[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill:01M7C52M0CTDW5DDP4452FJV0G":3,"skill-same:01M7C52M0CTDW5DDP4452FJV0G":236},{"skillNo":4,"name":5,"nameZh":6,"description":7,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":16,"categories":17,"keywords":18,"publishedAt":19,"linkOnly":20,"markdownContent":21,"markdownBody":22,"frontMatter":23,"refName":24,"commitSha":25,"skillDirectory":26,"forksCount":14,"repositoryHtmlUrl":27,"repositoryHomepageUrl":6,"ownerHtmlUrl":28,"ownerBlogUrl":29,"ownerDescription":6,"updatedAt":30,"install":31,"resources":44},"01M7C52M0CTDW5DDP4452FJV0G","tmp-block",null,"Guard denying durable writes to \u002Ftmp, \u002Fprivate\u002Ftmp, \u002Fvar\u002Ffolders, $TMPDIR; enforces \u003Crepo>\u002F.worktrees\u002F\u003Cname>. Triggers: tmp guard, TMP-BLOCK deny, WORKTREE-CONVENTION deny, bypass ledger. NOT for storage policy.","EtanHey\u002Fgolems","EtanHey","https:\u002F\u002Favatars.githubusercontent.com\u002Fu\u002F93338119?v=4","01M7C4SKC9HJC3B9TAX0D9MZ1E","Etan Heyman","01M7C4SKCT59Z49TTN4WWDDEYD",0,"Apache-2.0","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002FSKILL.md",[],[],"2026-11-15 14:19:17",false,"---\nname: tmp-block\ndescription: \"Guard denying durable writes to \u002Ftmp, \u002Fprivate\u002Ftmp, \u002Fvar\u002Ffolders, $TMPDIR; enforces \u003Crepo>\u002F.worktrees\u002F\u003Cname>. Triggers: tmp guard, TMP-BLOCK deny, WORKTREE-CONVENTION deny, bypass ledger. NOT for storage policy.\"\n---\n\n# tmp-block — durable staging only, NEVER \u002Ftmp\n\nHook-carried skill. The enforcing artifact is `hooks\u002Ftmp-block-pretooluse.py`\n(PreToolUse on `Write|Edit|NotebookEdit|Bash`); this page is its contract and\naudit guide.\n\n**Wired location:** `~\u002F.claude\u002Fhooks\u002Ftmp-block`, a symlink into the pinned, locked\n`golems\u002F.worktrees\u002Fhooks-live` tree (GO-5), never into a working tree, which would make\nthe live fleet-wide guard whatever branch happens to be checked out. Install with\n`scripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Cmbp|m1> --apply`; `--status` reports drift.\n**A merge does not deploy this hook; `--update --apply` moves the pin.** The legacy copy\ninstaller `scripts\u002Finstall.sh` is described in `hooks\u002FINSTALL.md`.\n\nBoth host manifests run this policy gate through the installed copy of\n`golems-fail-open.py --fail-closed` (#488). A missing, unreadable, syntax-broken\nor crashed hook produces the plain refusal: \"Blocked: a golems safety check couldn't run.\nAsk the lead to reinstall the hooks.\" Recovery uses\n`! bash ~\u002FGits\u002Fgolems\u002Fscripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Chost> --update --apply`.\nThe blocked agent cannot run Bash recovery. A human runs\nthe command from the prompt or an outside terminal (omit `!` in a terminal).\nThe installer clears the missing pin's locked registration before recreating\nit; `--update` also replaces a broken existing pin. Local damage inside\nhooks-live refuses until a human inspects it and re-runs with `--restore-live`. The prompt's `!` bypass is\nthe #411 assumption, not live-verified here. Failed-hook output\nis discarded; legitimate allow\u002Fdeny results and advisories pass through. This\ndoes not change the payload-level advisory rules below. Git-guardian's\n`pre_tool_use.py` and human-confirm are the other policy gates; other gates\nretain the default fail-open mode. A `\u002Fbin\u002Fsh` guard around the registered command denies\nwhen the pinned interpreter or the launcher cannot start; the registration\nitself must exist to enforce this boundary; lead installs through hooks-live after review\u002Fmerge.\nA hang, import-time included, is cut by the launcher's `--budget` watchdog\n(5 s manifest timeout, 4 s budget): the hook's process group is killed and the\ncall denied before the harness timeout, which would allow. Known residuals:\nregistration\u002Flauncher edits can disable enforcement, a hook subprocess that\nleaves the process group outlives the kill, and only Python-level\nstdout\u002Fstderr is captured.\n\n## Scope\n\nA PreToolUse guard. The deny covers Write\u002FEdit and Bash writes. The worktree location `\u003Crepo>\u002F.worktrees\u002F\u003Cname>` is the ratified convention.\n\n## Why it exists (S04, weave 2026-06-07 Phase-2 Fix-3)\n\nA worker ran `Write(\u002Ftmp\u002Forqi-tts-answer-msg.md)`; the era's \u002Ftmp guard\n**half-fired (validated then allowed)** — a hook validation error let the\ndurable write through, and Etan caught it live (\"Wait, why are we writing\nthose things in temp?\"). Class evidence: ≥24 \u002Ftmp anti-pattern findings,\n4 proven costs in 24h (reboot-wiped leak worksheet, a harness fix stranded in\na prunable \u002Ftmp worktree, Etan's own terminal burned by a \u002Ftmp-worktree-held\nbranch). Adversary verdict: KEEP, with the deny extended to the\n**canonicalized path-CLASS** and to **Bash writes**, \"else it's one directory\naway from useless.\"\n\n## The contract — two-valued (Etan, by voice, 2026-08-17)\n\n> *\"none of y'all would be able to write to temp, but also not ask me so we\n> don't get agent stuck.\"*\n\n**Provably outside the temp class -> ALLOW. Provably inside, or pointing at a\ntemp location -> DENY, with a reason the agent can act on. This hook NEVER\nemits a PreToolUse prompt.**\n\nGO-5 E2 (2026-09-25) softened the unknowns. A target the hook cannot read\nstatically, in a command that shows no temp hint (`mktemp`, `TMPDIR*`, `\u002Ftmp`,\n`\u002Fvar\u002Ffolders`), is ALLOWED with a `TMP-BLOCK advisory` in\n`hookSpecificOutput.additionalContext` (the channel the model reads), and so\nis a hook error on a payload with no temp hint. Denying those blocked honest\nwork (a conditional `cd … && P=\u003Cstatic path>` assignment, an unset variable)\nwithout evidence of a temp write.\n\nTemp hints (#226 r2): `mktemp`, `mkdtemp`, `tempfile`, `gettempdir`, `.tmpdir(` (Node\n`os.tmpdir()`), `DARWIN_USER_TEMP_DIR`, `TMPDIR*`, `$TMP`, `$TEMP`, `\u002Ftmp`, `\u002Fvar\u002Ffolders`.\n\n**Residual (documented, not closed):** the hint check is lexical. A command that builds\na temp path without naming one (`P=$(printf '\u002Ft''mp')`, a base64-decoded path, a value\nread from a file or from another command's output) gets the advisory, not a refusal.\nThe ledger and the temp-dir scan remain the detectors for that class; closing it\nstatically would mean refusing every unknown value again, which is what E2 removed.\n\nA prompt suspends the pane until a human answers it, and a headless Codex or\nCursor worker has no human in its pane at all — 2026-08-14\u002F17 lost hours to\n`~\u002FDocuments` probes stranding panes overnight. A deny comes back as a readable\nerror the agent reroutes around by itself, so the residual failure mode is the\nrecoverable one.\n\n| Surface | Behavior |\n|---|---|\n| `Write`\u002F`Edit`\u002F`NotebookEdit` into the class | DENY + redirect to the repo \u002F `docs.local\u002F` |\n| Bash output redirect (`>`, `>>`, `&>`), incl. heredoc+redirect | DENY |\n| Bash `tee` \u002F `tee -a` into the class | DENY |\n| `git worktree add` into the class | DENY + redirect to `\u003Crepo>\u002F.worktrees\u002F` |\n| `git worktree add` outside any `.worktrees\u002F` parent | DENY + the exact fixed command (Rule 2) |\n| `git worktree add` whose target can't be resolved | **DENY**, naming the exact resolution failure (golems#676's requirement, minus the prompt) |\n| Any target the hook cannot read statically, command shows a temp hint | **DENY** — rewrite it with a literal path, or a variable whose value has a literal head |\n| Any target the hook cannot read statically, no temp hint | **ALLOW + advisory** (GO-5 E2) |\n| Reads\u002Fdeletes (`ls`, `cat`, `grep`, `rm`, `worktree list`) | NEVER denied |\n| Hook\u002Fvalidation error on a payload that mentions a temp location | **DENY — fail CLOSED** (the S04 half-fire class) |\n| Hook\u002Fvalidation error, no temp hint in the payload | **ALLOW + advisory** (GO-5 E2) |\n| Escape hatch whose ledger can't be written | **DENY** (explicit; an unlogged bypass never proceeds) |\n| `CLAUDE_WORKER` | does NOT exempt (S04's violator was a worker) |\n| The harness session scratchpad | **ALLOW** — the one sanctioned temp location (below) |\n\n**Path-CLASS (canonicalized via realpath):** `\u002Ftmp`, `\u002Fprivate\u002Ftmp`,\n`\u002Fvar\u002Ffolders`, `\u002Fprivate\u002Fvar\u002Ffolders`, the live `$TMPDIR` value, and literal\n`$TMPDIR` tokens in Bash commands. The macOS `\u002Ftmp → \u002Fprivate\u002Ftmp` symlink is\nnot a route-around.\n\n### The one exception — the harness session scratchpad (2026-08-17)\n\nClaude Code's own system prompt hands every session a scratchpad and instructs\nit, verbatim: *\"IMPORTANT: Always use this scratchpad directory for temporary\nfiles instead of `\u002Ftmp` or other system temp directories\"*. The path it supplies\nis inside the class this guard denies:\n\n```\n\u002Fprivate\u002Ftmp\u002Fclaude-\u003Cuid>\u002F\u003Crepo-slug>\u002F\u003Csession-uuid>\u002Fscratchpad\u002F…\n```\n\nSo the harness said \"put temp files here\" and the guard said no. Observed live\n2026-08-17: brainlayerClaude took two consecutive denials arming a monitor, and\nskillcreatorClaude hit the same wall writing a test fixture the same day. Every\nagent walks into it, because they are **following instructions correctly** —\nthe same failure shape as the \"always notify on commits\" flood.\n\n**Etan's ruling, 2026-08-17: allowlist the harness session scratchpad, keep\ndenying every other temp path.** It is session-scoped and nothing durable\nbelongs there, so allowing it costs nothing Rule 1 was protecting.\n\nThe exception is structural and narrow — never a hardcoded uid or session id,\nand **not** a widening of the temp path-class. The whole\n`claude-\u003Cuid>\u002F\u003Cslug>\u002F\u003Csession-uuid>\u002Fscratchpad` chain must sit directly under a\ntemp root, and **every component is matched exactly**: `claude-\u003Cdigits>` (not\nany directory), a slug, a session UUID, and `scratchpad` (not `scratchpad-evil`\nor `myscratchpad`). The #727 review found both of those relaxations surviving\nthe suite, so all three are now pinned by deny-side tests that redden the\nmutation.\n\nIt applies to every Rule 1 write surface — `Write`\u002F`Edit`, redirects, appends,\n`tee`, heredocs — not to redirects only. Still denied:\n`\u002Fprivate\u002Ftmp\u002Fscratchpad\u002Fx.txt` (no session chain),\n`\u002Fprivate\u002Ftmp\u002Fclaude-501\u002Fx.txt` (no scratchpad component), every `$(mktemp)`\nform (a bare temp path with no chain at all), `\u002Ftmp\u002Ffoo.txt`.\n\n**`..` traversal is defended by the callers, not by the shape match.** The\npath is canonicalized *before* the shape is consulted: `in_temp_class`\nnormalizes each candidate (normpath **and** realpath) and only then asks\nwhether it is scratchpad-shaped, so `…\u002Fscratchpad\u002F..\u002F..\u002F..\u002F..\u002Fleak.txt`\narrives as `\u002Fprivate\u002Ftmp\u002Fleak.txt` and denies; `_literal_prefix_class` rejects\na literal `..` component outright before building a probe. The normalization\ninside the shape check itself is defense-in-depth for a direct caller — the\n#727 review's mutation M5 removed it and opened no hole.\n\n`scratchpad` is its own prefix-class verdict, decided independently of the\n`repo`\u002F`outside`\u002F`temp` proofs below.\n\n## Rule 2 — the worktree location convention (2026-08-09)\n\n**Ratified by Etan by voice, 2026-08-09:** the fleet worktree location is the\nin-repo **`\u003Crepo>\u002F.worktrees\u002F\u003Cname>`**.\n\nThe gap this closes is Etan's own catch, verbatim: *\"I thought the guard's job\nwas to guard it, so it actually goes the right way.\"* Rule 1's deny message\n**advised** `\u003Crepo>\u002F.worktrees\u002F` — but nothing **enforced** it, so\n`git worktree add $HOME\u002FGits\u002Ffoo.wt\u002Fbar` (durable, not temp) sailed straight\nthrough and **18 sibling `*.wt` directories** accumulated under `$HOME\u002FGits\u002F`.\n\n| Resolved target | Decision |\n|---|---|\n| has a `.worktrees` ancestor component | allow |\n| resolves anywhere else | **DENY**, naming the convention + `git -C \u003Crepo> worktree add \u003Crepo>\u002F.worktrees\u002F\u003Cname>` |\n| cannot be resolved (`$UNSET`, `$(…)`, relative after `cd`\u002F`git -C`) | **DENY**, naming the resolution failure |\n\nRule 2 silently allows a target whose literal or already-resolvable, repo-anchored\nprefix proves `.worktrees\u002F` containment before an unresolved variable suffix, such\nas `git worktree add .worktrees\u002F$wt` or `$HOME\u002FGits\u002Fgolems\u002F.worktrees\u002F$wt`. An\nunresolvable variable prefix still asks; a literal `..` escape still disqualifies\nthe proof; and lexical plus realpath containment checks still reject a\n`.worktrees` symlink that leaves the repo.\n\n**golems#676 applied.** git-guardian judged the *unexpanded literal* — it\nblocked a legitimate cleanup of a `mktemp -d` variable, then blocked the\n`gh issue create` that filed the bug because the title carried the pattern as\nprose. So Rule 2: expands `~`\u002F`$NAME` from the live environment and judges the\n**resolved** path (lexical *and* realpath, so a `.worktrees` symlink pointing\nat a sibling is not a route-around); fires only in **command position** (quoted\ntext and heredoc bodies run nothing); and **names the exact resolution failure\ninstead of blocking blind** when the target is unknowable. #676's requirement\nwas the actionable reason, not the prompt — the prompt was dropped on\n2026-08-17 and the reason kept.\n\nMigration-window hatch, mirroring Rule 1: `WEAVE_ALLOW_WT_MIGRATION=1` —\nallowed AND logged. It is location-scoped only; it never unlocks the temp class\n(that still needs `WEAVE_ALLOW_TMP`).\n\n**This is the backstop, not the only fix.** It sees `git worktree add` issued\nas a *Bash tool call*. Generators that create worktrees in-process never reach\nit — `cmuxlayer\u002Fsrc\u002Fworktree.ts:240` builds `join(homeGitsDir, ` `${repo}.wt`\n`, spec.name)`, which is where the 18 dirs actually came from, and is fixed in\nthe cmuxlayer\u002Fgolems lanes. Likewise a worktree add nested in a quoted payload\n(`bash -c \"git worktree add …\"`, an agent-spawn prompt string) is one token to\nthis parser and is not caught.\n\n## Escape hatch = bypass-detector seed\n\nOver-broad guards INDUCE route-arounds (adversary Attack 5.4), so\ngenuinely-ephemeral writes have a sanctioned path:\n\n```bash\nWEAVE_ALLOW_TMP=1 \u003Ccommand>              # temp path-class (Rule 1)\nWEAVE_ALLOW_WT_MIGRATION=1 \u003Ccommand>     # worktree location (Rule 2)\n```\n\nBoth are inline-prefix or session-exported, and both are scoped to their own\nBash simple command when used inline.\n\nEvery use is **logged** to the durable ledger\n`~\u002F.claude\u002Flogs\u002Ftmp-block-ledger.jsonl` (override: `TMP_BLOCK_LEDGER`). If the\nledger cannot be written, the bypass is DENIED — an unlogged bypass must not\nproceed. Audit the ledger during weaves: entries that look durable\n(`.md` notes, worksheets, worktrees) are route-arounds to flag.