Gotcontext Memory Hitl Honesty

gotcontext-memory HITL honesty

When to use

• Editing gotcontext-memory dream / store / review / MCP / doctor / corpus
• Claims about “same as JARVIS/omega dream”
• Auditing preference extraction or concurrent accept

For scale/digests load gotcontext-memory-streaming-digests.
For resurrection/prevalence load gotcontext-memory-claim-lifecycle.
Registry: docs/SKILLS.md.

Do not confuse

Parity target = HITL shape only, not LLM brain or auto-supersede.

Load-bearing rules (lessons L1–L20)

1. Fake-green tests — never catch ReferenceError as skip; only EPERM/EACCES/ENOTSUP.
2. Detect≠enforce — doctor caps must fail when over; share countIndexLines.
3. One lock key — always locks/<sha256(rel)>.lock.
4. Index under lock — regenerate MEMORY.md after target mutate, still locked; multi-process oracle.
5. MCP — memory_read = MEMORY.md|memory/**; memory_commit default-off (mcp.allowCommit).
6. Prior PASS ≠ closed — re-audit after security claims.
7. Alive ≠ useful — report staged content + accepts, not only run success.
8. Preference FP — require please remember|from now on; deny pong/ping/health.
9. Overlay types — deletes: [] not {}.
10. dream.enabled — enforce or stop pretending; CLI needs --force when false.
11. Scale — stream digests; never readFile whole multi-GB transcripts; truncated ≠ malformed.
12. No resurrection — TWO mechanisms, not one. REJECTED claims are suppressed by an immutable claimKey (loadSuppressedClaims reads proposals/rejected/, src/dream/run.ts). ACCEPTED claims are stopped by a SEPARATE settled-state check — if (storeHashes.has(targetPath)) continue;. Conflating them hides the real bug: accepting a re-proposed create over a materialised target silently overwrites any human edit made since. Never key claim identity on base_hash — it changes the instant the target exists, minting a "new" id for a settled claim (negative cache / tombstone; a tombstone that never expires is itself an anti-pattern).
13. Prevalence window — --max-sessions (default 400); sort by evidence strength.
14. Dirty-tree commits — stage only your paths; strip .tensor-grep/ before merge.
15. Cursor .vscdb — on the digest path with shared classifyText (BL-DRM-016 closed 2026-08-10).
16. Secret fixtures — never commit contiguous live-shaped tokens; split strings for scanners.
17. Stratified window — not newest-N alone (selectDigests).
18. Efficacy — score accepted patterns; thin windows → INSUFFICIENT_DATA only.

Authority

• docs/HONESTY.md, docs/LESSONS_2026-08-09.md, docs/BACKLOG.md, docs/CEO_UPDATE_2026-08-09.md
• Sibling: transcript-dream-hitl (omega/JARVIS)
• Companions: gotcontext-memory-streaming-digests, gotcontext-memory-claim-lifecycle

Verify

cd gotcontext-memory && npm test && npm run lint && npm run build

17. Prevalence is clustered by STRING, not by meaning. minePrevalence (src/dream/digest.ts) buckets by a normalised signalKey (lowercased; paths/hashes/digits stripped) and requires

    =2 DISTINCT sessions before proposing, reporting k/n sessions with line-numbered citations. Two phrasings of the same problem land in different buckets and are NOT merged. Say "counted", never "inferred" — this is not an LLM brain.

18. receiptCode names the cause. src/review.ts returns CAS_CONFLICT, SECRET_DETECTED, INDEX_CAP, TARGET_MISSING, INVALID_PROPOSAL, PROPOSAL_EXPIRED, PATH_VIOLATION, or INTERNAL_ERROR (fallback for genuinely unrecognised errors only). This replaced filing every unrecognised failure as INDEX_DRIFT_OR_CAS, which sent operators to reconcile an index that was never the problem. Read receipts/*.error.json when an accept fails.
19. Scale is a correctness property. Whole-file corpus reads die at the V8 string ceiling on a real machine; the digest path is load-bearing, not an optimisation. See ingest-a-corpus-that-exceeds-memory.