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).

Hook types and what the exit code means

Type	Fires	Exit 0	Exit 2	Other
PreToolUse	before a tool runs	allow	block the call (stderr → shown to model as guidance)	any other exit = "non-blocking error" → the call proceeds
PostToolUse	after a tool ran	quiet unless it prints a hookSpecificOutput JSON on stdout — that is how context injection works, and it happens at exit 0	feedback to the model (can't un-run the tool)	—
SessionStart	session begins	proceed	—	always exit 0 — never block a session
Stop (+ SubagentStop)	the model is about to finish responding	let it stop	block the stop — forces the model to keep going (stderr → fed back as the reason)	loop safety: the hook checks stop_hook_active (necessary, not sufficient — rule 7). The harness's consecutive-block ceiling (default 8) is not a general backstop — its counter resets on any continuation that executed tools, so it never arrives for a hook whose remediation involves tool calls, which is most of them (#27). Carry your own bound. All Stop hooks for an event run in parallel — one block round can carry several hooks' feedback

• PreToolUse is the workhorse — the only one that can stop an action. matcher selects the tool (Bash, Agent, WebFetch, …). 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.
• 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 output. 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
INPUT=$(cat)                                   # the JSON event on stdin
TOOL=$(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   # fast path: not relevant → allow
# ... 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

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']). Use the walker in references/hook_patterns.md verbatim rather than reaching for the one-liner. 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

• 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. Replace with: a native macOS dialog (osascript — model can't click) and/or a typed YES on /dev/tty (model can't type into the user's terminal); refuse/cancel/timeout = hard NO; log every prompt/bypass to an audit file. Pattern 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:

Guard class	Judges on	Correct exit	Why
Token matcher (is this a banned command form?)	the command text alone	2, block	TRIGGER is right there in the text; an unresolvable cd doesn't make it not-a-trigger, and if the guard goes quiet here it will also go quiet on cd ~/real-dir && TRIGGER
State deriver (does the repo's staged set span domains?)	state read from disk	0, allow	cd fails, && short-circuits, no commit ever happens — there is nothing to guard
Termination-state reader (has the remediation already happened?)	a receipt / counter file (rule 7)	0, allow — when the state file IS the termination condition	an unreadable receipt means the hook cannot know it already fired; failing closed here blocks forever with no remediation possible and no human-visible cause — that is the loop, and it is the one failure worse than a missed case. Inverted sub-case — read this before copying the row: when the state is only a budget on top of an independent predicate (the block still clears by doing the work), allow-on-unreadable silently disables the entire hook — one unwritable directory makes it mute for every input, forever, which is the worst failure shape there is. There, fail back to the behavior before the budget existed (keep evaluating the predicate), not to silence. Tell the two apart with one question: if the state vanished, would remediation still be possible? No → receipt case, allow. Yes → budget case, keep checking

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

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), not a cheap retry. And that same property is why the harness's 8-consecutive-block ceiling will not save you: its counter resets on every continuation that executed tools, so a remediation cycle made of tool calls keeps it pinned at 1 forever (measured — #27). Even where it does arrive, it is a backstop against a runaway session, not a design: the turn ends with the violation still standing, and the harness reports that turn as reason:"completed" — indistinguishable from genuinely finishing. "It eventually stops" is not termination in any sense you want, and here it does not even eventually stop. stop_hook_active does not save you here — that field covers exactly one layer of re-entry ("the stop I just blocked is being retried"). It says nothing about the cross-turn case, where the model genuinely goes off and does R (real work, many tool calls), then stops naturally: that is a brand new Stop, the field is false, and the hook fires again on the same grounds.

The test, borrowed from termination proofs in program verification — a loop variant / ranking function: write down a quantity V mapping into a well-founded order (usually just ℕ), and show that V strictly decreases across every trigger → remediate → re-check cycle. No V, no termination proof — don't register the hook.

V is a design-time obligation, not code — you never compute it in the hook. What ships is the predicate (the mechanisms below); V is the argument that the predicate converges. Put it where the next reader will trip over it — the script header:

# TERMINATION: V = 1 - exists(<receipt path>)
# decreased by: R writes the receipt; nothing R does afterwards can remove it.

