Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 14f99545e9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
There was a problem hiding this comment.
Pull request overview
Adds a research/design document proposing a typed MemoryFact schema plus a deterministic reconciliation + rendering pipeline intended to replace hand-maintained CURRENT-*.md / MEMORY.md views with generated outputs.
Changes:
- Defines a v0
MemoryFactschema (fields, invariants, canonical-key normalization). - Specifies reconciliation pseudocode, conflict surfacing, and rendering rules for generated views.
- Outlines a phased migration plan and CI hook ordering for eventual implementation.
Merged
4 tasks
AceHack
added a commit
that referenced
this pull request
Apr 24, 2026
…rows (Amara Govern-stage 1/2) Amara's 4th ferry (PR #221 absorb) named populating docs/CONTRIBUTOR-CONFLICTS.md as the Govern-stage action: the schema has existed since PR #166 but the Resolved table was empty despite multiple session-observed contributor-level disagreements that closed with evidence. Backfills three genuine contributor-level conflicts observed this session (narrow scope — not maintainer-directives, which are out-of-scope per the schema's contributor-level disagreement definition): - CC-001: Copilot (PR reviewer) vs Aaron on no-name-attribution rule scope (history-file exemption). Resolved in Aaron's favor via Otto-52 clarification; policy BACKLOG row filed in PR #210. - CC-002: Amara (4th ferry) vs Otto (pre-Otto-67 pattern) on Stabilize-vs-keep-opening-new-frames. Resolved in Amara's favor; Otto pivoted at Otto-68 to execute her roadmap; 3/3 Stabilize + 3/5 Determinize landed via PRs #222/#223/#224/#225/#226. - CC-003: Codex (PR reviewer) vs Otto (initial framing) on citing-absent-artifacts. Resolved in Codex's favor via fix commits 29872af/1c7f97d on PRs #207/#208; pattern now discipline (distinguish merged-on-main from proposed-in-PR-open). All three rows follow the schema's 8-column layout and include the full Resolution-so-far / Scope / Source cells the schema requires. No retroactive Aaron→human-maintainer sweep of prior rows; schema's rule 1 (resolutions are additive) honored. This is 1/2 of Amara's Govern-stage work. 2/2 is the authority-envelope + escalation-path ADR (deferred, M-effort). Part of Amara's 4-stage remediation roadmap (Stabilize → Determinize → Govern → Assure). Otto-75 tick.
AceHack
added a commit
that referenced
this pull request
Apr 24, 2026
…rows (Amara Govern 1/2) (#227) * govern: CONTRIBUTOR-CONFLICTS backfill — 3 resolved session-observed rows (Amara Govern-stage 1/2) Amara's 4th ferry (PR #221 absorb) named populating docs/CONTRIBUTOR-CONFLICTS.md as the Govern-stage action: the schema has existed since PR #166 but the Resolved table was empty despite multiple session-observed contributor-level disagreements that closed with evidence. Backfills three genuine contributor-level conflicts observed this session (narrow scope — not maintainer-directives, which are out-of-scope per the schema's contributor-level disagreement definition): - CC-001: Copilot (PR reviewer) vs Aaron on no-name-attribution rule scope (history-file exemption). Resolved in Aaron's favor via Otto-52 clarification; policy BACKLOG row filed in PR #210. - CC-002: Amara (4th ferry) vs Otto (pre-Otto-67 