Date: 2026-09-06. Decision: prepare an improved release of the existing tools. Do not merge the evidence-context research branch as the MVP.
The user needs a usable release and a finite delivery boundary after an expensive research/implementation run. The early capture/repair/sync fixes are separable: three existing commits apply cleanly to current main. They address missing or damaged conversation imports and make configured peer reconciliation usable without introducing a new context schema or distributed permission protocol.
The research branch contains useful source-grounded retrieval, interpretations, scope controls and lifecycle work, but also unfinished recovery and integration paths. Selecting an arbitrary later stage would inherit those dependencies. Disabling their safeguards to make a release would not resolve that problem.
This is deliberately a reliability release, not completion of the richer shared-memory product. Its value is recovering and retrieving conversation history with the existing StudyLoop integration. The rich-context work and its independently runnable lessons remain preserved on their separate branch.
- Branch:
codex/session-memory-mvp. - Base: current main at
f945d057. - Existing changes:
90869332,ebd71d49,66907922, cherry-picked asb08d4f06,a5304d1c,9e979e53. - Runtime/test/documentation change from those commits: 31 files, 3,568 additions and 199 deletions. This includes substantial tests and documentation; it is not a count of production code alone.
- No new runtime implementation in this audit. The additional verification script exercises the installed wheel; this file defines the release boundary. One existing five-line test-isolation fix was also copied from the research branch after the full regression exposed ambient credential fallback in secret tests.
- Existing schema version remains 30. No new context migrations are included.
- The main checkout's uncommitted work was not copied, changed or discarded.
- The research branch and its unfinished reconciliation changes remain intact.
| Area | Included in this candidate | Explicit limit |
|---|---|---|
| Capture | Existing parser fixes for Codex, Claude Code and Kiro; Grok Build exporter; stable conversation updates | Supported local transcript formats, not every desktop/cloud conversation or all native tool events |
| Repair | Inspect by default, explicit apply, consistent backup, transactional merge, retained history, repeatable reconciliation | Cannot recover missing original transcripts; no universal completeness guarantee |
| Retrieval | Existing session search/inspection and configured project aliases | Conversation retrieval, not proof that an agent's advice or validation claim is correct |
| Sync | Configured session-sync all, pushes before pulls, per-peer failure reporting, preserved message content and archived learner variants |
Existing trusted whole-database transport; conflicting databases are not promised to become identical |
| StudyLoop | Current main's installer, skill and doctor integration retained | No new standalone setup owner or comprehensive runtime capture-health system |
| Scope | Existing project filters | Filters are not enforced work/personal authorization; no per-peer scope protection |
| Forgetting | No new distributed forgetting claim | Deleted data may return through legacy sync/restore; no tombstone guarantee |
| Decisions | Original conversations can inform the agent | No automatic arbitration, source-bound decision acceptance or semantic truth certification |
The mentor-harness contract is six harnesses: Kiro CLI, Codex and Claude Code as core; OpenCode, pi and Grok Build as preview. Grok Build was restored to that contract after this candidate was cut, so the Capture row above describes its exporter only — the launch, install and doctor paths land separately.
Only use the legacy sync path where every participant is permitted to receive the entire selected database. This candidate is unsuitable for syncing a mixed work/personal store to a destination allowed to receive only one part. Preserving the current transport does not satisfy the richer product's scope requirements.
The exporter and the database must agree on the schema; studyloop doctor
reports exporter_schema. It is not a downgrade tool and does not preserve the modern
scope/retirement contract. Rehearse upgrades on a copy of an existing supported
database. No owner database or real peer was modified by this audit.
- Locked workspace dependency setup succeeded.
- Memory package and selected StudyLoop installer/harness/doctor tests:
1,182 passed, one failure caused by setting
CLAUDE_CONFIG_DIRwhile a default constructor test expects that variable to be absent. That exact test passed separately with the override unset. No runtime or test change was needed. - Workspace source and test type checking: zero errors/warnings.
- Memory package lint and format checks passed.
- Memory wheel built and installed in a fresh environment with StudyLoop absent.
- Installed-wheel verification: seven checks passed, covering native capture, continued conversation/idempotence, non-mutating repair preview, repair with pre-change backup, repeat repair, CLI retrieval, and database/source integrity.
- Full StudyLoop regression: 3,760 passed, four skipped, 704 deselected, two failures in secret-store tests due to ambient credential fallback. The existing fixture fix clears provider fallback variables at each secret test; all 41 secret-store tests then passed. The full suite was not repeated after this test-only change. Runtime code did not change.
- Memory package SAST passed with the repository's configured exclusions.
- All pre-commit checks on the three imported commits passed, including secrets, SAST, formatting, and workspace source/test type checking.
Operational notice: the secret-test failure printed an ambient API key in tool output. Rotate that credential. Its value is not included in this document or the verification script. Startup-only environment removal did not prevent its later presence during the suite; test-local isolation now covers the observed failures.
The installed check uses fictional native-format records and real installed package entry points. It does not launch a coding harness or exercise SSH. The audit made no model-provider or council calls.
Reproduce the small installed check after building/installing the memory wheel:
/path/to/fresh-memory-only-venv/bin/python scripts/verify-session-memory-mvp.py- Complete the default regression suite and normal release/CI checks on this candidate. Record failures explicitly; don't replace failing checks with a growing feature project.
- Exercise the existing installed StudyLoop/harness workflow and actual two-endpoint sync against disposable, same-trust databases. Include a continued conversation, repeat sync, a conflicting message and an unavailable peer. Existing local/helper tests do not replace this transport acceptance.
- Rehearse upgrade/repair on a disposable copy of a supported pre-research database, confirm retained history and backup recovery, then review the exact changes and release notes. Reconcile overlapping main-checkout work before merge. Use the normal version/release process only after acceptance.
These are acceptance gates for the frozen candidate, not permission to implement the deferred product. The candidate is not published and is not yet declared production-ready. If a gate reveals an architectural change rather than a bounded fix, stop and report the failed gate with a smaller alternative before spending more on implementation. No unbounded goal run, additional research stage, council loop or database redesign is authorized by this checklist.
The broader SessionWeave product remains a separate decision: explicit boundaries, code-owned provenance, rich source explanations, reviewed conflicts, forgetting across sync and restore, and shared installer/doctor ownership. Its existing implementation and lessons are retained for later selection. This reliability release neither discards them nor claims to deliver them.
See repair instructions and sync instructions. The user-facing conversation memory guide, setup, CLI reference, agent installation guide, shared session-memory skill and Unreleased changelog now describe this same candidate boundary. These source documentation updates do not publish a release or update installed user skills.