2ea5efc15172ff6ff39d8d3a50db5ee9cbe260fe
7 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Tom Boucher
|
2ea5efc151 |
enhance(#3911): hooks declare their crash policy (#3960)
* enhance(#3911): give hooks an exit seam that needs no build ADR-3889 Phase 7 foundation. The 19 shipped enforcement hooks hold 91 of the epic's 128 terminators and cannot reach `terminateNow` today. The obvious route — requiring `gsd-core/bin/lib/cli-exit.cjs`, as gsd-agent-isolation-guard.js already does for two other modules — is rejected. That precedent carries its own warning (#3582): those files are tsc output, gitignored and absent on a raw plugin-marketplace or git-clone install, so the hook must first call ensureRuntimeBuild() to self-heal. Making the module a hook needs IN ORDER TO TERMINATE depend on a build inverts the dependency, and its failure mode is precisely the fail-open this phase exists to remove: a guard that cannot terminate cannot deny. `lint-hooks-runtime-build-seam` already encodes that concern, and Design B would have had to add an ensureRuntimeBuild() call to all 19 hooks to satisfy it. So `hooks/lib/` becomes a third emit location for cli-exit and a fifth for the registry, preserving the invariant `src/cli-exit.cts`'s own header states: it imports nothing but node:fs and its sibling registry, and the generator dual-emits that sibling alongside each copy so a relative require resolves next to whichever copy loaded it. Shipping needed no change — build-hooks.js already declares HOOKS_SUBDIRS_TO_COPY = ['lib']. Proven, not asserted: the two files are copied into an otherwise-empty tmpdir and a child process requires them and terminates — PASS exits 0, HOOK_DENY exits 2 with the payload on both stdout and stderr. That test fails the moment the hooks copy gains a require reaching outside hooks/lib/. Also fixed inline: the registry's fifth target let any `--write` test overwrite the real committed hooks/lib/exit-code-registry.js, because the test helper derived only three of the other output paths. It now redirects all five, and a regression test asserts every committed artifact is byte-identical after a redirected write. Install-tree goldens pick up the two new shipped paths across 11 runtimes — insertions only, no removals. lint:ci was green while they were stale, so this was found by regenerating rather than by a gate. Verification runs on the remote runner. Refs #3911 * enhance(#3911): declare a crash policy, and migrate the write guard Adds `hooks/lib/hook-exit.js` — the hook-facing vocabulary over `terminateNow`, hand-written because the cli-exit copy beside it is generated: allow(payload) exit 0 deny(payload, stderr?) exit 2 crash(onCrash, payload) whichever the hook DECLARED `crash()` takes the policy as a required argument with no default, which is the whole mechanism: fail-open by accident stops being expressible. A hook must name ALLOW or DENY at the call site, and an unrecognized value terminates INTERNAL rather than guessing. Fail-open stays legal; fail-open by omission does not. `gsd-write-guard.js` is the first hook migrated, all 12 sites, and it exposed a gap in the seam. `terminateNow`'s doc comment justified its fd-2 write by citing this hook's `emitBlock` — but modeled it as sending the same bytes to both streams, when `emitBlock` actually sends full JSON to stdout and only the bare `reason` string to stderr, because Kimi's hook bus feeds stderr verbatim back to the model. Migrating as written would have turned a readable sentence into a JSON blob for Kimi-backed agents. #3911 requires both "all 19 hooks terminate through terminateNow" and "no hook's effective default changes". Those are jointly satisfiable only by teaching the seam to carry a distinct stderr payload, so `terminateNow` gains an optional third argument: omitted, behavior is byte-for-byte what it was; a string is written raw, which is exactly the Kimi case. The doc comment's inaccurate claim about emitBlock is corrected in place. Proven rather than asserted: the pre-migration file is reconstructed from HEAD and driven with the same catastrophic-shrink payload as the migrated one — exit code, stdout and stderr all byte-identical. Verification runs on the remote runner. Refs #3911 * enhance(#3911): all 19 hooks terminate through the seam Migrates the remaining 18 enforcement hooks onto allow/deny/crash. An AST walk now reports zero `process.exit(` call sites across every `hooks/*.js` — down from the 91 the census measured. Each hook with an outer catch declares its policy once, at module top, with the reason that policy is right for that specific guard: a read guard that cannot scan must not block the read; a statusline that renders every prompt must degrade rather than crash; an injection scanner must not retroactively block a result already returned. Those sentences are the deliverable — they are what turns fail-open-by-accident into fail-open-on-purpose. No hook's effective default changed. Wiring exposed two defects, both fixed here rather than noted. A SECOND stdout/stderr-splitting site turned up in `gsd-workflow-guard.js`'s `emitForceAddBlock`, matching the pattern already known from the write guard — full JSON to stdout, bare reason to stderr for the Kimi bus. It uses the `stderrPayload` argument added in the previous commit, which is now carrying its second real caller rather than one special case. More seriously, `terminateNow` emitted both streams inside ONE try, so a payload that failed to serialize aborted before the stderr write ever ran. The two windsurf guards write nothing to stdout on a block and only a reason string to stderr, so `deny(undefined, reason)` exited 2 with EMPTY stderr — a deny that silently loses its reason, which is the exact "fails with success" class this epic exists to close. The streams are now emitted independently, each with its own guard, and `undefined` means "nothing to write for this stream" rather than an error. Regression tests inject a throwing write on one fd and assert the other still receives its payload; they fail against the single-try version. Byte-identity was proven per hook, not assumed: each pre-change file is reconstructed from HEAD and driven side by side with the migrated one across its normal path, its deny path, malformed stdin and empty stdin — exit code, stdout and stderr compared. Verification runs on the remote runner. Refs #3911 * enhance(#3911): harden the three shell hooks, and pin every hook's policy `gsd-phase-boundary.sh`, `gsd-session-state.sh` and `gsd-validate-commit.sh` gain `set -euo pipefail`. The expected hazard did not materialize, and that is worth recording: every intentionally-non-zero command in all three is already the condition of an `if`/`elif`, which `set -e` never fires on, and none of them reads a possibly-unset variable or pipes through a grep that may legitimately match nothing. No `|| true` guards were needed. Each hook was still checked command-by-command before the flags went in rather than after. Twenty-one before/after cases across the three hooks — disabled and enabled, planning and non-planning, missing STATE.md, malformed JSON, the Kimi payload shape, quoted and unquoted `-m`, valid and over-long Conventional Commits — all match on exit code, stdout and stderr. The hardening is shown to actually fire, not merely added: with a stubbed `node` that fails at the JSON-emit step, phase-boundary and session-state go from silently exiting 0 with empty stdout to failing