Last updated: 2026-10-25

OpenCode full-worktree trial#

This is the original failed trial. The later ROCm context and harness follow-up records the backend fixes and subsequent acceptance results; it does not change these outcomes.

On 2026-10-25, OpenCode 1.19.4 was assigned a real Zinc feature in a complete Git worktree based on 09b802de: add loaded-model discovery to the minimal ROCm/CUDA HTTP server. The task required an OpenAI-compatible GET /v1/models response, basename-only model IDs, correct JSON escaping, CPU unit tests, and a documentation note.

Result: failed. Eight configurations produced 57 tool calls, zero feature source edits, and zero project test runs. The endpoint remained unimplemented. This was a supervised capability trial, including navigation hints, configuration changes, and a resumed session. It was not an unattended or one-shot success.

The runtime was Zig 0.17.0, Bun 1.4.4, ROCm 7.2.4, and a Radeon AI PRO R9700 with 32GB VRAM. The models were Qwen 3.8 27B Q4_K_M and Qwen 3.5 9B Q4_K_M. ZINC served the model directly; no proxy repaired responses or synthesized edits. Raw prompts, machine paths, node addresses, credentials, and transcripts are excluded from this report.

Configuration Observed outcome
27B, 32K context GPU allocation failure during the initial repository prompt
27B, 16K context Context overflow after source inspection
9B, 32K, general output limits Built-in file reads bypassed those limits and overflowed context
9B, 32K, Bash-only prototype Attention launch exceeded shared memory; output became corrupted
9B, 16K, early compaction Ended without an implementation
27B, shorter instructions, Bash-only Repeated inspection; stopped without edits
Resumed 27B, retained recent history Repeated inspection; stopped without edits
27B, normal tools, shorter instructions, small reads Illegal GPU memory access and corrupted output near the context boundary

The final server log identified attention_causal_batched_v2 with a 65,612-byte group segment at 16,080 prompt tokens, followed by illegal-memory errors and a rocBLASLt/Tensile error. Another run rejected naive_attention launches above 16K tokens. These are backend failures requiring attention-kernel validation beyond the earlier bounded recurrent-scan checks. They are not evidence that a larger configured context is usable.

The trial also found workflow problems. The original 31KB AGENTS.md consumed much of a small context window. Moving code maps and benchmark recipes into reference documents reduced it to about 10KB while retaining permission and GPU placement rules. Zero-history compaction discarded source that the model then read again. The Bash-only prototype bounded output, but did not demonstrate a completed coding task and was discarded. The public setup script retains its normal Read/Edit/Write tools.

The earlier nine-fixture result does not establish general repository-level coding reliability. A real task needs a source patch, independent acceptance checks, passing project tests, and a clean server log. A process exit code or number of tool calls is insufficient. All model servers owned by this trial were stopped, and the GPU was idle after cleanup.

Post-trial repository checks passed on macOS and Linux, along with 676 Bun tests and the 263-page site build. Six model-dependent Bun tests were skipped. These checks validate the documentation and comment cleanup; they do not turn the failed coding trial into a pass. Five pre-existing source-comment hygiene matches were corrected without changing engine behavior. The supervisor also fixed Markdown links in the website renderer and mobile wrapping of long inline code; these changes were not produced by the coding agent.

Reproduce an isolated worktree run.