"Show it decreases" is three concrete questions, and the answers go in that comment:

1. What does R change? Name the exact file / field / timestamp.
2. Is that thing an operand of T? If yes, and R moves it back toward "fire" → there is no V.
3. After R, what is the smallest input that makes T true again? If the answer is "the same input I just fired on" → there is no V. Redesign the predicate; do not retune the threshold.

A real counter-example. A Stop hook required an independent review before compounding artifacts (rule files, skills, other hooks) could be pushed:

• T (the condition that makes the hook fire) = "there are edits no review has covered", implemented as the timestamp comparison last_edit > last_review (last_edit = newest mtime across the artifact set, last_review = mtime of the review record — two single numbers, which is exactly what makes the comparison feel safe)
• R = dispatch an independent reviewer

But a review that is worth running has output: its findings get adopted by the same agent, immediately, before it next tries to stop → that produces new edits → last_edit moves past last_review → T is true again. (If a human adopted them later, out of band, there would be no loop — the loop needs the remediation and the re-check inside one agent's turn, which is exactly what a Stop hook guarantees.) There is no V — remediation doesn't decrease a quantity, it resets one. The only escape is "review, then change nothing," which is precisely the case where dispatching the reviewer was pointless. Observed: three consecutive rounds, each a complete review-and-adopt cycle, exited only by the user saying stop.

Two things make this hard to see. The comparison looks perfectly reasonable in isolation — "the review must be newer than the last edit" is exactly what you'd write. And that sentence is the pass condition — T is its negation. Copy it into your head as-is, without that negation, and you are reasoning about the wrong operand for the rest of the analysis; keep T oriented as the fire condition. (Writing the code as an early-exit guard clause — … && exit 0 — is normal shell style and not what this is about; the discipline is about which orientation you reason in. And note equality: same-second mtimes land on the pass side, i.e. fail-open, which matches what this rule requires of state reads below.) Run the checklist above and it falls out mechanically: R changes last_edit (Q1); last_edit is an operand of T (Q2); the smallest input that re-fires T is the remediation's own output (Q3) → no V.

A second failure form: the predicate can't see the remediation at all (observability gap). The counter-example above is a temporal predicate that remediation moves. A quieter failure of the same family: remediation happens, but the channel the predicate reads it through doesn't exist in this environment. Real case (2026-07-26, found by a full-fleet loop audit): a Stop hook detected "an independent review happened" by scanning tool_results for agentId: <hex> and reading subagents/agent-<hex>.jsonl — correct on the main profile. Team-mode sessions use a different schema entirely (spawn receipts agent_id: <name>@session-<uuid>, deliveries as teammate_id teammate messages, files agent-a<name>-<hex>.jsonl) — zero matches, ever, so last_review stayed None forever and every compounding-edit∧push turn re-fired the demand: a false-positive loop, bounded to one block per stop sequence but unbounded across turns, and its "2/2 fires" that session were both on fully-reviewed work. Same family, different medicine: the temporal loop needs a better predicate; the observability loop needs a better channel. Add a fourth question to the checklist — Q4: in every environment this hook will run in, can the predicate actually SEE R happen? For transcript-reading hooks that means parsing a real session from each profile/mode, not fixture-testing one schema. (The repair for the case above: multi-schema detection + teammate deliveries excluded from turn boundaries so they can't truncate the detection window — pitfall #20.)

Pick by axis first, then by order — these are not five strengths of one thing. 0 decides whether to block at all; 1 decides which event to hang it on; 2–4 are the predicate's shape (choose 1 and you still need one of 2–4). The 0→4 order is "how completely the loop is removed", and it runs inversely to how much you can enforce — so take the first one that still gives you the enforcement you actually need, not simply the first one.

