fix(server): refresh provider session lastSeenAt on turn and task activity

[kraptor23]

What

Refreshes provider_session_runtime.last_seen_at when the adapter emits turn or background-task lifecycle events, so the inactivity reaper measures idleness from the last completed work instead of the last explicit sendTurn.

Addresses the staleness half of #4198.

Why

last_seen_at is only written by startSession / recoverSessionForThread / sendTurn / stopSession. Turns delivered without a sendTurn call (queued follow-ups, plan-sourced turns) and background tasks never refresh it. Observed on 0.0.32-nightly.20260805.1006: one thread was reaped 6 times in ~12 hours, three of those 105 s - 4 m 53 s after a turn had actually completed — the reaper was measuring idleness from a last_seen_at several turns stale, exactly what the reporters in #4198 back-computed from their logs.

How

• ProviderSessionDirectory.touch(threadId) updates only last_seen_at for an existing binding via a new repository touchByThreadId (UPDATE ... WHERE thread_id). It never creates rows, and preserves every other column.
• ProviderService.processRuntimeEvent touches the binding on turn.started/completed/aborted and task.started/progress/completed, after event fan-out.
• Touches are throttled to one per thread per 60 s (in-memory map, cleared on stopSession) because this runs on the event drain fiber and task.progress can fire many times per second in multi-agent sessions; reaper thresholds are minutes, so minute-level granularity is sufficient. The gate timestamp is set before the write so a failing database is not retried per event.
• Touch failures log provider.session.activity-touch-failed and never disturb fan-out.

This deliberately does not add a busy-guard for in-flight background tasks (discussed in #4198) — that is a separate lifecycle change; this PR only fixes the idle clock.

Tests

• ProviderService.test.ts: every activity event type refreshes the binding's lastSeenAt; content.delta does not; a second boundary inside the throttle window does not write again.
• ProviderSessionDirectory.test.ts: touch refreshes lastSeenAt and leaves all other fields unchanged; touch on an unknown thread creates no row.
• vp test run on the five touched test files (97 tests) and tsgo --noEmit pass.

Prior art and non-goals

This addresses the liveness-staleness half of #4198. Back-computing the reaper's logged idleDurationMs against turn timestamps shows every incident reported there is an event-emitting case this PR covers: ordinary provider-started turns that never refreshed last_seen_at (NoorChasib's table — turns at 06:08 and 06:17 with last_seen_at stuck at 05:49), and background tasks killed seconds after their foreground turn settled (Mooned8, Zeus-Deus — task.started/turn.completed both now touch the binding).

Earlier attempts in this area (#4199, #3856, #4994) were closed for orchestration-V2 roadmap reasons rather than on the mechanism, and #2829 is still open — this PR is deliberately the smallest reviewable slice of that idea: one new directory method, one call site on the existing event drain, throttled, with no lifecycle behavior changes.

Non-goals, intentionally out of scope here:

• a busy-guard exempting sessions with outstanding-but-quiet background tasks (the remaining half of [Bug]: Idle session reaper kills in-flight background agent work (dynamic workflows / subagents) #4198) — planned as a small follow-up building on the existing backgroundLiveness thread-shell state;
• post-reap result re-delivery / respawn-on-completion ([Feature]: Don't reap sessions with pending background tasks (completion notifications are silently lost) #4265);
• reaper convergence for wedged threads ([Bug]: Deleting a thread's worktree while its session is "running" permanently wedges the thread (archive blocked, reaper never clears it) #4723, [Bug]: App restart during active turns leaves threads stuck on Working forever, with no startup reconciliation of dead provider sessions #4584) — only sensible after this PR makes lastSeenAt trustworthy.

Written by Claude Fable 5 via Claude Code.

---

Note

Medium Risk
Changes when sessions are considered idle for reaping (core provider lifecycle), but scope is narrow—targeted DB updates with throttling and non-blocking failure handling.

Overview
Provider session lastSeenAt now advances when adapters emit turn or background-task lifecycle events, not only on startSession, sendTurn, or stopSession. That aligns the inactivity reaper with last completed work (queued turns, plan-driven work, long-running tasks) instead of the last explicit user send.

A touch path updates only last_seen_at via repository touchByThreadId (UPDATE … WHERE thread_id); it never inserts rows and leaves other binding fields unchanged. ProviderService.processRuntimeEvent calls it after fan-out for SESSION_ACTIVITY_EVENT_TYPES (turn.*, task.*), with a 60s per-thread throttle and throttle state cleared on stopSession. Streaming events like content.delta do not touch; touch failures are logged and do not affect event delivery.

^{Reviewed by Cursor Bugbot for commit 0fadabd. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Refresh provider session lastSeenAt on turn and task activity events

• Adds a touch method to ProviderSessionDirectory and ProviderSessionDirectoryShape that updates only last_seen_at for an existing binding, with no-op behavior for unknown threads.
• Extends ProviderSessionRuntimeRepository with touchByThreadId, which performs a targeted UPDATE on last_seen_at without affecting other fields.
• In makeProviderService, emitting events from the new SESSION_ACTIVITY_EVENT_TYPES set (turn.started, turn.completed, turn.aborted, task.*) triggers directory.touch, throttled to at most once per 60 seconds per thread. The throttle map entry is cleared on stopSession.
• Behavioral Change: provider session lastSeenAt now updates during active turns and tasks, where previously it only updated on session creation or explicit upsert.

^{Macroscope summarized 0fadabd.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: b02ac06f-8f1a-4af5-ae12-8301f9034df2

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[macroscopeapp]

One finding: the new activity-touch warning log serializes the raw cause into the log payload instead of a normalized error tag.

^{Posted via Macroscope — Effect Service Conventions}

[macroscopeapp]

Approvability

Verdict: Needs human review

This PR changes session lifecycle behavior by updating lastSeenAt timestamps based on turn/task events, affecting how the inactivity reaper determines session idleness. While well-tested and throttled, this runtime behavior change to session management warrants human review.

^{You can customize Macroscope's approvability policy. Learn more.}