fix(#4652): confine every boundary that joins argv to a managed root

Phase 2 of epic #4636, absorbing #4327 and #4354. Implements ADR-4650
decision 3: containment is a boundary concern — the predicate runs where
external input enters, not at whichever interior call site remembered.

Four boundaries now validate against their managed root and reject with a
USAGE-shaped error before touching the filesystem:

  todo complete <name>                 -> todosDir(cwd)
  check predicate --phase-dir <dir>    -> projectDir
  check decision-coverage-plan <dir>   -> projectDir   (via resolvePath)
  check gap-analysis.plan-post <dir>   -> projectDir

#4327 understated its own severity. It reports that a traversal name
"resolves outside the todos root", which reads as an information leak.
Measured, it was destructive: the command exited 0, MOVED the outside file
into completed/, and unlinked the original. cmdTodoComplete ends in
fs.unlinkSync(sourcePath), so an unconfined name consumed across the
boundary rather than merely reading across it. Validation now precedes every
fs call — existsSync, readFileSync, ensureDir, writeSync, unlinkSync — and
both halves of the move are confined, so neither source nor destination can
land outside the root. --dry-run is rejected on the same terms; a preview
must not leak a resolved outside path either.

#4354 reproduces exactly: a BLOCKING gate returned block:false sourced
entirely from a SECURITY.md in a caller-chosen directory outside the project.

THE HARDER HALF, found by the isolated adversarial review of the first
attempt: validating a path and then using a DIFFERENT one closes nothing.
The first fix validated `--phase-dir` joined against `--cwd`, then passed the
RAW unjoined value into the predicate context. gate-predicate-evaluator uses
it as-is and findPhaseArtifact resolves a relative path against the REAL
process cwd — so validation and the read used two different roots whenever
process.cwd() differed from --cwd. Reproduced: running from a directory
holding a plan with `secret_field: LEAKED_VALUE`, a predicate declared
against an empty --cwd project exited 0 and returned "actual":"LEAKED_VALUE".

The rule now applied at all three router sites: **use the validated resolved
path, never the raw input.** Independently re-verified after the fix — the
lookup resolves in the --cwd project and no value leaks.

gate-predicate-evaluator.cts is untouched and still imports no fs. Confining
in the router is what keeps that pure-leaf contract intact AND covers
${PHASE_DIR} interpolation into command-exit-zero, which an evaluator-local
fix would have missed entirely.

Also fixed, same review: `todo complete .` and `..` passed containment
(they resolve to the pending dir, which IS inside the root) and then threw an
uncaught EISDIR with an absolute-path stack trace. Now a clean USAGE
rejection naming the real reason — "todo name is not a file" — rather than
borrowing the escape message, which would have stated something false.

Ripples discharged BEFORE the verification checkpoint rather than after, per
the Phase 1 retrospective: docs/reference/gate-predicates.md and
docs/CLI-TOOLS.md document the new constraints, CONTEXT.md records why
containment lives at the router rather than the evaluator, the changeset is
written, and the install-tree goldens were regenerated to confirm unchanged
(no new shipped file) rather than assumed.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This commit is contained in:
sim
2026-09-12 09:33:32 -04:00
parent 3925839f2a
commit 374300da17
9 changed files with 132 additions and 7 deletions

View File

@@ -81,6 +81,15 @@ the gate context; all others are left untouched for `sh` to interpret:
An undefined placeholder interpolates to the empty string.
**`--phase-dir` is confined to the project.** The value is validated to resolve
inside the project root before any predicate is evaluated; one that escapes is
rejected as a usage error rather than evaluated. This applies to both kinds —
`artifact-frontmatter-equals` resolves its artifact under that directory, and
`command-exit-zero` interpolates it into `${PHASE_DIR}` — so an unconfined value
would let a **blocking** gate return `block: false` on evidence from a directory
the caller chose (#4354). An absolute path inside the project is still accepted;
absolute is not a synonym for escaping.
**Sandbox.** cwd = project root; env = inherited from the GSD process; killed
(SIGTERM) on timeout. The command runs as the user, on the user's machine —
there is no sandbox boundary vs. the user's own shell. See ADR-2008 "Trust