fix(claude): fail closed on artifact absence claims - #290
fix(claude): fail closed on artifact absence claims#290Tarik Skalić (tarikskalic33) wants to merge 36 commits into
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Deploying with
|
| Status | Name | Latest Commit | Preview URL | Updated (UTC) |
|---|---|---|---|---|
| ✅ Deployment successful! View logs |
aegisomega | 54a74bd | Commit Preview URL Branch Preview URL |
Aug 22 2026, 02:49 AM |
|
Deployment failed for project hook-generator with the following error: Learn More: https://vercel.com/tariks-projects-e9198507?upgradeToPro=build-rate-limit |
|
Deployment failed for project platform-picker with the following error: Learn More: https://vercel.com/tariks-projects-e9198507?upgradeToPro=build-rate-limit |
|
Deployment failed for project hub with the following error: Learn More: https://vercel.com/tariks-projects-e9198507?upgradeToPro=build-rate-limit |
|
Repository-universe census is now mechanically established on full Git history. TDD RED: run 32515601975 / job 96876437555 failed at the preregistered fixture because GREEN implementation:
So the operator's A subsequent commit 4543b1f wires complete-history/ref hydration into Claude SessionStart. Claude now receives a repository-universe summary before forming repo-wide beliefs. If hydration fails, global absence claims are explicitly forbidden. Epistemic invariant: |
|
Repository reconciliation checkpoint: operator directed full organize/clean/reconcile/merge with no skipped lineage or evidence. A full design is now committed at |
|
Reconciliation Task 1 RED → GREEN checkpoint. VALID RED (after removing the unrelated pytest dependency failure):
GREEN implementation head:
No branch deletion, merge, admission, or other canonical mutation occurred. Discovery remains evidence-only; even complete discovery yields Next: Task 2 — stable WorkID lineage graph. It will be preregistered RED first on this same branch; no new branch. |
|
Deployment failed for project hub with the following error: Learn More: https://vercel.com/tariks-projects-e9198507?upgradeToPro=build-rate-limit |
|
Reconciliation Task 2 RED → GREEN checkpoint. A syntax-error run was explicitly rejected as invalid RED. Valid RED was then established at branch head
GREEN implementation head:
PR-backed work identity is now No canonical merge/deletion occurred. Next Task 3: fail-closed lineage resolver + branch-creation gate, again RED first on this same branch. |
|
Reconciliation Task 3 RED → GREEN checkpoint. VALID RED:
GREEN implementation head:
The branch-creation invariant is now executable: Claude SessionStart now attempts full Git + PR + WorkID hydration and fails closed if continuity cannot be established. UserPromptSubmit injects the branch-creation gate every prompt. No canonical merge or branch deletion occurred. Next: Task 4 — deterministic classification plus append-only replayable reconciliation decision chain, RED first on this same branch. |
|
Deployment failed for project hook-generator with the following error: Learn More: https://vercel.com/tariks-projects-e9198507?upgradeToPro=build-rate-limit |
Why
Claude sessions were producing false global negatives from a narrow scope (current checkout/default-branch search). A concrete example is
ProofPrism: the current GitHub state contains a branch namedclaude/proof-prism-v1-ltti67, but exact content search on canonicalmainfinds noProofPrismimplementation, and that branch is currently the head of unrelated cleanup PR #287. A branch-name hit and an implementation hit are different evidence classes.Change
.claude/hooks/artifact-locator.shto refresh and search all locally known Git refs, worktree paths/content, commit messages, and GitHub PR metadata;IMPLEMENTATION_EVIDENCE_FOUND,NAMED_REFERENCE_FOUND,INCOMPLETE, orNO_MATCHES_IN_COMPLETE_REPO_SCAN;.claude/skills/locate-artifact/SKILL.mdrequiring the locator before any negative existence claim;ABSENCE-CLAIM GATEon everyUserPromptSubmit, so the rule survives ordinary session drift/compaction;mainstate and existence elsewhere in stacked PRs, historical refs, Drive/corpus, or provider workspaces.Safety property
A miss in the current worktree can no longer justify the sentence
X does not exist.A matching ref name also cannot justify
X is implementedwithout inspecting that ref.If remote/PR discovery is incomplete, the locator fails closed and requires a scoped finding instead of a global claim.
Stack
Base:
feat/metacognitive-executive-v1/ PR #285 @955338f6cf2590c1c38b79ab57a9d09692fd560e.No constitutional runtime files are modified.