feat(integrations): unify Tau memory schemas and setup - #1493

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Manual acceptance test plan

Tester: Paul. Status: not yet run. Automated results above do not substitute for this live-model/interactive pass. Record pass/fail and the returned note paths below. Use non-sensitive test content and a disposable local Basic Memory project. Model synthesis incurs normal provider charges. Do not use a team/production destination for this pass.

1. Launch safely from the branch

Use the existing checkout below, or substitute your checkout of feat/tau-shared-memory. Run these in a shell, not as a Tau prompt:

export BM_SRC=/Users/phernandez/dev/basicmachines/basic-memory-tau-1487
# Confirm feat/tau-shared-memory and the current PR head before testing.
git -C "$BM_SRC" branch --show-current
git -C "$BM_SRC" rev-parse HEAD
uv sync --project "$BM_SRC/integrations/tau"

export BM_TEST_ROOT="$(mktemp -d)"
export TAU_BASIC_MEMORY_CONFIG="$BM_TEST_ROOT/basic-memory.json"
cat > "$TAU_BASIC_MEMORY_CONFIG" <<'JSON'
{
  "project": null,
  "auto_recall": false,
  "capture_knowledge": false,
  "checkpoint_on_compact": false,
  "summarize_on_shutdown": false,
  "capture_transcript": false
}
JSON
chmod 600 "$TAU_BASIC_MEMORY_CONFIG"
mkdir "$BM_TEST_ROOT/repo"
git -C "$BM_TEST_ROOT/repo" init -b main
git -C "$BM_TEST_ROOT/repo" -c user.name='BM Test' \
  -c user.email='bm-test@example.invalid' -c commit.gpgsign=false \
  commit --allow-empty -m 'Initialize disposable memory test'
cd "$BM_TEST_ROOT/repo"
uv run --project "$BM_SRC/integrations/tau" tau -e "$BM_SRC/integrations/tau"

The launch uses the pinned organization fork while preserving the test repo as cwd. Do not launch from the BM source directory when testing this repo's profile. Keep the same shell exports for subsequent launches. The separate config does not isolate provider credentials or Tau's normal session storage; use test conversations only. Do not load a second installed copy of the BM extension.

• /bm-status reports connected, no write destination, and all automatic controls off. No automatic notes are created.
• Normal chat and explicit BM read tools work. A configured model/provider is required for chat; do not change credentials as part of this test.

2. Test the setup interview

The skill need not be installed globally. In Tau, ask it to read the absolute file path shown by BM_SRC above, ending in:

integrations/tau/skills/basic-memory-setup/SKILL.md

Suggested prompt:

Read that setup skill and walk me through configuring this disposable checkout. Use the existing TAU_BASIC_MEMORY_CONFIG override, not ~/.tau/basic-memory.json. Ask before choosing the project or writing schemas. I want a coding profile and full continuity, with transcripts off. This test repo has no remote: ask me to confirm the test identity rather than inventing one.

Choose an existing disposable local BM project, or explicitly approve creating one as a separate step. For the test-only repository label, confirm bm-test/continuity. Confirm the actual Git root and retain project: null at the general level so unregistered directories remain tools-only.

• The assistant asks for the destination, coding/general choice, repository identity, capture policy, optional read sources, and schema writes. It explains model costs and privacy.
• It edits only the test config. The repository entry contains the actual absolute test root, kind: coding, repository: bm-test/continuity, and the selected project. Other config values are preserved.
• It offers Coding Session, Task, and Decision schemas. Approved missing schemas are written once; rerunning setup does not overwrite existing/customized schemas.
• After /reload, /bm-status shows the selected project, coding profile, expected read sources, and full-continuity controls. No extra assistant turn is triggered just by reload.

3. Verify structured knowledge, not just a saved summary

Send a small test task:

Our test decision is to paint the astrolabe blue rather than red because blue matches the existing instrument labels. We have not painted it or run any tests. Next action: inspect the paint samples. Preserve that distinction in memory.

Wait for the turn to settle, then inspect the confirmed checkpoint using BM tools or your note viewer.