visibly with the error surfaced. No such case could be constructed for `gsd-validate-commit.sh`, whose every statement already sits inside an if-condition — recorded as unproven rather than claimed. `tests/hooks-crash-policy.test.cjs` adds the per-hook coverage the issue asks for, table-driven over all 19 hooks rather than 76 hand-written cases: normal allow, deny where a deny path exists, crash-honors-the-declared-policy, and an unclosed-stdin case — the one `process.exitCode` structurally cannot serve. The deny assertions encode each hook's ACTUAL stream split rather than a uniform shape, since four of the six deliberately differ. A drift guard enumerates `hooks/*.js` and fails if a terminating hook is ever added without a row. Writing those tests surfaced two hooks that emit a block decision in their JSON body and exit 0. Both were checked rather than assumed, and neither is a fails-with-success: `gsd-read-injection-scanner.js` is PostToolUse, where the tool has already run and exit 2 has no meaning, and `gsd-cursor-subagent-start.js` follows Cursor's JSON-body protocol. They are deliberately left alone — a mechanical sweep to `deny()` would have broken exactly these two. Verification runs on the remote runner. Refs #3911 * fix(#3838): the commit validator says when it could not validate #3911 claims to subsume #3838. Measurement said otherwise, so this closes it for real rather than by assertion. `set -euo pipefail`, added earlier on this branch, does NOT fix #3838: bash exempts a command used as an `if` condition from `set -e`, and all three of the hook's swallow-and-pass sites are exactly that shape. Verified against the hardened hook with a node shim that fails only the classifier call — a non-conforming commit still exited 0 with empty stdout AND empty stderr, indistinguishable from "your commit conforms". That is the defect verbatim. All three sites named in #3838 now capture the real exit status instead of consuming it as a condition, and each distinguishes its genuine negative from "could not run": - the classifier: 0 = is a git commit, 1 = genuinely not one, anything else = could not classify. Its `node -e` now wraps the require and the call in try/catch and exits 3 on a throw, so a broken require chain can never be mistaken for `isGitSubcommand` legitimately returning false — which is the arm that matters, since `token-scanner.cjs` is a gitignored build artifact and a fresh checkout lands there. - the opt-in config read and the JSON command extraction get the same treatment. On "could not run" the hook emits a diagnostic to stderr naming which check failed and why, then exits 0. The issue confirms this is safe — it is a PreToolUse hook, so stderr does not disturb the JSON protocol — and ranks it the smallest sufficient fix. The gate still fails open, but it can no longer do so silently, which is the whole complaint: a validator that disables itself quietly costs more than one that is absent, because it is trusted. Both controls are unchanged and pinned by tests: a conforming commit still passes silently, a non-conforming one still exits 2 with its existing block payload. The defect test asserts stderr is non-empty and names the failure; it fails against the pre-fix hook. Verification runs on the remote runner. Refs #3911, #3838 * docs(#3911): document the hook crash-policy contract Reference and Explanation via a new docs/features fragment (FEATURES.md is generated from it), INVENTORY rows for the three new hooks/lib files, and an ARCHITECTURE note on the hooks section. How-To: docs/how-to/declare-a-hook-crash-policy.md, indexed from docs/README.md — a hook author now has to choose and declare a crash policy, which is more than one step and crosses into which harness protocol their hook speaks. It covers allow/deny/crash, writing an ON_CRASH reason that is actually useful, when a deny needs a distinct stderr payload, the two hooks whose harness reads a JSON-body decision and must NOT use deny(), and what to do when a check cannot run at all — with #3838 as the worked example. Refs #3911 * test(#3911): prove the seam actually ships, and stop hand-rolling temp cleanup Two review findings. The acceptance criterion 'hooks/dist/** stays in parity via the build seam (lint:hooks-runtime-build-seam)' was misstated and unmet: that lint checks something else — that a hook requiring a compiled gsd-core/bin/lib module also calls ensureRuntimeBuild(). Nothing exercised that the three new hooks/lib files reach hooks/dist/lib at all. That gap is not theoretical: #770 is a recorded ship-blocking bug where a new hook never shipped because a copy list missed it. The suite now builds dist through the repo's own ensureBuiltHooks(), byte-compares each shipped copy against its source, and spawns a child that requires the SHIPPED dist copy and denies — which is what catches a copy that exists but cannot resolve its sibling registry. gsd-validate-commit.sh hand-duplicated mktemp/run/rm three times; one idempotent trap on EXIT replaces them, guarded so cleanup cannot alter the exit status. Behavior-neutral across five cases, with temp-file counts taken before and after each run. Refs #3911 * fix(#3911): stage transitive hook lib requires, not just one level The remote run returned 7 failures across 3 real causes. The important one is a PRODUCTION bug this phase exposed rather than caused. `writeCursorHooksJson` scanned each hook script for `./lib/X` requires exactly one level deep and never re-scanned the lib files it staged for their own sibling requires. Nothing had a transitive lib dependency before, so the gap was invisible. Adding hook-exit.js -> cli-exit.js -> exit-code-registry.js made real Cursor installs ship a bundle that dies at require time with MODULE_NOT_FOUND. It now walks to a fixed point, and a real installed Cursor hook runs to completion. The staging harness in shared-hooks-dir-resolution hand-copied its fixture, so the injection scanner crashed at require time and its exit-1 was being read as a policy decision. Migrated to copyScriptWithDeps, which walks the require graph — the repo's recorded rule for this class, since adding another copyFileSync keeps it alive for the next person. The missing-lib-source test in cursor-hook-workspace-roots hardcoded which lib file it expected to be named in the abort message; the same throw now fires for a different file first. Its assertion is unchanged in substance — staging still must abort rather than ship a broken hook — only the name is no longer pinned. The last one was my own test asserting an uppercase reason code. Measured against origin/next: the pre-change hook emits the same lowercase 'config_unreadable', so the test was wrong, not the migration. Corrected to the real value rather than making the code match the test. Verification runs on the remote runner. Refs #3911 * chore(#3911): regenerate the cursor install-tree golden The staging fix means a Cursor install now correctly carries the two transitive lib files it was silently missing. Additive only — no path was removed. The golden diff is the evidence the packaging defect was real. Refs #3911 * chore(#3911): backfill the changeset PR number Refs #3911 * fix(#3911): a git probe that timed out is not a negative A macOS CI lane failed three deny cases at 2084ms, 2112ms and 2177ms — just past the 2000ms budget these hooks give their git probes. The three that passed took 72ms, 595ms and 651ms. Under shard contention `git rev-parse` overruns, the hook reads the non-zero result as "not a git repo", and allows with exit 0 and empty stdout AND empty stderr. Under load, the guards silently stop guarding. That is ADR-3889's thesis exactly, sitting inside the security hooks this phase is about. The repo had already recognized the class in one place — gsd-cursor-subagent-start.js fail-closed-denies on `git_timed_out` (#3045) — but nowhere else. `hooks/lib/git-probe.js` classifies a probe's outcome, distinguishing a real non-zero exit from ETIMEDOUT, a signal kill, and a spawn failure, rather than folding all four into `status !