0. Don't block — inject. If the demand is advisory (you want the model to consider R, not to be unable to finish without it), print it and exit 0. Nothing re-evaluates, so there is no loop to prove terminating. Right default for anything short of Tier-0, and the cost is honest: a reminder can be ignored, so say in the header that it is fail-open — rule 4's point stands, a gate the subject can walk past is not a gate. If you need a gate, use 2 and pay for the receipt. Injection channel: Pattern D. ⚠️ This option does not exist on Stop — and rule 7's main subject is Stop, so read this before reaching for it. On Stop, exit 0 means "let the turn end", so there is no later reasoning step for the text to land in; and hookSpecificOutput.additionalContext counts toward the same 8-block ceiling as exit 2 (see the hook-types section), i.e. it is also a block. Stop has exactly two modes: gate, or silence. Choosing mechanism 0 on Stop therefore means changing the event — hang the injection on the tool call that produced the artifact (PostToolUse, Pattern D) — or admitting you wanted a gate after all, and going to mechanism 2. ⚠️ "No loop" holds only if R isn't your own matcher's target. An injector on Bash that tells the model to run git ls-remote fires again on that very command, and re-injects. Same shape, softer — the model can ignore it, so there is no forced iteration, but it is broadcast-on-repeat rather than nothing. Check that the R you recommend is not an action this hook matches.

1. Move the check to the action boundary. If what you want to gate is an action — a push, a publish, a delete — guard the action with PreToolUse instead of guarding the turn with Stop. Stop-hook remediation loops are often action gates attached to the wrong event, and this is the concrete case of "Stop is the odd one out, and the one most often reached for by mistake" from the hook-types section. Be precise about what this buys. PreToolUse only re-fires when the model voluntarily retries the gated action, and the model can always decline and end its turn normally. So it guarantees the turn terminates — the worst case drops from "the turn can't end" to "this action doesn't happen". It does not make a non-converging predicate converge: take the counter-example above, move it to PreToolUse unchanged, and the loop survives intact (push → blocked → review → findings adopted → new edits → retry → last_edit is ahead again → blocked). That case is sick in its predicate, not in its event, so you still pick a shape from 2–4. Note also that PreToolUse has no harness backstop — the 8-block ceiling in the hook-types table is Stop-only — so a self-resetting predicate moved here has fewer safety nets, not more. ⚠️ Two shapes where this mechanism is the wrong answer: R has to be done with the very tool you gated (a guard on Edit demanding you fix a file header first — deadlock, nothing can ever satisfy it), and an action that recurs within one session (a git push gate in a session that pushes five repos = five full demands; that is the density problem in the war story below, and mechanism 1 doesn't exempt you from it).

2. Make "already remediated" an existence fact, not a temporal one — and key it on the thing that needed remediating. Have R land an artifact and test does it exist; the key is what makes this work:

   KEY=$(git rev-parse HEAD 2>/dev/null || printf 'nogit')   # or a hash of the
   RECEIPT="${TMPDIR:-/tmp}/my-guard.${KEY}.ok"              # reviewed content
   [ -f "$RECEIPT" ] && exit 0            # V = 1 - exists, for THIS key
   

   V = 1 - exists is per key: it decreases exactly once per key and can never be pushed back up for that key. New work mints a new key — that is a new demand, not a re-arm. Both naive keyings fail: one global path makes the hook fire once per machine and then sit dead forever with zero signal, and a time-based key is the temporal predicate this rule exists to forbid. A temporal predicate is almost always the wrong shape, because the remediation you demanded is usually what moves the operand you compare against. ⚠️ If the model can create the receipt, this is rule 4's retired GUARD_OK=1 escape hatch wearing a new hat. Have it written by something the model doesn't drive (the reviewer subagent's own output file, a git note), or accept that the hook is advisory and say so in its header.

3. A ceiling on repetitions. At most N reminders per session per target — session_id is the only stable key for this (it is on every event; see the JSON contract in Pattern references):

   SID=$(printf '%s' "$INPUT" | python3 -c "import sys,json;print(json.load(sys.stdin).get('session_id','nosid'))" 2>/dev/null || echo nosid)
   CNT="${TMPDIR:-/tmp}/my-guard.${SID}.count"
   N=$(cat "$CNT" 2>/dev/null || echo 0); N=$((N+1)); printf '%s' "$N" > "$CNT"
   [ "$N" -gt 3 ] && exit 0                # V = 3 - N, reaches 0 and stays
   

   Crude, and deliberately blind to whether R actually happened — but finite, which is the property that was missing. Do not substitute $$ or $PPID: each hook run is a fresh process, so those change every invocation and the counter never accumulates. Print the count ("reminder 2 of 3") — see the war story below for why that wording earns its place.

