claude-vibe-squad
CLAUDE.md220 starsChanged 16 days agoMITView on GitHub
Vibe Squad is markdown-first: The gemini lane name is historical: it dispatches through the agy CLI, which covers Gemini and Google's other products under OAuth. The separate API-key gemini CLI is retired and nothing here invokes it. Do not "fix" a gemini reference by pointing it at that binary. Chrono is the only controller and the only operator-facing voice. Model leads execute scoped markdown task packets. Source namespaces under departments/ locate specialist/role markdown only; they do not determine model choice.
In this repository179 other files it gives its agents.
- AGENTS.mdAGENTS.md
- chrono/CLAUDE.mdCLAUDE.md
- accessible-media-authoringSkill
- agentic-safety-auditSkill
- agent-prompt-engineeringSkill
- audio-event-map-authoringSkill
- auto-scaffoldSkill
- chain-constructSkill
- chain-construct-smart-contractSkill
- claim-verificationSkill
- code-reachability-auditSkill
- code-review-loopSkill
- color-theorySkill
- conversation-designSkill
- copy-refinementSkill
- cosmos-sdk-audit-checklistSkill
- cross-chain-bridge-auditSkill
- cross-file-relationship-synthesisSkill
- data-flow-traceSkill
- dedup-prior-art-checkSkill
- defi-invariant-checkSkill
- dependency-cycle-auditSkill
- dependency-health-triageSkill
- detection-as-codeSkill
- diff-aware-semgrep-scanSkill
- differential-reviewSkill
- dimensional-analysis-checkSkill
- dual-level-retrievalSkill
- durable-nonce-exploitationSkill
- erc1271-revert-data-checkSkill
- error-based-sstiSkill
- evm-audit-flowSkill
- figma-implement-designSkill
- findings-filterSkill
- forensic-timeline-authoringSkill
- game-design-fundamentalsSkill
- game-mechanics-balancingSkill
- head-tailSkill
- incident-response-runbookSkill
- interactive-audio-designSkill
- interface-ambiguity-checkSkill
- keyword-clusteringSkill
- knowledge-base-integrationSkill
- known-advisory-backport-checkSkill
- layered-analysis-loopSkill
- level-design-patternsSkill
- locale-adaptationSkill
- multi-stance-audit-fanoutSkill
- narrative-structureSkill
- osint-platform-auditSkill
- parser-differential-route-confusionSkill
- persuasive-technical-writingSkill
- platform-complianceSkill
- player-engagement-psychologySkill
- pre-audit-threat-modelSkill
- program-rubric-lookupSkill
- read-only-reentrancy-checkSkill
- requirements-elicitationSkill
- review-severity-ladderSkill
- rule6-rights-gateSkill
- The first 60 shown.
The file
We split this CLAUDE.md into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 48
No mode or external action starts without explicit operator consent.
- Distinctive: line 54
Write shared state atomically with temp + fsync/sync + rename.
- Distinctive: line 55
Verify before claiming done.
≈ “Common” is an estimate: at least 1 in 5 of the 26,980 CLAUDE.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Vibe Squad Root Instructions
Vibe Squad is markdown-first:
Operator -> Chrono -> gpt-codex | claude | gemini | grok -> specialistsThe gemini lane name is historical: it dispatches through the agy CLI, which covers Gemini and Google's other products under OAuth. The separate API-key gemini CLI is retired and nothing here invokes it. Do not "fix" a gemini reference by pointing it at that binary.
Chrono is the only controller and the only operator-facing voice. Model leads execute scoped markdown task packets. Source namespaces under departments/ locate specialist/role markdown only; they do not determine model choice.
Canonical Sources
- Routing:
shared/specialist-runtime-map.tsv - Dispatch protocol:
shared/protocol.md - Runtime rules:
shared/routing.md - Mode workflows:
shared/modes/*.md - Specialist behavior:
departments/*/specialists/*.mdandshared/specialists/*.md - Model lead prompts:
model-lanes/* - Durable memory: private markdown vault via
chrono-vault(record/recall); seeplugins/chrono-vault/README.md - Tool/skill triggers (when to reach for what):
docs/standards/tool-trigger-map.md - Operator-facing output (boxes carry the content; prose stays short):
docs/standards/operator-facing-output-standard.md— internal-only; withheld from the public export, with no public equivalent shipped
Generated adapters, stale handoffs, old specs, and runtime logs are not source of truth. The chrono-vault record/recall loop is the durable cross-session learning store; the legacy in-repo KG SQLite and the recall LIKE-stub it replaced are retired. The chrono-kg MCP namespace remains live as a compatibility alias backed by those canonical Markdown-vault operations; it does not revive SQLite.
