feat(session): support multiple PRs per session

[Pritom14]

Summary

Lets a single session own several pull requests (a root plus stacked children) instead of being capped at one. The SQLite schema was already 1-session→many-PR (pr.url PK, session_id a plain FK with idx_pr_session, no UNIQUE on session_id), so this is a behavioural change across the observe → persist → derive → react pipeline, not a migration.

State & user-action flow

The session's display status is never stored — it is derived at read time from durable PR facts. One full loop:

1. Agent opens PR(s). A root PR on the session branch (feat/x), optionally stacked children on the feat/x/... convention.
2. OBSERVE (observe/scm): each poll lists the repo's open PRs, attributes every PR whose source branch equals the session branch or descends from it to the owning session (longest-prefix wins), and persists the pr row including the source/target branch pair.
3. DERIVE (service/session/status.go, read-time): session status = worst-wins aggregate over the open PRs, by severity ci_failed > changes_requested > draft > review_pending > pr_open > approved > mergeable. A stacked child whose parent is still open is excluded from the aggregate (it cannot merge until the parent does). The stack model + per-PR facts are surfaced as prs[] on every session DTO, so the dashboard shows each PR's state.
4. ACT (lifecycle/reactions.go): per-PR nudges to the agent — CI failing → "fix CI" (with log tail); review changes/unresolved comments → "address feedback"; merge conflict → "rebase", but only the bottom of the stack; a stacked child blocked by an open parent is suppressed. Dedup signature is persisted per PR so a daemon restart never re-fires.
5. User action closes the loop. From the derived status + prs[] the user/agent fixes CI, addresses review, or rebases and pushes; the next poll re-observes and re-derives. No status is ever written back — only PR facts are.
6. Completion. The session terminates (→ status merged) only when no PR is open and at least one merged. A single PR merging while another stays open keeps the session alive.

What changed, by layer

• Observe (observe/scm/observer.go): one refresh subject per open tracked PR; discoverNewPRs + matchSession do the branch-prefix attribution (longest match wins so a child session claims its own stacked PRs).
• Derive (service/session/stack.go, status.go): worst-wins aggregation + stack model (B is a child of A iff B.target_branch == A.source_branch and A is open), surfaced as prs[].
• React (lifecycle/reactions.go): per-PR reactions, stacked-child nudge suppression, and sessionComplete (all closed/merged + ≥1 merged).
• API (httpd/controllers/dto.go, sessions.go, regenerated openapi.yaml + frontend/src/api/schema.ts): prs[] on the session views via SessionView.

Tests

• Unit coverage across stack, status, observer multi-subject discovery, and lifecycle per-PR reactions/completion.
• A real-SQLite store test (store/pr_facts_test.go) for the stacked-PR read path: ListPRFactsForSession returns every owned PR newest-first with state flags and branch pair preserved (the data the stack aggregator builds on).
• Functional end-to-end (integration/scm_observer_test.go, TestSCMObserverMultiPREndToEnd) drives the real sqlite.Store + lifecycle.Manager + observe/scm.Observer (canned provider only) through: (1) one session owning its root and stacked child from a single repo list with branch pairs persisted; (2) staying alive while a stacked PR is open and terminating only once all merge; (3) a stacked child blocked by an open parent suppressed while the stack bottom is nudged (asserted down to the persisted dedup signature).

Test plan

• go build ./... clean
• go vet ./... clean
• go test ./... green (53 packages, exit 0)
• TestSCMObserverMultiPREndToEnd passes (attribution / completion / nudge suppression)
• Live agent-driven run: spawning an agent that opens multiple real GitHub PRs on one session. Needs a real GitHub repo + agent credentials/environment not available in this workspace, so it is not exercised here — the functional proof above drives the real store/lifecycle/observer against a canned SCM provider instead.

🤖 Generated with Claude Code

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.

[greptile-apps]

Your free trial has ended. If you'd like to continue receiving code reviews, you can add a payment method here.