4. Hysteresis / a cool-down window (the control-theory answer to alert flapping): after firing, suppress re-evaluation for a window — a stamp file plus [ $(( $(date +%s) - <stamp mtime> )) -lt 900 ] && exit 0 (mtime is stat -f %m on BSD/macOS, stat -c %Y on GNU — as are the other snippets here). Right for conditions that oscillate around a threshold; wrong for conditions that remediation resets — those need 2 or 3. ⚠️ Hysteresis supplies no V — it is a rate limiter, not a termination proof. The loop ends only if the condition subsides on its own, and what ends it then is the world, not your hook. So its # TERMINATION: line has to name that external fact ("by the time the stamp expires, X has been resolved by <whom>"). If you can't write that line honestly, what you needed was 2 or 3. (Family resemblance worth seeing: mechanism 0 is the limit case of both — mechanism 3 with the ceiling set to 0, or mechanism 4 with the window set to ∞. They differ in enforcement, not in termination.)

Failure direction for the state itself (rule 5): fail open. If the receipt or counter can't be read or written — unwritable TMPDIR, sandbox, full disk — allow the stop. This is the one place in this skill where fail-open is mandatory rather than a judgement call: a termination mechanism that cannot read its own state and blocks anyway is the loop, now with no human-visible cause.

Prose in the demand text does not substitute for a converging predicate. A hook whose message says "if you judge this unnecessary, just finish again" still costs a full remediation cycle every round, because a model that has been told it must do X will usually do X. The escape hatch has to be in the predicate, not in the advice.

The testing requirement, and the easiest thing here to skip: the self-test needs an "after remediation" case — not just "fires when it should," but "stops firing once R is complete." Without it, non-termination is structurally invisible: every fixture is one isolated point-in-time judgment, while non-termination is a property of the sequence. A suite that only checks single points has zero coverage of convergence no matter how many cases it has — which is how a hook can ship with a green self-test and still loop on its first real encounter. The row pair that can see it (receipt absent → fires, receipt present → quiet, with the setup/teardown a plain run row can't express) is templated in scripts/test_hook.sh under "AFTER-REMEDIATION ROWS"; symptom → cause → fix is pitfall #16.

Termination proved ≠ it feels terminated (2026-07-25 war story). A Stop hook with a correct existence-fact V fired three times in one session — each fire a legitimate new push from a different completed task, the mechanism working exactly as designed — and the user's experience was still "why is this thing stuck in a loop?" (No contradiction with mechanism 2's "nothing R does can push it back up": V is per key — three distinct keys, three separate one-way decreases. That is also the diagnostic when you can't tell which situation you are in: if each fire carries a new key, the mechanism is right and the density is the problem; if repeated fires share the same key — or the predicate has no key at all because it compares timestamps — you are in the counter-example above and the predicate needs replacing.) Three independent remediation cycles back-to-back are indistinguishable from a loop from the outside. The variant-proof settles the mechanism; it says nothing about how many distinct remediations a session can demand. If your domain produces that density (compounding artifacts ship several times a day here), consider pairing mechanism 2 (the existence fact) with mechanism 3 (a session-scoped ceiling), or accept the optics deliberately and say so in the hook's output — "reminder 2 of at most N" reads as progress, an unadorned repeat reads as a loop. (2026-07-26 sequel: the same hook's fires that looked like this density problem turned out to be 100% false positives — its review channel was schema-blind in team mode; see the observability form above. Before accepting density as "legitimate", verify the fires are evidence-based at all.)

8. Waiting needs the same proof — notifications are advisory, polling must carry a budget

Rule 7 covers loops a hook creates. The same shape recurs with no hook involved: an agent polling for an asynchronous result — a subagent's report, a background task's completion notice, a CI status. Real session (2026-07-25): subagent completion notices arrive through a mailbox that can delay or drop them; three separate agents finished their work while the notification sat undelivered, and the waiting agent burned a dozen sleep 240 + nag cycles over ~40 minutes until the human asked what it was even doing. Nothing errored; the loop just had no variant.