• /bm-status has a confirmed checkpoint path and no failure. The file exists in the selected project's checkpoint folder.
• Frontmatter has type: coding_session, the selected project, repository: bm-test/continuity, actual repo_root/cwd, branch, SHA, timestamps, tau_session_id, and capture: summarized.
• Git values match git rev-parse --show-toplevel, git branch --show-current, and git rev-parse HEAD in the test checkout. With no PR, it does not invent PR metadata.
• The note has categorized observations and a clear next step. It does not claim the paint work or tests were completed.
• Ask Tau to validate that exact note with the advertised bm_schema_validate tool. Expect no required-field warnings/errors after the shared schema is seeded. If the tool is unavailable, record that stage as unverified rather than enabling server features silently.
• Ask for a search using note_types=["coding_session"] and metadata_filters={"repository":"bm-test/continuity"} in the selected project. The checkpoint is found without a Tau-specific receipt lookup.
• Try /bm-remember Record the astrolabe color decision with its rationale in our decision folder. Verify a deliberate decision note with its observation/rationale and expected placement, rather than another checkpoint-shaped dump. Read its returned path; do not accept an acknowledgement as proof of a save.

4. Reload, checkpoints, compaction, and shutdown

• Immediately after a settled turn, run /bm-checkpoint twice without another ordinary chat turn between them. Expect the same confirmed source-tip checkpoint, not repeated copies. Inspect note counts outside the chat or through the note viewer so inspection does not itself create a new public turn.
• /reload, then /bm-orient astrolabe. The prior work is inserted as historical reference, with the correct write destination and placement guidance. Neither command should cause an unsolicited extra assistant response.
• In this disposable session, run /compact after several turns. If Tau reports no compaction needed, record that and retry after more context; a no-op does not prove checkpoint restoration. On actual compaction, verify a confirmed checkpoint reference is restored. Then ask what remains unfinished: it should recover the paint-sample step without claiming completion.
• For a shutdown-specific check, temporarily set capture_knowledge: false while leaving summarize_on_shutdown: true, reload, add one new test fact, then quit normally. Read the final checkpoint after exit and verify that outstanding fact was saved. Restore the intended controls before other tests. A process kill is not a shutdown test.
• Resume the test session and verify its prior checkpoint can be recovered. No ambiguous or failed save should ever be presented as confirmed.

5. Cross-worktree and repository isolation

Exit the test Tau session normally first. From the same shell:

git -C "$BM_TEST_ROOT/repo" worktree add -b second-checkout "$BM_TEST_ROOT/worktree"

With approval, add a second repository entry to the test config, copying the explicitly approved project/read-source/placement choices, using root $BM_TEST_ROOT/worktree and the same repository: bm-test/continuity. Keep the first entry intact. Start a new Tau session from the second worktree using the absolute launch command in section 1.

• Startup recall or /bm-orient astrolabe finds the first checkout's coding checkpoint despite a different cwd and no inherited Tau receipts.
• In another disposable Git repo, configure a different identity, e.g. bm-test/unrelated, using the same BM project. Its coding-session recall must not present the astrolabe checkpoint as this repository's work. Project-wide tasks/decisions can still appear as shared project knowledge; they are not repository-isolated by this feature.
• Optional safety case: make an unregistered nested Git repo beneath the first root and launch from it. Expect a visible profile/root mismatch failure and no parent-repository checkpoint or transcript writes, not a silent fallback.

6. Optional read-only source and general-profile checks

• Choose a second disposable project containing a known open decision or active task. Explicitly approve it as a read_projects source and run /bm-orient. Its note content appears with read-only source labeling. Automatic checkpoints still go only to the primary project. Do not ask the agent to write to the secondary project during this check: explicit tool calls are not a read-only security boundary.
• For general use, work outside every configured coding root and explicitly choose a disposable general project in the test config. Create a short test conversation. Expect a session checkpoint, not coding_session, and no invented Git metadata. Seed/validate against the Session schema if testing this stage.

7. Finish and report

Quit test sessions normally before changing destinations or disabling capture: the old lifecycle can still write during shutdown/reload using its previous settings. Do not delete a worktree while its Tau session is running.

• Report the tested head, completed sections, any failures, relevant note paths, and whether live synthesis quality was acceptable. Redact secrets and private project content from screenshots/logs.
• Retain test artifacts until results are understood. Remove test notes, schema seeds, worktrees, or config only deliberately; no automatic cleanup of checkpoints/receipts. unset TAU_BASIC_MEMORY_CONFIG in the test shell restores the ordinary config choice for future launches; it does not alter an already-running Tau session.

Manual result: pending tester execution.