Outcome
Returning to a track gives the reviewer a short, evidence-backed account of what changed, what was checked, and what still needs judgment. It also lets a teammate take over without reading the entire transcript.
Proposed priority: near-term, alongside or after the attention inbox. Audience: implementing coding agent. Previews are already in progress; consume that capability when available rather than rebuilding it.
First useful version
Add a review summary with the original request, change summary, changed files/diff links, checks and their results, unresolved limitations, PR link, and preview link when available. Let the user request/refresh the summary; generation must not block prompt delivery or mutate repository files.
Keep agent-reported claims distinct from observed evidence. Link a claimed local test to its recorded command/result when available; otherwise label it unverified. Fetch GitHub check results for the actual pushed SHA. Missing checks do not mean passing checks.
Bind each summary to a conversation/turn boundary and a captured repository revision plus working-tree fingerprint, including relevant uncommitted/untracked changes. Detect changes during capture and retry or label the capture inconsistent. Mark the summary stale when the underlying work changes. A live preview must be labeled as live; it is not proof of the captured revision.
Implementation starting points
Read apps/switchyard/src/components/{Changes,Checks,Transcript,TrackView}.tsx, src/lib/tools.ts, and server/{tracks,repos,db}.ts. Reuse existing diff, event, and GitHub access paths. Persist the summary and evidence references server-side, with normal track authorization. Choose a bounded generation path after inspecting Fountain's current capabilities; do not send an unsolicited summarization turn into a busy editing conversation.
Completion checks
- A second project member can review the result and follow every evidence link.
- Edits after capture visibly stale the summary; old green checks cannot imply the new working copy passed.
- Failed, missing, and agent-claimed checks remain distinguishable.
- Refresh/restart preserves summaries and revocation blocks access.
- Tests exercise stale captures, evidence attribution, and authorization; Switchyard tests/build pass.
Automatic approval, merge/deploy controls, and a universal test runner are outside this issue.
Outcome
Returning to a track gives the reviewer a short, evidence-backed account of what changed, what was checked, and what still needs judgment. It also lets a teammate take over without reading the entire transcript.
Proposed priority: near-term, alongside or after the attention inbox. Audience: implementing coding agent. Previews are already in progress; consume that capability when available rather than rebuilding it.
First useful version
Add a review summary with the original request, change summary, changed files/diff links, checks and their results, unresolved limitations, PR link, and preview link when available. Let the user request/refresh the summary; generation must not block prompt delivery or mutate repository files.
Keep agent-reported claims distinct from observed evidence. Link a claimed local test to its recorded command/result when available; otherwise label it unverified. Fetch GitHub check results for the actual pushed SHA. Missing checks do not mean passing checks.
Bind each summary to a conversation/turn boundary and a captured repository revision plus working-tree fingerprint, including relevant uncommitted/untracked changes. Detect changes during capture and retry or label the capture inconsistent. Mark the summary stale when the underlying work changes. A live preview must be labeled as live; it is not proof of the captured revision.
Implementation starting points
Read
apps/switchyard/src/components/{Changes,Checks,Transcript,TrackView}.tsx,src/lib/tools.ts, andserver/{tracks,repos,db}.ts. Reuse existing diff, event, and GitHub access paths. Persist the summary and evidence references server-side, with normal track authorization. Choose a bounded generation path after inspecting Fountain's current capabilities; do not send an unsolicited summarization turn into a busy editing conversation.Completion checks
Automatic approval, merge/deploy controls, and a universal test runner are outside this issue.