[drewstone]

Problem. opencode retries a refused model call itself and sleeps the response's retry-after, capped only at 2^31 ms (session/retry.ts). opencode run --format json prints nothing during that wait: it never emits the session.status retry event. The free seat (opencode/glm-5.3-flash-free) answers a spent quota with 429 and retry-after: 47331 (13.1 h; probed 2026-10-19 10:51Z). discovery-lab's magnetar-rust-20261019b root logged the 429 at 06:28:45Z and then nothing until this bridge's CLI_TIMEOUT_MS (4 h) aborted the turn. agent-runtime received "turn aborted" with no status, classed it terminal, and ended a 48 h run at 4 h, with 0 workers spawned.

#264 does not cover this: the 429 never reaches the bridge as an error event while opencode is sleeping.

Change. In src/backends/opencode.ts:

• Each profiled turn writes a small plugin into its private OPENCODE_CONFIG_DIR, plugin/cli-bridge-retry-signal.js. Its event hook prints opencode's own session.status retry status as one stdout line, {"type":"cli-bridge.opencode-retry", attempt, message, next, timestamp}.
• The event loop reads that line. A wait longer than OPENCODE_RETRY_WAIT_MAX_MS (default 60 s; opencode's own header-less backoff stays at or under 32 s) ends the turn with error type upstream_unavailable. The message names the wait and reset at <ISO>. The finally stops the sleeping process.
• No HTTP status is sent: opencode's retry status names none, and the bridge never infers one. agent-runtime 0.326.0 pauses on upstream_unavailable without spending an attempt, and agent-runtime#1849 holds until the reset when it falls inside its 6 h window.
• A turn without an AgentProfile has no private config directory and keeps today's behaviour. Discovery's turns are all profiled.

Upstream: anomalyco/opencode#57049.

Proof. tests/opencode-retry-wait-real.test.ts runs the REAL opencode against a local endpoint that answers every call 429 with retry-after: 3600. Before this change the turn stays silent until killed. After it, the turn ends coded upstream_unavailable with the reset time, in under 2 minutes. The gate installs opencode 1.18.33 so the test does not skip.

Live bar: the next host free-seat run that meets a spent quota records a paused row instead of a 4 h abort.

Run ID: run_6f83ac468d6e44b1319f450ca91b8e57

[tangletools]

✅ Auto-approved PR — faba90da

Blanket team auto-approval is intentional. This is not a code review.
No automated review runs on this PR. This approval rests on the rule above alone.

_{tangletools · auto-approval · reason: blanket_auto_approve · 2026-10-19T11:01:53Z}