pattern) on Stabilize-vs-keep-opening-new-frames. Resolved in Amara's favor; Otto pivoted at Otto-68 to execute her roadmap; 3/3 Stabilize + 3/5 Determinize landed via PRs #222/#223/#224/#225/#226. - CC-003: Codex (PR reviewer) vs Otto (initial framing) on citing-absent-artifacts. Resolved in Codex's favor via fix commits 29872af/1c7f97d on PRs #207/#208; pattern now discipline (distinguish merged-on-main from proposed-in-PR-open). All three rows follow the schema's 8-column layout and include the full Resolution-so-far / Scope / Source cells the schema requires. No retroactive Aaron→human-maintainer sweep of prior rows; schema's rule 1 (resolutions are additive) honored. This is 1/2 of Amara's Govern-stage work. 2/2 is the authority-envelope + escalation-path ADR (deferred, M-effort). Part of Amara's 4-stage remediation roadmap (Stabilize → Determinize → Govern → Assure). Otto-75 tick. * govern: annotate CC-002/CC-003 Source cells — PR #221/#219 open, not yet on main Applies CC-003's own discipline (cite-as-open-not-landed) to CC-002 and CC-003 themselves. Both rows cited `docs/aurora/2026-04-23-amara-memory-drift-*` and `docs/aurora/2026-04-23-amara-decision-proxy-*` without the "not yet on main" marker — the files are added by PRs #221 / #219 which are still open. Drain for PR #227 review threads PRRT_kwDOSF9kNM59RFIx and PRRT_kwDOSF9kNM59RFJE (dangling file refs at lines 132, 133). * fix: markdownlint auto-fixes on governance doc Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
…y + smart-quote + pseudocode init + present-with-schema Six substantive Codex findings on memory-reconciliation algorithm doc: P0 (line 202) — retraction semantics inconsistency: reconcile() filtered by status == 'active' which masked the intent. Added explicit retraction-semantics docstring: - Facts transition via explicit FactRetracted / FactSuperseded events; never deleted, only marked. - reconcile() ignores retracted/superseded for liveness but STILL considers them when checking version-chain integrity. - Updated chain check to operate over ALL facts in the group (including retracted/superseded), not just active ones — chain integrity needs the full history. P1 (line 187) — stable fact identity vs grouping key: Distinguished fact ID (stable identity, unique) from (subject, predicate, canonical_key) grouping tuple (which multiple facts can share under invariant-2's collision case). Comment makes the distinction explicit. P1 (line 270) — MEMORY.md cap inconsistency: Default 30KB exceeded FACTORY-HYGIENE row #11 cap (24,976 bytes). Updated to 24,000 bytes — strictly under the hard cap with ~1KB headroom for header/annotation overhead. P2 (line 130) — smart-quote example ambiguous: Both sides showed plain ASCII ('"' / "'"). Replaced with explicit Unicode codepoint references (U+201C/D for double, U+2018/9 for single) so the rule is unambiguous in plain-ASCII source. P2 (line 186) — pseudocode by_key[k] used before init: Switched to defaultdict(list); added a comment noting the equivalence to 'if k not in by_key: by_key[k] = []' for non-Python implementers. P2 (line 216) — CONTRIBUTOR-CONFLICTS.md 'empty' wording: File is present and contains a schema; just unpopulated. Updated text to 'present-with-schema-but-unpopulated; this design starts populating it via the generator'.
