -, spaces, , and | is compared as an exact string (or a |/,-separated list of exact strings); anything else is treated as an unanchored JavaScript regex. Both directions bite silently — Edit.* also matches NotebookEdit (anchor it ^Edit$), while mcp__memory matches nothing because it is all exact-match characters and no tool is named exactly that (you want mcp__memory__.*). Matching is case-sensitive. Exit 2 blocks and the hook's stderr becomes the message the model sees — so put the why and the correct alternative there, not just "blocked".
• PostToolUse can't undo, but it can inject authoritative context so a later hallucination can't stand (e.g. re-read the real git HEAD after a commit and surface it — the model can't "believe it committed" against injected truth).
• SessionStart is for health checks of the guard rails themselves — silent when healthy, warn on breakage, always exit 0. Note why: it is not that a non-zero exit would block the session (it cannot), but that it would print a hook-error notice at every session start until someone fixes it — a check that cries wolf on startup is a check people learn to scroll past.
• set -euo pipefail vs set -uo pipefail — pick by contract, and know there are two ways to keep an always-exit-0 contract. A hook that may block (PreToolUse) wants -e: an unexpected failure aborting the script is survivable, because the caller treats a non-0/2 exit as "proceed". A hook whose contract is ALWAYS exit 0 (PostToolUse injectors, SessionStart checks) has two honest shapes: (a) drop -e and ||-guard every risky command — with -e on, one grep that legitimately finds nothing kills the hook mid-way and the CLI surfaces a bare Failed with non-blocking status code (pitfall #8, Pattern E's shape); or (b) keep -e and add trap 'exit 0' ERR so any failure still converts to exit 0 while -e keeps guarding the plumbing (git-commit-headcheck's production shape, Pattern D). Either is correct; what you cannot do is -e alone with no trap and no ||-guards. Rule of thumb: -e for hooks that decide; for hooks that report, drop -e or trap it (pitfall #8).
• Stop is the odd one out, and the one most often reached for by mistake: it's the only hook type that can react to what the model itself just generated (its own reply text). Every other hook type — including UserPromptSubmit, which sounds like a plausible place to police "what gets said" — only ever sees the user's input; it structurally cannot see the model's own current-turn output (that claim holds — this is still the right reason to route such a rule to Stop). That guarantee, however, does not extend to proving the .prompt field always originated from a keystroke: a background task-notification's own report text can populate it too, with nothing in the stdin JSON marking the difference — #30. A rule like "the model must not invent a shorthand name for something it hasn't verified" belongs on Stop; put it on UserPromptSubmit instead and it will (a) never once catch what it was built for, since that text never flows through that event, and (b) false-block the user's own unrelated typing whenever it happens to contain the trigger pattern. This is a category mistake, not a tuning problem — no amount of regex refinement on the wrong event fixes it. Full contract (last_assistant_message vs transcript_path, the anti-loop check) in Pattern E in references/hook_patterns.md.
• Stop has two block channels with identical loop protections — pick by intent, and make the first (only) block carry everything. decision: "block" + reason, or plain exit 2 + stderr, shows as a hook error — for hard gates ("this must not stand"). hookSpecificOutput.additionalContext shows as neutral "Stop hook feedback" with no error notification — for coaching and reminders the model should weigh, not gates. Both count toward the same consecutive-block ceiling from the table above, so the choice is tone, not safety. What that means for message design: a blocked retry round (stop_hook_active: true) is let through with whatever violations remain — so a Stop guard gets exactly one informed bite. (The ceiling reinforces this only when your remediation is "rewrite the reply"; if it involves tool calls the counter resets and the ceiling never lands — #27. Either way the one-bite conclusion holds, because it rests on the latch, not on the ceiling.) Report all findings in that one block (a guard that prints only the first loses the rest permanently — pitfall #17), and write the message as an escape manual naming the exact acceptable fix, not a verdict — the model converges in one round or it burns the cap guessing. v2.1.145+ inputs background_tasks / session_crons let a blocking hook tell "the session is done" from "the session is merely paused waiting for background work" — blocking a pause forces pointless continuations and wastes the same cap.