Rule 7's mechanisms map over — the first two directly; hysteresis has no analogue (a wait doesn't oscillate), and its slot is taken by a trap specific to waiting:

1. Poll the artifact, not the notification. If what you actually need is a result (a file, a git ref, an API state, a row in a DB), wait on that, not on "did it say it's done." The notification is a hint; the artifact is the fact. An existence check terminates the moment the fact lands, regardless of whether any message ever arrives.
2. Every wait carries a budget, chosen when the loop is written. Max rounds × interval (e.g. 3 × 4 min), and a degradation path that exists before the first sleep: do it yourself, ask the user, or mark it pending and move on to other work. "Wait indefinitely and see" is not a degradation path — it's the loop.
3. Delivery protocols are advisory, not mechanism. "Report back via SendMessage when done — silence counts as incomplete" is worth writing, but it governs whether the agent sends, not whether the mailbox delivers. Three agents with the protocol in their prompt all went silent in one session. Design the wait as if the notification may never arrive — because it may not.

Two adjacent traps, both paid for in the same session: TaskStopping a "stuck" agent that is actually mid-work — mailbox delay is not idleness; one reviewer doing 20 minutes of real corpus testing was killed as "stuck" minutes before delivering. And --dry-run-style probes of the wait itself: before concluding the other side is silent, confirm your own observation channel works (in that session, System Events window-counting returned a confident 0 for a dialog that was on screen — a permission failure masquerading as evidence).

9. Fixtures cannot tell you the false-positive rate — replay a real command corpus before you register

Rule 1 ranks which error is worse (a false block beats a missed one, because a guard people must bypass gets bypassed reflexively). Rule 2 makes you test before registering. Neither of them tells you how big your false-block surface actually is — and the test table cannot, because you wrote its inputs from the same mental model that produced the detector. Its cases carry the shapes you thought of; the shapes you didn't think of are, by construction, absent. That is not a coverage gap you can close by adding rows.

Measured, 2026-08-06. A PreToolUse/Bash guard passed a 26-case table with 5 mutations run against it, and was registered. Replayed afterwards against 11,903 deduplicated real commands harvested from 60 recent session transcripts: 143 produced candidates, the live hook blocked 46, and hand-checking all 46 found 10 wrong — 21.7% of everything it blocked (two borderline calls counted as correct; judged the other way, ~30%). Inside 39 minutes it had blocked 3 real sessions, one of them told to load a skill that cannot exist — the third defect below, surfacing as remediation guidance that points at nothing. It was removed the same hour. All three root defects sat in the parsing layer: a line-continuation token read as a separator, a segment scan that counted heredoc bodies and data arguments as execution, and a path classifier with no type filter.

Two conclusions that story does not license. It is not "mutation testing doesn't work": a later audit of that same suite found 5 pieces of the hook's logic that could be deleted with all 26 rows still green, and one row its author had annotated as "fixed from decorative" was still decorative — "I ran mutation testing" is not "the mutation testing was right", and the replay is what exposed both. And it is not "you could not have known": three of those defects were hand-rolled reimplementations of code this file already ships (split_shell_lines, the command-position walk, is_git_write's handling of -C and its argument). Rule 1's use the walker verbatim was the cheaper fix that got skipped; replay is the backstop, not the first line.

The method — four steps, and step 2 is the one that gets skipped:

1. Harvest. Session transcripts live at ~/.claude/projects/<encoded-cwd>/<session-id>.jsonl, plus any archives registered in ~/.claude/history-sources.json. Commands are .message.content[] entries with type == "tool_use" and name == "Bash", in field .input.command — not the hook event's .tool_input.command, which is a different shape. Dedupe, and take enough transcripts that your own recent working shapes are in there (that run used 60).
2. Pre-filter with the shipped detector, sliced out verbatim — never a hand-written equivalent, or you measure the agreement between two of your own guesses. The pre-filter exists only for cost (11,903 → 143); step 3 is what decides. If the detector is not a liftable block — shell, or split across a sourced library — make it one; a detector you cannot run standalone is also one you cannot unit-test.
3. Feed each candidate to the real hook, with its own real transcript and session_id. A guard that reads session state answers differently under a fabricated context. Event contract: references/hook_patterns.md. Two things will bite — run under a scratch TMPDIR, or rule 7's receipts and per-session counters write into real sessions and silence the guard partway through your own measurement; and if the hook has a human gate, drive it through Pattern B's forced-decline path rather than answering 143 dialogs.
4. Hand-check every block — the block list is the false-positive measurement. The allow list answers rule 5's question instead (did it go quiet because nothing was there, or because it could not see?), and a replay returning zero blocks makes the harness a suspect rather than a clean bill — pitfall #11 prescribes this same instrument in the under-firing direction.

