A multi-harness plugin from the not-my-job marketplace. Primary host: Claude Code. Also designed to work when the skill is loaded from Codex, Cursor CLI, Grok Build, OpenCode, or agy.

What it does

Delegates a task to peer AI CLIs in parallel. Second opinions on architecture, research across model families, security audits, design reviews.

Grounding (Skeptical Colleague + Goal Cards):

• Goal Cards at .braintrust/goal-cards/<slug>.md with a session_anchor
• 6-step protocol: restatement, assumptions, evidence, fidelity, honesty, GROUNDED / NOT GROUNDED
• Host actively curates context before delegating (not "whatever is in the chat")

Members (v1.15)

Slot	CLI	Notes
Anthropic	Claude	braintrust:peer agent (Opus 5.5, 1M) inside Claude Code; claude -p --model 'claude-opus-5-5[1m]' --safe-mode --settings '{"disableAllHooks":true}' from other hosts
Google	agy only	No Gemini CLI. Strongest Gemini High from agy models (gemini-4-argon-high as of 2026-10), falling back to the newest Flash High (gemini-3.8-flash-high); --output-format json. Loads user hooks (no switch)
OpenAI	Codex	gpt-6-astra (GPT-6 Astra; fallback gpt-6.1-sol); isolated CODEX_HOME + --ignore-user-config --ignore-rules; CLI 0.160.0 verified
xAI	Cursor CLI, else Grok CLI	Cursor: newest grok-X.Y-high (grok-4.7-high today), --mode ask. Grok CLI (optional fallback): grok models default, isolated GROK_HOME
Multi	OpenCode	User's configured/default model (expected zai-coding-plan/glm-5.3); --variant max for GLM-5.3; --pure

Up to five independent voices when everything is installed and authenticated. Availability is decided by scripts/bt_probe.sh, not by preference.

Every peer runs identity-isolated. The probe builds $TMPDIR/bt-codex-home and $TMPDIR/bt-grok-home (auth only), Cursor runs read-only (--mode ask, cmux hooks off, no MCP approval), OpenCode runs --pure, and the Claude peer skips CLAUDE.md and user hooks. The claude -p path runs --safe-mode, so it gets no CLAUDE.md, MCP servers or plugins. Ambient agent-bus hooks (hcom) in the real homes otherwise hijack headless answers. agy has no isolation switch for its hooks; braintrust watches its answers for bus chatter.

Commands

Command	What it does
/braintrust	Orchestrate a task across peer CLIs
/consult	Alias for /braintrust
braintrust:peer agent	Claude peer voice (Opus 5.5 1M, read-only) used by the Task tool

v1.15.1 Highlights

• Harness refresh dogfooded 2026-10-23: claude 2.1.309, agy 1.4.6, codex 0.160.1, cursor-agent 2026.10.22, opencode 1.18.34. Every pin holds; Gemini 4 Argon is still Google's newest model.

v1.15.0 Highlights

• Google → Gemini 4 Argon High (gemini-4-argon-high). The probe now pins the strongest gemini-*-high slug and falls back to the newest Flash High (bt_agy_fast_model). Argon is ~2.5x slower than Flash High.
• Claude claude -p peer runs --safe-mode. Hooks were already off, but the peer still loaded your CLAUDE.md, ~20 MCP servers and every plugin. Safe mode keeps keychain auth and cuts a one-word consult from ~62k tokens, 10s and $0.41 to ~15.6k tokens, 2.2s and $0.04.
• Harness refresh dogfooded 2026-10-16: claude 2.1.301, agy 1.4.0, codex 0.160.0, cursor-agent 2026.10.14, opencode 1.18.34. Pins unchanged elsewhere: gpt-6-astra, claude-opus-5-5[1m], grok-4.7-high, glm-5.3.
• New "never pass" flags documented (agy --remote-control, Codex --dangerously-bypass-hook-trust/--approve-for-me/-p, Cursor --worktree/--plan) plus agy's Gemini effort ceiling and OpenCode's glm-5.3-highspeed 429 retry loop.