\n\n## The three class proofs (`temp` \u002F `repo` \u002F `outside`)\n\nA target with an unreadable suffix is still judged by its **literal head**,\nbecause every value it can take starts with that head. The head proves one of\nthree classes: `temp` (deny), `repo` (allow), or — since 2026-08-17 —\n`outside`, an absolute head that is provably not temp and belongs to no\nrepository. `outside` is what makes `> ~\u002FDocuments\u002Fprobe_$$.txt` allow instead\nof stranding a pane; without it the guard could prove only two classes and had\nto ask about the entire home directory.\n\nThe head is read through variables too: `P=~\u002FDocuments\u002Fx_$$.txt; … > \"$P\"` is\njudged exactly like the same path spelled inline, and `P=\u002Fprivate\u002Ftmp\u002Fx_$$.txt`\nstill reaches the hard temp-class deny. A variable with **no** literal head\n(`P=$UNSET\u002Fx`, `P=$(mktemp)`) proves nothing and is refused. Rule 2 does not\nconsume `outside`: a worktree target must still prove repo membership.\n\n## Known limits (the detection frontier)\n\nStatically uncaught: inline-interpreter writes (`python3 -c \"open('\u002Ftmp\u002F…')\"`),\n`cp`\u002F`mv`\u002F`rsync` into the class, relative paths from a temp cwd.\n\n**The guard is no longer Claude-only (corrected 2026-08-19).** The 2026-08-17\nnote here said neither Codex nor Cursor had a hook configured and both reached\n`\u002Ftmp` freely. Measurement says otherwise on both halves; the findings are summarized in\n**Cross-agent coverage** below, and the raw probe record is kept out of git in\n`docs.local\u002Fplan\u002Ftmpblock-cross-agent\u002FMEASUREMENTS.md`.\n\n**The literal-head residual (#703, widened 2026-08-17; narrowed 2026-08-20).**\nThe suffix a head proof cannot read can still affect the final path. A literal\n`..` inside a command-substitution body is now disqualifying, just as it is in\nthe outer target, so `> \u002Fdurable\u002Fhead$(printf \u002F..\u002F..\u002Ftmp\u002Fleak)` denies while a\nnon-traversing suffix such as `$(date +%s)` remains allowed. Measured history:\n\n| head | `95964ba4` | now |\n|---|---|---|\n| repo-contained traversal — `\u003Crepo>\u002Fdocs.local$(printf \u002F..\u002F..\u002Ftmp\u002Fx)` | allow | **deny** |\n| non-repo traversal — `\u002Fworkspace\u002Fgolems$(printf \u002F..\u002F..\u002Ftmp\u002Fx)` | **ask** | **deny** |\n| non-traversing suffix — `\u003Crepo>\u002Fdocs.local\u002Fx-$(date +%s).log` | allow | allow |\n\nThe remaining accepted residual is a dynamic suffix whose output cannot be\nread and whose command text contains no literal parent traversal. The\ndiscriminator for this closure is the visible `..` component, not\nrepo-containment or whether a `\u002F` separates the head from the substitution.\nBoth deny and allow sides are pinned beside each other in\n`tests\u002Ftest_worktree_anchor.py`.\n\nThe ledger + a periodic temp-dir scan for agent-shaped artifacts are the\ndetectors for this whole frontier — detectors trail the evasion frontier by\nconstruction.\n\n## Cross-agent coverage (measured 2026-08-19)\n\nThe hook is a Claude-format `PreToolUse` script, and two of the three fleet\nharnesses read that format. What differs is the **tool name** each one sends,\nwhich is why \"the hook is installed\" and \"the hook is enforcing\" were not the\nsame sentence.\n\n| Harness | Loads this hook from | Shell arrives as | File write arrives as | Status |\n|---|---|---|---|---|\n| Claude Code | `~\u002F.claude\u002Fsettings.json` | `Bash` | `Write`\u002F`Edit`\u002F`NotebookEdit` | enforcing since 2026-06-07 |\n| Cursor | `~\u002F.claude\u002Fsettings.json` — read **unconditionally**, no opt-in | `Shell` | `Write` | `Write` was **already** enforcing; `Shell` was a silent no-op until this change |\n| Codex | `~\u002F.codex\u002Fhooks.json` | `Bash` | `apply_patch` | hooks-live installer supports wiring; verify each installed seat |\n\n**Cursor was never unguarded, and never fully guarded.** Its bundle\n(2026.08.11-e8db854) carries a literal `claudeUserConfigPath` pointing at\n`~\u002F.claude\u002Fsettings.json`, and maps Claude matchers through\n`{Bash: \"Shell\", Edit: \"Write\", …}`. A live probe with no local tap got the\nverbatim `⛔ TMP-BLOCK` reason back for `Write(\u002Ftmp\u002F…)` and the file was never\ncreated — so that half has been enforcing all along. A shell redirect,\narriving as `tool_name: \"Shell\"`, missed the `tool_name != \"Bash\"` gate and was\nallowed with exit 0. `TOOL_ALIASES` closes that.\n\n**Codex has a real hook surface, and `:workspace` is not a substitute for it.**\n`default_permissions = \":workspace\"` does **not** deny `\u002Ftmp` — Codex prints\nits own writable roots at startup, and they are\n`workspace-write [workdir, \u002Ftmp, $TMPDIR]`. The sandbox does enforce (a write\nto `~\u002FDocuments` under `codex sandbox` returns `Operation not permitted`), so\n`\u002Ftmp` is granted on purpose, not by omission. Two surfaces can close it:\n\n- **the hook** — `~\u002F.codex\u002Fhooks.json`, `matcher: \"Bash|apply_patch\"`, enabled\n  by default; this guard now speaks its dialect\n  (`hookSpecificOutput.permissionDecision`) and handles its `apply_patch`\n  envelope.\n- **the sandbox** — `[sandbox_workspace_write] exclude_slash_tmp` \u002F\n  `exclude_tmpdir_env_var`, or a `[permissions]` profile with a `deny` on the\n  `slash_tmp` \u002F `tmpdir` special paths. This is *stronger* than the hook: it\n  catches `python3 -c`, `cp`, `mv` and relative-from-temp-cwd — the entire\n  detection frontier above — because it is a kernel-level sandbox rather than a\n  static parse.\n\n**Codex uses a transport adapter, not the Claude hook command directly.**\nCodex 0.160.0 ignores stdout JSON at exit 2 unless there is a blocking reason\non stderr. `scripts\u002Fhooks\u002Fcodex-policy-hook.py` runs this existing gate and\ngit-guardian, translates deliberate denials into exit-0 Codex JSON, and denies\ntransport failures with a static repair hint. The registered shell fallback\nuses stderr + exit 2 if the interpreter or adapter cannot start.\n\n`scripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Cmbp|m1> --update --apply` registers both\ngates from the selected pinned manifest into `$CODEX_HOME\u002Fhooks.json` (default\n`~\u002F.codex\u002Fhooks.json`), backs up\nchanged hook configuration, and preserves `config.toml`. It never creates trust\nhashes or overrides an existing disabled-hooks setting. Review `\u002Fhooks` from\n**plain `codex` with no `--profile`** so trust persists in base config.toml.\nrepoGolem's per-launch profile is deleted on exit, taking trust given there\nwith it. `--apply` prints this instruction; `--status` reports wiring and\nbase-config trust as missing\u002Fdisabled\u002Fpresent-unverified, with nonzero status for\nmissing\u002Fdisabled trust. A stored hash needs native review; positional hook keys\nchange when groups are reordered. Profile\u002FCLI overrides and managed requirements\ncan change effective enablement and are outside this base-config status check.\n\nA synthetic local Responses fixture exercises the real 0.160 CLI, including\ndenial before execution and permitted commands. It does not prove an installed\nauthenticated fleet seat; the lead must run that proof after deployment.\nOuter hook timeouts and skipped\u002Funtrusted hooks remain runtime limits. Codex\nomits per-call `workdir` from Bash hook input: a relative write executed in a\ntemp-class workdir can pass when session cwd is ordinary. `write_stdin` into a\npersistent shell has no hook payload, so a force push typed there is unhooked.