What to do with the number. A false block whose remediation guidance is wrong or impossible → do not register at all: that shape is the manufacturing process for the reflexive bypass rule 1 exists to prevent. Otherwise treat the block list as a fix list and re-replay. Expect the false positives to cluster on whatever you were doing while you wrote the guard — half of that run's landed on hook-development files, because its author was building hooks that week. And scan the block list specifically for ops actions (edits under the hooks dir, bash -n on a hook, the guard's own SSOT): a guard that blocks its own removal cannot be switched off from inside a session (#25 — that one blocked the very command that unregistered it).

Sizing, so this doesn't read as a research project: one harvest plus one loop, minutes of wall time.

Build order (in sequence)

1. Confirm it's a real recurrence, not hypothetical — else don't build it. If the hook will demand a remediation rather than just block, write its termination variant V into the script header as a # TERMINATION: line (rule 7) before any logic — and first check whether the thing you're gating is an action, in which case a PreToolUse guard on that action removes the loop instead of taming it. Can't name a quantity that strictly decreases per trigger → remediate → re-check cycle? The design is non-terminating — fix the design, not the regex.
2. Write the script in the SSOT dir; chmod +x.
3. Detection with shlex token-level matching (rule 1), keyed on a fact the world can answer rather than your own rendering or a naming convention (rule 6).
4. bash -n + test_hook.sh with trigger AND healthy-lookalike cases (rule 2) — do not register until green. Include the shapes that carry an unexpanded path (cd ~/elsewhere && …, rule 5); if the hook has a human gate, a forced-decline row (Pattern B, "Make the gate testable"); and if it demands remediation, the after-remediation row pair — fires without the receipt, quiet with it (template in scripts/test_hook.sh; rule 7 — point-in-time fixtures structurally cannot see non-termination).
5. Replay a real command corpus and hand-check every block (rule 9) — this measures the false-block surface, which the fixture table in step 4 structurally cannot. Slice the shipped detector out verbatim to pre-filter; feed each candidate to the real hook with its own real transcript and session_id, under a scratch TMPDIR.
6. Symlink into ~/.claude/hooks/ (rule 3).
7. Register in main settings.json + converge profiles (rule 4).
8. For a Tier-0/irreversible action, add the human-confirmation release gate (rule 4).
9. Persist: commit the SSOT to its private repo. Optionally add a CLAUDE.md line (prose says why + the alternative; the hook enforces).

Known pitfalls (read before debugging a misfiring hook)

Full catalog with symptom → cause → fix: references/hook_pitfalls.md. Headliners: stdin consumed by a python3 - <<PY heredoc (hook silently allows everything), awk-split false-blocks (rule 1), corrupted hook poisoning the session (rule 2), a quote or backtick inside a Python comment silently corrupting a python3 -c "…" block with no syntax error (pitfall #9 — use the quoted-heredoc form from Pattern E instead), static env escape hatch (rule 4), multi-profile under-registration, a commit message reaching the walker as pseudo-command-text and false-blocking your own fix commit unless git write segments are exempted (#7), and a path parsed from command text keeping its literal ~ so the guard fails open with no symptom at all (#10 — the one you cannot wait to notice, because silence is its only sign), a branch reading the hook's own truncated display string (#12) or keyed on a naming convention this repo doesn't follow (#13) — both invisible while the suite asserts only exit codes (#14) — command text that merely contains a redirect counted as a write (#15), and a hook whose demanded remediation re-arms it, looping with a green self-test because point-in-time fixtures structurally cannot see non-termination (#16, rule 7).

