f7df920681f233ae0fe064ee659550bdf41ff708
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Tom Boucher
|
9e4f0e99ad |
fix(#3631): exclude only __pycache__-resident bytecode from the consent digest (#3650)
* test(3631): failing-first coverage for bytecode-cache in the consent hash
bundleContentHash digests a walk with no exclusion, so a routine 'python3 -m unittest'
inside a Python-backed capability bundle writes __pycache__ under the bundle, the
recomputed hash stops matching the consent record, and the capability silently goes
inactive — no error, no warning, and loop render-hooks then omits its step and gate.
Two distinct triggers, and the second is the sharper one: collectBundleEntries pushes a
{kind:'dir'} entry for EVERY directory and the digest emits a TAG_DIR marker for it, so an
EMPTY __pycache__/ flips the hash before a single .pyc is written. A fix filtering only
*.pyc would leave that live. Verified by execution against the built lib: 5 of 7 probe
rows diverge from intent today, including the empty-directory row.
The anti-regression rows are the point of the shape: editing a real scripts/m.py and
adding node_modules/pkg/index.js must BOTH still change the hash. node_modules is
deliberately not excludable — its contents are required at runtime, so dropping it from
the digest would stop consent binding executable content. The symlink row pins ordering:
exclusion must apply after the lstat fail-closed rejection, never before.
Refs #3631
* fix(3631): exclude derived bytecode caches from the consent digest
RED proven at e5ba8f1fe on the remote runner: 8 failures, exactly the rows predicted to
fail, with the four anti-regression rows already green.
collectBundleEntries now skips a hardcoded, gitignore-independent set from the DIGEST:
basenames __pycache__, .pytest_cache, .DS_Store, and any .pyc/.pyo file. Matching is
byte-exact on the raw Buffer name (the walk never utf8-decodes) and case-sensitive, so the
digest does not vary with how a name happens to be spelled on a case-insensitive volume.
Three properties were preserved deliberately, each pinned by a test:
- The filter runs AFTER the lstat symlink/non-regular fail-closed rejection. Filtering
first would have turned the exclusion into a way to smuggle a symlink past the check;
a symlink named x.pyc still throws.
- Excluded entries still count toward BUNDLE_MAX_FILES and BUNDLE_MAX_TOTAL_BYTES. The
caps guard the WALK; the digest answers a different question, and exclusion must not
become an unbounded-bytes hole.
- An excluded DIRECTORY is neither emitted as a TAG_DIR marker nor recursed into. The
directory marker was the sharper half of this bug: an empty __pycache__ flipped the
hash before any .pyc existed, so a *.pyc-only filter would have left it live.
The issue proposed either a gitignore-aware walk or a list including node_modules. Both
are rejected. A consent binding must not delegate its scope to a .gitignore the bundle
author does not control — one line there would drop arbitrary executable content out of
the hash. And node_modules holds code that is required at runtime; excluding it would stop
consent binding executable content, turning a usability bug into a supply-chain hole. What
makes __pycache__ different is that CPython validates each .pyc against its sibling
source, which remains hashed, so a real code change still invalidates consent.
Docs: CONTEXT.md's 'EVERY regular file AND directory' claim is corrected in place.
ADR-2363's residual-gap section said the walk had 'no exclusions' — per
docs/adr/README.md ('ADRs are append-only') that is corrected by a dated amendment rather
than an in-place edit. Its D4 argument is unaffected: skill bodies are .md and stay bound.
Fixes #3631
* fix(3631): narrow the digest exclusion after two isolated security reviews
The first cut of this fix passed the full suite and was still wrong. Both orthogonal
reviews rejected it, and the second one found a hole that has nothing to do with Python.
HIGH — an excluded DIRECTORY was 'continue'd before recursion, so its whole subtree was
permanently outside the digest. Declared hook script paths allow '_', '.' and '/' with no
directory or extension rule, so hooks:[{script:'__pycache__/run.js'}] installed, executed
via node, and its bytes could be rewritten forever without moving the hash. Ship benign
v1, collect consent, then own the machine. No Python involved.