\nBoth holes are proven on real 0.160; this adapter cannot recover missing context\nor intercept that input. Seven-second policy checks and deduplicated patch paths\nstay within the ten-second native timeout; large\u002Ftimed-out calls ask to split\nthe patch. Indented patch headers use the shared parser, and temp deletes remain\nallowed. See [version-pinned research](..\u002F..\u002F..\u002Fdocs\u002Fcodex-policy-hooks.md).\n\n## Install & evals\n\n- Install: `hooks\u002FINSTALL.md` (symlink + settings.json PreToolUse entry).\n- Evals: `hooks\u002Ftests\u002Ftest_tmp_block.py` (**280 cases**, incl. the verbatim S04\n  replay), `tests\u002Ftest_worktree_anchor.py` (**78 cases**, static anchor and\n  literal-head unit coverage), and `hooks\u002Ftests\u002Ftest_cross_agent.py`\n  (**26 cases**, the measured Cursor and Codex payload shapes and the refusal\n  dialect) — **384 together**. Rule 1: RED against the prior guard stack\n  (15\u002F17 pass-through), GREEN. Rule 2: RED 12 failing \u002F 51 passing against the\n  advise-only guard (every enforcement case exited 0 with `{}`; the\n  `.worktrees\u002F` and prose cases were already silent), GREEN.\n  Run both: `python3 -m pytest hooks\u002Ftests\u002F tests\u002F -q`.\n","# tmp-block — durable staging only, NEVER \u002Ftmp\n\nHook-carried skill. The enforcing artifact is `hooks\u002Ftmp-block-pretooluse.py`\n(PreToolUse on `Write|Edit|NotebookEdit|Bash`); this page is its contract and\naudit guide.\n\n**Wired location:** `~\u002F.claude\u002Fhooks\u002Ftmp-block`, a symlink into the pinned, locked\n`golems\u002F.worktrees\u002Fhooks-live` tree (GO-5), never into a working tree, which would make\nthe live fleet-wide guard whatever branch happens to be checked out. Install with\n`scripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Cmbp|m1> --apply`; `--status` reports drift.\n**A merge does not deploy this hook; `--update --apply` moves the pin.** The legacy copy\ninstaller `scripts\u002Finstall.sh` is described in `hooks\u002FINSTALL.md`.\n\nBoth host manifests run this policy gate through the installed copy of\n`golems-fail-open.py --fail-closed` (#488). A missing, unreadable, syntax-broken\nor crashed hook produces the plain refusal: \"Blocked: a golems safety check couldn't run.\nAsk the lead to reinstall the hooks.\" Recovery uses\n`! bash ~\u002FGits\u002Fgolems\u002Fscripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Chost> --update --apply`.\nThe blocked agent cannot run Bash recovery. A human runs\nthe command from the prompt or an outside terminal (omit `!` in a terminal).\nThe installer clears the missing pin's locked registration before recreating\nit; `--update` also replaces a broken existing pin. Local damage inside\nhooks-live refuses until a human inspects it and re-runs with `--restore-live`. The prompt's `!` bypass is\nthe #411 assumption, not live-verified here. Failed-hook output\nis discarded; legitimate allow\u002Fdeny results and advisories pass through. This\ndoes not change the payload-level advisory rules below. Git-guardian's\n`pre_tool_use.py` and human-confirm are the other policy gates; other gates\nretain the default fail-open mode. A `\u002Fbin\u002Fsh` guard around the registered command denies\nwhen the pinned interpreter or the launcher cannot start; the registration\nitself must exist to enforce this boundary; lead installs through hooks-live after review\u002Fmerge.\nA hang, import-time included, is cut by the launcher's `--budget` watchdog\n(5 s manifest timeout, 4 s budget): the hook's process group is killed and the\ncall denied before the harness timeout, which would allow. Known residuals:\nregistration\u002Flauncher edits can disable enforcement, a hook subprocess that\nleaves the process group outlives the kill, and only Python-level\nstdout\u002Fstderr is captured.\n\n## Scope\n\nA PreToolUse guard. The deny covers Write\u002FEdit and Bash writes. The worktree location `\u003Crepo>\u002F.worktrees\u002F\u003Cname>` is the ratified convention.\n\n## Why it exists (S04, weave 2026-06-07 Phase-2 Fix-3)\n\nA worker ran `Write(\u002Ftmp\u002Forqi-tts-answer-msg.md)`; the era's \u002Ftmp guard\n**half-fired (validated then allowed)** — a hook validation error let the\ndurable write through, and Etan caught it live (\"Wait, why are we writing\nthose things in temp?\"). Class evidence: ≥24 \u002Ftmp anti-pattern findings,\n4 proven costs in 24h (reboot-wiped leak worksheet, a harness fix stranded in\na prunable \u002Ftmp worktree, Etan's own terminal burned by a \u002Ftmp-worktree-held\nbranch). Adversary verdict: KEEP, with the deny extended to the\n**canonicalized path-CLASS** and to **Bash writes**, \"else it's one directory\naway from useless.\"\n\n## The contract — two-valued (Etan, by voice, 2026-08-17)\n\n> *\"none of y'all would be able to write to temp, but also not ask me so we\n> don't get agent stuck.\"*\n\n**Provably outside the temp class -> ALLOW. Provably inside, or pointing at a\ntemp location -> DENY, with a reason the agent can act on. This hook NEVER\nemits a PreToolUse prompt.**\n\nGO-5 E2 (2026-09-25) softened the unknowns. A target the hook cannot read\nstatically, in a command that shows no temp hint (`mktemp`, `TMPDIR*`, `\u002Ftmp`,\n`\u002Fvar\u002Ffolders`), is ALLOWED with a `TMP-BLOCK advisory` in\n`hookSpecificOutput.additionalContext` (the channel the model reads), and so\nis a hook error on a payload with no temp hint. Denying those blocked honest\nwork (a conditional `cd … && P=\u003Cstatic path>` assignment, an unset variable)\nwithout evidence of a temp write.\n\nTemp hints (#226 r2): `mktemp`, `mkdtemp`, `tempfile`, `gettempdir`, `.tmpdir(` (Node\n`os.tmpdir()`), `DARWIN_USER_TEMP_DIR`, `TMPDIR*`, `$TMP`, `$TEMP`, `\u002Ftmp`, `\u002Fvar\u002Ffolders`.\n\n**Residual (documented, not closed):** the hint check is lexical. A command that builds\na temp path without naming one (`P=$(printf '\u002Ft''mp')`, a base64-decoded path, a value\nread from a file or from another command's output) gets the advisory, not a refusal.\nThe ledger and the temp-dir scan remain the detectors for that class; closing it\nstatically would mean refusing every unknown value again, which is what E2 removed.\n\nA prompt suspends the pane until a human answers it, and a headless Codex or\nCursor worker has no human in its pane at all — 2026-08-14\u002F17 lost hours to\n`~\u002FDocuments` probes stranding panes overnight. A deny comes back as a readable\nerror the agent reroutes around by itself, so the residual failure mode is the\nrecoverable one.\n\n| Surface | Behavior |\n|---|---|\n| `Write`\u002F`Edit`\u002F`NotebookEdit` into the class | DENY + redirect to the repo \u002F `docs.local\u002F` |\n| Bash output redirect (`>`, `>>`, `&>`), incl. heredoc+redirect | DENY |\n| Bash `tee` \u002F `tee -a` into the class | DENY |\n| `git worktree add` into the class | DENY + redirect to `\u003Crepo>\u002F.worktrees\u002F` |\n| `git worktree add` outside any `.worktrees\u002F` parent | DENY + the exact fixed command (Rule 2) |\n| `git worktree add` whose target can't be resolved | **DENY**, naming the exact resolution failure (golems#676's requirement, minus the prompt) |\n| Any target the hook cannot read statically, command shows a temp hint | **DENY** — rewrite it with a literal path, or a variable whose value has a literal head |\n| Any target the hook cannot read statically, no temp hint | **ALLOW + advisory** (GO-5 E2) |\n| Reads\u002Fdeletes (`ls`, `cat`, `grep`, `rm`, `worktree list`) | NEVER denied |\n| Hook\u002Fvalidation error on a payload that mentions a temp location | **DENY — fail CLOSED** (the S04 half-fire class) |\n| Hook\u002Fvalidation error, no temp hint in the payload | **ALLOW + advisory** (GO-5 E2) |\n| Escape hatch whose ledger can't be written | **DENY** (explicit; an unlogged bypass never proceeds) |\n| `CLAUDE_WORKER` | does NOT exempt (S04's violator was a worker) |\n| The harness session scratchpad | **ALLOW** — the one sanctioned temp location (below) |\n\n**Path-CLASS (canonicalized via realpath):** `\u002Ftmp`, `\u002Fprivate\u002Ftmp`,\n`\u002Fvar\u002Ffolders`, `\u002Fprivate\u002Fvar\u002Ffolders`, the live `$TMPDIR` value, and literal\n`$TMPDIR` tokens in Bash commands. The macOS `\u002Ftmp → \u002Fprivate\u002Ftmp` symlink is\nnot a route-around.