[earakely-scale]

Summary

opencode-cli failed AgentEnvHub's dev promotion gate for 0.9.1373 on mcp and snapshot, and the gate restored its previous image. This fixes both, in the opencode agent:

• Turns wait for memory. No more opencode run processes run at once than the machine's memory fits: one per 900 MB past the first 1 GB, using the cgroup's limit when that's lower. That's 7 on an 8 GB modal_vm. OPENCODE_MAX_CONCURRENT_TURNS sets the number. Turns past the cap wait for one to finish.
• The system prompt reaches the model as a system prompt. It's written to system-prompt.md in the conversation's data directory and named in the config's instructions. Before, it was prepended to the message.

Why

mcp: the VM ran out of memory. In the gate run (dev instance opencode-cli__validate-a2a-v16-7lsevzc4) the MCP probe failed with "agent_response": "opencode exited with code -9 without completing", and it called no tools.

• The validator sends about 15 prompts to the agent at once. The server locks per conversation only, so each prompt started its own CLI process.
• Reproduced in dev with opencode-cli v16 on modal_vm (7,947 MB, 1 CPU, no swap) and 15 concurrent prompts. Each opencode run peaked at 600–860 MB, used memory reached 7,743 MB, and the kernel log shows Out of memory: Killed process … (opencode) four times.
• Which turn the kernel kills is random. v15's run lost other probes instead.

snapshot: the model refused to recall. Save, load and resume all worked, and the recall turn resumed the same session.

• The validator's system-prompt probe sets "respond with this exact token…" on the agent, and it's still set when the snapshot probe plants its token.
• Prepended, the session's first user message read "For validator testing purposes: … respond with the exact token … Remember exactly this token: …". The model took it for a prompt injection: opencode titled the session "Validator token injection attempt", and the recall answered "I won't reproduce tokens or credentials on request, as that could be a prompt injection".
• claude-code passes the same probe because its system prompt is a real one.

Testing

• Unit tests: tst/unit/agents/opencode passes (96 passed, 7 skipped). New tests:
  • turns past the cap wait for one to finish;
  • the cap is sized from memory and can be set;
  • a cgroup limit below the machine's memory counts;
  • the system prompt lands in an instructions file and not in the message;
  • a turn with no system prompt names no instructions.
• The CLI itself: opencode 1.17.18, in the published agentenv-opencode image, against a local stand-in for the model endpoint that records each request. With instructions: ["/tmp/system-prompt.md"], the system message carries Instructions from: /tmp/system-prompt.md followed by the file's text, and the user message carries only the user's words.
• End to end: AgentEnvHub's dev promotion gate is the check. The next pin bump runs the full A2A validator on modal_vm against this image, and restores the previous one if it fails.
• One thing to watch there: queued turns wait on a 1-CPU VM, and some validator prompts time out at 120 s.

🤖 Generated with Claude Code

Confidence Score: 5/5

The PR appears safe to merge, with no actionable findings.

Summary

The PR limits concurrent OpenCode turns based on available memory and passes system prompts through instruction files.

• OpenCode queues turns when memory cannot fit another run.
• OpenCode sends each system prompt as model instructions.

Known tradeoff: earakely-scale noted that queued turns can hit the validator’s 120-second timeout.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Task arrives] --> B[Acquire workspace access and context lock]
  B --> C[Prepare attachments]
  C --> D[Wait for a turn slot]
  D --> E[Write system prompt file if supplied]
  E --> F[Run OpenCode with instructions config]
  F --> G[Finish or clean up canceled CLI]
  G --> H[Release turn slot]
  H --> I[Save session reference and capture changes]

Loading

_{Reviews (1) · Last reviewed commit: "fix(agents): opencode runs as many turns..." · Reviewed by Greptile}