Distillation

Date: Sep 16, 2026

• 🟡 (20:11) User asked to implement all accepted recovery correctness/security findings in /home/byk/Code/opencode-lore-recall-recovery, working in additive child change pqvltmkr over immutable parent spyyzvyy/3f3ae768.
• 🔴 [enforced-workflow] (20:11) User authorized file edits but prohibited commit, push, rebase, bookmark operations, server operations, and production DB access.
• 🔴 [enforced-workflow] (20:11) User required reading AGENTS.md, quality/REVIEW.md, and scripts/, and required edits via apply_patch.
• 🔴 [requested-tests] (20:11) User mandated TDD: deterministic real-path tests must fail on the current parent behavior before source fixes and then pass.
• 🔴 [requested-mutation-tests] (20:11) User required independently removing each relevant guard, proving its named test fails, and restoring the guard exactly.
• 🔴 (20:11) Recovery finding 1: malformed non-JSON event: codex.rate_limits must never be generically forwarded in either principal or continuation flows; it must be suppressed or rejected before opaque forwarding, with no sentinel or quota-accounting escape.
• 🔴 (20:11) Recovery finding 2: the byte-cap fixture must cross 16 KiB before 64 events using max-schema payloads; deletion of the byte guard must fail a test; coverage must include a live principal-plus-continuation request-wide case.
• 🔴 (20:11) Recovery finding 3: a failed continuation retry must restore the request-wide public quota ledger baseline so quota updates from the failed attempt never consume successful retry cap.
• 🔴 (20:11) Recovery finding 4: disguised SSE served as application/json must receive recovery-specific strict successful-completion validation for OpenAI, Anthropic, and Gemini; hostile text, role, failure, missing identity, missing model, or missing usage must produce a fixed 502, rollback, and no persistence.
• 🔴 (20:11) Recovery finding 5: successful OpenAI JSON recovery must require numeric prompt_tokens and completion_tokens; usage:{} must be rejected.
• 🔴 (20:11) Recovery finding 6: non-OK error-body reads must use a short, dedicated, bounded timeout; even when hostile read/cancel never settles, the caller proceeds to recovery; cancellation must be fire-and-forget, detail must be empty, and logging must be fixed.
• 🔴 (20:11) Recovery finding 7: assertSuccessfulResponsesCompletion must use a terminal type/status registry rather than blanket completed; valid failed function_call and image_generation_call output items must remain accepted in ordinary non-stream parsing, while recovery usability must still reject failed tool handoffs.
• 🔴 (20:11) Recovery finding 8: strict-envelope guard discrimination must expand across Anthropic, Vertex, OpenAI Chat, Responses/Codex, and Gemini, independently covering ID, model, usage, role, cardinality, content, tool status, and terminal status guards; tests should use concise tables.
• 🔴 (20:11) User required preserving cache isolation, exactly one synthesis, no warning-only success, stream/consumer coupling, atomic usage/quota handling, EOF-only persistence, and fixed logs/privacy.
• 🔴 [requested-tests] (20:11) User required running the exact formerly failing foreground-body-limit cases, all new focused tests, the full affected recovery matrix, gateway typecheck, and formatting.
• 🔴 (20:11) User required the final report to include fail-first counts, mutant counts, passing counts, changed files, and residual risks.
• 🔴 (20:11) If a recall-context warning occurs, user directed using lore recall --project /home/byk/Code/opencode-lore --scope project --json or the local UI and continuing; never return only warning/empty.
• 🔴 (20:16) User-provided jj-guide establishes that the repository must use jj rather than git when .jj/ exists; interactive flags are forbidden, descriptions/commits require -m "msg", mutations must be verified with jj st and jj log, change IDs are preferred, and immutable commits must never be rebased or described.
• 🔴 (20:16) User-provided jj-guide states that jj never fails on conflict; rebase, new, and squash can record conflicts in the resulting commit, after which jj st must be checked and files resolved manually rather than with interactive jj resolve.
• 🟡 (20:16) jj-guide recovery commands documented: jj undo reverses the last operation, jj op log displays operation history, jj op restore <op-id> rewinds repository state, and jj workspace update-stale repairs stale working-copy errors.
• 🟡 (20:17) AGENTS.md identifies the project as Lore, a TypeScript pnpm monorepo using Node.js >=22.5 and SQLite with WAL mode plus FTS5; packages include packages/core/, packages/gateway/, packages/opencode/, and packages/pi/.
• 🔴 (20:17) Project review policy in AGENTS.md requires quality/REVIEW.md regression-test discipline, adversarial-order setup, fan-out registry coverage, recurring bug-class batteries, and a two-reviewer workflow; every non-trivial PR requires an adversarial correctness review before merge.
• 🔴 (20:17) Project tests use a temporary SQLite database through packages/core/test/setup.ts, never the production DB.
• 🔴 (20:17) Project release policy forbids manually bumping versions or editing CHANGELOG.md; Sentry Craft handles both.
• 🟡 (20:17) Root package.json uses pnpm@10.28.0, Volta Node 24.16.0, and scripts including pnpm test, pnpm run typecheck, pnpm run lint, pnpm run format, pnpm run format:check, pnpm run build, and Stryker mutation testing via pnpm run mutation.
• 🟡 (20:17) packages/gateway/package.json identifies @loreai/gateway version 0.40.0; its scripts include typecheck: tsc --noEmit, build: tsx script/build.ts, bundle: tsx script/bundle.ts, and start: tsx src/index.ts.
• 🟡 (20:17) Repository scripts/ contains exactly 12 entries: bump-version.sh, cch-scan.c, check-cc-version.ts, check-docs-links.mjs, check-patches.ts, check-preview-links.mjs, check-social-meta.mjs, extract-cch-seed.ts, generate-config-docs.ts, generate-env-docs.ts, generate-release-checksums.mjs, and smoke-install.mjs.
• 🟡 (20:20) Repository inspection showed a clean working copy: @ is empty change pqvltmkr at commit f4c25568, described fix(recall): harden quota metadata; parent @- is spyyzvyy at commit 3f3ae768, bookmark fix/recall-failure-synthesis, described fix(recall): validate recovery metadata.
• 🟡 (20:20) jj reported a conflicted bookmark named fix/responses-private-references; the working-copy files themselves had no changes.
• 🟡 (20:20) Code search located Codex quota limits in packages/gateway/src/codex-rate-limits.ts: MAX_CODEX_RATE_LIMIT_EVENTS = 64, MAX_CODEX_RATE_LIMIT_BYTES = 16 * 1024, event validation around lines 68-85, event-count guard around line 102, and encoded-byte guard around line 107.
• 🟡 (20:20) Code search located assertSuccessfulResponsesCompletion in packages/gateway/src/stream/openai-responses.ts around line 255, codex.rate_limits handling around lines 618 and 2380, and terminal validation around line 2032.
• 🟡 (20:20) Code search located recovery entry points in packages/gateway/src/recall.ts: isUsableRecallContinuation around line 1191 and runRecallRecovery around line 1769.
• 🟡 (20:20) Code search located recovery and response-limit paths in packages/gateway/src/pipeline.ts, including continuation usability checks around lines 7634, 11239, 18517, 18637, and 18650; Codex rate-limit handling around lines 11083 and 11490; MAX_FOREGROUND_RESPONSE_BYTES = 4 * 1024 * 1024; MAX_FOREGROUND_ERROR_BYTES = 64 * 1024; JSON completion validation around line 12163; and runRecallRecovery usage around line 18481.