\n\n### The one exception — the harness session scratchpad (2026-08-17)\n\nClaude Code's own system prompt hands every session a scratchpad and instructs\nit, verbatim: *\"IMPORTANT: Always use this scratchpad directory for temporary\nfiles instead of `\u002Ftmp` or other system temp directories\"*. The path it supplies\nis inside the class this guard denies:\n\n```\n\u002Fprivate\u002Ftmp\u002Fclaude-\u003Cuid>\u002F\u003Crepo-slug>\u002F\u003Csession-uuid>\u002Fscratchpad\u002F…\n```\n\nSo the harness said \"put temp files here\" and the guard said no. Observed live\n2026-08-17: brainlayerClaude took two consecutive denials arming a monitor, and\nskillcreatorClaude hit the same wall writing a test fixture the same day. Every\nagent walks into it, because they are **following instructions correctly** —\nthe same failure shape as the \"always notify on commits\" flood.\n\n**Etan's ruling, 2026-08-17: allowlist the harness session scratchpad, keep\ndenying every other temp path.** It is session-scoped and nothing durable\nbelongs there, so allowing it costs nothing Rule 1 was protecting.\n\nThe exception is structural and narrow — never a hardcoded uid or session id,\nand **not** a widening of the temp path-class. The whole\n`claude-\u003Cuid>\u002F\u003Cslug>\u002F\u003Csession-uuid>\u002Fscratchpad` chain must sit directly under a\ntemp root, and **every component is matched exactly**: `claude-\u003Cdigits>` (not\nany directory), a slug, a session UUID, and `scratchpad` (not `scratchpad-evil`\nor `myscratchpad`). The #727 review found both of those relaxations surviving\nthe suite, so all three are now pinned by deny-side tests that redden the\nmutation.\n\nIt applies to every Rule 1 write surface — `Write`\u002F`Edit`, redirects, appends,\n`tee`, heredocs — not to redirects only. Still denied:\n`\u002Fprivate\u002Ftmp\u002Fscratchpad\u002Fx.txt` (no session chain),\n`\u002Fprivate\u002Ftmp\u002Fclaude-501\u002Fx.txt` (no scratchpad component), every `$(mktemp)`\nform (a bare temp path with no chain at all), `\u002Ftmp\u002Ffoo.txt`.\n\n**`..` traversal is defended by the callers, not by the shape match.** The\npath is canonicalized *before* the shape is consulted: `in_temp_class`\nnormalizes each candidate (normpath **and** realpath) and only then asks\nwhether it is scratchpad-shaped, so `…\u002Fscratchpad\u002F..\u002F..\u002F..\u002F..\u002Fleak.txt`\narrives as `\u002Fprivate\u002Ftmp\u002Fleak.txt` and denies; `_literal_prefix_class` rejects\na literal `..` component outright before building a probe. The normalization\ninside the shape check itself is defense-in-depth for a direct caller — the\n#727 review's mutation M5 removed it and opened no hole.\n\n`scratchpad` is its own prefix-class verdict, decided independently of the\n`repo`\u002F`outside`\u002F`temp` proofs below.\n\n## Rule 2 — the worktree location convention (2026-08-09)\n\n**Ratified by Etan by voice, 2026-08-09:** the fleet worktree location is the\nin-repo **`\u003Crepo>\u002F.worktrees\u002F\u003Cname>`**.\n\nThe gap this closes is Etan's own catch, verbatim: *\"I thought the guard's job\nwas to guard it, so it actually goes the right way.\"* Rule 1's deny message\n**advised** `\u003Crepo>\u002F.worktrees\u002F` — but nothing **enforced** it, so\n`git worktree add $HOME\u002FGits\u002Ffoo.wt\u002Fbar` (durable, not temp) sailed straight\nthrough and **18 sibling `*.wt` directories** accumulated under `$HOME\u002FGits\u002F`.\n\n| Resolved target | Decision |\n|---|---|\n| has a `.worktrees` ancestor component | allow |\n| resolves anywhere else | **DENY**, naming the convention + `git -C \u003Crepo> worktree add \u003Crepo>\u002F.worktrees\u002F\u003Cname>` |\n| cannot be resolved (`$UNSET`, `$(…)`, relative after `cd`\u002F`git -C`) | **DENY**, naming the resolution failure |\n\nRule 2 silently allows a target whose literal or already-resolvable, repo-anchored\nprefix proves `.worktrees\u002F` containment before an unresolved variable suffix, such\nas `git worktree add .worktrees\u002F$wt` or `$HOME\u002FGits\u002Fgolems\u002F.worktrees\u002F$wt`. An\nunresolvable variable prefix still asks; a literal `..` escape still disqualifies\nthe proof; and lexical plus realpath containment checks still reject a\n`.worktrees` symlink that leaves the repo.\n\n**golems#676 applied.** git-guardian judged the *unexpanded literal* — it\nblocked a legitimate cleanup of a `mktemp -d` variable, then blocked the\n`gh issue create` that filed the bug because the title carried the pattern as\nprose. So Rule 2: expands `~`\u002F`$NAME` from the live environment and judges the\n**resolved** path (lexical *and* realpath, so a `.worktrees` symlink pointing\nat a sibling is not a route-around); fires only in **command position** (quoted\ntext and heredoc bodies run nothing); and **names the exact resolution failure\ninstead of blocking blind** when the target is unknowable. #676's requirement\nwas the actionable reason, not the prompt — the prompt was dropped on\n2026-08-17 and the reason kept.\n\nMigration-window hatch, mirroring Rule 1: `WEAVE_ALLOW_WT_MIGRATION=1` —\nallowed AND logged. It is location-scoped only; it never unlocks the temp class\n(that still needs `WEAVE_ALLOW_TMP`).\n\n**This is the backstop, not the only fix.** It sees `git worktree add` issued\nas a *Bash tool call*. Generators that create worktrees in-process never reach\nit — `cmuxlayer\u002Fsrc\u002Fworktree.ts:240` builds `join(homeGitsDir, ` `${repo}.wt`\n`, spec.name)`, which is where the 18 dirs actually came from, and is fixed in\nthe cmuxlayer\u002Fgolems lanes. Likewise a worktree add nested in a quoted payload\n(`bash -c \"git worktree add …\"`, an agent-spawn prompt string) is one token to\nthis parser and is not caught.\n\n## Escape hatch = bypass-detector seed\n\nOver-broad guards INDUCE route-arounds (adversary Attack 5.4), so\ngenuinely-ephemeral writes have a sanctioned path:\n\n```bash\nWEAVE_ALLOW_TMP=1 \u003Ccommand>              # temp path-class (Rule 1)\nWEAVE_ALLOW_WT_MIGRATION=1 \u003Ccommand>     # worktree location (Rule 2)\n```\n\nBoth are inline-prefix or session-exported, and both are scoped to their own\nBash simple command when used inline.\n\nEvery use is **logged** to the durable ledger\n`~\u002F.claude\u002Flogs\u002Ftmp-block-ledger.jsonl` (override: `TMP_BLOCK_LEDGER`). If the\nledger cannot be written, the bypass is DENIED — an unlogged bypass must not\nproceed. Audit the ledger during weaves: entries that look durable\n(`.md` notes, worksheets, worktrees) are route-arounds to flag.