Full runnable skeletons: references/hook_patterns.md.

The skeleton (PreToolUse Bash guard)

#!/usr/bin/env bash
set -euo pipefail
IFS= read -rd '' INPUT || true                 # builtin; NOT $(cat) — see below
# 0-fork fast path: a builtin `case` on the raw JSON, BEFORE paying for python3.
# Your guard runs on EVERY matching tool call, so the irrelevant path is the one
# that has to be cheap. Keep this filter BROADER than what you actually block and
# never flag-level — it answers "is this even about X", nothing finer (#22).
case "$INPUT" in *TRIGGER*) ;; *) exit 0 ;; esac
TOOL=$(printf '%s' "$INPUT" | python3 -c "import sys,json;print(json.load(sys.stdin).get('tool_name',''))" 2>/dev/null||echo "")
[ "$TOOL" != "Bash" ] && exit 0                # only guard the tool you mean to
CMD=$(printf '%s' "$INPUT" | python3 -c "import sys,json;print(json.load(sys.stdin).get('tool_input',{}).get('command',''))" 2>/dev/null||echo "")
[ -z "$CMD" ] && exit 0
printf '%s' "$CMD" | grep -qw 'TRIGGER' || exit 0   # precise relevance check
# ... precise detection here ...
if <command actually does the banned thing>; then
  echo "BLOCKED: ... WHY ... USE INSTEAD: ..." >&2   # stderr = the guidance shown
  exit 2
fi
exit 0

Why the first two lines are not stylistic. INPUT=$(cat) plus each printf … | python3 -c … costs forks on every call this hook matches, including the ones it has nothing to say about. A fleet of ~13 Bash-matcher hooks × parallel sessions × sub-second tool cadence turned that into a sustained 40–200 forks/sec of pure guard overhead and put Gatekeeper at the top of an all-day CPU ranking with no runaway process anywhere — the fleet was fine; the irrelevant path's per-call cost was the bug (#22, with the per-guard conversion recipe and its measured floor).

Three caveats before you copy the case line anywhere else — the first one is the difference between a fast path and a bypass:

• A coarse filter must be a SUPERSET of what you block, and a raw substring test is not one. TRIG''GER -x runs TRIGGER — bash splices the quotes away before execution — but the raw event text contains no TRIGGER substring, so a bare case "$INPUT" in *TRIGGER*) exits 0 and the guard never sees it. Measured: drop this exact line into the shipped Pattern A and TRIG''GER -x flips from exit 2 to exit 0, a full bypass — while scripts/test_hook.sh still reports 21 pass / 0 fail, because no row carries a spliced trigger. Pattern A already carries the fix and the reason ("de-splice — strip quotes and backslashes — and check again; a false negative is a full bypass"); a coarse filter placed before that de-splice makes it unreachable. Two safe shapes, in order of preference: filter on something the splice cannot touch — a JSON key or a tool name (case "$INPUT" in *'"tool_name":"Bash"'*)), since quote-splicing lives in the command text and cannot rewrite the event's own structure; or de-splice inside the filter before testing (strip ", ' and \ from a copy of the input, then match). Prefer the first: it needs no escaping gymnastics, and a filter whose own quoting you have to get right is a filter you can get wrong silently. The skeleton above is safe as written only because its own detection is likewise a plain word match; the moment the guard below the filter is smarter than the filter, the filter decides.

• This skeleton is fail-open on irrelevance (grep -qw … || exit 0), so a coarse filter in front of it changes cost, not semantics. A fail-closed guard is different: a bare substring filter silently converts its contract from block-unknown to allow-unknown (measured — 'not json' sailed straight through the first cut of that fix), and a *tool_name* marker alone re-opens the same hole from the other side. Read #22's gate requirements before fitting a fast path to a guard that is supposed to block on malformed input.

