phrazzld/roster

/roster

You are not the only frontier agent on this machine. These CLIs are installed and each runs headlessly. They are options, not obligations: native subagents remain the default delegation path, and a peer harness earns a lane only when you can name what it adds — usually a decorrelated model family or a genuinely fresh context.

When a peer harness beats a native subagent

• Adversarial critique of your own work. A reviewer from a different model family has decorrelated failure modes and no loyalty to your reasoning. Give it the artifact (diff + oracle) only — never your reasoning trail.
• Second opinion on a contested judgment — architecture call, risk assessment, "is this idiomatic" — where one model's taste shouldn't decide alone.
• Competing attempts at the same bounded problem, graded blind.
• Wide bench — high-stakes adversarial review (a P0, a security or data-loss surface, a pre-ship "be exhaustive" pass). Fan the artifact across several distinct open-model families via Pi/OpenRouter, each critic on its own lens; distinct families surface distinct real findings instead of confirming each other. See Adversarial bench below.

A native subagent is still better for exploration, scoped builds, and anything where harness identity doesn't matter — it shares your tools, needs no cold start, and the orchestrator is trained on it.

The CLIs

Verified installed and probed 2026-06-14. Each row is the headless form; add the prompt as the argument or via stdin.

| CLI | Stack | Headless invocation | |---|---|---| | codex | OpenAI Codex (gpt-5.5) | codex exec "<task>" (--model, --config model_reasoning_effort=) | | pi | Pi over OpenRouter (Kimi, DeepSeek, …) | pi -p --no-extensions --provider openrouter --model <id> "<task>" | | goose | Goose over OpenRouter | goose run --no-session --quiet --provider openrouter --model <id> --text "<task>" | | opencode | OpenCode over OpenRouter | opencode run --model openrouter/<id> --format json "<task>" | | claude | Claude Code (Opus/Fable) | claude -p "<task>" (--model, --effort) | | cursor-agent | Cursor (composer) | cursor-agent -p "<task>" | | grok | xAI Grok | grok -p "<task>" (--model, --effort) | | agy | Antigravity (Gemini) | agy --print "<task>" | | hermes | Hermes agent | hermes -z "<task>" (-m <model>) | | oracle | Oracle browser consult | npx -y @steipete/oracle --engine browser --model gpt-5.5-pro -p "<task>" --file <paths> | Current model ids, pricing, context windows, and freshness dates: references/model-provider-harness-index.md. Open-model facts rot in days — check the review-due date before quoting them.

Adversarial bench

High-stakes review is a model-family spread, not one second opinion: distinct families each catch distinct real bugs (in one P0 pass, one family found a startup-bricking blocker, another a null-timestamp escape, another a write-path gap — no overlap).

• 3–5 distinct families (Kimi/Moonshot, DeepSeek, MiniMax, Qwen, GLM/Zhipu when listed). Same-family variants don't decorrelate — a wide bench of one family is waste.
• Discover slugs live: pi --provider openrouter --list-models <family>. Slugs rot in days; substitute when a family isn't listed.
• One distinct lens per critic (correctness/data-loss, durability, security, perf, API-shape — harnesses/shared/references/lenses.md). Optionally load that lens or a domain skill into the critic: pi --skill <path>.
• Cold, bounded, artifact-only: inline diff + oracle + context, run pi -p --no-extensions --no-tools --provider openrouter --model <slug> "<prompt>", background the lanes, synthesize the verdicts yourself. Add one or two native critics on their own lenses (a further family), then re-review the delta after you fix.

Scale to stakes — a routine diff gets one well-aimed cross-family critic, not the full bench.

Judgment

• One well-aimed critic beats three vague ones. Aim it at the claim that would embarrass us in production, with explicit "ignore style/naming" bounds.
• Peer output is evidence, not authority. You weigh it, accept or reject it, and own the result.
• A failed or rambling lane is a result too — report it, don't re-roll silently.
• Oracle is browser-mode only in Harness Kit. Use it for high-context consults through a signed-in ChatGPT session; do not use Oracle API mode or multi-model API panels from the Harness Kit roster default.
• For a bounded lane whose evidence should outlive the session, harness-kit-checks dispatch-agent wraps the invocation and writes a delegation receipt (.harness-kit/traces/delegations.jsonl). Optional for quick second opinions; useful when the lane's verdict feeds a ship decision.
• Heavy, long-running, or isolation-needing lanes run on sprites (/sprites) regardless of which CLI executes them.

Gotchas

• Peer CLIs run cold: no conversation history, no local skills unless the harness loads them itself. Inline everything the lane needs.
• Match the bench to the stakes (see Adversarial bench). A routine second opinion is one well-aimed cross-family critic; fanning wide there is theater. The real waste is a wide bench of the same family — variants don't decorrelate — not a wide bench of distinct ones, which a P0 or security/data-loss surface earns.
• Auth rots independently per CLI. A lane failing instantly with an auth error means re-login locally, not a provider verdict.

Prompting frontier lanes

For Fable/GPT-5.5-class lanes (and the lead's own operation), load harnesses/shared/references/prompting-frontier.md: goal-not-steps briefs fenced by house rules, executable bars (delegate metric-invention when the bar is fuzzy), builder-never-grades verification against the REAL output, loop-until-the-bar with a live status artifact, prior traces as fuel, and budgets instead of permission-asks. Cheaper models still earn more mechanism — tier the prescription to the model.