FALSE RATIONALE — the justification I wrote into the code, CONTEXT.md, the ADR amendment
and the changeset claimed CPython validates a cached .pyc against its sibling source, so
the source staying hashed kept consent honest. That is not true, and I proved it by
execution rather than argument: default timestamp invalidation compares only the source's
mtime and size, both settable by anyone who can write the bundle. A forged pyc ran while
the .py was byte-identical.
Also wrong: '*.pyc' matched anywhere, but a legacy sourceless scripts/x.pyc IS importable,
so excluding it was a live vector.
Narrowed to what is actually defensible:
- a DIRECTORY named __pycache__/.pytest_cache has only its TAG_DIR marker suppressed;
the walk still recurses and hashes every non-excluded child.
- .pyc/.pyo are excluded ONLY when the parent basename is exactly __pycache__.
- a regular FILE named __pycache__, and a DIRECTORY named x.pyc, stay bound.
- declared hook paths containing a __pycache__/.pytest_cache segment or a .pyc/.pyo
basename are now rejected in both validator copies — a file named .pyc can contain
perfectly valid JavaScript, so the exclusion must not be reachable from a declared
surface.
Accepted residual risk, stated plainly in ADR-2363 and CONTEXT.md instead of explained
away: a forged __pycache__/mod.pyc matching an unmodified, still-hashed mod.py executes
without moving the digest. Before this change that write was detected. It is accepted to
stop routine bytecode caching from silently deactivating capabilities, and it is bounded —
the attacker needs post-consent write access, everything outside __pycache__/*.pyc stays
hashed, and no declared surface can point into the excluded space.
Known limitation, not papered over: .pytest_cache CONTENTS still move the digest. Only the
directory marker is suppressed. Excluding that subtree would reopen the HIGH finding.
Refs #3631
* fix(3631): drop the .DS_Store exclusion and pin what the caps actually bind
Second round of isolated review findings. The hardening closed the two original holes —
both re-reviews confirmed that by execution — but it introduced a new one of the same
shape, and left three claims unbacked.
HIGH, self-inflicted: .DS_Store was excluded from the digest at any depth, but the hook
path validator was hardened only for __pycache__/.pytest_cache/.pyc/.pyo. So
script:'hooks/.DS_Store' was ACCEPTED, runnableHookCommand emits the bare quoted path for
a non-.js name (the branch .sh hooks already use), and capability-source copies it with
its mode bit intact. Ship it +x with a benign shebang, take consent, then rewrite it
forever — the digest never moves. Fixed by DELETING the .DS_Store exclusion rather than
teaching the validator about it: .DS_Store has nothing to do with this issue's Python
bytecode symptom, and an excluded filename is a permanently unhashed name. The narrower
the exclusion, the smaller the hole.
The residual-risk bound in ADR-2363 and CONTEXT.md claimed declared surfaces cannot reach
excluded space. That is false and is now stated correctly: node resolves an unregistered
extension through the default .js handler, so a hashed, consent-covered hooks/run.js that
requires '../__pycache__/mod.pyc' reaches it in one hop. The validator guard raises the
bar for DECLARED surfaces; it does not contain the risk. The two bounds that are real —
post-consent write access required, everything outside __pycache__/*.pyc still hashed —
are kept.
The BUNDLE_MAX_FILES boundary test had gone vacuous: it padded with root-level *.pyc,
which the hardening made non-excluded, so it no longer proved anything about excluded
entries while the ADR claimed the caps were test-pinned. It now pads __pycache__/f{i}.pyc,
with the arithmetic re-derived by execution (capability.json + the still-counted
__pycache__ dir + N). BUNDLE_MAX_TOTAL_BYTES had zero coverage at all and is now pinned by
a sparse 32 MiB __pycache__/big.pyc that must still trip the size cap — the test that
proves exclusion did not become an unbounded-bytes hole.
Added the parity assertion CLAUDE.md's Generative Fix Divergence rule requires for the two
isSafeHookScriptPath copies, and proved it can fail: mutating one BUILT copy to drop .pyo
made the parity check report the divergence. Also pinned semantics that were correct but
untested and would have survived mutation — __pycache__/sub/x.pyc stays hashed (the parent
resets to sub, which is the recursion threading itself), .pytest_cache/y.pyc stays hashed,
and .pyo in both directions, which was a free surviving mutant.
