Arch Agent: retry a dropped model connection, and say why it failed

[asaficontact]

Spec: Tawab, in the 0.1.0 desktop app on his Mac, asked "Draw this project's system overview…". The transcript showed "looked at 1 places", then "Connection error.", and no answer.

Cause:

• pi-ai 1.0.2 turns the OpenAI and Anthropic SDKs' own retries off (maxRetries: 0) and runs its own retry loop, which defaults to 0. archdraw never set it, and pi-agent-core has no option for it.
• So a single dropped connection on the request carrying the tool result ended the turn.
• pi-ai also keeps only the error's message. The real reason (err.cause.cause: UND_ERR_SOCKET, ECONNRESET, ENOTFOUND…) was thrown away, and the app keeps no log, so the failure could not be diagnosed.

Change:

• streamWithRetries (app/core/src/models.ts) sends each model request with 2 retries, the SDKs' own default.
  • Retries happen before a response starts streaming, so text is never duplicated.
  • For the OpenAI and Anthropic APIs (and OpenRouter through the OpenAI one), it passes a fetch that records why a connection failed.
• A lasting failure now reads e.g. Connection error (other side closed, UND_ERR_SOCKET).
• The scheduled change check and its drafts get the same retries.
• CHANGELOG has an Unreleased entry.

Evidence:

• app/core/test/agent-network.test.ts runs a real Conversation over a real socket against a local Responses API stand-in:
  • request 2 dropped once: it is retried, and the agent answers;
  • request 2 dropped every time: the error names the cause, after 3 attempts.
• Both cases failed before the fix with the bare "Connection error.", as in Tawab's transcript.
• The same two drop modes were reproduced and then fixed in the bundled server for OpenAI and Anthropic, and in the Electron desktop bundle for OpenAI (Linux, xvfb).
• scripts/ci-local.sh passed on this head, which is rebased on main 25ac207.

Not tested:

• real OpenAI or Anthropic endpoints;
• real macOS;
• Google. Its adapter refuses a custom fetch, so it gets the retries but not the cause text.

Breaks: if retries ever duplicated a reply, the transcript would show it twice. agent-network.test.ts guards the retry path.

Rollback: git revert the commit.

— Banna, one of Tawab's agents

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
archdraw	Ready	Preview	Oct 6, 2026 1:47pm UTC