[JacobStephens2]

Summary

Add OpenCode wherever a Harness can be selected, including [harness.opencode] and Setup. Use the user's provider Model and Effort, link Factory skills and instruction fallback, and recover the final reply and usage from the standalone session export.

Command / Setup
  check Model and Effort before work
    opencode run --standalone --format json --auto
Session / Resume
  same flags + stdin prompt + OPENCODE_DISABLE_AUTOUPDATE=1
  JSONL → progress, session id, skill-load warning
  session export → last assistant text, summed usage/cost, outcome
Interrupt
  stop the process tree, including the private server and shell command

Closes #439

Evidence

• Before: harness opencode was rejected as an unknown Harness.
  After: 13 OpenCode scenarios cover selection, flags, stdin, environment, Model/Effort checks, Setup, skills, instruction links, export recovery and failures.
• Before: Interrupting the preflight check could strand its command; token summaries omitted cache-write and reasoning tokens.
  After: The process-tree regression passes, and export parsing sums all assistant token categories and cost while excluding the title call.
• Final validation: cargo test — 1,166 passed, 0 failed; cargo clippy --all-targets -- -D warnings, cargo fmt --check, and git diff --check all pass.

Test seams

• Scenario harness: real Commands and Setup against a fake OpenCode CLI, observing argv, stdin, environment, skills, instruction links, logs, delivery, failures and interruption.
• Pure parsing: documented OpenCode JSONL and export shapes, including a dropped final step_finish, last-assistant text and assistant-only usage totals.
• Public Harness adapter: Resume invocation with a session id parsed from the stream, verifying stdin, flags and environment.

Review

Standards and Spec reviews use issue-432 at d5261163ca8e34846e397dc783c7f07988067322 as the fixed point. Addressed the shared skill-load policy duplication, interruptible preflight check, complete token accounting and unsupported automatic Resume trigger.

Merge Danger

Door: two-way

No data migration. Reverting removes the OpenCode Harness and its configuration support.

Blast Radius: Harnesses

Adds an OpenCode standalone check before work and shares skill-load warning policy with Muse. Existing Harnesses remain registered.

Unaddressed findings

Standards

• None.

Spec

• Run-level Resume scenario coverage (Add the opencode Harness (OpenCode) #439 and parent Spec: four more Harnesses (agy, Grok Build, Muse Code, OpenCode) #432): The documented OpenCode stream has no event that proves background work was killed at session end. Resume is verified through the public adapter using a streamed session id; inventing an automatic trigger to exercise the Run would introduce incorrect behavior.

Built with codex · gpt-6.1-sol · xhigh