v1.14.0 Highlights

• Harness refresh dogfooded 2026-10-07: claude 2.1.292, agy 1.3.1, codex 0.160.0, cursor-agent 2026.10.01, opencode 1.18.34.
• Codex fallback → gpt-6.1-sol (new catalog priority 1). Astra stays primary as the frontier model.
• agy hooks documented. hcom and herdr hooks now load from ~/.gemini/config/hooks.json; agy has no switch, so consults watch for bus chatter.
• cli-contracts.md gains a Cursor contract and current versions; all references agree on Opus 5.5 and grok-4.7.
• Doc fallbacks for isolated homes use ${TMPDIR:-/tmp}, matching the probe.
• Grok CLI marked optional and unverified since 1.0.30.

v1.13.0 Highlights

• xAI slot via Cursor CLI (cursor-agent) with the newest Grok High model it lists (grok-4.7-high today). Grok CLI stays as the fallback; the probe picks one (bt_xai_via).
• Cursor uses your normal login and runs --mode ask (read-only, no shell) with cmux hooks off and no MCP approval. Its hcom notice and user rules still reach context; ask mode keeps them from becoming actions.
• Claude → Opus 5.5 1M (claude-opus-5-5[1m]) for braintrust:peer and claude -p.
• Grok CLI fallback default grok-4.7.

v1.12.0 Highlights

• Codex → GPT-6 Astra (gpt-6-astra), fallback gpt-5.6-sol. Probe reports usage-limit errors verbatim (bt_codex_error) instead of "empty/timeout".
• agy → newest Flash High (gemini-3.8-flash-high today); probe auto-selects the newest gemini-*-flash-high slug.
• Claude → Opus 4.8 1M (claude-opus-4-8[1m]) via the new braintrust:peer plugin agent (read-only, no CLAUDE.md) and claude -p --settings '{"disableAllHooks":true}'.
• Grok isolated GROK_HOME (hooks/MCP/compat scanning off) + --no-subagents; fixes hcom agent-bus hijack of headless consults. --no-auto-update replaced by GROK_DISABLE_AUTOUPDATER=1.
• OpenCode --pure verified to drop config plugins; expected model zai-coding-plan/glm-5.3 with --variant max.
• Harness versions dogfooded 2026-09-16: claude 2.1.273, agy 1.2.3, codex 0.154.0, grok 1.0.30, opencode 1.18.31.

Earlier (v1.11)

• Claude consult default opus. agy pin gemini-3.7-flash-high. Grok grok-4.6. OpenCode --variant max when the resolved model contains glm-5.3.
• Probe reads Grok's advertised Default model:. agy --output-format json. Codex JSONL checks error / turn.failed.

Earlier (v1.10)

• Codex primary model: GPT-5.6 Sol (gpt-5.6-sol). Explicit -m because consults use --ignore-user-config.
• Probe prefers Sol, falls back to product default if Sol is unavailable, and records bt_codex_model.
• Documented minimum Codex CLI 0.144.0+ for Sol (npm i -g @openai/codex@latest).

Earlier (v1.9)

• Hybrid always-on skill (eval-backed); Gemini CLI removed; OpenCode user-default model discovery
• Grok default grok-4.5; Codex clean CODEX_HOME + --ignore-user-config
• Multi-host matrix; durable evals under evals/

Requirements

All peers optional; braintrust uses whatever the probe authenticates.

• Claude Code
• Antigravity CLI (agy) — curl -fsSL https://antigravity.google/cli/install.sh | bash
• Codex CLI — 0.160.0 verified for GPT-6 Astra (npm i -g @openai/codex@latest)
• Cursor CLI — curl https://cursor.com/install -fsS | bash, then cursor-agent login (xAI slot)
• Grok Build — optional xAI fallback, curl -fsSL https://x.ai/cli/install.sh | bash
• OpenCode — with an authed provider; set "model" in ~/.config/opencode/opencode.json so headless matches your TUI default

jq required for Codex / Cursor / Grok / OpenCode / agy JSON parsing.

Install

claude plugins install braintrust@not-my-job

Refresh / dogfood / evals

bash scripts/bt_probe.sh
bash evals/run_eval.sh matrix all agy,codex,cursor,opencode
# deeper harness drift notes: skills/braintrust/references/self-improvement.md

License

MIT