== 0`. Three guards route their eight git probes through it. The resolution is the same shape #3838 took, and the same one that issue endorsed as smallest-sufficient: fail open, but loudly. **No exit code changes on any path** — a developer on a loaded machine is still not blocked, which keeps #3911's declaration-pass contract intact for exit codes. What changes is that the hook now says on stderr which probe could not answer, instead of presenting silence as a clean verdict. Scope was checked across every hooks/*.js, not just the three that failed: gsd-agent-isolation-guard spawns no git; gsd-statusline's two probes gate only a cosmetic display segment, not an allow/deny decision, and are left alone. The C2 deny assertion was a real-race test — it demanded exit 2 while a slow git legitimately yields 0. It now requires the hook to either deny, or allow with a diagnostic naming the probe that could not run; a silent allow still fails, so the assertion is not vacuous. A deterministic regression stubs git on PATH to sleep past the budget rather than waiting for load to reproduce it. Verification runs on the remote runner. Refs #3911 * test(#3911): a PATH shim cannot intercept the hooks' git spawn on Windows The deterministic timeout regression stubbed git on PATH and asserted the guard reports rather than silently allows. It passes on Linux and macOS and failed on Windows in 83ms and 176ms — the stub was never invoked at all. Mechanism: the hooks call spawnSync('git', args) with no shell:true, so on Windows CreateProcess resolves git.exe only and never a PATH .cmd shim. The git.cmd branch could not have worked and is removed rather than left implying a Windows path that does. Adding shell:true to the hooks to serve a test would change product behavior and widen an injection surface, so the case is skipped on win32 only, with the mechanism written into the skip reason so a future reader does not 'fix' it that way. Linux and macOS keep the coverage, and macOS is where the underlying fail-open was actually caught. Refs #3911 --------- Co-authored-by: sim <sim@local> |
||
|
Tom Boucher
|
268ca7e32d |
fix(#3504): harden hook injection patterns and force-add guard (#3510)
* test(#3504): add failing-first parity, fail-closed, and bypass suites * fix(#3504): harden hook injection patterns and force-add guard * test(#3504): stage the scanner lib dependency in shared-hooks fixture * chore(#3504): backfill changeset pr number * test(#3504): build parity samples from fragments for the ci scan --------- Co-authored-by: sim <sim@local> |
||
|
Tom Boucher
|
0c413bbc9c |
chore(#3059): close the ESLint glob-coverage escape and guard it (#3277)
* chore(#3059): close the ESLint glob-coverage escape and guard it 62 tracked source files matched no `files:` glob in eslint.config.mjs, so ESLint skipped them entirely while `eslint .` still exited 0 — including all 26 files under hooks/, the enforcement machinery itself. Covers 56 of them (eslint-rules/, hooks/, bin/lib/, pi/, examples/, vscode/, the plugin shims, root *.mjs) and allowlists the 6 deliberate must-not-compile brand-typing fixtures with a recorded reason each. hooks/** is covered with n/no-process-exit deliberately off: a hook's whole contract is its exit code, several exits are load-bearing stdin-timeout guards where nothing else terminates the process, and ADR-0012/0174 scope the no-process-exit convention to the Command Routing Hub. bin/lib/ui-safety-gate.cjs is dual-mode, so it keeps the rule live and takes two targeted disables in its require.main===module tail instead. Adds scripts/lint-eslint-glob-coverage.cjs + a node:test drift guard so the class cannot regrow: allowlist entries require a non-empty reason, the list ratchets down only (a stale entry fails), and a tracked-count floor means a broken `git ls-files` fails rather than reporting clean. Closes #3059 * chore(#3059): apply review findings — correct the changeset count, add parser properties Isolated adversarial review, confirmed by rebuilding a byte-for-byte replica of the pre-change eslint.config.mjs: the changeset claimed 44 previously- unlinted files. The real figure is 56 (56 covered + 6 allowlisted = 62). That was user-facing CHANGELOG text and was wrong; corrected, along with three consequential figures in the design record. CLAUDE.md requires a fast-check property test for parsers and budget limits, and listTrackedSourceFiles is a parser. The standards review called this "satisfied in spirit"; it is not. Adds three properties driving the real exported parser through an injected execFile: extension totality/soundness including a trailing terminator, backslash-normalization totality, and CRLF/LF equivalence — the invariant the repo's recurring CRLF defect class breaks. Also de-duplicates the anchor rows onto one shared resolver, kept deliberately independent of the guard's own resolveFileCoverage so an anchor still fails if that resolution regresses, and records in the guard's header why the bin/install.js family is NOT allowlisted: it resolves to 2 rules under ADR-1703, so an entry would trip the allowlist_stale ratchet. * fix(#3059): make the coverage guard's git call container-safe The remote runner reported the guard degrading to `git_failed` on both Node lanes: fatal: detected dubious ownership in repository at '/work' The runner executes in a container where the repo is owned by a different UID, so git refuses to operate on it. The guard's degraded-verdict path worked exactly as designed — it reported the failure instead of throwing or falsely reporting clean — but a guard that cannot run in CI is not a gate. `git ls-files` is now invoked as `git -c safe.directory=* ls-files`. `-c` scopes the override to the single invocation and mutates no config file, and the wildcard is appropriate because this command only enumerates tracked paths in the repository it is already executing inside. Adds a regression test that captures the argv through the injected execFile seam and asserts `-c safe.directory=*` precedes `ls-files`, so the container case is pinned behaviorally rather than by reading the script's source. * chore(#3059): backfill changeset PR number pr:0 placeholder replaced with the real number now that #3277 exists. --------- Co-authored-by: sim <sim@local> |
||
|
0xdhx
|
a8b40fa53f |
fix(#2547): fail closed on crashing and path-shadowing Kimi payloads (#2595)
* fix(#2547): fail closed on a malformed Kimi edit list in normalizeKimiPayload
`normalizeKimiPayload` rebuilt old_string/new_string with
`String(e.old ?? '')`. `??` guards the value, not the dereference, so a
nullish entry in a Kimi `edit` list threw a TypeError at the top of the
handler, before any tool dispatch. Each guard's outer
`catch { process.exit(0) }` swallowed that crash and emitted the same exit
code as "nothing to report" — turning a should-BLOCK call into a silent
allow.