AceHack
force-pushed
the
research/memory-reconciliation-algorithm-design-amara-determinize-l
branch
from
14f9954 to
16b6ccf
Compare
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 16b6ccf67d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ize L-effort item) Amara's 4th ferry (PR #221 absorb) centerpiece proposal: replace hand-maintained CURRENT-*.md distillations with generated views over typed memory facts. Her sketch was ~40 lines of Python; this is the design that downstream implementation follows. ~380 lines covering: - MemoryFact record schema (id / subject / predicate / object / source_kind / source_path / source_anchor / timestamp_utc / supersedes / priority / status / confidence / tags) - 6 schema invariants (at-most-one-active-per-canonical-key + monotone-timestamps-on-chain + retraction-leaves-trail + ...) - Canonical-key normalization rules (7 apply; 3 deliberately NOT applied to preserve distinctions) - Reconciliation pseudocode (group by canonical key, detect conflicts, follow supersession chains) - Conflict output format → CONTRIBUTOR-CONFLICTS.md rows - Rendering rules for CURRENT-<maintainer>.md + MEMORY.md - 5-phase incremental migration (schema adoption → generator prototype → mechanical backfill → cutover → LLM extraction) - CI integration hooks composing with rows #58, #59, #12 - Worked examples (MF-2026-04-23-001 "Aaron endorses deterministic reconciliation"; MF-2026-04-23-004 "Aaron grants full GitHub access") - 5 open questions for Phase 1 PR design decisions Composes with: - Otto-73 retractability-by-design foundation — MemoryFact status (active / superseded / retracted) is the retraction- native primitive at the memory substrate - PR #222 decision-proxy-evidence — consulted_memory_ids can now reference MemoryFact.id directly - PR #225 memory-reference-existence CI (row #59) — generated output preserves the invariant by construction - Zeta's ZSet algebra — MemoryFact records ARE Z-set entries at the memory layer; same primitive, different surface Addresses MEMORY.md cap-drift (Otto-70 snapshot-tool surfaced 58842 bytes vs. 24976-byte cap): a generated index can be bounded by construction (top-N most-recent, archive the rest). Not implementation. Research doc only. Downstream arc: schema adoption (S) → generator prototype off-CI (S-M) → mechanical backfill (M) → cutover with retractability (M) → LLM-assisted extraction (L research). Amara Determinize-stage: 3/5 (with this PR). ✓ Live-state-before-policy (PR #224) ✓ Memory reference-existence lint (PR #225) ✓ Memory reconciliation algorithm design (this PR) Remaining: - Generated CURRENT-*.md views (L; this doc's Phase 2) - Memory duplicate-title lint enforcement (partial via AceHack PR #12; graduates via batch-sync) Per Aaron Otto-73 retractability foundation: the design itself embodies the thesis — supersession + status + retraction make the memory layer's reconciliation deterministic, same primitive as Zeta's data layer. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
…y + smart-quote + pseudocode init + present-with-schema Six substantive Codex findings on memory-reconciliation algorithm doc: P0 (line 202) — retraction semantics inconsistency: reconcile() filtered by status == 'active' which masked the intent. Added explicit retraction-semantics docstring: - Facts transition via explicit FactRetracted / FactSuperseded events; never deleted, only marked. - reconcile() ignores retracted/superseded for liveness but STILL considers them when checking version-chain integrity. - Updated chain check to operate over ALL facts in the group (including retracted/superseded), not just active ones — chain integrity needs the full history. P1 (line 187) — stable fact identity vs grouping key: Distinguished fact ID (stable identity, unique) from (subject, predicate, canonical_key) grouping tuple (which multiple facts can share under invariant-2's collision case). Comment makes the distinction explicit. P1 (line 270) — MEMORY.md cap inconsistency: Default 30KB exceeded FACTORY-HYGIENE row #11 cap (24,976 bytes). Updated to 24,000 bytes — strictly under the hard cap with ~1KB headroom for header/annotation overhead. P2 (line 130) — smart-quote example ambiguous: Both sides showed plain ASCII ('"' / "'"). Replaced with explicit Unicode codepoint references (U+201C/D for double, U+2018/9 for single) so the rule is unambiguous in plain-ASCII source. P2 (line 186) — pseudocode by_key[k] used before init: Switched to defaultdict(list); added a comment noting the equivalence to 'if k not in by_key: by_key[k] = []' for non-Python implementers. P2 (line 216) — CONTRIBUTOR-CONFLICTS.md 'empty' wording: File is present and contains a schema; just unpopulated. Updated text to 'present-with-schema-but-unpopulated; this design starts populating it via the generator'.
…tired groups P1 (line 210) — chain-HEAD liveness, not 'any active in group': The reconcile filter marked a key live whenever any record in the group had status==active. That's wrong — a key with active(t=1) → retracted(t=2) has an earlier active record but the HEAD of the supersession chain is retracted, so the key is not live. Fix: `follow_supersession_to_head(group)` walks supersedes-pointers to find the most-recent record; liveness keyed on its status == active. P2 (line 224) — chain integrity for fully retired groups: The chain-integrity check looped over `accepted.items()`, which only included keys with at least one active record. Retired groups (all members retracted/superseded) could have broken chains and we'd silently miss them. Fix: loop over `by_key.items()` (all groups, including fully retired ones). Chain integrity is independent of liveness.
