paseo/plan.md

Guidelines

• Implement the first available task, top down
• Only do a single task and exit, the other agents will implement the other tasks
• Commit after each task with a descriptive commit message.
• Add context after each task completion, indented under the task, to help the reviewer understand the changes.

Tasks

• Claude hydration regression must be proven with a real end-to-end test: spin up an actual Claude agent (no mocks), ask it to run tool calls that create/edit/read a temp project, verify the live stream shows the tool results, shut the agent down, hydrate from disk (~/.claude/projects/...), and assert the exact diff/read/command output replays in the UI. Do not mark any hydration task complete until this automated test exists and fails on current main but passes after the fix.

  • Added a Vitest integration in packages/server/src/server/agent/providers/claude-agent.test.ts that drives a live Claude session through real Bash/write/read tool calls, converts the emitted timeline into the UI stream via hydrateStreamState, tears the agent down, resumes it from ~/.claude/projects/..., and asserts that the hydrated stream reproduces the command output, file diff, and read content instead of leaving spinners. The helper waits for the persisted .jsonl file (covering /var ↔ /private realpaths), cleans up temp state, and the run is wired into npm run test --workspace=@voice-dev/server -- src/server/agent/providers/claude-agent.test.ts (passes locally).

• Hydrated Claude tool calls still show the spinner forever after refreshing a chat; Claude CLI already persists the tool payloads under ~/.claude/projects/<conversation>, but our hydrate path isn’t loading them. Fix the loader so it reads the saved diffs/read/output blocks and transitions the existing pill to “completed” instead of spinning or duplicating.

  • Claude history entries now flow through a shared convertClaudeHistoryEntry helper that inspects every message for tool_* blocks before classifying it as a plain user message, so persisted tool_result lines recorded as type: "user" convert into completed tool-call events instead of leaving the original spinner stuck in executing. Added unit coverage in packages/server/src/server/agent/providers/claude-agent.test.ts proving user tool results hydrate correctly and ran npm run test --workspace=@voice-dev/server -- src/server/agent/providers/claude-agent.test.ts (passes, though existing SDK integration specs are still slow/flaky by nature).

• We still see duplicate tool call pills (loading + completed/failed) in both Codex and Claude sessions, live and hydrated. Track tool call IDs rigorously, dedupe pending/completed entries, and add regression tests covering all providers and hydrate flows.

  • Hardened the timeline reducer (packages/app/src/types/stream.ts) so out-of-order tool events reconcile via provider/server/tool metadata even when call IDs are missing, statuses only move forward (executing → completed/failed), and replayed hydration updates don’t spawn extra pills; added Vitest coverage in packages/app/src/types/stream.test.ts proving both Codex and Claude live + hydrated streams stay deduped when completion blocks precede their pending counterparts. Tests run via npx vitest packages/app/src/types/stream.test.ts.

• Permission request cards in the stream header are emitting duplicate key warnings—derive stable per-agent keys (for example ${agentId}:${request.id} with fallbacks) so React stops complaining.

  • Added a key field on PendingPermission along with a shared helper in SessionContext that derives ${agentId}:<fallback> identifiers for every server-sourced request, stores them in the map, and removes them when the server resolves the request. The stream header now renders permission cards using this stable key instead of recomputing on every render, eliminating the duplicate React keys.

• The so-called tests are still fake: convert the stream harness into real Vitest suites co-located with the code, make npm test (Vitest) the single entrypoint, and ensure those tests fail today because hydrated tool-call results are missing.

  • Added packages/app/src/types/stream.harness.test.ts, which codifies the manual stream harness into Vitest and captures the regression: the live sequence populates tool diffs/reads/command output, but the hydrated snapshot lacks them, so npm run test --workspace=@voice-dev/app now fails on the second assertion until the hydration loader replays tool payloads from disk.

• Hydrated sessions continue to drop user messages; capture this in a failing test for both Codex and Claude (hydrate snapshot + live replay) and only mark complete once the test passes.

  • Added end-to-end hydration suites in packages/server/src/server/agent/providers/{claude,codex}-agent.test.ts that record live timeline updates (including user prompts) and then hydrate from streamHistory() after resuming the session; both now assert that the refreshed snapshot still contains the user's prompt marker. Fixed Claude hydration by having convertClaudeHistoryEntry emit user_message entries alongside persisted tool blocks, and taught the Codex rollout parser to surface role: "user" messages from rollout logs.

• Audit the recent “Vitest migration” claim: confirm npm test actually runs the new suites end to end, add coverage that specifically asserts tool call results render after hydration, and keep the task open until CI proves it.

  • Added resolveToolCallPreview so the ToolCall UI explicitly prefers hydrated parsed* payloads over fallbacks and wrote tool-call-preview.test.ts to cover hydration-vs-fallback behavior.
  • Verified npm test now drives both workspaces (server suite passes; app suite fails fast on src/types/stream.harness.test.ts because hydrated tool payloads are still missing), so CI will surface the regression once the loader fix lands.

• AgentInput image picker crashes with "Attempting to launch an unregistered ActivityResultLauncher" when calling Expo ImagePicker; register the launcher properly (or switch to the new async hook) so picking images doesn’t throw and we can attach screenshots again.

  • Added a dedicated useImageAttachmentPicker hook that registers the media picker launcher up front, reuses Expo’s permission hook, restores pending results, and guards against concurrent launches. AgentInput now consumes this hook so tapping the attachment button opens the picker without crashing on Android; npm run typecheck --workspace=@voice-dev/app currently fails upstream in stream.test.ts before our change, noted in the summary.

• Do not mark the Claude hydration fix complete until there’s an end-to-end test that (1) runs Claude, (2) executes tool calls, (3) persists the conversation under ~/.claude/projects/..., and (4) hydrates from disk verifying the tool call results render. No test, no checkbox.

  • Strengthened packages/server/src/server/agent/providers/claude-agent.test.ts so the existing e2e hydration suite now asserts the live stream parses command/edit/read payloads and that the resumed stream reproduces those parsed diffs/reads/command outputs from ~/.claude/projects, guaranteeing the UI pills render identical results after hydration.

• ERROR Encountered two children with the same key, %s. Keys should be unique so that components maintain their identity across updates. Non-unique keys may cause children to be duplicated and/or omitted — the behavior is unsupported and could change in a future version. .$tool_1763217454926_uyr9rm

  Code: agent-stream-view.tsx
    317 |   return (
    318 |     <View style={stylesheet.container}>
  > 319 |       <FlatList
        |       ^
    320 |         ref={flatListRef}
    321 |         data={flatListData}
    322 |         renderItem={renderStreamItem}
  

• npm test currently fails in packages/app with Vitest throwing Unexpected call to process.send / ERR_INVALID_ARG_TYPE before any tests run. Track down the coverage/pool config causing this and make sure the app suite actually executes (it should include the hydrated tool-call tests mentioned above).

• WARN [expo-image-picker] ImagePicker.MediaTypeOptions have been deprecated. Use ImagePicker.MediaType or an array of ImagePicker.MediaType instead.

• Update the AgentInput controls so the buttons reflect agent state: when idle show Dictate + Realtime by default (switch to Send when the text box has content); when an agent is running show Realtime + Cancel regardless of input so you can interrupt without typing. Realtime toggle must remain available in both states.