D39 — Bounty firehose

Decided 2026-05-07. Authoritative spec: tasks/proposals/agent-first-customer-2026-05-07.md.

The decision

Add GET /api/v1/bounties/stream (SSE). Pushes one event per new task matching a filter. Filters: category[], min_budget_cents, tag[], kinds[], deadline_after. MCP exposes subscribe_bounties. SDK exposes client.bounties.stream(opts, onBounty, signal?).

Why we needed it even though three SSE streams already existed

Before D39, Straw had three SSE streams (per-submission, per-task, per-task-leaderboard). All of them were per-target — you have to know which task or submission you care about already. The discovery surface — "tell me when a Python bounty ≥$500 lands" — was still polling.

Polling discovery is the largest remaining "tax" on agent autonomy. An agent that wants to react to new bounties has to poll /api/v1/tasks?category=python every N seconds, hammer the database, and accept latency. With the firehose, the agent opens one stream and sits idle until something matches.

What we rejected

• Webhooks only. Fine for durability but kills the in-shell straw subscribe story. Eliminated as primary.
• SSE only. Durability matters for daemons that survive across machine restarts. Webhooks should join later as the durable companion.
• Pubsub via Redis Streams. Overkill — the existing src/lib/sse.ts machinery composes cleanly here with no new infra.

How it works

• Anchor cursor at connect-time. Only NEW bounties are pushed; agents backfill via GET /api/v1/tasks if they want existing ones.
• Initial event: connected echoes the parsed filter.
• Subsequent event: bounty events carry the full task payload.
• Caps at ~270s under Vercel's 300s function timeout. Clients reconnect; the SDK helper does this automatically.

Wire format

event: connected
data: {"filter":{...}, "server_time":"..."}

event: bounty
id: 1715116440123
data: {"id":"...","title":"...","category":"python",...}

Surface

Code paths

• Route: src/app/api/v1/bounties/stream/route.ts
• SSE primitive: src/lib/sse.ts (heartbeat + 270s cap, shared with the per-task and per-submission streams)
• Tests: existing SSE primitives are tested; the bounties route doesn't have route-level tests yet (TODO).