[XY-code1]

Summary

• add one shared, read-only PiAdapter for Pi and Oh My Pi v3 JSONL sessions
• follow the recorded tool homes for Pi and OMP, read cwd from the session header, and keep only the active parent-linked branch
• normalize text, thinking, images, tool calls/results, model changes, OMP's leading title slot, and later title changes
• enable Pi/OMP session scoring in TOOLS, every README matrix, and the generated support docs

Refs #1418 (Pi / Oh My Pi row only).

Verification

• npm run lint
• npx tsc --noEmit
• npm run build
• npx vitest run src/__tests__/pi-session-adapter.test.ts src/__tests__/session-platform-facts.test.ts src/__tests__/readme-matrix.test.ts src/__tests__/tool-support.test.ts src/__tests__/agent-session-list.test.ts --maxWorkers=2
  • 59 passed, 1 skipped
• npx vitest run tool-support -u

Representative built-CLI run against a Pi v3 fixture store:

$ node dist/index.js session platforms
pi: ✓ installed

$ node dist/index.js session migrate 123e4567-e89b-12d3-a456-426614174099 --source pi --target claude-code --cwd <repo> --yes --json
Session:   Verify Pi CLI migration (123e4567...)
Messages:  2
Fidelity:  100.0% (Mode A)
✓ Migration successful
CLI_EXIT=0

I did not label this as a real authenticated Pi/OMP session run: the local machine has no configured Pi/OMP provider session. The fixture uses the documented v3 record graph and the adapter tests cover both variants, including OMP's title slot.

Shared-state review

• Native writers: Pi and OMP continue to own their JSONL stores.
• New reader: PiAdapter recursively reads .jsonl files only.
• TeamAI writers/deleters: none; writeSession throws ToolUnavailableError and deleteSession returns false.
• Registry/store resolution: sessionStoreDir remains the single resolver, so recorded member tool roots and OMP profiles keep their existing behavior.

Real-CLI record on real Pi and Oh My Pi sessions (added by a reviewer)

Real-CLI record (teamai-cli at b1664c0c, npm ci && npm run build; baseline 3da56ad2 built the same way)

Stores: Pi 1.4.0 (@earendil-works/pi-coding-agent@1.4.0) and Oh My Pi 18.12.6
(@oh-my-pi/pi-coding-agent@18.12.6, run under Bun 1.4.0). Each wrote its own session
file through a local OpenAI-compatible stub that returns a fixed reply (no provider
account). The judge was a stub claude that prints a fixed label, so the score lines
prove the path, not a model judgment.

Isolation: env -i; HOME=<fixture home>; PATH=<node dir>:/usr/bin:/bin;
GIT_CONFIG_GLOBAL=<empty file>; GIT_CONFIG_NOSYSTEM=1; TEAMAI_NONINTERACTIVE=1;
TMPDIR=<tmp>.

$ node dist/index.js session platforms                       # head
  pi: ✓ installed                                             (exit 0)
$ node dist/index.js session push --source pi --all --scrub -y --cwd <proj> --repo-root <team clone>
  ✓ Pushed 3 session(s) from pi                               (exit 0)
$ node dist/index.js score --scorer <file> --session <pi id> --source pi
  fixture-quality: good (score 1, passed)                     (exit 0)
$ node dist/index.js session push --source omp --all --scrub -y --cwd <proj> --repo-root <team clone>
  ✓ Pushed 1 session(s) from omp                              (exit 0)
  archived title: "Fix the OMP parser" (OMP title slot empty, so the first prompt)
$ node dist/index.js score --session <omp id> --source omp
  fixture-quality: good (score 1, passed)                     (exit 0)
$ node dist/index.js session migrate <pi id> -s pi -t omp --yes
  Error: TeamAI cannot write omp sessions: omp resumes a conversation only from a store of its own. Pick another target platform.
                                                              (exit 1; no file in the Pi home changed)

Baseline 3da56ad2: each command above fails with Unknown platform "pi" / "omp" (exit 1).

[github-actions]

• [P2 non-blocking] src/session-flow/adapters/pi.ts:232: Later model changes are ignored once any assistant message has a model. For an assistant message using model A followed by a parent-linked model_change selecting model B, readSession() still returns A in session metadata because the message lookup wins before model-change entries are considered. Resolve both record types in branch order so the latest model selection wins.

The PR description documents sufficient testing: the representative built-CLI fixture migration satisfies the real-CLI requirement. An authenticated Pi/OMP run or full provider × agent matrix is not required. No commands from the PR were run.

[XY-code1]

Fixed the Windows matrix failure in b1664c0: the sessions locate matrix now covers Pi/OMP default, moved-home, profile, project, and PI_CODING_AGENT_SESSION_DIR stores; sessionStoreDir also honors the explicit Pi-family session-dir override. Local Windows-safe suite: 33 files, 4,876 tests passed. Lint, tsc, and focused session tests pass.

[github-actions]

• [P2 non-blocking] src/session-flow/adapters/pi.ts:232: The earlier finding remains unresolved. An assistant message using model A followed by a parent-linked model_change selecting model B still yields session metadata with A, because assistant-message models take precedence over all model-change records. Resolve both record types in branch order so the latest selection wins.