Two hard blocks were bypassable:
* gsd-worktree-path-guard's cross-git-root write block (#260) — a
StrReplaceFile write whose path resolves to a different git root is
correctly blocked with a well-formed edit list, and silently allowed
with `edit: [null]`.
* gsd-workflow-guard's force-add block on agent-* branches — a Shell
payload carrying a spurious `edit: [null]` field walks past it. The
Bash path never reads `edit`; the field only has to be present to
trigger the crash.
Fixed with `e?.old` / `e?.new`, landed identically in all five copies so
tests/kimi-guard-normalization-parity.test.cjs's byte-identity assertion
still holds.
The crash boundary is nullish specifically, not "non-object": `('x').old`
and `(7).old` are legal reads yielding undefined, so string/number entries
never threw. Both are kept as controls proving the fix did not change
their behaviour.
Regression coverage is folded into the owning suites per CONTRIBUTING.md
(no new bug-* files). Negative-controlled: the nullish cases exit 0
against pre-fix guards and exit 2 after, with positive controls (the
equivalent well-formed payload blocks) and negative controls (in-worktree
writes and benign commands still pass) alongside.
Refs #2547
* test(#2547): exercise the production Kimi payload shape in read-guard tests
The `#2304: Kimi tool vocabulary engages the read guard` cases send
payloads with no `session_id`, and runHook injects none. A live Kimi turn
always carries one — kimi-cli's hooks/events.py `_base()` sets it
unconditionally, and soul/kimisoul.py calls `set_session_id()` at the top
of every turn before tool dispatch, so the ContextVar's `default=""` never
reaches a tool call.
gsd-read-guard treats any non-empty `data.session_id` as "Claude Code
already enforces read-before-edit, skip" (#2520). So the advisory those
tests assert fires only for a shape production never sends: the tests were
green, and the guard was dormant on Kimi. A sibling #2520 case in the same
file asserts the skip when `session_id` IS present — both passed, and the
production shape hits the skip.
Two changes, test-validity only:
* Retitle the #2304 block to say what it proves — the tool VOCABULARY is
normalized through to the Write/Edit branch — with a comment warning
not to read it as production evidence.
* Add a #2547 block asserting behaviour against the production shape
(session_id populated), including a case that pins the delta directly:
the same payload fires without session_id and is silent with it.
The #2547 block characterizes a known gap; it does not endorse it.
Redesigning how the guard discriminates runtimes is explicitly out of
scope for #2547. If a later change makes the advisory fire on Kimi these
tests are supposed to fail — update them then rather than dropping the
coverage.
Refs #2547
* docs(#2547): scope the Kimi guard-engagement claim to what Kimi enforces
#2518 engaged the guards' Kimi matchers and the release notes describe the
result as "All seven guard hooks now engage on Kimi", singling out the
prompt-injection read scanner as "the security-relevant guard" taken "from
silently dormant to engaged". That is not achievable for the scanner at
the emit layer.
gsd-read-injection-scanner.js is a PostToolUse hook, and kimi-cli's
dispatch never inspects PostToolUse hook results: src/kimi_cli/soul/
toolset.py awaits PreToolUse and honours `result.action == "block"`, but
fires PostToolUse via asyncio.create_task() and returns the ToolResult
without awaiting it — the done_callback only retrieves the task's own
exception. So no output shape the scanner emits can block or flag a Kimi
tool call, and `security.injection_blocking` cannot take effect there.
Reshaping the scanner's output would not change this; the enforcement gap
is in kimi-cli's PostToolUse handling, which is out of scope here.
This corrects the claim rather than the code — there is no gsd-core emit
fix that would make it true:
* .changeset/2304-kimi-guard-tool-name.md — the fragment is unreleased,
so it would otherwise ship this as a CHANGELOG security claim.
Headline narrowed to "normalize Kimi's payload shape" and a scope
paragraph added naming what actually blocks on Kimi (the two
PreToolUse blocks) versus what cannot.
* docs/migration/kimi-to-kimi-code.md — the scanner was listed under
"Every GSD `PreToolUse` guard"; it is PostToolUse. Corrected, and the
"What about the dormant guards?" section now splits enforceable from
not-enforceable instead of saying Phase 0 "fixed all seven".
* hooks/gsd-read-injection-scanner.js — the same scope note in the
file's own Kimi rationale comment, where the next contributor to touch
the normalization will actually read it. Comment only; the shared
normalization block is untouched and byte-identity still holds.