Changeset rewritten: it still described the rejected wholesale-exclusion semantics.
Refs #3631
* chore(3631): backfill changeset PR number (#3650)
---------
Co-authored-by: sim <sim@local>
|
||
|
Tom Boucher
|
b183317abd |
docs(#3256): ratify ADR-2363 to Accepted (#3260)
Both phases of epic #2363 have landed and the epic is closed as completed, so ADR-2363's own stated ratification bar is met: #3248 merged, the consent summary renders instruction surfaces, and a passing test pins the D4 signature behavior. Adds the dated Ratification section the corpus requires, naming the files, symbols and tests for each of D1-D5, and restores the reciprocal back-link on ADR-1244 - owed only on ratification, which is why the premature flip in 4d26887e correctly withdrew it. Records the judgment call the ratification rests on rather than burying it: D3 classifies instruction surfaces as skills and agents, and the mechanism discloses skills only. Third-party agents are never staged into the instruction context, so disclosing them would have named a surface that does not exist. Everything actually staged is disclosed, which is what the decision requires; whether agents should be staged is recorded in D5 as an open maintainer question. Docs-only. No behavior change. Closes #3256 Co-authored-by: sim <sim@local> |
||
|
Tom Boucher
|
f96cb44f85 |
enhance(#3248): disclose capability skills as an instruction surface (#3253)
* test(#3248): failing-first suite for instruction-surface disclosure 28 matrix rows from 50-test-matrix.md. Rows requiring the new Disclosure.instructionSurfaces field fail today; rows 18-20/23-25 (the ADR-2363 D4 signature invariants) pass today by construction because the current code never reads skills/agents at all, and stand as regression guards for the implementation commit. Refs #3248 * feat(#3248): disclose capability skills and agents as an instruction surface ADR-2363 D5. A capability whose only contribution was skills disclosed nothing at install: summarizeDisclosure early-returned "ships no executable surfaces (declarative only)" because hasExecutable was false, while each SKILL.md body landed verbatim in the agent's instruction context. discloseExecutableSurfaces gains a fifth, NON-executable class, instructionSurfaces, collecting declared skills/agents stems through the same safeCollect wrapper as the four existing collectors, so a hostile value degrades only this class and the function stays total for any manifest shape. Nothing existing is edited: the collectors, hasExecutable, disclosureSignature and missingArtifacts are untouched. get_impact rates the symbol CRITICAL at 196 affected, which is why the design is strictly additive. D4 is implemented by omission and pinned rather than left incidental: adding, changing or removing skills/agents leaves disclosureSignature byte-identical, so no stored consent record is perturbed and no spurious re-consent fires. ADR-2782's conditional-append trick is deliberately NOT reused - it worked because no manifest could declare a reviewer body before that class existed, whereas skills predate this one, so a conditional append would re-sign every already-consented skill-bearing capability. The renderer is extracted as summarizeInstructionSurfaces and called from BOTH branches of summarizeDisclosure. Appending only at the end would never render for skill-only capabilities - the ones that need it - since those take the early return. That branch's "declarative only" claim is now conditional on there being no instruction surface either. The renderer iterates rather than spreading into push, so an unbounded stem count cannot throw RangeError, and tolerates the bare {} the CLI edge passes via `res.disclosure || {}`. Scope note: #3248's prose says "skill stems"; ADR-2363 D3 classifies instruction surfaces as "skills, agents". Shipping skills alone would leave an ADR deliverable owned by no phase, and the epic has no Phase 2. Agents are the same shape at no extra cost. Narrowing back is a two-line change. Ratifies ADR-2363 (Proposed -> Accepted) and adds the owed ADR-1244 back-link. Closes #3248 * fix(#3248): escape consent-prompt values and narrow disclosure to skills Two review findings, both of which made the previous commit wrong. BLOCKER (isolated adversarial review). Every manifest-supplied value interpolated into a consent-prompt line was rendered unescaped. Those lines are joined with \n and written RAW to stderr on the needs-consent path (capability-command-router -> cli-exit runMain), so a stem carrying a newline forged additional lines indistinguishable from genuine GSD disclosure text, and an ANSI escape could clear or rewrite lines already printed. That defeats the informed-consent guarantee this change exists to provide, and is a prompt-injection vector against any agent that reads the stderr text to decide whether