• There is a cheaper layer above the script. A hook handler can carry an if field in its registration — permission-rule syntax such as "Bash(git *)" — and the hook command does not run at all when it doesn't match: zero forks, because zero processes. It is best-effort by design (the docs say it fails open, running your hook anyway, when the Bash command can't be parsed), so treat it as a cost optimization and never as the gate — the in-script check still decides. Three sharp edges: it holds exactly one rule (no &&/||), it is only evaluated on tool events, and a hook that sets if on a non-tool event never runs at all.

Rules that separate a working guard from a session-poisoning one

Not style preferences — each is a specific failure we shipped and traced back.

1. Match at the token level with shlex, never awk-split the raw string

A guard that false-blocks a healthy command is worse than one that misses — a guard people must bypass gets bypassed reflexively, and then it protects nothing (the core discipline: 误杀健康输入比漏报更糟). The recurring cause of false-blocks is matching on the raw command string.

• Wrong: awk '{gsub(/&&|\|\||;|\|/,"\n")}' to split into segments — awk doesn't understand shell quoting, so grep -E "a|TRIGGER|b" gets split at the | inside the quoted regex, TRIGGER becomes a phantom command, and the guard blocks a plain grep. (Shipped 2026-07-21; the guard's very first real use was a false-block on my own grep.)
• Right: tokenize the whole command with the shlex.shlex class, not the shlex.split() function — split() only treats | ; & < > as separators when they are space-separated, so ls|TRIGGER x tokenizes to ['ls|TRIGGER', 'x'] and your command-position check never sees TRIGGER at all (measured; the class with punctuation_chars=True yields ['ls', '|', 'TRIGGER', 'x']). Copy a shipped walker verbatim rather than reaching for the one-liner — but copy the one that passes scripts/test_hook.sh, which is Pattern A's. The walker section is a compact form and says so: it omits the per-wrapper valued-flag tables, so it misses a target riding a valued-flag wrapper — measured, it returns "not in command position" for timeout 5 TRIGGER, sudo -u root TRIGGER and nice -n 10 TRIGGER, while Pattern A's version catches all three. (Bare sudo TRIGGER is fine in both — it is the wrapper's own flag taking an argument that the compact table doesn't know to skip.) Only one of those shapes is in the shipped harness, so the run you actually see is 20 pass / 1 fail on wrapper-timeout against 21/0 for Pattern A's; the other two fail silently because no row covers them. A quoted "a|TRIGGER|b" stays one token, so a regex argument is never mistaken for a command. Then check whether your target is in a command position (token[0], or right after a ;/&&/||/| separator, skipping VAR=val env-assignment prefixes). Command-position walker in references/hook_patterns.md.
• Corollary: echo "…TRIGGER…", grep TRIGGER, # TRIGGER, man TRIGGER must all pass. Your test set MUST include these mention-not-execute cases.
• Corollary — exempt git write segments before they reach the walker. A commit message is arbitrary data, and the whole message text reaches your command-position walk as pseudo-command-text — git commit -F - <<EOF with a body quoting foo|TRIGGER lands TRIGGER in command position, and the guard blocks its own fix commit (pitfall #7 is exactly this, shipped). Any Bash guard that inspects command strings must skip segments whose head is git + commit/rebase/tag/am/cherry-pick — and do it at the whole-command level, before any line splitting (Pattern A shows the order; the production version is lib-git-commit-detect's adjacency check).
• Corollary — the walker is two-stage for a reason. whitespace_split=True treats newlines as ordinary whitespace, so a multiline block (cd /x\ngit add\nTRIGGER -y) collapses into one segment headed by cd and the trigger is never in command position — replayed trigger rate 0 on real transcripts (pitfall #11). Split into lines shell-aware first (quote state and backslash continuations honored, so quoted multiline strings don't fragment), then shlex-walk each line — both Pattern A and the walker section ship that splitter (split_shell_lines, production-proven in qlmanage-guard). What even it cannot parse is a heredoc body (not quote syntax); when to accept that residual is #11's call.
• But shlex isn't a silver bullet, and what you detect changes whether fail-open is safe. shlex.split() itself throws ValueError on an unbalanced quote — a multi-line git commit -m "… message with a # or an unclosed quote is the classic trigger. The except ValueError: cmd.split() fallback then allows, which is right when you're detecting a banned modifier (does this carry --no-verify? — missing it errs safe, Rule 1's direction), but dangerous when you're detecting whether the command IS your target at all (is this a git commit? — a ValueError there means the guard never recognises the commit and silently doesn't fire; a real cross-domain commit shipped with no confirmation dialog this way). For the is-this-the-command decision, prefer a narrow regex (git and commit as separate words, any flag tokens between) that's immune to multi-line-quote breakage; reserve the shlex walker for the command-position / modifier checks where fail-open is the safe direction. (The boundary: regex when the predicate is "is this a specific common command at all" — git commit, git push — whose own message/arguments are what breaks tokenizing; walker when the predicate is "is a banned command or modifier in command position" — there the banned thing is rare and a ValueError fail-open errs safe, Rule 1's direction.)

