[wilke]

Follow-up to #67/#71, from the second live refiner build on Argo (2026-10-08, build_request_df9ee3a5…, Run run_7645d122…).

What happened

• The Code Implementer ran 98 steps: 96 run_command, 3 read_file, 1 list_files, and no source edit.
• The conversation grew to ~530k characters (~130k tokens). Per-step latency rose from a median of ~6 s to 180 s and 81 s on the last calls.
• Argo then dropped the stream mid-response:

  peer closed connection without sending complete message body (incomplete chunked read)
  

• Add a chat-completions coding adapter for Implementer (Argo / OpenAI-compatible routes) #67 treats any provider error as fatal (api_error → CodingTurn.error → ModelCallFailed), so the build ended interrupted with no edits.

Change (agent/transports/refinement_chat.py)

• Retries: transient transport failures are retried on the same route up to 3 times, with 5/20/60 s backoff. "Transient" uses the existing auxiliary_client._is_transient_transport_error, which covers dropped streams, connection errors, 5xx and 408. Retries are reported as activity events, not api_error, because the Implementer interrupts the turn on api_error. Non-transient errors, and transient ones still failing after the retries, surface as before.
• Smaller context:
  • history budget 600k → 240k characters;
  • per-tool output 40k → 20k;
  • elided old tool output keeps its first 300 characters, so the model still sees what each command was.
• Writing nudge: after 40 steps without write_file/edit_file (and every further 40), the model is told to write the assigned source or record the specific blocker in .openchia-implementation.json.

Tests: 14 pass, 4 new:

• a dropped stream is retried and the turn succeeds;
• a persistent drop still surfaces;
• the nudge fires after N steps without a write;
• compaction keeps the head of old tool outputs.

Context

Most of the exploration reads library sources and earlier packages outside the workspace. #74 (container shell plus read-only /ctx/openchia) and #73 stage 3 (curated context) should shorten it further.

🤖 Generated with Claude Code