Conversation
A deep review of v0.1.2..v0.2.0 surfaced 15 defects; an adversarial re-review of the fixes surfaced 8 more. All 23 are fixed and verified. - Add lib/memory-paths.mjs as the single memory-dir resolver (explicit flag, then config.json probed in .ownmem/.memory, then the layout on disk), replacing five divergent copies of the rule. Legacy .memory installations now work across every command. - ownmem embed pins root/memory-dir/index-dir correctly; the bare command prints usage; embed ab gains --cases and an actionable error instead of a raw ENOENT leaking the install path. - ownmem recall rejects unknown flags loudly (they were silently searched as query text) and routes --stdio/--runtime and the other runtime flags to the snapshot resident runtime. - ownmem hook status|enable|disable|serve are reachable and documented. - compile validates the corpus directory before taking the index lock, so it can no longer create an empty .ownmem/ that poisons resolution; the dashboard embedding endpoints share the CLI's snapshot directory. - Optional-layer detection fires only when the layer entry itself is missing; transitive module errors surface unchanged. - ownmem init rejects a memory directory the resolver could never find. - Structure gate: git ls-files --cached -z with a graceful non-git fallback, dynamic lazy-import scanning, directory-driven skill-mirror and schema-id checks, and a shipped-scripts-vs-files invariant. - The npm package ships test/ and benchmarks/ so the self-test and the locked benchmark reproduce from the release artifact. - engines raised to the honest >=20.6.0 (import.meta.resolve floor). - CI: pack gate on all three platforms, cancel-in-progress only for pull requests, benchmark on a single lane. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGkc3Poso3PUfsPR8xc3sH
Temporary observation setup: hook_enabled=true, claude added to the adapters, gates/compiler added to the layers. Revert this commit before release if the repository should not carry the dogfood state. The generated root files (CLAUDE.md, MEMORY_INSTRUCTIONS.md, .claude/) stay untracked -- the locked root layout is a release contract -- and are listed in .git/info/exclude; regenerate them on any machine with ownmem init --update --hook --hosts claude,generic. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGkc3Poso3PUfsPR8xc3sH
…anch Dogfood: the 0.2.1 observation-period runbook becomes a real .ownmem topic (observation_period_bootstrap, routed in the General L2, snapshot recompiled), so any fresh clone of this branch wakes an agent with the full bootstrap and acceptance context through the repository's own memory. The generated agent entry files (CLAUDE.md, MEMORY_INSTRUCTIONS.md, .claude/) are committed so a new machine gets the recall entry point without any manual setup; the locked root layout in the structure gate gains the three entries. Revert together with the dogfood commit before release if the repository should not keep them. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01QGkc3Poso3PUfsPR8xc3sH
The instrumentation health check folded three different things into one "zero variance (a placeholder, or the producer is broken)" warning, and acting on that warning writes false data. recall.completed.cache_hit is always null on purpose: production recall has no result cache, and the producer says so in a comment. "Fixing" it to false would report a measured cache miss that nothing measured. Three names on EXPECTED_MEMORY_COMPONENTS have no emitter anywhere in this package, so every run reported them as code paths that had never run: memory-recall-runtime is named only by the list itself, while memory-read and memory-maintenance belong to CLIs this package does not ship (recall.consumed is emitted by memory-observe, and maintenance.completed has a consumer in features/report.mjs but no producer at all). A permanent false alarm is how a genuinely silent component gets lost in the noise. EXPLAINED_CONSTANT_FIELDS now carries the constants whose second value is impossible, and they are still printed with their reason rather than suppressed: hiding one is exactly how the next reader makes a contractual constant vary. detectConstantPayloadFields keeps its existing `ignore` semantics untouched, so explicit suppression stays explicit and non-default. Two drift guards keep the mapping honest: every payload field the event schema pins to a const must be declared, and every expected component must be named by some source file. Both were verified in reverse -- re-adding the ghost fails the suite with the exact assertion. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01A6ZiTen2S5DGy8g1JLeM7z
A half-finished incremental build let stale vectors read as fresh, with
no visible degradation.
checkpointEmbeddingBatch rewrote the whole artifact from the new corpus
on every batch, so the new corpus_snapshot_id landed on disk from the
very first one. An incremental build starts from {...existing.entries},
so at that moment the artifact claims the new snapshot while carrying
the previous vectors for everything not yet embedded.
staleState in memory-embedding-channel.mjs takes a fast path whenever
the artifact's snapshot id equals the planner's, and that path compares
presence only -- it never compares content_sha256. Every carried-over
entry therefore read as fresh, the channel reported
degraded_reason: null, and ranking ran on vectors it knew were stale.
That inverts the position the channel takes everywhere else: fall back
to the deterministic ranking with a named reason rather than answer
quietly from something unmeasured.
The build now keeps the previous snapshot id for the whole incremental
run and finalizes the new one only after the last batch lands, which
forces the reconciling path that compares content hashes. A full
rebuild carries nothing forward, so presence alone is already correct
there and it keeps the new id from the first batch. An aborted build
never reaches the finalize, leaving the previous id in place; resume is
unaffected because reconciliation skips already-embedded batches by
content hash.
This package had no embedding tests at all, so the suite from the
upstream working repository is added here and wired into npm test,
giving CI 18 cases over build, reconciliation, resume and schema
retirement. The regression case was verified in reverse: reverting the
fix fails it with the exact assertion.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A6ZiTen2S5DGy8g1JLeM7z
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Observation-period WIP for the next release — do not merge until dogfood acceptance completes. Supersedes #1 (auto-closed by an accidental, immediately-reverted push to main; all context and the machine-bootstrap comment live there).
All 23 review defects are fixed and verified; CI was green on all 5 lanes (run 32498277237). This branch now also carries the observation runbook as a repo-owned
.ownmemtopic plus the generated agent entry files (CLAUDE.md,MEMORY_INSTRUCTIONS.md,.claude/), so a fresh clone bootstraps an agent with full context through the repository's own memory — the product dogfooding itself.New-machine bootstrap: clone this branch, then follow
.ownmem/observation_period_bootstrap.md(or #1 (comment)).🤖 Generated with Claude Code
https://claude.ai/code/session_01QGkc3Poso3PUfsPR8xc3sH