QA Evidence - PR #6127 sisyphus-junior GPT variant/reasoningEffort overrides

What was tested

• Unit RED->GREEN on createSisyphusJuniorAgentWithOverrides + getSisyphusJuniorPromptSource (worktree at PR head, then merged with dev + corrected impl).
  • RED 1 (PR head 0a43f7922e): 3 failures - reasoningEffort-only override fabricates variant; variant-only override contaminates reasoningEffort; gpt-5.6 misroutes to generic "gpt" prompt (accidental revert).
  • RED 2 (post dev-merge, PR impl kept): 3 failures - both cross-contamination cases + Claude variant override ignored.
  • GREEN (corrected impl): 62 pass / 0 fail; bunx tsgo --noEmit -p packages/omo-opencode exit 0.
  • Adjacent suites: agents/ 375 pass, plugin-handlers/ 227 pass.
• REAL OpenCode surface (AGENTS.md mandate): runtime bundle dist/index.js (bun build + require-shim patch) loaded into real opencode serve v1.18.7 in an isolated XDG sandbox; two projects under $HOME:
  • override project .omo/omo.jsonc: agents.sisyphus-junior = { model: openai/gpt-5.6-sol, variant: xhigh, reasoningEffort: xhigh }
  • control project: same model, no variant/reasoningEffort.
  • Asserted via GET /config?directory=... (merged agent config) and GET /agent?directory=....

What was observed

• override: variant: "xhigh", reasoningEffort: "xhigh", model openai/gpt-5.6-sol in merged agent config (real-opencode-config.json); /agent shows variant: "xhigh" reaching the runtime agent surface.
• control: NO variant key, reasoningEffort: "medium" (default preserved - negative control proving no cross-contamination and unchanged default behavior).
• GPT-5.6 prompt routing restored: rendered Sisyphus-Junior prompt begins "...based on GPT-5.6 Sol" (gpt-5-5 model-aware family), not the generic GPT prompt.
• Isolation: server ran with sandboxed XDG_*; real ~/.local/share/opencode/opencode.db session count 21932 before, 21932 after (unchanged). Server pid terminated; sandbox dirs removed.

Why it is enough

The change is confined to createSisyphusJuniorAgentWithOverrides output. The unit RED->GREEN pins each override field independently (distinct values, so cross-contamination cannot pass), plus gpt-5.6 routing and Claude-variant honoring. The real-OpenCode probe proves the exact user-visible contract end to end: a user's .omo/omo.jsonc override lands in the merged agent config OpenCode serves, and the default stays "medium" when unset. Adjacent agent/plugin-handler suites cover regression risk in the assembly path.

What was omitted

• No secrets involved; sandbox served with a throwaway password, never logged. Raw serve logs discarded with the sandbox (contained only startup noise).
• reasoningEffort is not surfaced by GET /agent options on opencode 1.18.7 (empty for every agent incl. Hephaestus/Momus); the merged-config endpoint is the authoritative surface for it and was used instead.

Generated: 2026-07-29T11:37:30.801487