paranoia: Instructions file for Codex

Repository agent instructions

Issue #139 follows docs/lfs-proposal-cleanliness-139-plan.md. Proposal cleanliness may recognize canonical extension-free v1 Git LFS pointers through exact size and SHA-256 of streamed checkout bytes plus the cached lfs attribute. Never execute clean/process filters. Preserve exact committed proposal preimages and all existing HEAD/index, filesystem, sparse-checkout, submodule and post-response checks.

Issue #138 follows docs/proposal-disposition-138-plan.md. Caller proposal disposition is bounded lineage-bound accounting only, never repair or clearance authority. Preserve the original settled review on supplemental receipt failure. Executing agents SHOULD use the proposed diff after inspection and validation. Give a concrete one-line reason for every addressed target repaired differently; send the exact proposal audit basename and disposition on the next tracked round. Omission remains API-compatible and visibly reports none-recorded, never acceptance.

Issues #131–133 follow docs/issues-131-133-plan.md. Consolidation receives canonical bare-anchor manifests; derive output rationales from validated source summaries, remedies and evidence instead of inventing transport-metadata findings. Repeat this contract on the existing validation retry. Historical settled metadata debt is not auto-closed: normal reviewer correction owns withdrawal. Bound plan rebut may cite any coordinate within the latest stored reviewed-plan bound, retaining the original finding and separate concession. Arbitration diagnostics name the decisive citation and failed substantiation condition without weakening carried grounding. Existing historical source-diff allowance metadata may be refreshed for the changed bytes; preserve original provider evidence and do not claim renewed native acceptance.

Tiered review is the default on main (promoted 2026-10-02 from codex/beta-tiered-review; wire release name tiered-review-beta-1, kept stable). docs/beta-tiered-review-plan.md is its design record and remains the bounded contract for this behaviour; it supersedes the opt-in proposal default below. Omitted propose_patch means automatic eligibility (never coerced to a boolean); false makes zero proposal calls; explicit true keeps its preflight rejection. Proposals run only after a blocked census or blocked cold final, resuming that source session with its actual model/effort. Structural roles route by durable phase (review_model_policy, default tiered: census/final strongest, correction gpt-6.1-sol/Opus 5.5 at high; gpt-6.1-sol needs Codex CLI 0.159.3 or later); effort resolves per model family (effort_by_model, then global effort, then release family default); claim roles, query, rebut and arbitrate are not routed. Every tiered acceptance needs a strongest-model cold final bound by the durable acceptance record to the exact snapshot, including after a clean census; legacy or custom-override clear never qualifies. Historical acceptance records change only existing allowance hash/scope metadata for changed files (plan §10.8); never add entries or alter provider evidence. Do not claim savings or unchanged quality: a 2026-10-02 controlled comparison (Claude engine, one frozen target, n=3 per arm) found review quality not distinguishable from the prior default and cost per round roughly equal, which supports neither claim.

Historical opt-in contract: reviewer-authored patch proposals follow docs/reviewer-patch-proposal-plan.md and were delivered as an explicit opt-in on main. The user declined the expensive paired campaign and approved that narrower opt-in merge; do not claim campaign-proven speed/cost benefit. Inspect every candidate rather than applying it blindly, and pass false for architectural/authority gaps or when a complete verified repair is already prepared. Proposal generation is supplemental after durable settlement: ordinary proposal-local preparation, provider, validation, logging and rendering failures must return the original review and exact trailer with an inert UNAVAILABLE diagnostic; retain every completed proposal attempt and rejected reply in the bounded supplemental failure audit, including when validation retry prompt or schema preparation fails before another dispatch. Branch admission requires a clean caller tree; pinned blob sizes and one shared retry-surviving source allowance gate reads, complete Git tree inventories retain directories, and returned suitability rechecks ref and preimages. Plan review keeps its historical universal-newline text/digest/snapshot semantics while proposals bind exact captured bytes. Source-processing and localized-discovery-omission claim debt are never factual rewrite targets without a current full semantic adjudication. Reject a replacement group when its final bytes equal its pinned original, while preserving valid adjacent edits and the existing single validation retry for both branch and plan. Branch proposal cleanliness must compare HEAD, index, untracked paths and raw checkout bytes without invoking repository-selected clean/process filters, both before dispatch and after response. Require real-directory checkout ancestors for every tracked leaf, preserving sparse-checkout and submodule handling. Validate create paths as one prospective case-folded inventory, including new/new and existing ancestor file/directory collisions. Retain non-ASCII tracked names in that collision inventory while keeping proposed paths portable ASCII. Treat an absent leaf, an absent ancestor, or an empty real directory at a tracked uninitialized submodule as clean, recursively validate initialized submodules, and mark create suitability stale for ignored destinations or non-directory checkout ancestors using lstat. If such debt is the only remaining claim blocker, report proposal UNAVAILABLE rather than NOT-NEEDED only while claim verification is active; disabled verification leaves retained claim history inactive for supplemental status and targeting. Runner telemetry excludes both the requested author handle and returned proposal handle when a supplemental receipt is missing. Known proposal-only sessions are not rebut authority even with an explicit engine and cannot be promoted by a rebut continuation; an independent completed critique or query remains authoritative.