Versions — three numbering systems, all real
Confusing these is the single most common source of "is this doc out of date?" in this repo.
V2.0.0is the current release version. Git tags runv1.1.0throughv1.1.6;v1.1.7onward are tagged by Chrono on the PUBLISHED public commit as each release is cut, so they are disjoint from privatemain—git merge-base --is-ancestor v1.1.8 HEADcorrectly returns false. The next patch would beV2.0.1. This is the version of the system as a product. Release v2.x is unrelated to the V3/V4 architecture generations.V3/V4are architecture generations, not old release numbers. Git tagsv3-finalandv4-baseline-2026-08-07. The dispatcher still runs a live V3 compatibility bridge, soshared/protocol.mdsaying "the staged V4 boundary" is current, not stale. Renaming these toV1.1.1would destroy a real distinction.contract/v1,dispatch-preflight/v1and friends are schema identifiers that code parses. Never touch them.
Bounty mode's internal v2 → v3 (shared/modes/bounty.md) is a fourth, mode-local axis
describing why that mode dropped its gates. Also not a release number.
Hard Rules
- No mode or external action starts without explicit operator consent.
- Chrono chooses mode, specialist, write scope, model, and review gate.
source_namespacechooses the specialist (role) markdown location, not the mailbox — every task is transported through the canonical mailbox root (departments/coding);to_modelchooses the model/CLI that runs it. (Home:shared/protocol.md§ board-native transport, enforced bydispatch_context_builder.pyCANONICAL_MAILBOX_ROOT.)- Model leads do not talk to the operator directly.
- Reviewers are read-only unless Chrono serializes a later write packet.
- These held categories require explicit operator approval as policy:
cleanup,credential_change,delete,live_outreach,malware_detonation,offensive_execution,paid_media,production_mutation, andpublic_release.production_mutationmeans mutating a live production system that is not itself a public release (operator-ratified 2026-07-13). Ordinary workers are not given declared authority for these held categories: the supervisor denies that authority at admission rather than asking at action time. This admission check does not itself constrain later tool calls; deletion has a separate Git-integration gate.scripts/python/tests/test_held_action_gate.pypins this policy list to the controller constant and the canonicalshared/lane-policy.tsvvocabulary. Seeshared/protocol.md§ Held-category authority and logical scopes for the enforced boundary. - Write shared state atomically with temp + fsync/sync + rename.
- Verify before claiming done. No fabricated citations or unverifiable provider claims. Vendor-provided benchmark numbers may be cited as vendor claims, but may not be used as planning assumptions until reproduced on at least one Vibe Squad-controlled benchmark or explicitly labeled unverified.
- Capability is proven by a live probe, never by a config file. Declared (capability source) ≠ delivered (generated adapter) ≠ actual (what the runtime exposes) — only actual counts. Before any routing/architecture decision that rests on what a lane or specialist can do, dispatch a bounded probe that reports the literal command and its literal result. Agreement between documents that share an origin is not corroboration. A capability is not available until a real board dispatch demonstrates it (2026-07-30: three config sources agreed a lane had no shell; the probe found shell plus 42 working tools).
- One fact, one home. When you copy a file, or restate a fact that already lives somewhere, either delete the original in the same change or record which copy wins and what keeps the copies in agreement. An unenforced copy is not a backup — it is a second answer that ages independently, and the next reader cannot tell which one is true. A duplicate is legitimate only when a validator enforces the identity and a named file states the winner. Duplication accumulates from caution, not carelessness — a backup before an edit, a mirror before a cutover, a tracked copy so a worker could read it — which is why it hides (2026-08-10: nine plan files, four claiming to describe current work and three of those wrong; a standard copied to a reachable path with the unreachable original left tracked).
Session Resume
There is a single resume contract, and it is Chrono's to execute — no board worker runs it.
At session start Chrono regenerates, then reads the bounded capsule _state/chrono/resume.md
(one step, never separated) as the primary source. The capsule itemises live and deferred work
with a loud marker for any status the partition does not know, so nothing owed is ever silently
absent (canary-tested). A capsule read without regenerating is stale by construction; never
bulk-read the multi-MB _state/active-tasks.json monolith — the capsule extracts the live slice, and
chrono/current.md is an archive, not a resume source.
- Operational steps (regenerate command, read order, direct-query filter):
chrono/CLAUDE.md§ Start Of Session. - How the contract was settled, and what its two rejections cost:
chrono/resume-contract-decision-record.md.
docs/handoffs/, old plans/specs, and _state/*report* files are historical unless current state
references them.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

