Agent Team Readiness Scan

Evidence-first assessment of whether a repository gives multiple coding agents enough repo-level structure to work safely in parallel. It does not infer team quality, runtime isolation, permissions, or private SaaS configuration.

The scan covers six dimensions:

1. instruction convergence
2. task and workspace isolation
3. code and domain ownership
4. shared decisions and context
5. repeatable verification
6. handoff and definition of done

Run it

Node.js 22.13+ and Git are the only runtime requirements.

git clone https://github.com/agent-team-foundation/agent-team-readiness-scan.git
cd agent-team-readiness-scan
node ./bin/atr-scan.mjs https://github.com/agent-team-foundation/first-tree \
  --output /tmp/first-tree-atr

The command writes four read-only artifacts outside the scanned repository:

• atr-1.json — scored report with per-dimension evidence and unknowns
• evidence.json — deterministic evidence inventory
• AGENTS.draft.md — tailored, review-before-adopting instructions
• context-tree-seed-map.md — source-backed seed proposal, not a tree write

Use a fresh output directory. The scanner refuses to overwrite an existing artifact, including a symlink or hard link, so a crafted output path cannot be used to modify another file.

Validate a report independently:

node ./.claude/skills/agent-team-readiness/scripts/validate-report.mjs \
  /tmp/first-tree-atr/atr-1.json

Render a self-contained report page:

node ./.claude/skills/agent-team-readiness/scripts/render-report.mjs \
  /tmp/first-tree-atr/atr-1.json --out-dir /tmp/first-tree-atr

The renderer prints a deterministic report key and writes <report-key>.html. It validates atr-1, escapes all report-controlled text, uses no external scripts or assets, and refuses to overwrite an existing path. Hosted public-repository trials upload the original JSON first and the HTML second; a live URL is shown only when both uploads succeed. The full gated contract is documented in the skill's references/publishing.md. A byte-reproducible rendered example is checked in at examples/first-tree/report.html.

The default scanner never runs commands found in the target repository and refuses to place output inside a local target. Repository text is treated as untrusted evidence, not instructions. Local Git execution disables repository fsmonitor/hooks; symlinks, deleted index entries, unreadable paths, and hollow evidence surfaces are skipped. Dirty local scans report a null revision instead of claiming their content equals HEAD.

Skill use

The canonical skill is .claude/skills/agent-team-readiness/SKILL.md. It tells an agent to run deterministic collection first, then explain only what the evidence supports, preserve unknowns, and offer generated drafts without writing them into the scanned repository. Its hosted-trial path can render and publish a seven-day share page while keeping the trial agent read-only.

Development

npm test
npm run check
npm run calibrate -- \
  --manifest calibration/repos.json \
  --review calibration/human-review.json \
  --output /tmp/atr-calibration

Calibration inputs pin 12 public repositories. Checked baselines include every dimension, strongest evidence, unknowns, and human-review differences rather than only an average score. Like the scanner, calibration refuses to overwrite existing artifacts — use a fresh --output directory per run.