Record Tutti Agent Session Replay

Produce evidence in this order:

case -> deterministic scenario -> real record -> cassette audit -> real replay -> case artifact update

Do not report a cassette as recorded before both the recording and a fresh Replay succeed.

Prepare

1. Read the root and closest AGENTS.md.
2. Read docs/architecture/agent-session-replay.md.
3. For AgentGUI interactions, read docs/architecture/agent-gui-node.md and packages/agent/gui/AGENTS.md.
4. For lifecycle semantics, read packages/agent/host/README.md and keep those semantics in Host.
5. Inspect git status --short. Preserve pre-existing work.
6. Read the case repository's case.json, scenario.mjs, fixtures, current Cassette, and evidence. The repository is the case source of truth.

Use CDP through the repository runner. Do not use Computer Use unless the user explicitly requests it.

Design the scenario

Add or update cases/<scenario-id>/scenario.mjs in the QA case repository. Do not add a case registry or case-specific scenario to the Tutti repository.

Make the scenario deterministic:

• Use one exact prompt with stable markers.
• Use an exact final token when the case expects an assistant reply.
• Use a harmless, reversible command for approval cases.
• Set every behavior-affecting composer default explicitly: permission mode, plan mode, model, and reasoning effort as applicable.
• Match the real accessible label, test ID, or semantic DOM state. Do not depend on layout coordinates.
• Wait for the interaction to exist before acting.
• Submit each approval, answer, or plan decision once.
• Require the expected terminal state and absence of enabled stale controls.
• Cassette directory name is {scenarioId}_{provider} (for example c01_codex), derived by defineRecordScenario.

For question cards, trigger the Provider's real user-input request. For Codex plan cases, enable Plan Mode, wait for the completed plan and implementation card, then drive the real plan decision or feedback action.

Record

Run:

pnpm e2e:agent-gui -- \
  --record .tmp/cassettes/<scenario-id>_codex \
  --scenario <scenario-id> \
  --scenario-file ../tutti-agent-session-replay-cases/cases/<scenario-id>/scenario.mjs \
  --agent-target-id local:codex \
  --keep-runtime \
  --timeout-ms 300000

Omit --headless to show the Electron window while debugging; add it for unattended runs. Keep --keep-runtime while developing. Inspect:

• artifacts/record-agent-gui.png
• logs/desktop.log
• state/logs/tuttid.log
• state/tuttid.db
• the incomplete cassette and Provider frames

Fix root causes. Do not weaken transport matching, semantic validation, or terminal-state assertions to make a cassette pass.

Audit

Run the bundled deterministic audit:

node .codex/skills/tutti-record-agent-session-replay/scripts/audit-cassette.mjs \
  .tmp/cassettes/<scenario-id>_codex

Require:

• cassette inventory and hashes verify;
• Provider manifest is complete;
• global and per-connection frame sequences are continuous;
• the expected interaction or plan decision appears once;
• tool execution count and exit status match the case;
• all expected Turns are terminal;
• activity sequences are contiguous and intent→effect causality satisfies packages/agent/session-replay/activity-contract.json: every effect references an earlier intent with a compatible correlationId and a declared effect type, every requiresEffect intent is answered, and direct-stimulus events carry no cause;
• the final assistant token or expected no-text state matches the case.

Inspect a small decoded Provider-frame window around the interaction. Do not dump the entire Provider stream or expose account data.

Replay

Run a fresh isolated Replay:

pnpm e2e:agent-gui -- \
  --replay .tmp/cassettes/<scenario-id>_codex \
  --keep-runtime \
  --timeout-ms 300000

Omit --headless to show the Electron window; add it for unattended runs. Require the runner's replay passed result. Inspect artifacts/replay-agent-gui.png and confirm the case's UI terminal state.

Provider transport remains fail-closed. Only repository-declared observer-only probes may be absent or yield to causal traffic.

Update the case artifacts

Only after record and Replay pass, replace the case repository's cases/<scenario-id>/cassette and evidence directories and update its case.json status when required. Do not read or write a remote catalog.

Finish

Run the repository validation selection policy from docs/conventions/testing.md.

Report:

• cassette names and artifact links;
• record and Replay evidence per case;
• Provider frame, interaction, tool, Turn, and final-response summaries;
• case repository artifacts changed;
• implementation changes and why they changed;
• changed-line distribution, separating pre-existing dirty work;
• documentation impact;
• remaining failed gates or unimplemented scope.