galaxy/specs/UNTRIAGED-tombstone-non-oz/PRODUCT.md

Tombstone metadata for cloud agent runs — Product Spec

Linear: placeholder. Figma: none.

Summary

For cloud agent runs (Oz and non-Oz harnesses like Claude Code), make the conversation-ended tombstone show Run time, Credits used, and the artifact buttons row (plans, PRs, files, screenshots) from the AmbientAgentTask when one is present, matching the conversation details panel for the same run.

Today the tombstone reads run time, credits, and artifacts off the in-memory AIConversation. For non-Oz cloud runs there is no AIConversation, so all three render empty. For cloud Oz runs the conversation values diverge from the details panel: run time excludes queue/sandbox lifecycle, credits exclude compute cost, and artifacts can lag the task-side copy.

Goals

• Non-Oz cloud tombstones render Run time, Credits used, and the artifact buttons row populated from the AmbientAgentTask.
• Cloud Oz tombstones converge on the same task-derived values, matching ConversationDetailsData::from_task numerically and on artifacts.
• Hide the Continue locally button (tombstone + details panel) for tasks whose harness is non-Oz, since forking a non-Oz cloud run into a local Warp conversation is unsupported.
• Client-only change: no harness, server, schema, or endpoint work.

Non-goals

• Directory: / Branch: segments on non-Oz tombstones.
• Implementing Continue locally for third-party harnesses (button is hidden, not wired up).
• Splitting credits into Inference + Compute sub-rows like the details panel does. The tombstone keeps a single aggregated Credits used: X.X segment because the •-separated row has no room.

Behavior

The tombstone metadata row segments behave as follows. Source-of-truth rules mirror ConversationDetailsData::from_task.

• Run time (Run time: <human-readable>):
  • Cloud (task present): AmbientAgentTask::run_time(). Omitted if started_at missing.
  • Local (no task): unchanged. Exchange-derived.
• Credits used (Credits used: <formatted via format_credits>):
  • Cloud (task present): AmbientAgentTask::credits_used() (inference + compute). Omitted if request_usage missing.
  • Local (no task): unchanged. conversation.credits_spent().
• Source / Skill: unchanged (already enriched from the task).
• Working directory: Oz unchanged (AIConversation::initial_working_directory()); non-Oz not rendered.
• Artifacts row (plan, branch, PR, screenshot, file buttons):
  • Cloud (task present, non-empty task.artifacts): task.artifacts. Wins over conversation artifacts.
  • Cloud (task present, empty task.artifacts): falls back to whatever from_conversation set (avoids blanking Oz when the task is partially populated).
  • Local (no task): unchanged. conversation.artifacts().
• Continue locally button (tombstone desktop + details panel):
  • Hide rule is purely on harness: hide iff harness == Some(non-Oz). None (unknown / not yet loaded / plain conversation) and Some(Oz) both show the button.
  • Local (no task): unchanged — harness stays None, shown when AI is enabled.
  • Cloud Oz task: shown both before and after task load.
  • Cloud non-Oz task (Claude, Gemini): button shows briefly until task fetch resolves, then hides.
  • Tombstone wasm Open in Warp button: unchanged. Opens the same conversation in the desktop client, where the same hide rules then apply.

Empty-row, error, and snapshot/transcript-viewer behaviors are unchanged.

Success criteria

• For any cloud run with an AmbientAgentTask (Oz or non-Oz), the tombstone's Run time and Credits used strings equal what ConversationDetailsData::from_task produces for the same task.
• Non-Oz cloud tombstones with a populated task.artifacts render the same artifact buttons as the details panel for the same run.
• Cloud Oz tombstone numbers will visibly change: both increase. New values match the details panel and Oz task list.
• Cloud non-Oz tombstones and details panels do not render Continue locally once the task is loaded.
• Cloud Oz tombstones and details panels still render Continue locally.
• Local Oz tombstone behavior is unchanged.
• Non-Oz tombstones remain feature-flagged behind FeatureFlag::AgentHarness.

Validation

• Unit tests on enrich_from_task covering: task wins for run time / credits / artifacts when present; falls back to from_conversation values when the corresponding task field is missing.
• Unit tests on enrich_from_task covering harness extraction: snapshot absent → harness stays None; snapshot without explicit harness → defaults to Oz; snapshot with explicit harness propagates.
• Manual: cloud Claude Code run shows non-empty Run time, Credits used, and artifact buttons matching the details panel, and Continue locally is hidden in both surfaces; cloud Oz run shows updated numbers and artifacts also matching the details panel, with Continue locally shown; local Oz run unchanged.

Follow-ups

• Wasm parity (cloud tombstones on web don't currently fetch the task; pre-existing gap).
• Directory: / Branch: on non-Oz, via a future POST /harness-support/report-context sibling endpoint that piggy-backs in parallel with the existing transcript upload.
• Possible Download conversation button surfacing the harness JSON.