squads

Agent Squads for Pancake — installable bundles that deploy one or more proactive sub-agents into a Pancake pod in a single command.

This repo is the home of Pancake's official squads, the source the Pancake marketplace seeds its catalog from, and the canonical public reference for anyone building their own squad.

What's in here

• squads/ — the official squads (one directory per squad bundle).
• template/ — a complete, valid skeleton bundle to copy from.
• docs/ — the full contract and authoring guides.
• scripts/validate.mjs — a zero-dependency validator that mirrors marketplace ingestion.
• scripts/test-validator.mjs — self-tests for the validator: builds known-bad bundles in a temp dir and asserts each is rejected with the expected error. CI runs it alongside the validator so a weakened check turns red.
• manifest.schema.json + agent.schema.json — JSON Schemas for editor validation of manifest.json and per-agent agent.json.
• .claude/skills/ — Claude Code skills to author and validate squads.

Official squads

Each squad publishes a catalog of outcome-typed workflows and reports only through the company task board — the co-founder delegates by matching intent to a workflow, and relays.

Squad	What it does	Agents	Workflows
ai-seo-squad	AI-SEO / GEO — daily citation audits, blog posts, and GEO engineering PRs (self-merging).	geo-agent	seo.audit_citations, seo.write_article, seo.ship_technical_fix, seo.weekly_report
community-squad	Community presence — Reddit channel: monitors subreddits, drafts value comments, runs a multi-account karma strategy.	reddit-agent	reddit.scan_and_draft, reddit.monitor_keywords, reddit.account_health, reddit.post
outreach-squad	Daily outbound — finds leads, runs sequences, handles replies, files a digest on the board.	outreach-agent	outreach.run_campaign, outreach.find_leads, outreach.triage_replies, outreach.weekly_report
paid-ads-squad	Paid advertising — Google Ads + Meta Ads operators, each running one account end to end; holds spend flat, surfaces budget raises on the board.	google-ads-agent, meta-ads-agent	google.optimize_account, google.daily_digest, google.root_cause, google.scale_budget, meta.daily_operations, meta.daily_digest, meta.weekly_review, meta.investigate
product-squad	Product management & discovery — triages Notion product ideas into researched PRDs with build/reject verdicts; harvests daily user feedback from configured sources into a Notion feedback DB.	pm-agent, feedback-agent	product.triage_ideas, product.write_prd, product.harvest_feedback
eng-squad	Engineering squad — GitHub issue triage (classifies issues P0–P3, sweeps daily, reports weekly) plus end-to-end implementation of specified ideas/issues/PRDs as reviewable PRs.	triage-agent, build-agent	eng.triage_issue, eng.sweep_open_issues, eng.weekly_report, eng.implement_idea
analytics-squad	Product analytics (PostHog-backed) — daily DAU/WAU/MAU + north-star digest, activation-funnel debugger, weekly recap, ad-hoc reports.	analytics-agent	posthog.daily_digest, posthog.weekly_recap, posthog.debug_funnel, posthog.adhoc_report

How squads work

An Agent Squad installs proactive specialist sub-agents — their identity, skills, crons, and seed memory — under a pod's main co-founder agent, with no manual file editing. Start with docs/how-squads-work.md for the concepts and the install lifecycle.

Build your own

Read docs/creating-a-squad.md, then either:

• In Claude Code: run the create-squad skill — it interviews you, scaffolds the bundle, and validates it.
• By hand: copy template/ and fill it in.

The exact file contract is in docs/bundle-reference.md.

Validate

node scripts/validate.mjs                      # every squads/* bundle and template/
node scripts/validate.mjs squads/<bundle-name> # one bundle

Zero dependencies — just Node. It mirrors marketplace ingestion exactly, so a bundle that passes here passes ingestion. CI runs it on every push and pull request, alongside node scripts/test-validator.mjs which self-tests the validator against negative fixtures (forbidden files, wrong heartbeat shape, etc.) — both must pass for a merge.

Replay-eval workflows

scripts/validate.mjs checks what a bundle declares. scripts/eval.mjs checks what each workflow actually does at the squad↔board contract level — the qualified id the cofounder stamps, the tool calls the squad agent makes, the terminal it closes with, the digest it writes. No LLM is invoked; this is the deterministic tier between schema validation and live-LLM e2e.

node scripts/eval.mjs                                  # every recorded trace
node scripts/eval.mjs squads/<bundle-name>             # one bundle
node scripts/eval.mjs path/to/case.trace.json          # one trace

Traces live at squads/<name>/evals/replay/<workflow-id>/*.trace.json. Every official workflow ships a happy-path.trace.json; recorded production bugs are committed as negative-case traces (expected: "FAIL" + expected_failures: [...]) so a future refactor can't silently un-catch them. Full per-bundle docs in squads/<name>/evals/README.md.

CI runs node scripts/eval.mjs alongside the validator on every push/PR — both gates must pass.

Publish

See docs/publishing.md.

• Official squads (Pancake team) are added to this repo as a squads/<name>/ directory.
• Third-party squads are self-hosted: external authors keep their bundle in their own public GitHub repo and submit the URL to the marketplace. This repo is Pancake-curated and not open to outside squad PRs — it is an example to copy, not a place to contribute.

Repo layout

squads/                          ← this repo
├── README.md
├── LICENSE                       MIT
├── CLAUDE.md                     orientation for Claude Code sessions
├── CONTRIBUTING.md               curation policy
├── manifest.schema.json          JSON Schema for manifest.json
├── agent.schema.json             JSON Schema for agents/<id>/agent.json
├── .github/                      CI workflow (validator + replay evals) + PR template
├── .claude/skills/               create-squad, validate-squad
├── docs/                         how-squads-work, bundle-reference, creating-a-squad, squad-scorecard, publishing
├── lib/eval-runner.mjs           shared replay-eval check engine, zero deps
├── scripts/validate.mjs          zero-dependency validator
├── scripts/eval.mjs              replay every squad's recorded *.trace.json
├── template/                     a complete, valid skeleton bundle
└── squads/                       official squad bundles, one directory each
    └── <name>/evals/             happy-path + negative-regression traces per workflow

License

MIT — Copyright (c) 2026 Basalt AI.