\n\n## The three class proofs (`temp` \u002F `repo` \u002F `outside`)\n\nA target with an unreadable suffix is still judged by its **literal head**,\nbecause every value it can take starts with that head. The head proves one of\nthree classes: `temp` (deny), `repo` (allow), or — since 2026-08-17 —\n`outside`, an absolute head that is provably not temp and belongs to no\nrepository. `outside` is what makes `> ~\u002FDocuments\u002Fprobe_$$.txt` allow instead\nof stranding a pane; without it the guard could prove only two classes and had\nto ask about the entire home directory.\n\nThe head is read through variables too: `P=~\u002FDocuments\u002Fx_$$.txt; … > \"$P\"` is\njudged exactly like the same path spelled inline, and `P=\u002Fprivate\u002Ftmp\u002Fx_$$.txt`\nstill reaches the hard temp-class deny. A variable with **no** literal head\n(`P=$UNSET\u002Fx`, `P=$(mktemp)`) proves nothing and is refused. Rule 2 does not\nconsume `outside`: a worktree target must still prove repo membership.\n\n## Known limits (the detection frontier)\n\nStatically uncaught: inline-interpreter writes (`python3 -c \"open('\u002Ftmp\u002F…')\"`),\n`cp`\u002F`mv`\u002F`rsync` into the class, relative paths from a temp cwd.\n\n**The guard is no longer Claude-only (corrected 2026-08-19).** The 2026-08-17\nnote here said neither Codex nor Cursor had a hook configured and both reached\n`\u002Ftmp` freely. Measurement says otherwise on both halves; the findings are summarized in\n**Cross-agent coverage** below, and the raw probe record is kept out of git in\n`docs.local\u002Fplan\u002Ftmpblock-cross-agent\u002FMEASUREMENTS.md`.\n\n**The literal-head residual (#703, widened 2026-08-17; narrowed 2026-08-20).**\nThe suffix a head proof cannot read can still affect the final path. A literal\n`..` inside a command-substitution body is now disqualifying, just as it is in\nthe outer target, so `> \u002Fdurable\u002Fhead$(printf \u002F..\u002F..\u002Ftmp\u002Fleak)` denies while a\nnon-traversing suffix such as `$(date +%s)` remains allowed. Measured history:\n\n| head | `95964ba4` | now |\n|---|---|---|\n| repo-contained traversal — `\u003Crepo>\u002Fdocs.local$(printf \u002F..\u002F..\u002Ftmp\u002Fx)` | allow | **deny** |\n| non-repo traversal — `\u002Fworkspace\u002Fgolems$(printf \u002F..\u002F..\u002Ftmp\u002Fx)` | **ask** | **deny** |\n| non-traversing suffix — `\u003Crepo>\u002Fdocs.local\u002Fx-$(date +%s).log` | allow | allow |\n\nThe remaining accepted residual is a dynamic suffix whose output cannot be\nread and whose command text contains no literal parent traversal. The\ndiscriminator for this closure is the visible `..` component, not\nrepo-containment or whether a `\u002F` separates the head from the substitution.\nBoth deny and allow sides are pinned beside each other in\n`tests\u002Ftest_worktree_anchor.py`.\n\nThe ledger + a periodic temp-dir scan for agent-shaped artifacts are the\ndetectors for this whole frontier — detectors trail the evasion frontier by\nconstruction.\n\n## Cross-agent coverage (measured 2026-08-19)\n\nThe hook is a Claude-format `PreToolUse` script, and two of the three fleet\nharnesses read that format. What differs is the **tool name** each one sends,\nwhich is why \"the hook is installed\" and \"the hook is enforcing\" were not the\nsame sentence.\n\n| Harness | Loads this hook from | Shell arrives as | File write arrives as | Status |\n|---|---|---|---|---|\n| Claude Code | `~\u002F.claude\u002Fsettings.json` | `Bash` | `Write`\u002F`Edit`\u002F`NotebookEdit` | enforcing since 2026-06-07 |\n| Cursor | `~\u002F.claude\u002Fsettings.json` — read **unconditionally**, no opt-in | `Shell` | `Write` | `Write` was **already** enforcing; `Shell` was a silent no-op until this change |\n| Codex | `~\u002F.codex\u002Fhooks.json` | `Bash` | `apply_patch` | hooks-live installer supports wiring; verify each installed seat |\n\n**Cursor was never unguarded, and never fully guarded.** Its bundle\n(2026.08.11-e8db854) carries a literal `claudeUserConfigPath` pointing at\n`~\u002F.claude\u002Fsettings.json`, and maps Claude matchers through\n`{Bash: \"Shell\", Edit: \"Write\", …}`. A live probe with no local tap got the\nverbatim `⛔ TMP-BLOCK` reason back for `Write(\u002Ftmp\u002F…)` and the file was never\ncreated — so that half has been enforcing all along. A shell redirect,\narriving as `tool_name: \"Shell\"`, missed the `tool_name != \"Bash\"` gate and was\nallowed with exit 0. `TOOL_ALIASES` closes that.\n\n**Codex has a real hook surface, and `:workspace` is not a substitute for it.**\n`default_permissions = \":workspace\"` does **not** deny `\u002Ftmp` — Codex prints\nits own writable roots at startup, and they are\n`workspace-write [workdir, \u002Ftmp, $TMPDIR]`. The sandbox does enforce (a write\nto `~\u002FDocuments` under `codex sandbox` returns `Operation not permitted`), so\n`\u002Ftmp` is granted on purpose, not by omission. Two surfaces can close it:\n\n- **the hook** — `~\u002F.codex\u002Fhooks.json`, `matcher: \"Bash|apply_patch\"`, enabled\n  by default; this guard now speaks its dialect\n  (`hookSpecificOutput.permissionDecision`) and handles its `apply_patch`\n  envelope.\n- **the sandbox** — `[sandbox_workspace_write] exclude_slash_tmp` \u002F\n  `exclude_tmpdir_env_var`, or a `[permissions]` profile with a `deny` on the\n  `slash_tmp` \u002F `tmpdir` special paths. This is *stronger* than the hook: it\n  catches `python3 -c`, `cp`, `mv` and relative-from-temp-cwd — the entire\n  detection frontier above — because it is a kernel-level sandbox rather than a\n  static parse.\n\n**Codex uses a transport adapter, not the Claude hook command directly.**\nCodex 0.160.0 ignores stdout JSON at exit 2 unless there is a blocking reason\non stderr. `scripts\u002Fhooks\u002Fcodex-policy-hook.py` runs this existing gate and\ngit-guardian, translates deliberate denials into exit-0 Codex JSON, and denies\ntransport failures with a static repair hint. The registered shell fallback\nuses stderr + exit 2 if the interpreter or adapter cannot start.\n\n`scripts\u002Fhooks\u002Finstall-hooks.sh --host \u003Cmbp|m1> --update --apply` registers both\ngates from the selected pinned manifest into `$CODEX_HOME\u002Fhooks.json` (default\n`~\u002F.codex\u002Fhooks.json`), backs up\nchanged hook configuration, and preserves `config.toml`. It never creates trust\nhashes or overrides an existing disabled-hooks setting. Review `\u002Fhooks` from\n**plain `codex` with no `--profile`** so trust persists in base config.toml.\nrepoGolem's per-launch profile is deleted on exit, taking trust given there\nwith it. `--apply` prints this instruction; `--status` reports wiring and\nbase-config trust as missing\u002Fdisabled\u002Fpresent-unverified, with nonzero status for\nmissing\u002Fdisabled trust. A stored hash needs native review; positional hook keys\nchange when groups are reordered. Profile\u002FCLI overrides and managed requirements\ncan change effective enablement and are outside this base-config status check.\n\nA synthetic local Responses fixture exercises the real 0.160 CLI, including\ndenial before execution and permitted commands. It does not prove an installed\nauthenticated fleet seat; the lead must run that proof after deployment.\nOuter hook timeouts and skipped\u002Funtrusted hooks remain runtime limits. Codex\nomits per-call `workdir` from Bash hook input: a relative write executed in a\ntemp-class workdir can pass when session cwd is ordinary. `write_stdin` into a\npersistent shell has no hook payload, so a force push typed there is unhooked.\nBoth holes are proven on real 0.160; this adapter cannot recover missing context\nor intercept that input. Seven-second policy checks and deduplicated patch paths\nstay within the ten-second native timeout; large\u002Ftimed-out calls ask to split\nthe patch. Indented patch headers use the shared parser, and temp deletes remain\nallowed. See [version-pinned research](..