[munchist]

What broke

In the live test of #123, Helmsman's first turn on Gemma 4 26B failed three times, each attempt at almost exactly 300 s (11:16:29, 11:21:42, 11:26:58). Each time it was reported as Ollama is not reachable at http://127.0.0.1:11434: fetch failed, and Ollama was running throughout.

Why: Node's fetch (undici) abandons a response whose headers take 5 minutes (headersTimeout), and a body that stays silent for 5 (bodyTimeout). Ollama sends no headers until it has read the whole prompt, and Helmsman's prompt on this 24 GB machine took longer than that. The platform's own 15-minute idle window for local models (AI_LOCAL_MODEL_IDLE_TIMEOUT_MS, §9 G5) never got the chance to apply.

The cause is reproduced with a probe: an undici Agent with a 1 s headersTimeout fails as fetch failed, cause UND_ERR_HEADERS_TIMEOUT, and with 0 it waits. That holds on this machine's Node 22.16, which bundles undici 6.21, with undici 7 passed in as the dispatcher.

The fix

• No transport limit on local-runtime requests. localRuntimeFetch sends every Ollama request (chat, embed, warm, status reads, pull, delete) through an undici Agent with neither limit. The caller's signal is the only limit: the idle window for a turn, the lease for a download. The pull stream needed this as well, because Ollama is silent while it checks a large layer's digest.
  • undici@^7 is now a declared dependency of @aflow/ai-client. 8.x requires Node ≥ 22.19, and this machine runs 22.16.
• "Not reachable" only when nothing answers. That message is kept for connection causes (ECONNREFUSED, ENOTFOUND, EHOSTUNREACH, …). Any other broken request names its cause.
• No retry after a silent idle window. When Ollama stays silent through the whole platform idle window, the turn fails non-retryable, with "Ollama sent nothing for 15 minutes…" and the remedy: a smaller window or a smaller model. A retry would read the prompt again from the start. The agent-turn "keep this turn short" advice is appended only to retryable timeouts.
• Spill wording. The 32K estimate now says the model will spill as its conversations fill the window. While loaded, the model sits fully on the GPU.

Not in this PR

• A long Helmsman prompt is still slow on a local model; that is the planned slimmer local prompt. After this fix, the first turn waits for the prompt rather than failing at 5 minutes.
• The run surface still shows "internal platform error" rather than the stored provider message. That is a separate surface.

Testing

• ollama.test.ts: four tests replace the old network-failure one:
  • nothing answering at the URL;
  • a request that broke off, reported by its cause;
  • requests carry the dispatcher;
  • silence through the idle window is not retried.
• 40 tests across the three affected files pass, and yarn typecheck passes.
• yarn test:changed: 16250 pass, 1 fails. The failure is the host checkoutReplacement test, which also failed on Plan 323 P3 + P3b — models on this computer, diagnosed, managed, and changed by Helmsman with approval #123 and imports nothing touched here.
• Not yet run live. Once this is in live, a Helmsman turn in Walkthrough should run past 5 minutes and answer.

🤖 Generated with Claude Code