The harness is the hidden variable — use scripts/test_hook.sh, don't hand-roll one. Every hand-rolled failure mode below produces the same output as a clean pass, so it reads as success (2026-07-22, three in one sitting while fixing a Stop hook's whitelist):

1. Wrong event shape. A Stop hook reads last_assistant_message / transcript_path, not tool_name/tool_input. Feed a PreToolUse-shaped event and it finds no text → exits 0 → "no false blocks!"
2. JSON quoting. '{\"a\":1}' inside single quotes emits a literal backslash-quote; json.loads throws, the hook's 2>/dev/null || exit 0 swallows it, every case "passes".
3. A test case the rule legitimately exempts. The baseline string used a string the rule deliberately exempts (the guard flagged coined nicknames of the form <name> Group, but exempted the ordinary phrases in the group / group chat — and the baseline row happened to use one of those). The one row meant to prove the guard still bites didn't bite, and the whole suite read green.

And if the hook's product is its message, exit codes cannot test it. A blocking hook's contract is mostly its exit code, so run rows cover it. But a hook that exists to say something — a PreToolUse explanation of the correct alternative, a Stop reminder — has a second output channel the codes never see: break the wording, invert a conditional paragraph, let a heredoc swallow a section, and the exit code stays exactly 2 while every row passes. Add says <label> <event> <pattern> <yes|no> rows from scripts/test_hook.sh, asserting both polarities across two fixtures (present for the input it targets, absent for the lookalike it skips) — a lone want=no passes vacuously when the hook prints nothing at all, so it only means something beside a want=yes row proving the hook speaks. Match fixed strings, not regexes: the phrases worth asserting often contain brackets, and as a BRE [skill] is a character class matching any text with an s, k, i or l in it. Then mutate to prove the rows can die: copy the hook, inject the exact bug each row claims to catch, and confirm that row goes red. A green suite carries zero information until you have watched it fail for the right reason — two real bugs once survived a fully green 24-case suite because every row looked only at exit codes (pitfall #14).

The common shape: all-cases-agree is a smell, not a green light. test_hook.sh catches shapes 1 and 2 above structurally — it asserts an explicit expected-exit per run row (not "did it print something") and forces trigger rows alongside healthy-lookalike ones, so a trigger row that returns 0 fails loudly instead of blending in. It cannot catch #3: whether a row's content accidentally lands in an exemption is a property of what you wrote, and no harness knows your rule's intent. That one is caught only by the habit — assert a known-good trigger first, and when it doesn't fire, suspect the row before the hook.

Reference material

• references/hook_patterns.md — runnable skeletons for every hook type covered here, the shlex command-position walker, and the JSON event contract.
• references/hook_pitfalls.md — every real failure mode with symptom → cause → fix.
• scripts/test_hook.sh — end-to-end test harness; copy it next to any new hook.
• scripts/test_hook.group-name-guard.sh — a worked harness instance for a real Stop guard (event shapes, says rows, both polarities).

Maintenance — where new content goes

New incident backports land outside this file, in the place that already holds their kind: a fresh pitfall or failure anatomy → references/hook_pitfalls.md; a reusable skeleton or pattern → references/hook_patterns.md; a worked harness instance (a test script) → scripts/. This file only takes contract-level rules: content every blocking hook consumes (a new hook type, a changed exit-code contract, a new rule in the ## Rules that separate a working guard from a session-poisoning one series). The loaded-at-trigger surface stays stable while the knowledge base keeps growing; depth lives one pointer away. (The "Known pitfalls" headliners above are a highlights list, not an index — they have not been extended since pitfall #16; the numbered catalog in hook_pitfalls.md is the SSOT, and a new pitfall does not owe a headliner.)

Why this is written down (2026-08-02): backports have in fact always gone to references — what grew this file 10k→50k chars in one week was rules prose (rules 5–8 landing inline), which this policy deliberately keeps here. The policy's job is to make the default explicit for the next session holding a fresh incident, so future growth stays limited to contract-level rules. A four-frame design review (cost / SSOT / architecture / evidence, cross-examined) chose this over a structural split of the eight rules that existed then. Restart-the-split criteria, for the next time someone proposes one: a measurement (not a vibe) showing the main file's size degrades rule compliance, or the whole skill's churn settling (30 consecutive days with no new rule or backport landing anywhere).