to retry with --yes. The hole was not unique to the new class - hook event/script, command family/module/router, every MCP field, and every reviewer-lane field were equally unescaped. Fixing only the new one would have created the generative-fix divergence this repo tracks, so renderValueForPrompt is applied to all five classes through one helper, guarded by a parity test that fails if a future class skips it. Escaping is identity for ordinary names, so no well-formed manifest's output changes. The disclosure OBJECT stays verbatim - only the rendered LINE is escaped - because the signature and every consumer reasoning about identity depend on the declared value. NARROWED to skills only. The previous commit also collected agents, arguing ADR-2363 D3 classifies instruction surfaces as "skills, agents". Verified against staging: stageSkillsForRuntimeAsSkills takes a registry and unions third-party skills in via readInstalledCapabilitySkill, while stageAgentsForRuntimeWithConverter takes only a source directory and has no registry-aware path. Third-party agents are never staged into the instruction context, so disclosing them would have put a false claim in a security prompt - worse than the scope creep two reviewers flagged it as. D3's classification stands; D5 now records that Phase 1 implements the skills half and that whether agents should be staged at all is an open maintainer question. Also reverts the premature ADR-2363 ratification. The previous commit flipped it to Accepted and asserted "#3248 merged" while this branch IS #3248 and is unmerged. Status returns to Proposed, and the ADR-1244 back-link - owed only on ratification - is withdrawn. Adds the fast-check property suite CLAUDE.md requires and the direct precedent (reviewer-trust-disclosure) already had: totality, D4 signature invariance, D3 hasExecutable invariance, and renderer totality over adversarial manifests. Refs #3248 * chore(#3248): correct changeset scope claim and backfill pr number The fragment was written against the pre-narrowing commit and still advertised 'skills and agents'. 4d26887e narrowed disclosure to skills only - third-party agents are never staged into the instruction context - but did not touch the fragment, so the release notes would have carried a claim the code does not implement. Also backfills pr:0 -> 3253 and names the prompt-escaping fix, which is user-visible and was absent from the original body. Changeset-only; no code or test changed, so the gsd-test pass recorded for 4d26887e still describes this tree's behavior. Refs #3248 --------- Co-authored-by: sim <sim@local> |
||
|
Tom Boucher
|
bc5619dd27 |
docs(#3247): record the capability instruction-surface trust model (#3249)
* docs(#3247): record the capability instruction-surface trust model ADR-2363 records the trust posture for third-party capability SKILL.md bodies, which #2322/#2340 made agent-invocable without any content-level control. The path-level protections that fix shipped are all present; no content scanner exists, and external-descriptor-trust.cts never had one. Nothing was bypassed - the control did not exist and the boundary was never written down. D1 records the posture: skill bodies are trusted, unscanned agent instructions. D2 rejects content scanning on Kerckhoffs (a shipped rule set is readable by the adversary who installs it), on threat-model non-transfer from ADR-1577 (there, instructions are anomalous inside data; here they are the payload's legitimate form), and on Goodhart (a scanned-OK line displaces the judgment the consent prompt exists to provoke). D3 replaces the executable/non-executable binary with three classes, adding instruction surface. D4 keeps instruction surfaces out of the v1 disclosureSignature. The signature is NOT the activation binding - hasProjectConsent compares contentHash only, and a global install carries no consent record at all. What re-encoding would do is perturb the signature of every skill-bearing capability and fire a spurious re-consent prompt on its next upgrade, which is what ADR-2782 D4 rule 5 already forbids. If instruction surfaces ever need to be signature-bound, that lands as a versioned v2 signature with a migration, never an in-place re-encoding. Corrects capability-trust-model.md, which claimed skills get lighter consent because they do not execute code - true, and not the relevant property, since the agent is the interpreter. Adds the author-side boundary to develop-a-capability.md and links it from publish-a-capability.md. Both state that per-skill disclosure at the consent prompt lands with #3248 and does not happen today. Docs-only. No behavior change; no consent record perturbed. D5's mechanism is Phase 1 (#3248), which is why the ADR is Proposed. Refs #2363 * chore(#3247): backfill changeset pr number to 3249 --------- Co-authored-by: sim <sim@local> |