Close memory durability and retrieval audit by RMANOV · Pull Request #14 · RMANOV/sqlite-memory-mcp

RMANOV · 2026-08-02T22:03:41Z

Summary

close audit 924b7796a000 findings in retrieval/search fallbacks, service liveness, and debate-pump durability
enforce a directory-wide dead-process log budget
make backup publication collision-safe and WAL recovery verifiable, including concurrent same-process staging
add isolated regression coverage for every repaired seam
carry the two preceding schema reconciliation commits already used by the live hardening branch

Local verification

python3 -m pytest -q: 1538 passed, 19 skipped
ruff check .: passed
manifest tests: 22 passed
registry smoke: passed (9 tools)

Draft only. No merge or deployment is authorized by this PR.

chatgpt-codex-connector · 2026-08-02T23:17:17Z

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

Copilot

Pull request overview

This PR closes a durability/retrieval audit by tightening “live vs finished” semantics in task querying/lookup, hardening backup/logging durability behaviors, and adding broad regression coverage to lock the fixes in across the MCP servers.

Changes:

Make query_tasks live-by-default (unless an explicit status is requested or include_completed=True) and ensure find_by_title keeps finished rows reachable while consistently ranking live work first.
Improve search fallback behavior/performance (FTS overfetch, literal top-up gates) and refine ready-context classification with word-boundary + inflection-aware marker matching.
Add extensive regression tests for liveness, search fallbacks, backup collision/WAL verification, directory-wide log budgets, schema migrations, and test-suite filesystem isolation.

Reviewed changes

Copilot reviewed 30 out of 30 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
unified_server.py	Updates tool instructions to document the new liveness defaults and lookup ranking behavior.
tray_dialogs.py	Passes a live DB connection into tray search to enable FTS and avoid expensive in-memory fallback work.
tests/test_task_server_liveness.py	New regression suite for `query_tasks` liveness defaults and `find_by_title` liveness-first ordering.
tests/test_server_imports.py	Fixes logger fallback test to patch the handler type actually used (RotatingFileHandler).
tests/test_schema_task_field_versions.py	Adds integrity tests for orphan pruning + FK cascade behavior around task field versions.
tests/test_schema_role_binding_migration.py	Adds regression coverage for deterministic retirement of duplicate active bindings before creating unique index.
tests/test_recover_literal_matches_thin.py	Adds differential tests ensuring recovery passes don’t suppress each other for thin answers.
tests/test_ready_context_word_boundary.py	Adds large regression suite for ready-context marker matching, intent scoping, and ordering invariants.
tests/test_query_tasks_search_shape.py	Ensures `query_tasks(search=...)` preserves recall and row shape in core-only installs.
tests/test_log_directory_budget.py	Adds tests for directory-level log budget enforcement without harming live processes.
tests/test_fts_token_selection.py	Verifies longest-first token selection so FTS caps don’t drop non-Latin content (Cyrillic).
tests/test_find_by_title_reachability.py	Ensures finished rows remain reachable on every `find_by_title` path and rank below live work.
tests/test_debate_pump_worker_census.py	Adds coverage for correct worker census identity (topic+role+session) and downstream projections.
tests/test_db_utils_durability.py	Adds durability + logging + DF-stopword tests for backups, bounded logs, and tokenization behavior.
tests/test_backup_generation_collision.py	Adds tests for collision-safe backup publishing and WAL recovery retry semantics.
tests/fixtures/retrieval_eval_corpus.json	Updates expected retrieval evaluation fixture outputs.
tests/conftest.py	Introduces session-wide HOME/filesystem isolation to prevent tests touching operator’s live state.
task_tray.py	Clarifies bounded index admission policy and intent (fresh rows stay reachable via on-disk/substring passes).
task_server.py	Implements `include_completed`, liveness filtering, and liveness-first ordering for lookup results.
task_search.py	Adds FTS overfetch + literal recovery/top-up behavior and bounded-index recency admission.
smart_retrieval.py	Reworks marker matching with compiled regex + inflections and reorders ready-state precedence rules.
schema.py	Adds migrations for orphan pruning, role/session binding reconciliation, FK-baseline handling, and log noise reduction.
pyproject.toml	Bumps project version to 3.13.2.
memory_thread_clustering.py	Switches FTS token cap selection to longest-first for non-Latin fairness.
link_suggestions.py	Switches safe FTS query token selection to longest-first for non-Latin fairness.
hooks/debate_pump.py	Fixes worker census identity to avoid undercounting across topics/roles; projects to bare ids where needed.
entity_server.py	Switches token cap selection for overlap computation to longest-first for non-Latin fairness.
db_utils.py	Adds crash-consistent backups, bounded rotating logs with directory budget sweeping, and DF-based stopwords.
CHANGELOG.md	Documents fixes for v3.13.1–v3.13.2.
init.py	Bumps `__version__` to 3.13.2.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+    merged: list[dict] = []
+    seen: set[str | None] = set()
+    for row in (*projected, *_scored_fallback(tasks, query, limit)):
+        task_id = row.get("id")
+        if task_id in seen:
+            continue
+        seen.add(task_id)
+        merged.append(row)
+        if len(merged) >= limit:
+            break
+    return merged