2. Test with bash -n + a real JSON event, end-to-end, BEFORE registering

A corrupted or wrong-logic PreToolUse hook poisons the entire session — every later Bash call gets truncated / duplicated / falsely-failed / looks hallucinated-executed, and you'll blame "the environment" when it's the hook you just installed. (2026-07-05: a [^;&|] regex broke in one edit, ;& became a bash case-fallthrough token, poisoned half a session until bash -n found it.) "My tests passed at deploy" isn't enough — the file can corrupt in a later edit.

Gate before registering ANY hook:

bash -n hook.sh                                # syntax
printf '%s' '{"tool_name":"Bash","tool_input":{"command":"<trigger case>"}}'    | ./hook.sh; echo "exit=$?"  # want 2
printf '%s' '{"tool_name":"Bash","tool_input":{"command":"<healthy lookalike>"}}'| ./hook.sh; echo "exit=$?"  # want 0

Bundle the harness: scripts/test_hook.sh runs a whole table of trigger/allow cases. Self-block gotcha: once the hook is live in the session you cannot test it by putting the trigger string in your own Bash command — the live hook blocks your test command. Put the cases in a script file and run bash test_hook.sh; the outer command doesn't contain the trigger, so it isn't self-blocked.

Once a hook has caused one real incident (a false-block or a silent miss), solo re-reading the code is not enough — a same-day rewrite of a Stop-hook guard was itself re-broken twice by the author while fixing the first bug (a quote inside a Python comment, invisible on re-read, only surfaced by running the actual failing JSON case). The escalation is a multi-lens agent-team review where every finding must be reproduced by executing a real payload against the live script, not by reading the code and agreeing — this is the general Counter Review methodology (skill-creator's skill-development-methodology reference, Phase 6), applied to a hook instead of a skill. In one such pass, 3 lenses (matching logic / shell-embedding safety / event-contract robustness) surfaced 13 confirmed, independently-reproduced bugs and 1 finding whose own cited evidence turned out to be a hallucinated doc quote — caught only because the verifier was required to curl the raw source and grep for the exact string rather than trust the citation.

3. SSOT + symlink so a reinstall can't silently disarm the guard

Real script in a version-controlled dir, symlinked into the hooks dir Claude reads:

~/scripts/claude-hooks/<name>.sh      # SSOT (this setup: a private git repo)
~/.claude/hooks/<name>.sh             # symlink → SSOT

# install / recover:
ln -s ~/scripts/claude-hooks/<name>.sh ~/.claude/hooks/<name>.sh

A ~/.claude reinstall wipes the hooks dir; the symlink target survives, and recovery is one ln -s. A dangling symlink disables a Tier-0 guard with zero signal — which is why a SessionStart health check exists (rule 4; runnable skeleton: Pattern C in references/hook_patterns.md).

4. Registration is per-profile — converge ALL profiles, release via a HUMAN gate

• There are two registration surfaces, and the health check probably only watches one. Besides the profile settings below, a repo's own .claude/settings.json (and settings.local.json) registers hooks for sessions in that repo. Those live outside ~/.claude/hooks/, so a guard-rail check that walks that directory covers none of them — syntax, path, --selftest, nothing. Two silent failures shipped in one such file and survived two months under a SessionStart health check built to prevent exactly that (#38, #39). Register project hooks by absolute path — a relative one resolves against the session cwd and breaks the first time someone starts claude from a subdirectory — and extend the health check to walk up from the event's cwd for project settings.
• A hook in ~/.claude/hooks/ protects nothing if the active profile's settings.json doesn't call it. Multi-profile users ran with zero guards until every profile was converged. Register in the main profile's settings (~/.claude/settings.json in this setup; the Registration section of references/hook_patterns.md has the exact jsonc shape) — PreToolUse → matcher Bash → your hook — then converge the rest (this setup uses sync-profile-settings.py --all, owned by the claude-switch-models-setup skill). A SessionStart health check greps each profile for the Tier-0 guards to catch drift. Settings edits are picked up by the CLI's file watcher (official hooks docs), so registration is live without a restart — confirm by watching the guard fire on a safe probe, or at next session's health-check line.
• If the guard needs a release valve, make it a human gate, not an env var. A static GUARD_OK=1 escape hatch is no gate — the model can set the env var itself. Use a native macOS dialog (osascript — the model can't click); refuse/cancel/timeout = hard NO; log every prompt/bypass to an audit file. Pattern in references/hook_patterns.md.
  • ⚠️ /dev/tty is not a second channel — the docs say hooks cannot open it. This file used to prescribe a typed YES on /dev/tty alongside the dialog. The official reference is explicit: hooks "run in their own session without a controlling terminal", and "the hook process and any child processes can't open /dev/tty" (terminalSequence is the documented replacement for writing to it). So a "two-channel" gate built that way is one channel plus dead code, and on a box with no GUI session the gate can never be approved by anyone. Consistent with local observation, though read the boundary carefully: in one setup's shared audit log — 1,801 entries, several guards writing to it, three of which implement a tty channel — 360 lines carry a channel tag (236 dialog confirmations, 124 declines or timeouts) and not one line of any kind names the tty channel. That means the tty branch was never entered, which on macOS is what you would predict anyway, because the dialog answers first and short- circuits it. So the log shows nothing here ever depended on tty; it is the documentation, not this measurement, that establishes tty cannot work at all. Keep the dialog; if you need a non-macOS gate, you need a channel this file does not yet have a verified answer for.
  • ⚠️ A human gate that outlives the hook timeout fails OPEN. Hook command timeout defaults to 600s (30s on UserPromptSubmit), and a timed-out hook does not block the tool call — so an unanswered dialog does not become a "no", it becomes an allow. Bound your wait well under the timeout and make no-answer resolve to block yourself, before the harness resolves it for you.
  • The docs also carry an in-UI channel — PreToolUse hookSpecificOutput permissionDecision: "ask", which prompts through Claude Code's own interface. It is worth knowing about, but unverified here under bypassPermissions / auto-accept, which is precisely the mode a Tier-0 gate must survive; the dialog is prescribed because it does not depend on permission mode.
  • Below Tier-0, where a model-serviceable escape is allowed, make it the correct usage rather than a bypass flag. The rule above is absolute for Tier-0 and does not bend here — this is about the correctness guards that fall short of it, which still need a way out for the legitimate case the detector cannot distinguish. The question is what you make that way out be. A SKIP=1 / --force env escape trains exactly the reflex rule 1 warns about, and under deadline it is indistinguishable from a bypass. Prefer an escape that is the thing you wanted them to do anyway, so taking it improves the command instead of disarming the guard: pipe-fallback-guard exits 0 the moment the command mentions pipefail / PIPESTATUS / pipestatus, because an author who wrote any of those has already demonstrated they understand pipeline exit codes — the guard has nothing left to teach them. The test to apply: if someone takes my escape hatch, is the resulting command better, or merely unblocked? If the honest answer is "merely unblocked", you have a bypass flag with a nicer name. Sibling principle for the Tier-0 case, where the override exists only so the gate can be tested: the escape hatch may only make the gate stricter — see the GIT_GUARD_TEST discussion in references/hook_patterns.md.

5. Decide the failure direction, and test that — not just the happy path

Rule 1 ranked detection-tuning errors: given that the guard ran, false-blocking a healthy command beats missing a rare bad one, because a guard people must bypass gets bypassed reflexively. This rule is about a different axis — the guard's machinery not running at all — so "which is worse" is not being reversed here; the two rankings never meet. A tuning miss costs you one case; this costs you the guard, silently, on every input of that shape.

The failure: the guard cannot obtain the thing it judges on — a parse throws, a path doesn't resolve, a dependency is missing, a subprocess times out — and the very 2>/dev/null || true that stops the hook from crashing quietly converts "I could not check" into "nothing to report." The hook exits 0. That output is identical to a real pass, which is why this survives for weeks.

So at every point where the hook obtains something (parses the command, reads staged files, queries a service), decide explicitly: if this comes back empty, does that mean allow or block? — and write the answer next to the branch. Fail-open is often right for a modifier check (does this carry --no-verify? missing it costs you one case). Fail-closed is usually right for the is-this-even-the-thing check (is this a cross-domain commit? an empty answer means the guard never fired at all).

Then test the direction, not the happy path: hand it an unresolvable path or an unparseable command on purpose and assert it still does what you decided. A suite where every row passes because the hook silently allowed everything is indistinguishable from a suite that passes.

Read those results carefully — the same input has opposite correct answers for different guard classes. Take cd ~/no-such-dir && TRIGGER:

So decide which class your hook is before writing the row, and the harness's unresolvable path template row expects 2 because that template targets the token-matcher class. Getting this backwards produces a confident FAIL against a correct guard. For a state-deriving guard the failure you are hunting is: the command would really have run and the guard didn't see it — an unbalanced quote makes tokenizing throw, the fallback allows, and a genuine cross-domain commit ships with no dialog (rule 1's ValueError note). Ask of every allowed row: would this command actually have done the thing? If no, the allow is correct.

Running this exact probe against a real state-deriving guard returned two allows on the first pass: one was correct (the short-circuit above) and one was a genuine fail-open. The probe finds things; you still have to classify what it found — which is why the class table above comes before the rows.

Real case (2026-07-22): a scope guard read staged files via git -C "$REPO_DIR" with REPO_DIR parsed out of the command text — so cd ~/repo && git commit handed it a literal ~/repo, git -C failed, staged came back empty, and the guard concluded "no cross-domain files, allow." Every cross-repo commit went unguarded and nothing ever looked wrong. Anatomy + the shared-library twist: pitfall #10.

That parser has a second failure direction, and it is the nastier one. Once you add a fallback so it stops failing open, the fallback becomes correct for one reason and wrong for another — and both print the same line. git push (no explicit target) legitimately falls back to the event's cwd; git -C "$R" push names a target the hook cannot resolve, falls back to the same cwd, and then renders a confident ✅ about a different repository. Those two cases render byte-identically (measured, MD5-equal), so neither the hook nor the reader can tell the honest verdict from the misbound one. A fallback value must carry the reason it was chosen, and only "no explicit target" earns a verdict. Full anatomy, the confused-deputy framing, and why fixtures with literal paths never catch it: pitfall #28.

6. Judge on a fact the world can answer — never on your own rendering, never on a naming habit

Rules 1 and 5 are about how you match and which way you fail. This one is about where the thing you match on came from, and it has two failure shapes that both go silent:

• Never branch on a string you formatted for a human. If the hook builds a report — sorted, joined, truncated to the first N with a (+M more) tail — and then pattern-matches its own decision against that report, the branch inherits the rendering's losses. Items past the cutoff simply do not exist to it, so the branch works on every small fixture and stops firing on exactly the large sessions it was built for. Emit the machine fact on its own channel (one untruncated KINDS:a,b,c line) and match that. A rendering is an output, not a data source (pitfall #12).
• Prefer a checkable fact over a naming convention. Classifying by path shape (/skills?/[^/]+/references/) encodes one directory layout; a repo laid out any other way is classified None — silently, forever. The fix is not to widen the pattern, which trades a silent miss for machine-wide false positives (rule 1 forbids exactly that trade); it is to ask a question the filesystem can answer — is there a SKILL.md beside this references/ directory? Facts survive layout changes; conventions do not. (When the candidate is a SKILL.md, there is no sibling to ask about — classify by basename; #13 explains why that is a spec-defined fact and not the naming habit this rule warns against.) A checkable fact can still be the wrong fact — anchor the question and filter by type. test -f SKILL.md is true in a downloads folder too, and one guard that walked ancestors looking for exactly that swallowed an entire home directory, then told a real session to load a skill named after it — a name that cannot exist (rule 9's incident). Anchor to a sibling of a specific directory or to a known install path; an unanchored ancestor walk is a convention wearing a fact's clothes.

The tell for both: a branch that has never once fired in production while its tests are green. Print the raw pre-formatting classification and you will see which of the two you have.

7. If the hook demands remediation, prove the loop terminates

A hook that blocks (exit 2) until X is done — Stop hooks especially, since they re-fire on every subsequent stop — is not a check, it's a feedback loop. (A hook that merely injects a demand and exits 0 has no loop at all: nothing re-evaluates. That is mechanism 0 below, and it is the right default more often than people reach for it.)

condition T is true → hook demands remediation R → model performs R → T checked again

Write the Loop Contract before the first cycle — for hook-enforced loops and agent-driven review / wait / retry loops alike:

LOOP KEY: immutable logical target / lineage + one failure axis
FIRE T: the condition that starts another cycle
REMEDIATION R: the exact action one cycle performs
VARIANT V: the well-founded quantity that strictly decreases for this key
BUDGET: maximum cycles, fixed before cycle 1
SUCCESS EXIT: the observable that proves the axis is clear
CAPPED EXIT: what is left blocked / unshipped / pending when the budget ends

No completed contract means no blocking Stop hook and no repeated reviewer or polling loop. Freeze the key before cycle 1. A remediation snapshot, commit, or reviewer name stays inside that same lineage and cannot mint a new budget. A new, unrelated finding is a new key: record it separately; it does not reset this loop's budget. A cycle that cannot name a new falsifying experiment or a smaller V adds no evidence and stops.

For an agent-driven independent-review loop, the default budget is one initial review plus one narrowly scoped re-review after substantive fixes. A third reviewer is not automatic. If the re-review still reproduces a BLOCKER or MAJOR on the same axis, leave the hook unregistered / artifact unshipped, report the blocked state, and require a new user-authorized task whose Loop Contract declares its budget before cycle 1. An agent-declared budget cannot authorize itself. Inside that authorized task, name the concrete safety or business failure caused by stopping now; optional polish does not qualify.

Filled review-loop example:

LOOP KEY: <initial frozen commit>'s review lineage + termination-contract fidelity
FIRE T: fresh review reports a same-axis BLOCKER / MAJOR
REMEDIATION R: reproduce that finding, apply one bounded fix, run its narrow check
VARIANT V: 2 - completed review cycles
BUDGET: 2 cycles total (initial review + one re-review)
SUCCESS EXIT: no same-axis BLOCKER / MAJOR
CAPPED EXIT: artifact stays unregistered / unshipped; report remaining findings

Every repair descendant of the initial frozen commit remains in this key. The current snapshot changes so the reviewer can inspect the fix; the lineage and its remaining budget do not.

Nothing mechanically enforces this hookless budget — it holds only while the agent follows the Skill. That limitation is why the capped exit must be visible and must never be reported as “completed.”

If completing R can make T true again, the loop does not converge. Nothing errors, nothing crashes; it burns round after round until a human interrupts — which is what usually happens, because each round is a complete remediation cycle (dispatch, wait, adopt, edit

…(truncated)