\u002F..\u002F..\u002Fdocs\u002Fcodex-policy-hooks.md).\n\n## Install & evals\n\n- Install: `hooks\u002FINSTALL.md` (symlink + settings.json PreToolUse entry).\n- Evals: `hooks\u002Ftests\u002Ftest_tmp_block.py` (**280 cases**, incl. the verbatim S04\n  replay), `tests\u002Ftest_worktree_anchor.py` (**78 cases**, static anchor and\n  literal-head unit coverage), and `hooks\u002Ftests\u002Ftest_cross_agent.py`\n  (**26 cases**, the measured Cursor and Codex payload shapes and the refusal\n  dialect) — **384 together**. Rule 1: RED against the prior guard stack\n  (15\u002F17 pass-through), GREEN. Rule 2: RED 12 failing \u002F 51 passing against the\n  advise-only guard (every enforcement case exited 0 with `{}`; the\n  `.worktrees\u002F` and prose cases were already silent), GREEN.\n  Run both: `python3 -m pytest hooks\u002Ftests\u002F tests\u002F -q`.\n",{"name":5,"description":7},"master","3400bc971628f00ab7bf1063993477e211a5508a","skills\u002Fgolem-powers\u002Ftmp-block","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems","https:\u002F\u002Fgithub.com\u002FEtanHey","etanheyman.com","2026-11-15 03:05:08",{"source":8,"skillName":5,"commands":32,"prompt":42,"zipAvailable":43},[33,36,39],{"tool":34,"command":35},"npx","npx skills add EtanHey\u002Fgolems --skill tmp-block",{"tool":37,"command":38},"bunx","bunx skills add EtanHey\u002Fgolems --skill tmp-block",{"tool":40,"command":41},"pnpm","pnpm dlx skills add EtanHey\u002Fgolems --skill tmp-block","安装技能，技能地址：https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002FSKILL.md",true,[45,51,57,61,65,69,73,77,81,85,89,93,97,101,105,109,113,117,121,125,129,133,137,141,145,149,153,157,161,165,169,173,177,181,185,189,193,197,201,205,209,213,217,220,224,228,232],{"path":46,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":49,"sourceHtmlUrl":50},"conftest.py","SCRIPT","脚本",257,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fconftest.py",{"path":52,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":55,"sourceHtmlUrl":56},"hooks\u002FINSTALL.md","REFERENCE","参考文档",8111,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002FINSTALL.md",{"path":58,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":59,"sourceHtmlUrl":60},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002F__init__.py",56,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002F__init__.py",{"path":62,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":63,"sourceHtmlUrl":64},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases01.py",20351,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases01.py",{"path":66,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":67,"sourceHtmlUrl":68},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases02.py",17516,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases02.py",{"path":70,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":71,"sourceHtmlUrl":72},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases03.py",16412,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases03.py",{"path":74,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":75,"sourceHtmlUrl":76},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases04.py",23640,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases04.py",{"path":78,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":79,"sourceHtmlUrl":80},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases05.py",20022,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases05.py",{"path":82,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":83,"sourceHtmlUrl":84},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases06.py",10779,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases06.py",{"path":86,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":87,"sourceHtmlUrl":88},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases07.py",17979,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcases07.py",{"path":90,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":91,"sourceHtmlUrl":92},"hooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcommon.py",10711,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002F_test_tmp_block_parts\u002Fcommon.py",{"path":94,"resourceType":53,"resourceTypeDesc":54,"sizeBytes":95,"sourceHtmlUrl":96},"hooks\u002Ftests\u002Ffixtures\u002Fvariable-state-differential.json",20640,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ffixtures\u002Fvariable-state-differential.json",{"path":98,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":99,"sourceHtmlUrl":100},"hooks\u002Ftests\u002Ftest_authoritative_path.py",1224,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_authoritative_path.py",{"path":102,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":103,"sourceHtmlUrl":104},"hooks\u002Ftests\u002Ftest_callback_seams.py",4374,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_callback_seams.py",{"path":106,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":107,"sourceHtmlUrl":108},"hooks\u002Ftests\u002Ftest_cross_agent.py",13484,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_cross_agent.py",{"path":110,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":111,"sourceHtmlUrl":112},"hooks\u002Ftests\u002Ftest_impl_layout.py",12412,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_impl_layout.py",{"path":114,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":115,"sourceHtmlUrl":116},"hooks\u002Ftests\u002Ftest_ledger_isolation.py",2228,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_ledger_isolation.py",{"path":118,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":119,"sourceHtmlUrl":120},"hooks\u002Ftests\u002Ftest_tmp_block.py",548,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_tmp_block.py",{"path":122,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":123,"sourceHtmlUrl":124},"hooks\u002Ftests\u002Ftest_variable_state_differential.py",4777,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftests\u002Ftest_variable_state_differential.py",{"path":126,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":127,"sourceHtmlUrl":128},"hooks\u002Ftmp_block_impl\u002F__init__.py",6539,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002F__init__.py",{"path":130,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":131,"sourceHtmlUrl":132},"hooks\u002Ftmp_block_impl\u002Fanchors.py",8545,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fanchors.py",{"path":134,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":135,"sourceHtmlUrl":136},"hooks\u002Ftmp_block_impl\u002Fassignments.py",7587,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fassignments.py",{"path":138,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":139,"sourceHtmlUrl":140},"hooks\u002Ftmp_block_impl\u002Fbypass.py",6713,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fbypass.py",{"path":142,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":143,"sourceHtmlUrl":144},"hooks\u002Ftmp_block_impl\u002Fchain_status.py",1692,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fchain_status.py",{"path":146,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":147,"sourceHtmlUrl":148},"hooks\u002Ftmp_block_impl\u002Fcompounds.py",12402,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fcompounds.py",{"path":150,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":151,"sourceHtmlUrl":152},"hooks\u002Ftmp_block_impl\u002Fpolicy.py",4459,