+def _live_worker_session_ids(topic_id: str) -> set[tuple[str, str, str]]:
+    """Resolve live derived workers from their durable real-spawn receipts.
+
+    Returns ``(topic_id, role, worker_session_id)`` claim keys — the identity
+    ``debate_worker_claims`` is actually unique on. Callers that need bare
+    session ids for a single topic project the third element."""
    con = sqlite3.connect(DB_PATH)


Both defects were found by Copilot's review of #14, after that PR had already merged to main and been deployed here. 1. `_live_worker_session_ids` silently changed shape. The census fix needed the full claim key `(topic_id, role, worker_session_id)` and the helper's return type was changed in place to supply it — but callers treat the result as a set of bare ids: * `tests/test_debate_windows_adapter.py` asserts `_live_worker_session_ids("T1") == {"cc-x-W1"}` * `recover_stale_worker_claims()` scans one topic and matches ids The Windows test is `@windows_only`, so it skips on Linux and the suite stayed green while the contract was broken — one of the 19 skips was covering the regression. Split instead of overload: `_live_worker_session_ids()` returns `set[str]` as it always did, and `_live_worker_claim_keys()` returns the triple for the cross-topic census. Within one topic the id alone is already unique, so the projection is exact. `test_debate_zero_paste_recovery.py` patched the bare-id helper and passed for the wrong reason: the census unions whatever the stub returns, so a set holding one string counted as one worker even though the census's identity is the triple. It now patches the claim-key helper, which is what the census actually calls. 2. `_project_and_top_up` evaluated `_scored_fallback()` eagerly. Spelling the merge as `(*projected, *_scored_fallback(...))` computes the fallback before the loop runs, so every query paid an O(n_tasks) lower-casing pass over the whole pool even when the ranked hits already filled the quota and the first `break` threw the result away. Costliest on core-only installs, where this is the only search path. Measured: quota filled by FTS -> fallback called 0 times (was 1); quota short -> called once; output byte-identical to the eager version in both cases. Full suite 1538 passed, 19 skipped. Ruff clean. Operator log inodes unchanged across the run. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_014Po7GZbvPSVjr16bSBBmhA

RMANOV added 4 commits August 2, 2026 13:33

fix(schema): reconcile legacy role binding duplicates

58e435c

fix(schema): prune orphan task field versions

f870789

fix: close memory durability audit

28f1cf9

fix: preserve core-only search recall

bdc3fbf

RMANOV marked this pull request as ready for review August 2, 2026 23:17

Copilot AI review requested due to automatic review settings August 2, 2026 23:17

Copilot started reviewing on behalf of RMANOV August 2, 2026 23:17 View session

RMANOV merged commit 74ded79 into main Aug 2, 2026
3 checks passed

Copilot AI reviewed Aug 2, 2026

View reviewed changes

Navigation Menu

Close memory durability and retrieval audit - #14

Close memory durability and retrieval audit#14
RMANOV merged 4 commits into
mainfrom
agent/memory-firstwave

RMANOV commented Aug 2, 2026

Uh oh!

chatgpt-codex-connector Bot commented Aug 2, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

RMANOV commented Aug 2, 2026

Summary

Local verification

Uh oh!

chatgpt-codex-connector Bot commented Aug 2, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants