agent-scripts

Canonical home for generic agent policy skills, routers, and workflow recipes, spun out of ohyeh/tmux-agent-tools (which retains only tmux worker lifecycle mechanics and its narrow router).

This repository is the canonical home (ADR-0001) for the fleet's generic agent policy skills, routers, workflow recipes, and shared rule files; machines are deployed copies only. Built incrementally per a frozen, second-model-reviewed implementation plan. Key contents:

• skills/using-design-skills/ — the Gate R2 CLEAN-reviewed design router (repaired, 16 regression cases in evals/evals.json).
• skills/using-workflows/ — the workflow meta-router and its 12 canonical recipes (skills/using-workflows/workflows/), migrated from ohyeh/tmux-agent-tools. This bundle is the single canonical copy; machines deploy it to ~/.claude/workflows/ via the install script.
• skills/delegation-templates/, skills/unknowns-discovery/, and skills/shared-memory-intake/ — generic policy skills. These are canonical skill directories; the .agents/rules/ exclusion below covers only same-named rule Markdown files, not skills.
• .agents/rules/ — the shared routed-rule files (model-dispatch.md, judgment-rubrics.md, maintenance.md, session-titles.md, harness-diagnosis.md, agent-environment-provisioning.md, LETTER-TO-FUTURE-SESSIONS.md). delegation-templates.md and unknowns-discovery.md are deliberately not carried over as rule files — their skills above are canonical instead. lessons.md stays machine-local and is git-ignored; it is never committed here.
• scripts/scrub.sh — the pre-push secret/path/hostname/Tailscale-IP/commit-metadata scrub that must PASS before any push to this repo's remote.

Release channel: immutable tag (primary), protected main (fallback), per the frozen ADR governing this spinout's repo boundary, release policy, and per-skill fleet cutover invariant.

Local execution-frontier check

Use the zero-dependency harness to compare observable execution outcomes without conflating single-model reasoning effort, repeated sampling, and multi-agent topology:

node scripts/execution-frontier.mjs \
  --fixture evals/execution-frontier/fixtures/local-smoke.json \
  --output-dir .workflow/execution-frontier-local
node scripts/test-execution-frontier.mjs

The runner writes per-attempt JSONL plus a JSON summary. Case summaries report assertion_pass_rate, not model task success. Provider-only metrics such as hidden reasoning tokens remain explicitly unavailable; the harness does not estimate them. A fixture may name one comparison axis, and validation fails if its profiles differ on any other axis. Every comparison profile must execute the same workload_id with the same command and expectations; unused profiles and mismatched workloads fail validation. Commands receive EXECUTION_PROFILE_ID, EXECUTION_REASONING_EFFORT, and EXECUTION_PROFILE_JSON; the sample fixture verifies this plumbing but does not invoke a model.

Deploy (clone-free, one command)

Canonical deploy is a single raw-GitHub bootstrap — no clone, no working copy left on the machine. It downloads the current main tarball to a scratch dir and deploys all four runtime layers (global files → ~/.claude/~/.codex, rules → ~/.agents/rules/, workflow recipes → ~/.claude/workflows/, and the skills-lock.json skill set), each layer printing PASS/FAIL with hash/diff evidence and aborting fast on the first failure:

curl -fsSL https://github.com/ohyeh/agent-scripts/main/scripts/deploy.sh | bash

The skills-only restore below is Layer 4 of that script, kept for the skill-set-only case.

Fleet skill restore (skills-lock.json)

skills-lock.json at the repo root is the fleet's canonical skill set (101 entries, snapshot 2026-07-27, union contract — every machine converges to this same set): every skill installed on the fleet that the npx skills CLI manages, merged from the machines' ~/.agents/.skill-lock.json files. Two manual-only items are deliberately excluded: commit-commands (a Claude Code plugin, not a skill) and note (hand-copied, no tool-resolvable source).

Any skill named in the global rules belongs in this lock — otherwise the rule points at something a freshly deployed machine does not have. stop-slop was that gap until 2026-07-27: CLAUDE.md loads it for writing-heavy work, but it survived on the fleet only as a hand-copied folder.

To restore all skills on a (new) machine — the CLI restores into .agents/skills/ relative to the cwd, so run it from $HOME for a global install:

cd ~
curl -fsSL https://github.com/ohyeh/agent-scripts/main/skills-lock.json -o skills-lock.json
npx -y skills experimental_install
rm skills-lock.json

Format verified against the CLI (one-entry restore, exit 0). The $HOME-cwd global-install path is inferred from the CLI's cwd-relative behavior — UNCONFIRMED on a fresh machine until first used. Before running npx skills update/upgrade -g afterwards, clear ghost lock records (entries whose folder no longer exists) or deleted skills resurrect.