multi-harness-research

name	multi-harness-research
description	Activate only when the user explicitly invokes `/omg:multi-harness` or explicitly names the `multi-harness-research` skill. It sends one confirmed read-only research task directly to four fixed harnesses; ordinary research, review, comparison, or model-delegation requests never auto-activate it.

Multi-harness research

This is an explicit, read-only research surface. It directly orchestrates four independent harnesses; it does not create a GJC team, task, goal, session, worktree, or worker model.

Locked intent

• artifact:research-docs — publish four fixed-position lane documents and a factual summary.md; failed lanes keep a fixed placeholder and ledger entry.
• surface:explicit-command — only an explicit multi-harness-research skill request or /omg:multi-harness invocation may run it.
• integration:four-harnesses — run exactly the four lanes below, in order.
• constraint:read-only — workers can inspect the canonical target only; only the project-external XDG artifact directory is writable.
• constraint:same-task — every lane receives byte-identical normalized task bytes plus the same safety/output suffix, recorded by one SHA-256 digest.
• constraint:single-suite — all active files and runtime ownership use oh-my-gajae-code; no additional marketplace plugin is involved.

Activation and task confirmation

An explicit argument is the canonical task. Without one, preview a one-sentence goal, the research questions, and expected output; obtain confirmation before running. Do not infer a task from ordinary conversation, start a background run, or silently split one request into multiple tasks.

Use the direct multi-harness-research.mjs runner installed by the native user-scope installer. Execution requires its trusted mode-0600 binding and private runtime snapshot at ~/.gjc/agent/runtimes/multi-harness-research/. A project-scope installation or a project-local runner never authorizes execution. Missing, malformed, changed, or unsafe runtime state fails closed; do not install, update, migrate, set up, or log in on the user's behalf.

Fixed lanes and isolation

The runner fans the one normalized task to exactly these lanes in this order; it never adds a fifth model, substitutes a selector, retries through a fallback provider, or asks for a winner, majority, vote, consensus, ranking, recommendation, or final verdict.

1. gjc-opus — GJC 0.11.x with --model anthropic/claude-opus-4-8 --thinking max.
2. gjc-sol — GJC 0.11.x with --model openai-codex/gpt-5.6-sol --thinking xhigh.
3. codex-sol — Codex CLI with --model gpt-5.6-sol, -c 'model_reasoning_effort="xhigh"', and exec --ephemeral.
4. claude-ultracode — Claude Code with -p --no-session-persistence --effort ultracode.

Linux and an executable bwrap are required. The runner binds the canonical target read-only at the same path and gives every lane private HOME/XDG/TMP state. It exposes neither the target's .gjc, Git mutable state, host home, nor a target snapshot. Workers cannot edit, delete, chmod, create symlinks, write notebook data, run browser automation, use MCP/hooks/extensions, or access skills/rules.

GJC lanes have GJC_NOTIFICATIONS=0 GJC_SDK_DISABLE=1, --no-session, --no-extensions, --no-skills, and --no-rules; their allowlist is read/search/find plus provider-native public web/package-metadata tools. They have no built-in Bash, write, or edit capability. Claude's allowlist is exactly Read,Grep,Glob,WebSearch,WebFetch; it has no Bash, Write, Edit, Notebook, browser, MCP, hook, plugin, or persistence capability. Codex uses a private ephemeral state, read-only workspace profile, provider-native web search only, and shell_network=false; it inherits no shell environment and has no MCP/apps/hooks/browser/plugins. Codex receives no OMO prompt or resolution. Current Codex OAuth 401 is pending-environment, not a successful lane or a reason to fall back.

Credential-bearing input is limited to these read-only regular files after canonical, owner/mode/link, descriptor-identity, and no-symlink validation. There is no credential discovery, parent-directory bind, broad home bind, alternate layout, or writable refresh path.

Lanes	Exact host source	Private sandbox destination
gjc-opus, gjc-sol	${XDG_DATA_HOME:-$HOME/.local/share}/gjc/auth.json	the identical canonical path in the lane-private HOME/XDG view
codex-sol	${CODEX_HOME:-$HOME/.codex}/auth.json	$CODEX_HOME/auth.json in private CODEX_HOME
claude-ultracode	$HOME/.claude/.credentials.json	$HOME/.claude/.credentials.json in private HOME

Artifacts, lane truth, and output limits

The orchestrator alone atomically publishes new artifacts under ${XDG_DATA_HOME:-$HOME/.local/share}/oh-my-gajae-code/multi-harness/<repo-id>/<run-id>/; the fallback is $HOME/.local/share. Artifact directories are mode 0700, files are mode 0600, creation is no-follow and exclusive, and publication uses same-directory fsync plus atomic rename. Raw tasks, raw child streams, credentials, tokens, auth state, and target snapshots are never persisted.

Preserved historical artifact data: the former ${XDG_DATA_HOME:-$HOME/.local/share}/oh-my-gjc/multi-harness/<repo-id>/<run-id>/ root is read-only historical data. Never write, migrate, or clean it.

Each lane is independent. Valid Markdown is non-empty UTF-8, no more than 1 MiB, secret-clean, and contains ## Conclusion, ## Evidence, and ## Uncertainties with a verifiable path+line or external URL. Raw stdout and stderr each stop at 8 MiB; output is never truncated into success. The closed lane error classes are preflight, spawn, timeout, nonzero_exit, invalid_output, and contract_breach. A failed lane never cancels a peer or discards a valid peer document.

After all four lanes reach terminal state, phase 1 seals the factual base summary. Four valid lanes produce COMPLETE with runner exit 0; mixed valid/failed lanes produce INCOMPLETE with exit 10; no valid document or a run-level fatal produces exit 1. The bounded user surface may show only overall lane state, per-lane status/error class, absolute summary/success paths, and a short comparison—never full documents, raw stderr, task bytes, credentials, or a nonce.

Two-phase comparison

The factual base is sealed first with immutable lane facts, fixed-order placeholders, failure ledger, comparison_status: pending, task digest, and final phase-1 exit. Only after that does the current GJC leader read successful lane documents and author a bounded, explicitly non-authoritative list of commonalities, differences, and uncertainties. The leader does not change lane facts and is not a fifth synthesis model.

The leader sends only that bounded prose and the protected one-use receipt path to the no-model finalize-comparison runner mode through protected mode-0600 stdin/env transport. The runner reads the receipt itself; never put the task, comparison, receipt nonce, credential, or token into shell source, argv, logs, or chat. The finalizer revalidates the nonce digest, base digest, immutable facts, held summary descriptor, and current summary path identity immediately before atomic publication. It may replace only the comparison placeholder and status.

Finalizer exit 0 is FINALIZED. Its independent exit 20 is FINALIZATION_FAILED with only one bounded class: leader_input_invalid, authorization_failed, base_changed, or publication_failed. It does not reinterpret finalizer failure as a lane failure or change the sealed lane outcome, run exit, lane artifacts, or factual base summary. A controlled retry is permitted only with the same unconsumed receipt and byte-identical base.

Never do

• Do not use gjc team, worktrees, arbitrary shell brokers, a fifth model, winner selection, consensus, voting, ranking, recommendation, or fallback.
• Do not install, update, migrate, configure, set up, or log in to any provider or CLI.
• Do not permit target writes, target .gjc access, mutable Git access, built-in Bash in GJC or Claude lanes, Codex shell networking, or credential/config directory mounts.
• Do not claim the Codex 401 pending-environment condition passed.