Refs #2547
* chore(#2547): regenerate golden install-parity fixtures for the guard fix
The golden install-parity fixtures record a content hash per installed
file, so changing the five guard hooks changes their hashes across every
runtime's fixture. Regenerated with the full sweep (build, gen:golden,
size:baseline) rather than a single generator — running gen:golden alone
leaves tests/workflow-size-baseline.json stale and loses CI jobs to a
regeneration that looked complete.
The size baselines came out unchanged (no workflow or agent bodies
touched) and the hash delta is confined to exactly the five guards:
gsd-prompt-guard, gsd-read-guard, gsd-read-injection-scanner,
gsd-workflow-guard, gsd-worktree-path-guard.
Refs #2547
* fix(#2547): guard the String() coercion in normalizeKimiPayload too
Found by adversarial review of the first commit, then reproduced against
pristine next: `e?.old` closes the nullish dereference but leaves a second
route to the same crash-to-allow.
`{"toString": null}` is valid JSON, and coercing it throws
`TypeError: Cannot convert object to primitive value` — so an edit entry
that IS a well-formed object still crashes normalization, still lands in
the outer `catch { process.exit(0) }`, and still downgrades a should-BLOCK
call to a silent allow. Confirmed on both hard blocks:
{"tool_name":"Shell","tool_input":{
"command":"git add -f secret.env",
"edit":[{"old":{"toString":null},"new":"x"}]}} -> exit 0 (was)
{"tool_name":"StrReplaceFile","tool_input":{
"path":"<main-repo>/src/index.ts",
"edit":[{"old":{"toString":null},"new":"x"}]}} -> exit 0 (was)
Both exit 2 now.
The coercion is wrapped rather than replaced with a `typeof === 'string'`
test on purpose. Degrading only the non-coercible entry keeps
stringification identical for every value that CAN coerce — numbers,
arrays, plain objects — which matters because gsd-prompt-guard scans
new_string for injection patterns, and a `typeof` test would silently stop
scanning content that reaches that scan today (e.g. `new: ["ignore all
previous instructions"]` currently stringifies and is scanned). Verified:
zero behaviour change across string, number, bool, null, array-of-strings,
nested array, plain object and `__proto__`-keyed input; only the throwing
case changes, from crash to ''.
Regression cases are negative-controlled against the previous commit: the
four new coercion-trap tests fail with only the `e?.old` fix in place and
pass with this one.
Refs #2547
* chore(#2547): cover the String() coercion vector in the changeset
The release note described only the nullish-dereference route. Both routes
reach the same fail-open, so both belong in the changelog entry, along with
why the coercion is wrapped rather than type-tested.
Refs #2547
* chore(#2547): point the changeset fragment at the real PR number
The fragment has to exist before `gh pr create` runs, so it carried the
issue number as a placeholder. Corrected to 2595 now that the PR is open.
Refs #2547
* fix(#2547): make Kimi's `path` authoritative over a model-supplied `file_path`
normalizeKimiPayload copied Kimi's `path` into `file_path` only when
`file_path === undefined`, so any `file_path` the model chose to include won
outright. Every guard reads `file_path`; kimi-cli executes on `path`. The guard
therefore inspected one file while the write landed on another.
This bypass needs no crash. A payload pairing a cross-root `path` with a
spurious `file_path: ""` left gsd-worktree-path-guard reading an empty string
and exiting 0, while the identical write without the extra key blocked — the
same cross-root write the #260 block exists to catch. The shadowing also
preserved a non-string `file_path` (`[]`), which threw inside that guard's
path.isAbsolute() and reached its outer `catch { process.exit(0) }`: the same
crash-to-allow the rest of #2547 closes, reached through the guard's own read
rather than through normalization.
Reachability is not speculative. kimi-cli's soul/toolset.py json-parses the
model's raw tool arguments and passes that dict verbatim as tool_input to
PreToolUse, performing typed validation only later inside tool.call() — after
the hook has already decided. So the model controls extra keys in tool_input at
the moment the guard runs. kimi-cli's file tools carry no `file_path` field at
all (src/kimi_cli/tools/file/write.py, replace.py), so a `file_path` in a Kimi
payload is always model-supplied.
`path` now wins outright. Overwriting can only ever narrow what a guard inspects
to the path that will actually be written, so it cannot under-block.
Normalization returns early for non-Kimi tool names, so the native Claude Code
contract (file_path governs) is untouched.
Landed identically across all five inlined copies; the byte-identity assertion
in tests/kimi-guard-normalization-parity.test.cjs enforces that.
* test(#2547): cover the file_path-shadowing bypass in the #260 guard suite
Four cases, each exiting 0 (bypass) against the pre-fix guards: a spurious
empty-string file_path, an in-worktree decoy file_path, and non-string
file_path values (array and object) that additionally crashed
path.isAbsolute() into the outer catch.
Two controls that are not bypass cases and matter as much:
- an in-worktree write carrying a cross-root DECOY file_path must still exit
0. Pre-fix this blocked, because the decoy won; the guard now follows the
path kimi-cli executes on in both directions, so the fix narrows what is
inspected without over-blocking.
- a native Claude Edit (no `path` field) must still block on file_path alone.
normalizeKimiPayload returns early for non-Kimi tool names, and this pins
that the non-Kimi contract did not move. It passes both pre- and post-fix
by design.
Negative-controlled: run against the pre-fix hooks, the four bypass cases and
the decoy control fail, and the native-Claude control passes.
* test(#2547): back the totality claim with property tests over fc.anything()
This PR claims the fix "makes normalization total over the inputs JSON can
express" — a for-all guarantee — while the tests backing it are example-based,
each shape added reactively after a crash was found by hand (the String()
coercion trap was itself found by adversarial review after the first commit
shipped). Example-based tests cannot substantiate a for-all claim; they record
the counterexamples someone happened to think of.
Four properties over fc.anything(), which is exactly the JSON-expressible
domain the claim names:
(a) totality over any tool_input
(b) totality over any edit list — the crash surface both #2547 fixes targeted
(c) `path` always wins over any model-supplied `file_path` (the review blocker
invariant: a guard reading file_path can never be aimed at a file other
than the one kimi-cli writes)
(d) a non-Kimi tool_name passes through untouched — the native Claude contract
normalizeKimiPayload is inlined per hook with no runtime binding, so there is
nothing to require. The block is extracted from hook source and evaluated via
the SAME extraction contract kimi-guard-normalization-parity.test.cjs uses, so
a source edit that breaks one breaks both instead of silently testing a stale
block. An extraction floor test fails loudly if the extraction yields a no-op.