The PR description documents sufficient testing: the representative built-CLI fixture migration satisfies the real-CLI requirement. An authenticated Pi/OMP run or full provider × agent matrix is not required.

No additional confident findings. Only the specified diff was read; no PR code was run, built, or installed.

[SaulMoro]

This was generated by AI during triage.

Reviewed at b1664c0 against #1418's Pi and Oh My Pi rows. The adapter works on real stores: I wrote sessions with Pi 1.4.0 and Oh My Pi 18.12.6 (through a local stub provider) and ran the built CLI. session push and score --session work for both, OMP's empty title slot falls back to the first prompt, and session migrate --target omp fails with the read-only message and writes nothing. main fails each of these with Unknown platform. I added that record to the PR body, since the issue asks for push and score on a real session.

Two things to fix before merge:

• [P1 blocking] CI is red at this head. Lint & Test on Node 20, 22 and 24 (ubuntu) and E2E shard 3/3 fail in three cases that use pi as the example of a tool whose sessions teamai can't read. This PR makes that false. Switching them to copilot (still unreadable on main) fixes it; grep -rn "'pi'" src/__tests__ finds any others.
• [P1 blocking] The Pi sessionDir setting is in the TOOLS note but is not read. With {"sessionDir": "<dir>"} in ~/.pi/agent/settings.json, Pi 1.4.0 writes there, and teamai session push --source pi --all archives only the default store, exits 0 and says nothing about the missing sessions. Either read the setting in sessionStoreDir (src/locate.ts:1177), resolved like the env var, or drop it from the note.

Non-blocking:

• [P2 non-blocking] An unexpanded ~ in PI_CODING_AGENT_SESSION_DIR (from an env file or a quoted value) is resolved against the working directory; Pi expands it. expandHomeOn (src/tool-home.ts:289) handles it.
• [P2 non-blocking] The CI review's model_change finding: agreed, reproduced on a real Pi file. Only the session-level metadata.model is affected.
• [P2 non-blocking] docs/guide/dashboard.md:145 and its zh-CN page list the platforms teamai session covers and don't include pi and omp.
• [P3 nit] A dangling activeLeafId returns no messages, where OMP falls back to the newest entry; the comment at platform-facts.ts:209 still names only PI_CODING_AGENT_DIR; listConversations and readSession in pi.ts repeat the same normalization.

@jeff-r2026 Your call on this one once the two blocking items are fixed: it adds Pi and Oh My Pi as readable session platforms.

[XY-code1]

Addressed the blocking and follow-up findings in f74bfb9:\n\n- replaced the three stale tests that used Pi as an unreadable-session example with Copilot\n- made sessionStoreDir honor Pi's settings.json sessionDir, after the env override and with ~/relative expansion via �bsoluteOn\n- fixed PI_CODING_AGENT_SESSION_DIR tilde expansion\n- resolve assistant models and model_change records in active-branch order, so the latest event wins\n- fall back to the newest entry for a dangling �ctiveLeafId\n- added Pi/OMP to both dashboard docs\n- added regression tests for settings resolution, latest model changes, and dangling leaves\n\nLocal verification: lint, tsc, build, Windows-safe suite (33 files / 4,894 tests), focused adapter/store tests, and the dashboard E2E all pass.

[github-actions]

No new confident findings in the specified diff.

• Resolved: The earlier model-ordering finding is fixed in src/session-flow/adapters/pi.ts:83. Assistant messages and model-change records now update metadata in active-branch order, with a regression test covering a later model change.
• Testing: The PR description documents sufficient real-CLI verification. The representative built-CLI fixture migration satisfies the requirement; the additional Pi/OMP records provide further coverage. No full provider × agent matrix is required.

Only the specified diff was read. No PR code was run, built, or installed.

[SaulMoro]

This was generated by AI during triage.

Re-checked at f74bfb9 with real Pi 1.4.0 and Oh My Pi 18.12.6 sessions. Resolved:

• The three tests now use Copilot as the unreadable example; no CI job fails at this head (some macOS and one E2E shard were still running).
• sessionDir in ~/.pi/agent/settings.json is read: a session Pi writes there is archived by session push --source pi --all, ~ and relative paths resolve as in Pi, and PI_CODING_AGENT_SESSION_DIR still wins.
• An unexpanded ~ in PI_CODING_AGENT_SESSION_DIR works, the latest model_change wins, a dangling activeLeafId falls back to the newest entry, and the dashboard pages list pi and omp.
• push and score --session still work on a real Pi and a real OMP session.

One left:

• [P1 blocking] Pi's project-level sessionDir is not read. Pi also reads sessionDir from <project>/.pi/settings.json, over the global file. With one set, Pi writes the project's sessions there, and teamai session push --source pi --all exits 0 and archives an older session from the default store instead. Reading the project file first fixes it; narrowing the TOOLS note to the global setting would make this a known limit rather than a bug.

Nits:

• [P3 nit] A last entry with activeLeafId: null now shows the newest branch; OMP treats a null leaf as no leaf, and the previous head showed none.
• [P3 nit] The win32 assertion change in the dashboard e2e test is unrelated to Pi, and dashboard-session-sharing-view.test.ts still has unreadTools: ['pi'].

@jeff-r2026 your call once the project-level sessionDir is handled.