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fpolicy.py",{"path":154,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":155,"sourceHtmlUrl":156},"hooks\u002Ftmp_block_impl\u002Fprefixes.py",8119,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fprefixes.py",{"path":158,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":159,"sourceHtmlUrl":160},"hooks\u002Ftmp_block_impl\u002Fresolution.py",1262,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fresolution.py",{"path":162,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":163,"sourceHtmlUrl":164},"hooks\u002Ftmp_block_impl\u002Fruntime.py",398,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fruntime.py",{"path":166,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":167,"sourceHtmlUrl":168},"hooks\u002Ftmp_block_impl\u002Fscope.py",3964,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fscope.py",{"path":170,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":171,"sourceHtmlUrl":172},"hooks\u002Ftmp_block_impl\u002Fshell_targets.py",7311,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fshell_targets.py",{"path":174,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":175,"sourceHtmlUrl":176},"hooks\u002Ftmp_block_impl\u002Fshell_words.py",1852,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fshell_words.py",{"path":178,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":179,"sourceHtmlUrl":180},"hooks\u002Ftmp_block_impl\u002Ftool_targets.py",5365,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Ftool_targets.py",{"path":182,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":183,"sourceHtmlUrl":184},"hooks\u002Ftmp_block_impl\u002Fvariable_builtins.py",4587,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fvariable_builtins.py",{"path":186,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":187,"sourceHtmlUrl":188},"hooks\u002Ftmp_block_impl\u002Fvariables.py",13534,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fvariables.py",{"path":190,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":191,"sourceHtmlUrl":192},"hooks\u002Ftmp_block_impl\u002Fwords.py",7975,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fwords.py",{"path":194,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":195,"sourceHtmlUrl":196},"hooks\u002Ftmp_block_impl\u002Fworktree_args.py",2624,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fworktree_args.py",{"path":198,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":199,"sourceHtmlUrl":200},"hooks\u002Ftmp_block_impl\u002Fworktrees.py",8151,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fworktrees.py",{"path":202,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":203,"sourceHtmlUrl":204},"hooks\u002Ftmp_block_impl\u002Fwrite_targets.py",13516,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp_block_impl\u002Fwrite_targets.py",{"path":206,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":207,"sourceHtmlUrl":208},"hooks\u002Ftmp-block-pretooluse.py",30116,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fhooks\u002Ftmp-block-pretooluse.py",{"path":210,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":211,"sourceHtmlUrl":212},"scripts\u002Finstall.sh",640,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fscripts\u002Finstall.sh",{"path":214,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":215,"sourceHtmlUrl":216},"scripts\u002Fprobes.sh",1997,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Fscripts\u002Fprobes.sh",{"path":218,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":59,"sourceHtmlUrl":219},"tests\u002F_test_worktree_anchor_parts\u002F__init__.py","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Ftests\u002F_test_worktree_anchor_parts\u002F__init__.py",{"path":221,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":222,"sourceHtmlUrl":223},"tests\u002F_test_worktree_anchor_parts\u002Fcases01.py",23654,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Ftests\u002F_test_worktree_anchor_parts\u002Fcases01.py",{"path":225,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":226,"sourceHtmlUrl":227},"tests\u002F_test_worktree_anchor_parts\u002Fcases02.py",5230,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Ftests\u002F_test_worktree_anchor_parts\u002Fcases02.py",{"path":229,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":230,"sourceHtmlUrl":231},"tests\u002F_test_worktree_anchor_parts\u002Fcommon.py",4472,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Ftests\u002F_test_worktree_anchor_parts\u002Fcommon.py",{"path":233,"resourceType":47,"resourceTypeDesc":48,"sizeBytes":234,"sourceHtmlUrl":235},"tests\u002Ftest_worktree_anchor.py",276,"https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ftmp-block\u002Ftests\u002Ftest_worktree_anchor.py",[237,244,251,258,265,272],{"skillNo":238,"name":239,"nameZh":6,"description":240,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":241,"categories":242,"keywords":243,"publishedAt":19},"01M7C53JEF6A6SS20Q0D2YJYM8","yash","Push dependency bugs upstream. Triggers: dependency bug, workaround, monkey-patch, fork\u002Ffix. NOT app bugs.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fyash\u002FSKILL.md",[],[],{"skillNo":245,"name":246,"nameZh":6,"description":247,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":248,"categories":249,"keywords":250,"publishedAt":19},"01M7C53EPMNSZKHK47VY7WQDP0","whats-new","Diff Claude\u002FCodex\u002FCursor\u002FWispr\u002Fagy changelogs vs local setup. Triggers: whats new, changelog, version bump.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fwhats-new\u002FSKILL.md",[],[],{"skillNo":252,"name":253,"nameZh":6,"description":254,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":255,"categories":256,"keywords":257,"publishedAt":19},"01M7C53C4TVQ9G7AB3X57DACPB","weave","Orchestrator convergence: mine JSONLs, cite findings, route actions. Triggers: weave, run weave.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fweave\u002FSKILL.md",[],[],{"skillNo":259,"name":260,"nameZh":6,"description":261,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":262,"categories":263,"keywords":264,"publishedAt":19},"01M7C5342EDDN7YH53KKW511V2","unslop","Cut AI tells while keeping every receipt, number, path, and hedge intact. Triggers: unslop, cut the slop, wall of text, too verbose, rambly, make this readable. NOT for delivery cadence (see i-have-adhd), channel choice, or artifact length governance.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Funslop\u002FSKILL.md",[],[],{"skillNo":266,"name":267,"nameZh":6,"description":268,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":269,"categories":270,"keywords":271,"publishedAt":19},"01M7C52HN4Q9E1Z74Z11K23E2H","t3-visuals","Show\u002Fvisualize\u002Fchart\u002Fdiagram\u002Fcompare\u002Fstatus board in T3 with compact interactive HTML. Not standalone dashboards.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Ft3-visuals\u002FSKILL.md",[],[],{"skillNo":273,"name":274,"nameZh":6,"description":275,"descriptionZh":6,"repositoryFullName":8,"ownerLogin":9,"ownerAvatarUrl":10,"ownerAccountNo":11,"ownerDisplayName":12,"repositoryNo":13,"starsCount":14,"licenseSpdxId":15,"licenseFileUrl":6,"sourceHtmlUrl":276,"categories":277,"keywords":278,"publishedAt":19},"01M7C52CNEAH6HZ5YN3TBMS6ZY","stitch-design","Use Google Stitch for design-to-code prototypes\u002Ftokens, then implement. Triggers: Stitch, screen prototypes.","https:\u002F\u002Fgithub.com\u002FEtanHey\u002Fgolems\u002Fblob\u002Fmaster\u002Fskills\u002Fgolem-powers\u002Fstitch-design\u002FSKILL.md",[],[]]