Non-vacuous, and checked rather than assumed: against pristine pre-#2547 `next`,
(a), (b) and (c) all FAIL and (d) passes. (a) needed the fix that makes it
meaningful — a bare fc.anything() for tool_input passed even against the live
defect, because arbitrary generation essentially never invents the `edit` key
the crash lives behind, so the generator is biased onto the keys normalization
actually reads and unioned back with unbiased input.
* chore(#2547): cover the shadowing vector in the changeset and regen goldens
Golden install-parity churn is hash-only, on exactly the five hook files this
round changed. gsd-phase-boundary.sh is deliberately unchanged.
* test(#2547): make the property test able to kill the coercion mutant
Review Major 1: the generative test added to stop the NEXT counterexample
could not kill the one it was written for. Reproduced the reviewer's matrix
independently — against the shipped generator, a mutant reverting `editText`
to the unguarded `String(v ?? '')` passed all four properties.
Cause, confirmed by measurement: the edit-array ENTRIES were bare
`fc.anything()`, which essentially never invents an `old`/`new` key, so
`e?.old` was always undefined and `String(undefined ?? '')` never coerced
anything. That is the same vacuity the file's own comment describes one level
up, reproduced one level down.
The review's prescribed fix — bias the entry onto `{old, new}` — is necessary
but NOT sufficient, and this is the part worth recording: measured over 20,000
draws, bare `fc.anything()` yields a non-coercible value 3 times (0.015%). At
`numRuns: 200` an `old` key holding a hostile value essentially never
co-occurs, and the mutant survives the entry bias too. Both levels need
biasing — the entry onto the keys normalization reads, and the VALUE onto the
shape that actually throws.
`{"toString": <non-function>}` is that shape and stays inside the
"JSON-expressible" domain the claim names (JSON.parse produces it verbatim);
`fc.anything({withNullPrototype: true})` would also kill the mutant but widens
the domain past what the PR asserts, so it is not used.
Verified: M1 now dies at every seed tried (1/7/42/99/4242/31337, failing
within 3-31 cases) while HEAD stays green at all of them.
Also closes three coverage gaps the review listed as nits — properties (e)
totality over any JSON value as the WHOLE payload, (f) the tool_output →
tool_response mapping (including that an existing tool_response is not
clobbered), and (g) an empty edit list reconstructing nothing.
Property (e) required a one-line fix in the normalizer itself: `JSON.parse
('null')` is null, and null/primitive payloads threw on the `data.tool_name`
read — falsifying the "total over the inputs JSON can express" claim. Harmless
in practice (the throw landed in the same fail-open catch as the exit 0 it now
takes deliberately), but the claim should be true as stated. Landed
byte-identically across all five copies.
Refs #2547
* fix(#2547): make old_string/new_string authoritative, not merely defaulted
Review Major 2: the shadowing class the prior round's BLOCKER closed for
`file_path` survived one field over. `old_string`/`new_string` were still
reconstructed only `if (input.<field> === undefined)`, so a model-supplied
value won.
The argument for making `path` authoritative applies verbatim here.
kimi-cli's StrReplaceFile schema is `path` + `edit` only
(src/kimi_cli/tools/file/replace.py @ 4a550ef) and carries no
`old_string`/`new_string` at all, so either key appearing in a Kimi payload is
always model-supplied — exactly like `file_path`.
Verified end-to-end against the reviewer's payload: a cross-root write
carrying `new_string: ""` alongside an injected `edit[].new` left
gsd-prompt-guard reading '' and returning at its `if (!content)` guard, so the
injection advisory never fired and the reconstructed content was never
scanned. `new_string: null` behaved identically. Negative-controlled: both
produce empty output against pre-fix source and fire the advisory after.
Chose unconditional reconstruction over the offered `typeof` alternative
deliberately. A type test closes `""`/`null` but leaves the interesting case
open — a benign NON-EMPTY decoy (`new_string: "chore: tidy"`) shadows just as
effectively and passes any type test. The new suite includes that case
specifically; it is what discriminates between the two candidate fixes.
Also pins the kimi-cli SHA in the authoritative-path comment, as requested —
it cited file names with no version while the issue pins 4a550ef.
Landed byte-identically across all five inlined copies; the parity test's
byte-identity assertion holds.
Refs #2547
* fix(#2547): close the non-string file_path crash-to-allow at every read site
Review Major 3: the crash-to-allow was closed only as a side effect of `path`
masking the bad value, while the changeset read as though it were closed
outright. Confirmed both of the review's reachability claims: `[]`/`{}`/`42`
are truthy, survive the `!rawFilePath` early-out, and throw inside
path.isAbsolute() into the outer `catch { process.exit(0) }`; and normalization
returns early for native Claude Code payloads (KIMI_TOOL_NAMES has no 'Edit'
entry), so `{"tool_name":"Edit","tool_input":{"file_path":[]}}` reached it
untouched — this guard's original #260 surface.
Reproduced on a real fixture: string cross-root path exits 2, the identical
payload with `[]` or `{}` exits 0.
Swept the class rather than the instance. Five more untyped read sites across
four other hooks, each one line from a type-strict or method-dependent call.
Census of what each can actually do:
gsd-worktree-path-guard.js:173 BLOCKS -> live bypass (the review's finding)
gsd-prompt-guard.js:128 scanner -> silenced the injection scan, the
same outcome as Major 2 by another
route; verified empirically
gsd-workflow-guard.js:206 advisory only (its exit-2 is the Bash
force-add path, which reads
`command`, not `file_path`)
gsd-read-guard.js:141 advisory only
gsd-read-injection-scanner.js:213 advisory only
gsd-windsurf-pre-write.js:75 ALREADY TYPED — the shape adopted here
All six now read typed. The workflow-guard site keeps its truthiness fallback
(`(typeof x === 'string' && x) || ...`) because a bare type test would let an
empty `file_path` shortcut the `path` fallback.
Also declares one swept hit NOT fixed: `gsd-workflow-guard.js:175` reads
`command` untyped on a genuinely blocking path. Same shape, but not
exploitable — unlike file_path/path there is no second field carrying the
executable value, so a non-string command cannot smuggle a real `git add -f`
past the block. Left alone rather than widen this PR into the Bash path.
The regression gate is a SOURCE-level invariant, not a behavioural one, and
that is deliberate: the fixed read and the crashing read are black-box
identical — both end at exit 0, one via the catch and one via the early-out.
A test asserting exit 0 on a non-string payload passes against the unfixed
code, which is the same false-green the review flagged in the existing
`['non-string file_path (array)', []]` cases. Repeating it one level up would
be no better. tests/kimi-guard-typed-payload-reads.test.cjs fails if any hook
regresses to an untyped read (negative-controlled: it reports all five
pre-fix sites with correct file:line).
The behavioural cases requested — non-string file_path with NO `path` key —
are added to worktree-safety.test.cjs and labelled honestly as documenting the
explicit fail-open rather than detecting a revert.
Also states the relative-path premise (review Minor 5) at the early-out that
depends on it: "always safe" holds only while every runtime reaching there
resolves relative paths against the tool CWD. Claude Code satisfies it by
requiring absolute paths; kimi-cli's resolution behaviour is NOT verified here
and is recorded as an unverified premise rather than an asserted bypass.
Refs #2547
* docs(#2547): correct the changeset's closed-claim and fold the misattributed note
Review Major 3 also flagged the fragment: it said the non-string vector "threw
inside that guard's path.isAbsolute() ... `path` now wins outright", which
reads as closed when it was closed only conditionally. Rewritten to state what
is now true — closed unconditionally at all six read sites — and extended with
the Major 2 finding.
Review Minor 6 (the #2547 scope note living in a `pr: 2518` fragment) turns out
to understate the problem. Rendering the changelog and re-parsing it shows the
note is not merely misattributed — it is DROPPED. serializeChangelog emits each
fragment as a single `- ` bullet, and parseChangelog terminates a bullet at the
first non-continuation line, so everything after a blank line is lost on
re-parse. Audited all 44 fragments: exactly one was lossy —
2304-kimi-guard-tool-name.md, losing 656 of 1730 characters, i.e. precisely
that second paragraph. Folding it into this PR's fragment fixes the
attribution and the silent loss together; all 44 now round-trip losslessly.
That same mechanism is why the remaining nit — reformat this fragment's
~2,000-character paragraph for readability — is NOT applied. A paragraph break
or a bullet list would silently truncate the entry at the first blank line
(verified for both). The single-paragraph form is load-bearing under the
current serializer, not an authoring preference. Worth its own issue; noted in
the PR thread rather than worked around here.
Refs #2547
* chore(#2547): regenerate golden install parity after rebase onto next
Rebased onto `next` @
|
||
|
Tom Boucher
|
7e905aa137 |
feat(#2505): Phase 0 — Kimi PreToolUse guard vocabulary normalization (precondition; carries PR #2326 forward) (#2518)
* fix(#2304): normalize Kimi tool vocabulary in PreToolUse guard payload checks
The Kimi [[hooks]] registrations translate the matcher to Kimi's tool
vocabulary (WriteFile|StrReplaceFile) but the guard scripts early-exit
unless the payload's tool_name is a Claude name (Write/Edit/MultiEdit),
so every guard was dormant on Kimi: the matcher fired, the script saw
WriteFile, and exit(0)'d.
Normalize the payload's tool_name at the top of each guard
(WriteFile -> Write, StrReplaceFile -> Edit; bare or module-qualified
kimi_cli.tools.file:* forms) before the check. Inlined per guard rather
than a hooks/lib/ helper because hook scripts are staged as standalone
files on every hook surface, and a sibling require is a staging
dependency that can fail silently.
Regression tests pipe Kimi-vocabulary payloads at each guard and assert
it engages (typed fields: exit status, decision, hookSpecificOutput) —
verified red against the pre-fix scripts, green after.
* fix(#2304): normalize Kimi tool_input fields and route block reasons to stderr
Cross-AI review of the initial fix, verified against kimi-cli source,
found the tool_name normalization alone leaves the guards dormant on a
real Kimi runtime: kimi-cli forwards tool_input verbatim
(src/kimi_cli/hooks/events.py), and its tool schemas
(src/kimi_cli/tools/file/{write,replace}.py) use path/content and
edit.old/edit.new (single Edit or list) — not Claude's
file_path/old_string/new_string. The guards read file_path, got '',
and exited 0 past the now-open tool_name gate.
Extend the per-guard normalization to the payload fields
(path -> file_path, edit -> old_string/new_string with list flattening),
and write the worktree guard's block reason to stderr as well as the
stdout JSON — Kimi feeds stderr, not stdout, back to the model on
exit 2 (docs/en/customization/hooks.md exit-code table).
Regression tests rewritten to Kimi's actual payload shapes (plus an
edit-list case and a stderr-reason assertion) — verified red against
the name-only fix, green after.
* fix(#2304): join all edit[] entries into old_string, matching new_string
Review nit on #2326: old_string took only edits[0].old while new_string
joined the whole list. Symmetric join removes the latent trap for any
future consumer sizing before/after content (e.g. the #2255 write guard).
* fix(#2304): normalize Kimi ReadFile vocabulary in read-injection scanner
Review Major 2 on #2326: gsd-read-injection-scanner.js had the identical
dormancy — its Kimi matcher fires on 'ReadFile' but the SCANNED_TOOLS
check only knew 'Read', so injected content in read files was never
flagged on Kimi installs.
Folds the same inlined normalization block into the scanner and extends
the shared KIMI_TOOL_NAMES map with ReadFile:'Read' in all four copies so
they stay byte-identical. Harmless in the three write guards: a
normalized 'Read' falls out of their Write/Edit allowlist exactly as the
unmapped name did. Field mapping verified against kimi-cli upstream
(src/kimi_cli/tools/file/read.py Params.path); the existing
path->file_path copy covers the scanner's file_path read.
