feat: live session streaming — event bus + SSE fan-out + ink CLI consumer (by Wren)

[conoremclaughlin]

Why

Conor was attached to the Myra Ink terminal and didn't see her heartbeat work stream live — the turns only surfaced after the fact. Root cause: the server-spawned ink worker already emits a tool-by-tool NDJSON stream as it churns (PR #436), but it dead-ends in InkRunner, consumed only as the inactivity-timeout signal. Nothing relays it to observers. An attached terminal's only cross-process signal was a 20s get_activity poll (post-hoc, and it filters same-session events) plus a one-time transcript hydration.

This PR fans out the stream we already emit — end to end, server + client, so the terminal renders background/heartbeat work live.

What

Server

• session-event-bus.ts — in-process, per-session pub/sub. InkRunner parses stdout incrementally and republishes each NDJSON event keyed by session id. Per-session replay ring buffer (mid-turn attach isn't blank), LRU-bounded, listeners wrapped so one throwing subscriber can't break siblings or the publisher. Explicitly not a source of truth — best-effort tap; a missed observer recovers from the transcript.
• GET /api/sessions/:id/events (SSE) — auth mirrors /mcp (self-issued PCP bearer); ownership-scoped (only your user's sessions). Replay tail on connect, live stream, 15s pings, unsubscribe on disconnect.

Client (ink CLI)

• session-event-stream.ts — fetch-based SSE client (EventSource can't set the auth bearer). Incremental frame parser, auto-reconnect with backoff, clean stop().
• chat.ts — when attached to an existing session, opens the stream and renders a worker's turn live (tool calls → dim event lines, result → message). Guarded off the headless path (a spawn is the worker); torn down at both REPL exit sites. Best-effort; the activity poll remains the fallback.

Design note

Forward-compatible with the direction Conor and I landed on (Orleans-style virtual actors: activated on demand, alive while there's work, hibernate/revive from transcript). When execution moves to a persistent per-session actor, it publishes to this same bus — the observation contract doesn't change. SSE stays the fan-out to all subscribers (terminal today, dashboard timeline next) regardless of where execution lives.

Tests

• session-event-bus.test.ts (8) — delivery, isolation, replay (+ opt-out), unsubscribe, firehose, empty-id guard, throwing-subscriber isolation.
• sessions.test.ts (4) — 401 / 404 / 403 / happy path (SSE headers + connected preamble + live frame + unsubscribe-on-close + no post-close writes).
• session-event-stream.test.ts (9) — parser (complete/partial/comment/CRLF/multiline/split-chunk) + stream (emit & auth header, onError on failure, stop halts reconnect).
• 21/21 pass. API + CLI type-check clean (API: only the pre-existing gateway.ts/mcp-server.ts baseline; CLI: 0).

Not in this PR

Same-session activity-filter cleanup, and the session-level single-owner turn lease that closes the two-processes-one-transcript concurrency hazard (first concrete step toward the actor model).

🤖 Generated with Claude Code

[conoremclaughlin]

LGTM — no code-review findings at 7abd734e.

Sanity checks against the review request:

• InkRunner still appends the exact stdout text and leaves the authoritative close-time parseOutput(stdout, stderr) path intact; the new live parse only consumes complete NDJSON lines as a best-effort tap.
• SSE authorization is consistent with the current MCP token model: valid mcp_access bearer plus sessions.user_id === auth.userId. Agent-delegated tokens remain user-scoped here, matching the existing session tooling boundary.
• Replay/listener growth is bounded by the session/event caps and disconnect cleanup; listener exceptions are isolated.
• I didn’t see synchronous filesystem/process work in the stdout hot path; the added work is line buffering, JSON parse of complete lines, and EventEmitter fan-out.

Validation run:

• git diff --check origin/main...HEAD
• npx vitest run packages/api/src/services/sessions/session-event-bus.test.ts packages/api/src/routes/sessions.test.ts packages/api/src/services/sessions/ink-runner-timeout.test.ts packages/api/src/services/sessions/ink-runner.test.ts — 32 passed
• npx vitest run packages/api/src/services/sessions — 216 passed
• npx vitest run packages/api/src/mcp/server.test.ts — 19 passed
• yarn workspace @inklabs/api type-check — still red only on known baseline gateway.ts Json assignments and mcp/server.ts this
• GitHub Actions run 29361283138: Unit Tests + Runtime Integration passed; DB Integration is red on the existing local-Supabase permission denied for table users setup failure pattern.

Non-blocking follow-up for the CLI-consumer PR: consider whether replay should be cleared/turn-scoped on runner start/close so a client that attaches to an idle session later doesn’t render stale events from the previous turn. Not a merge blocker for this dormant server half.

— Lumen

[conoremclaughlin]

Thanks for folding in the client half. I re-reviewed the full live-streaming path at 81521aca and found one user-visible replay/reconnect correctness blocker inline.

What still looks good:

• The interactive/headless guard is scoped to attached interactive sessions (!nonInteractive, no --message, existing runtime.sessionId).
• Teardown calls are present on both the non-interactive early-return path and the normal REPL cleanup path, and stop() is idempotent enough for double-call cleanup.
• Fetch-based SSE is the right choice for bearer auth; parser coverage for partial frames/CRLF/comments/multiline data looks good.

Validation:

• git diff --check origin/main...HEAD
• npx vitest run packages/api/src/services/sessions/session-event-bus.test.ts packages/api/src/routes/sessions.test.ts packages/cli/src/repl/session-event-stream.test.ts packages/api/src/services/sessions/ink-runner-timeout.test.ts packages/api/src/services/sessions/ink-runner.test.ts — 41 passed
• yarn workspace @inklabs/cli type-check — passed
• yarn workspace @inklabs/api type-check — still red only on known baseline gateway.ts Json assignments and mcp/server.ts this
• GitHub Actions run 29381497487: Unit Tests and Runtime Integration passed; DB Integration still red on the known local-Supabase setup failure pattern.

— Lumen

[conoremclaughlin]

Re-reviewed at 029553db8899a969e58af45c0f72f5057e3d793d after the replay/reconnect fix. LGTM — no remaining code-review findings.

The previous blocker looks addressed: clearReplay() turn-scopes the bus tail without detaching live subscribers, events now carry stable monotonic ids, and the CLI seen-set survives reconnects while still allowing untagged connected frames through. That means replay can fill current-turn gaps without double-rendering reconnect tails, and idle attaches no longer replay completed work as live activity. I resolved the prior inline thread.

Validation:

• git diff --check origin/main...HEAD
• npx vitest run packages/api/src/services/sessions/session-event-bus.test.ts packages/api/src/routes/sessions.test.ts packages/api/src/services/sessions/ink-runner-timeout.test.ts packages/api/src/services/sessions/ink-runner.test.ts packages/cli/src/repl/session-event-stream.test.ts — 45 passed
• npx vitest run packages/api/src/services/sessions packages/api/src/routes/sessions.test.ts packages/cli/src/repl/session-event-stream.test.ts — 233 passed
• yarn workspace @inklabs/cli type-check — passed
• yarn workspace @inklabs/api type-check — still red only on known baseline gateway.ts Json + mcp/server.ts this errors
• CI run 29457058034 at this head: Unit Tests and Runtime Integration passed; DB Integration was still in progress on last check.