Skip to content

fix(agentic-engineering): require complete identity-bound forge observations - #475

Merged
devantler merged 2 commits into
mainfrom
codex/forge-observations-468
Oct 4, 2026
Merged

devantler merged 2 commits into
mainfrom
codex/forge-observations-468

Conversation

@devantler

Copy link
Copy Markdown
Contributor

🤖 Generated by the Agentic Engineer

Why

Portable forge observations could erase an unresolved review or failed workflow conclusion from ambiguous JSON. The survey examples could also report a clear candidate from an incomplete response, a different issue or an unfinished census. Blank routing references could appear verified.

What

Require unique decoded JSON keys before semantic parsing, complete GraphQL envelopes, exact requested issue identities, integer counts and terminal census pagination. Require visible routing verification references. Ship a pinned shared JSON observer with consumer onboarding in agentic-engineering 6.1.2.

Validation: new counterexamples fail before repair and pass afterward. Tests execute the installed readers and the actual filters extracted from the shipped agent. Existing reader/routing suites, full ShellCheck inventory, manifests, generated digests and release gates pass. All fixtures are offline; no routing or inference is enabled.

Fixes #468
Fixes #469
Fixes #470
Fixes #471
Fixes #472

@devantler devantler left a comment

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🤖 Generated by the Agentic Engineer

Self-review (fallback — CodeRabbit, Codex and Cursor Bugbot unavailable)

Reviewed commit: de0e1ac

All three provider sources and this PR's complete reviews, conversation, threads and native provider checks were freshly read at 2026-10-04T17:25:44Z. This PR has no provider verdict, request in flight or unresolved finding:

  • CodeRabbit: #473 (comment) — authenticated included-review limit, updated 2026-10-04T17:13:55Z. The account's included allowance is exhausted; the reply states recovery in 28 minutes (17:41:55Z). No usage-based review billing is authorized.
  • Codex: devantler-tech/ksail#7433 (comment) — authenticated account-wide code-review usage limit, unchanged since 2026-10-04T02:11:02Z. No reset is stated; recovery requires renewed allowance or account action. No paid credits are authorized.
  • Cursor Bugbot: devantler-tech/ksail#7481 (comment) — authenticated user/team usage or spend limit, unchanged since 2026-10-04T12:21:20Z. No timed reset is stated; recovery requires renewed allowance or administrator action.

Reviewed the complete change and final repair for correctness, input boundaries, copied-consumer behavior and unnecessary complexity. The shared observer rejects repeated decoded JSON paths before semantics, including nested objects, escaped-equivalent keys and later pages. It preserves valid independent documents and paginated arrays. Both retained readers use the same observed bytes for raw and semantic parsing; failure cannot become an empty success. The library is included in onboarding and its executable permission and digest match the declared runtime asset. No native CI access boundary is widened.

The actual shipped survey projections require an absent errors field or exactly [], exact independently selected issue identities, integer counts with consistent bounds, and terminal complete census pagination with stable totals and unique issue/cursor identities. They preserve known empty and positive results. The advisory routing evaluator refuses whitespace, control, format or default-ignorable-only evidence references, preserves visible international references and still reports executionAdmitted:false. References remain caller reports, never authenticated admission evidence.

User-path evaluation exercised the actual counter/classifier against valid single/multipage observations and repeated-key malformed observations; exercised copied guards/readers with their shared library, including missing-library refusal; and extracted and executed the actual published survey filters against complete empty/multipage controls and malformed envelopes, foreign issue identities, fractional counts and incomplete pagination. The real routing CLI rejected five invisible-only references while retaining the international positive. The full required bundled suite passed all 15 suites, plus Go checks, ShellCheck, manifests, generated digests, version/cache-key parity and changelog gates. Independent read-only review findings were reproduced and repaired, and its final review found no remaining issue. Hosted CI remains mandatory before promotion; its lint job is still running at this head.

Verdict: no P0/P1 findings

@devantler
devantler marked this pull request as ready for review October 4, 2026 17:32
@devantler
devantler merged commit d88451d into main Oct 4, 2026
46 checks passed
@devantler
devantler deleted the codex/forge-observations-468 branch October 4, 2026 17:33
@devantler

Copy link
Copy Markdown
Contributor Author

🤖 Generated by the Agentic Engineer

Post-merge user-path verification passed from freshly fetched main at d88451d. The actual raw observer, retained review readers, published survey projections and advisory routing CLI all pass their positive and negative controls. Repeated decoded keys, malformed/partial envelopes, foreign identities, fractional or incomplete counts/pagination and invisible-only references stay UNKNOWN or HOLD. Valid empty/multipage observations and visible international references remain accepted. Issues #468–#472 are closed and every card was read back as Done.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment