diff-rules hardening: fail closed on an unparseable hunk header, share the blocking threshold, bound the park detail to the store's cap #1561

Closed
opened 2026-10-27 13:40:53 +00:00 by sam · 0 comments
Owner

Follow-ups from the review of the gate.diff-rules lane that are not P0/P1 fixes.

  1. Fail closed on a hunk header the walker cannot parse.
    internal/domain/workflow/diff_rules.go: hunkNewStart's three failure returns are
    swallowed (inHunk = ok), so every added line of that hunk is silently dropped and the
    rules report no violation for a diff the parser did not understand. The rest of the
    step fails closed (a settings mistake refuses the definition, a Diff read failure
    fails the stage), so this path should too: propagate a parse failure as a stage error
    and add a malformed-header case to the test table. Blast radius today is theoretical
    (both Trees implementations hand the walker go-git's UnifiedEncoder output), which is
    why it is not fixed in the lane.

  2. Share the blocking threshold instead of re-inlining it.
    diffFinding carries no Blocking(); finding.Level == DiffRuleLevelError is written
    twice (runDiffRules and logRuleFinding). The sibling finding type in the same package
    answers the same question with ReviewFinding.Blocking (review_findings.go:142), and the
    commit that fixed this area before (daeb1f61 "share the blocking threshold") states the
    intent: one source of truth for park-vs-log. Give diffFinding a Blocking() method and
    call it from both sites.

  3. Bound the park detail against the store's real cap.
    diffRuleDetailBytes = 4000 and clip() appends a 15-byte truncation marker beyond that,
    while the store clips a park reason at exactly 4000 (internal/store/store.go), so an
    overflowing detail is re-cut by the store. The comment claims "headroom" that does not
    exist (reviewDetailBytes copies the same idiom). Pick a number below the cap and tie
    the two with a test.

  4. Correct comments that claim more than the code does.

    • diffTargetPath's comment says a path may be quoted/escaped by git, but the code only
      trims quote runes and decodes nothing; the only producer (go-git's unified encoder)
      never quotes. Either drop the git-quoting clause or decode with strconv.Unquote.
    • repoHooks' comment says "the step types replacing the repository-authored hooks"
      (plural) while it contributes exactly one step type and .archie/stages/*.go still has
      no equivalent (see the .archie/stages decision bead).
  5. Two assertions in the new registration test restate the constants they claim to pin
    (diff_rules_test.go: the stage-name comparison and DiffRulesStepName == "gate.diff-rules"),
    which AGENTS.md forbids: they cannot fail for an interesting reason. Keep the Run and
    settings-refusal halves and drop or replace the vacuous ones.

Close reason

Items 2-3 fixed (diffFinding.Blocking; clip now keeps marker within n). Items 1, 4, 5 obsolete: the walker is the shared prreview parser, diffTargetPath/repoHooks and the vacuous assertions no longer exist.


Imported from beads archie-core-s1jz · GitHub #1145 · type task · priority P2 · created 2026-09-21 · closed 2026-10-16

Follow-ups from the review of the gate.diff-rules lane that are not P0/P1 fixes. 1. Fail closed on a hunk header the walker cannot parse. internal/domain/workflow/diff_rules.go: hunkNewStart's three failure returns are swallowed (inHunk = ok), so every added line of that hunk is silently dropped and the rules report no violation for a diff the parser did not understand. The rest of the step fails closed (a settings mistake refuses the definition, a Diff read failure fails the stage), so this path should too: propagate a parse failure as a stage error and add a malformed-header case to the test table. Blast radius today is theoretical (both Trees implementations hand the walker go-git's UnifiedEncoder output), which is why it is not fixed in the lane. 2. Share the blocking threshold instead of re-inlining it. diffFinding carries no Blocking(); `finding.Level == DiffRuleLevelError` is written twice (runDiffRules and logRuleFinding). The sibling finding type in the same package answers the same question with ReviewFinding.Blocking (review_findings.go:142), and the commit that fixed this area before (daeb1f61 "share the blocking threshold") states the intent: one source of truth for park-vs-log. Give diffFinding a Blocking() method and call it from both sites. 3. Bound the park detail against the store's real cap. diffRuleDetailBytes = 4000 and clip() appends a 15-byte truncation marker beyond that, while the store clips a park reason at exactly 4000 (internal/store/store.go), so an overflowing detail is re-cut by the store. The comment claims "headroom" that does not exist (reviewDetailBytes copies the same idiom). Pick a number below the cap and tie the two with a test. 4. Correct comments that claim more than the code does. - diffTargetPath's comment says a path may be quoted/escaped by git, but the code only trims quote runes and decodes nothing; the only producer (go-git's unified encoder) never quotes. Either drop the git-quoting clause or decode with strconv.Unquote. - repoHooks' comment says "the step types replacing the repository-authored hooks" (plural) while it contributes exactly one step type and .archie/stages/*.go still has no equivalent (see the .archie/stages decision bead). 5. Two assertions in the new registration test restate the constants they claim to pin (diff_rules_test.go: the stage-name comparison and DiffRulesStepName == "gate.diff-rules"), which AGENTS.md forbids: they cannot fail for an interesting reason. Keep the Run and settings-refusal halves and drop or replace the vacuous ones. ### Close reason Items 2-3 fixed (diffFinding.Blocking; clip now keeps marker within n). Items 1, 4, 5 obsolete: the walker is the shared prreview parser, diffTargetPath/repoHooks and the vacuous assertions no longer exist. --- _Imported from beads `archie-core-s1jz` · GitHub #1145 · type task · priority P2 · created 2026-09-21 · closed 2026-10-16_
sam added the mediumtask labels 2026-10-27 13:40:53 +00:00
sam self-assigned this 2026-10-27 13:40:53 +00:00
sam closed this issue 2026-10-27 13:40:53 +00:00
Sign in to join this conversation.
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: sam/archie#1561