* test(#2304): parity test binding the four inlined Kimi normalization copies
Review Major 1 on #2326: KIMI_TOOL_NAMES + normalizeKimiPayload is
deliberately inlined in four hook scripts (staging-dependency rationale,
unchanged), with the inverse table in bin/install.js — five
hand-maintained surfaces and nothing binding them.
Static binding, zero runtime coupling:
- the four inlined blocks must be byte-identical;
- each guard-map entry must be the value-inverse of
convertKimiToolName() for its Claude name;
- every guard-relevant Claude tool (Write/Edit/MultiEdit/Read) must have
a reverse entry — a vocabulary rename or extension that updates the
installer without updating the guards now fails in CI instead of
leaving a guard silently dormant (the #2304 recurrence door).
Negative-controlled: diverging one copy or dropping a map entry fails
the suite against the fixed code.
* test(#2304): regenerate golden parity fixtures for guard hook changes
CI red on #2326: all 10 golden-parity failures were the staged guard
hooks drifting from their fixtures. Regenerated with npm run gen:golden
(after npm run build) under throwaway HOME/CLAUDE_CONFIG_DIR; diff
verified to change exactly the four PR-touched guard entries per
surface, nothing else.
* test(#2304): regression tests for Kimi ReadFile engaging the scanner
Mirrors the per-guard Kimi vocabulary tests the PR added for the three
write guards: bare and module-qualified ReadFile produce the advisory,
path exclusions still apply post-normalization, unknown Kimi names stay
fail-open. Negative-controlled against the pre-fold scanner (the two
positive cases fail there; exclusion/fall-through correctly pass on
both sides).
* fix(#2304): normalize Kimi Shell vocabulary in workflow guard
Withdraws the disclosed out-of-scope split: verification showed the
Bash->Shell case needs NO different mapping — kimi-cli's Shell.Params
names its field `command` (src/kimi_cli/tools/shell/__init__.py), same
as Claude's Bash — and the guard's write branch (Write/Edit/MultiEdit
allowlist) was ALSO dormant on Kimi under its Shell|WriteFile|
StrReplaceFile matcher. Same defect class as the other four hooks.
Folds the identical inlined block into gsd-workflow-guard.js and
extends the shared map with Shell:'Bash' in all five copies (harmless
outside the workflow guard: a normalized Bash falls out of the other
guards' checks as before). Parity test now binds five copies and adds
Bash to the dormancy alarm. New workflow-guard test file exercises the
observable block (force-add on a worktree-agent branch): Shell bare and
module-qualified block with WORKTREE_AGENT_FORCE_ADD_FORBIDDEN, benign
Shell passes, Claude Bash unchanged — negative-controlled against the
pre-fold guard (the two Kimi cases fail there). Golden parity fixtures
regenerated; diff verified to change exactly the five guard entries per
surface.
* fix(#2304): map Kimi tool_output and route workflow-guard block to stderr
Third-party review (cross-AI verifier) caught two gaps in the revision:
1. Kimi PostToolUse events carry `tool_output`, not `tool_response`
(kimi-cli src/kimi_cli/hooks/events.py post_tool_use()), so the
read-injection scanner — which reads data.tool_response — was STILL
dormant on real Kimi payloads; the earlier tests passed because they
sent Claude-shaped payloads. The shared normalization block now maps
tool_output -> tool_response (inert in PreToolUse guards, where the
field is absent), and the scanner's Kimi tests send the real shape.
2. The workflow guard's force-add block wrote its reason to stdout only.
Kimi's exit-2 protocol feeds stderr back to the model — the exact
fix this PR already applied to the other blocking guard — so the
newly-awakened block would have been a silent denial. Reason now
also routed to stderr, asserted in the test.
Also: the scanner's "unknown name" test now uses a genuinely unmapped
name (FetchURL) — Shell stopped qualifying when it entered the map —
and the workflow guard's write branch (WriteFile advisory,
StrReplaceFile .planning pass) gains behavioral coverage. All five
copies stay byte-identical (parity test green); golden fixtures
regenerated, diff verified to the five guard entries per surface.
Negative-controlled: 3 new assertions fail against the pre-fix hooks.
* docs(#2304): update changeset to cover the full five-guard fix
Review round 2 (2026-07-18) flagged the changeset as stale: it was
written for the first commit and still described only the three guards
named in the issue. The shipped diff grew to five guards plus two
payload dimensions the original body never mentioned. The body now
names gsd-read-injection-scanner and gsd-workflow-guard, the ReadFile
and Shell vocabulary entries, the tool_output -> tool_response mapping,
and the workflow guard's stderr block-reason routing.
* test(#2304): regenerate kilo golden fixture after #2305 landed on next
The branch's fixture sweep predates
|
||
|
Jeremy McSpadden
|
8d29ecd02f |
fix: add missing 'act as' injection pattern to prompt guard hook (#1697)
Fixes #1696 The gsd-prompt-guard.js hook was missing the 'act as a/an/the' prompt injection pattern that security.cjs includes. Adds the pattern with the same (?!plan|phase|wave) negative lookahead exception to allow legitimate GSD workflow references. |
||
|
Tom Boucher
|
62db008570 |
security: add prompt injection guards, path traversal prevention, and input validation
Defense-in-depth security hardening for a codebase where markdown files become LLM system prompts. Adds centralized security module, PreToolUse hook for injection detection, and CI-ready codebase scan. New files: - security.cjs: path traversal prevention, prompt injection scanner/sanitizer, safe JSON parsing, field name validation, shell arg validation - gsd-prompt-guard.js: PreToolUse hook scans .planning/ writes for injection - security.test.cjs: 62 unit tests for all security functions - prompt-injection-scan.test.cjs: CI scan of all agent/workflow/command files Hardened code paths: - readTextArgOrFile: path traversal guard (--prd, --text-file) - cmdStateUpdate/Patch: field name validation prevents regex injection - cmdCommit: sanitizeForPrompt strips invisible chars from commit messages - gsd-tools --fields: safeJsonParse wraps unprotected JSON.parse - cmdFrontmatterGet/Set: null byte rejection - cmdVerifyPathExists: null byte rejection - install.js: registers prompt guard hook, updates uninstaller Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> |