AceHack
force-pushed
the
research/memory-reconciliation-algorithm-design-amara-determinize-l
branch
from
16b6ccf to
32721c0
Compare
AceHack
deleted the
research/memory-reconciliation-algorithm-design-amara-determinize-l
branch
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 32721c01b8
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
P1 (line 222) — invariant-2 winner via priority tie-break: The reconcile pseudocode picked chain_head as the winner in the invariant-2 violation path. Per invariant 6, simultaneous actives must use priority tie-break (max priority, then max timestamp). chain_head determines liveness but NOT winner-among-actives. Updated: actives includes the chain_head + any sibling active records; winner = max by (priority, timestamp_utc). chain_head wins only if it has the highest tuple. P1 (line 128) — markdown formatting delimiters not raw chars: The canonicalization rule removed * _ ` as raw characters, which corrupted identifiers like _internal_var or __private. Reworded to 'strip formatting delimiters' — unwrap text from PAIRED markdown spans, not remove every occurrence. Specifically: **text** / *text* / _text_ / `text` → text, where _text_ stripping requires the contents to match [A-Za-z0-9-]+ so identifiers survive. Single occurrences and unpaired delimiters preserved. P2 (line 296) — MEMORY.md dedup by source_path: Phase 3 backfills multiple typed facts from a single prose memory file. Without dedup, each MEMORY.md index row would emit per-fact rather than per-file, multiplying the index. Updated: dedup by source_path, picking the highest-priority fact per file as the row's representative; row description = first-sentence of that fact; tags = union of all facts from that source_path.
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
Codex post-merge findings on PR #226 (after the first follow-up #433 merged): P1 (line 244) — CONTRIBUTOR-CONFLICTS.md is already populated: The doc said the file is 'present-with-schema-but- unpopulated'. Wrong: CC-001..CC-003 are filled in (no-name-attribution-rule scope, Stabilize-vs-keep-opening frames, absent-artifact-citation discipline). Updated text to reflect the populated state. P1 (line 270) — Conflict-row format mismatch: Proposed format used '### CONF-<date>-<nnn>' + bullet fields. Actual file uses CC-### IDs in a markdown table with columns (Conflict ID | Date | Question | Parties | Positions | Resolution | Scope | Source). Replaced the proposed format with a template that matches the existing table schema. CC-### counter continues from the highest existing ID (next auto-detected = CC-004). Generator preserves manually-curated rows and only appends auto- detected ones.
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
P1 (L269) — clarify Open-table targeting: Replaced "appending machine-generated rows" (ambiguous given the file has separate Open / Resolved / Stale tables) with "inserting machine-generated rows into the Open table (or a delimited autogenerated subsection)". Without this, a literal append-to-EOF implementation wouldn't land in Open. P1 (L290) — specify idempotent generator strategy: Added explicit canonical-key → CC-NNN mapping. Generator updates in-place when key already matches; allocates new CC-NNN only when unmapped. Avoids unbounded growth / duplicate entries on repeated CI runs. Documents both strategies (in-place update + delimited autogenerated subsection). P2 (L275) — placeholder labels match column headers: Renamed schema placeholders from "Parties:" / "Resolution:" to "Between:" / "Resolution-so-far:" matching the actual headers in docs/CONTRIBUTOR-CONFLICTS.md (verified via grep). Reduces drift between the design doc and the live schema.
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
) * drain(#226 follow-up 2: CC log schema + populated state) Codex post-merge findings on PR #226 (after the first follow-up #433 merged): P1 (line 244) — CONTRIBUTOR-CONFLICTS.md is already populated: The doc said the file is 'present-with-schema-but- unpopulated'. Wrong: CC-001..CC-003 are filled in (no-name-attribution-rule scope, Stabilize-vs-keep-opening frames, absent-artifact-citation discipline). Updated text to reflect the populated state. P1 (line 270) — Conflict-row format mismatch: Proposed format used '### CONF-<date>-<nnn>' + bullet fields. Actual file uses CC-### IDs in a markdown table with columns (Conflict ID | Date | Question | Parties | Positions | Resolution | Scope | Source). Replaced the proposed format with a template that matches the existing table schema. CC-### counter continues from the highest existing ID (next auto-detected = CC-004). Generator preserves manually-curated rows and only appends auto- detected ones. * drain(#226 follow-up 2): fix Copilot P1/P2 — schema + idempotency P1 (L269) — clarify Open-table targeting: Replaced "appending machine-generated rows" (ambiguous given the file has separate Open / Resolved / Stale tables) with "inserting machine-generated rows into the Open table (or a delimited autogenerated subsection)". Without this, a literal append-to-EOF implementation wouldn't land in Open. P1 (L290) — specify idempotent generator strategy: Added explicit canonical-key → CC-NNN mapping. Generator updates in-place when key already matches; allocates new CC-NNN only when unmapped. Avoids unbounded growth / duplicate entries on repeated CI runs. Documents both strategies (in-place update + delimited autogenerated subsection). P2 (L275) — placeholder labels match column headers: Renamed schema placeholders from "Parties:" / "Resolution:" to "Between:" / "Resolution-so-far:" matching the actual headers in docs/CONTRIBUTOR-CONFLICTS.md (verified via grep). Reduces drift between the design doc and the live schema.
This was referenced Apr 25, 2026
AceHack
added a commit
that referenced
this pull request
Apr 25, 2026
…e) (#463) Otto-268 follow-on: drain-log for the 3-finding first cascade PR #433 (post-merge follow-up to #226 memory reconciliation algorithm design v0). Captures three orthogonal algorithm-correctness improvements. Per Otto-250 training-signal discipline. Pattern observations: 1. chain_head-liveness vs priority-tie-break-for-winner distinction: two invariants conflated at high level (Invariant 2 vs Invariant 6) distinguished. Spec-correctness findings benefit from per-invariant orthogonal-checks reasoning. 2. Paired-delimiter-vs-raw-character is a normalization-rule precision class. Same shape as #206's K-relations subset-vs-superset precision error. `_internal_var` should be preserved (single delimiters); `_text_` should be unwrapped (paired). 3. Index-rendering dedup is its own correctness class — multiple typed facts from same source_path → ONE index row. 4. Memory-reconciliation algorithm spec benefits from multiple cascade waves: #226 → #433 → #434 walked through schema + invariants + normalization + dedup + CC alignment + idempotent generator across three iterations. Each wave catches a different class of gap.
Open
4 tasks
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
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.
Summary
Amara's 4th-ferry centerpiece: replace hand-maintained
CURRENT-*.mddistillations with generated views over typed memory facts. ~380-line v0 design for downstream incremental implementation.What landed
docs/research/memory-reconciliation-algorithm-design-2026-04-24.md:MemoryFactschema (13 fields, 6 invariants, canonical-key normalization rules)CURRENT-*.md+MEMORY.mdgenerated outputComposes with Otto-73 retractability-by-design
MemoryFact.status(active / superseded / retracted) IS the retraction-native primitive at the memory substrate — same primitive as Zeta's ZSet algebra, different surface. Aaron's Otto-73 foundation and Otto-74 broader-human-trust addendum both flow through.Addresses known drift
Otto-70 snapshot tool surfaced MEMORY.md at 58,842 bytes vs. 24,976-byte cap. Generated index can be bounded by construction (Phase 4 cutover archives older entries).
Amara Determinize progress
What this PR is NOT
Test plan
🤖 Generated with Claude Code