fak
CLAUDE.md47 starsChanged 11 days agoApache-2.0View on GitHub
Not the human contributor guide. This file is operating instructions for automated contributors working inside the maintainers' shared checkout. Humans want README.md for what fak is and CONTRIBUTING.md for how to contribute; nothing on this page applies to you. The canonical agent instructions for this repo are in AGENTS.md — read it first for build/test/run, the repo map, and the rules. The six that will bite you if you skip them: - Work directly on the trunk (main). Never open…
- Installs packages
In this repository150 other files it gives its agents.
- AGENTS.mdAGENTS.md
- .github/copilot-instructions.mdCopilot instructions
- llms-full.txtllms.txt
- llms.txtllms.txt
- agent-readinessSkill
- appeal-scoreSkill
- bottleneck-mapSkill
- ci-repairSkill
- claim-repro-scoreSkill
- clear-out-wipSkill
- commit-cleanSkill
- conflation-scoreSkill
- cross-validateSkill
- curate-clusterSkill
- debt-cleanSkill
- debt-orchestratorSkill
- disambiguation-scoreSkill
- dojo-rsi-scoreSkill
- dos-dispatch-loopSkill
- dos-dispatchSkill
- dos-next-upSkill
- dos-replanSkill
- dos-witness-claimSkill
- field-borrowSkill
- fleet-waveSkill
- git-subagent-syncSkill
- goalSkill
- guard-rsi-scoreSkill
- harness-creatorSkill
- harness-gardenSkill
- industry-scoreSkill
- issue-orchestratorSkill
- issue-queueSkill
- issue-triageSkill
- lightgap-scoreSkill
- memory-compactSkill
- milestone-scoreSkill
- modularizeSkill
- negframe-scoreSkill
- opencode-waveSkill
- operator-heaviness-scoreSkill
- persona-scoreSkill
- phased-planSkill
- plan-auditSkill
- quality-scoreSkill
- question-loopSkill
- refresh-cachedoc-numbersSkill
- refresh-readmeSkill
- releaseSkill
- resume-watchdog-auditSkill
- run-it-all-nightSkill
- score-2xSkill
- scorecardSkill
- scout-loopSkill
- signal-firstSkill
- skill-lifecycleSkill
- skill-overlapSkill
- skill-scoreSkill
- slop-scoreSkill
- sota-checkSkill
- The first 60 shown.
The file
We split this CLAUDE.md into the rules it gives an agent
2 rules. About 0 are common, 2 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
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
Use guarded headless agents or equivalent isolated workers for investigation, implementation, tests, and review.
- Distinctive: line 75
Treat unmet targets as empirical data that eliminates a variable and always formulate the concrete next checkable step.
≈ “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
CLAUDE.md
Not the human contributor guide. This file is operating instructions for automated contributors working inside the maintainers' shared checkout. Humans want
README.mdfor what fak is andCONTRIBUTING.mdfor how to contribute; nothing on this page applies to you.
The canonical instructions
The canonical agent instructions for this repo are in AGENTS.md —
read it first for build/test/run, the repo map, and the rules.
The six hard rules
The six that will bite you if you skip them:
- Work directly on the trunk (
main). Never open a feature branch or new worktree — the trunk guard refuses off-trunk commits (OFF_TRUNK). The one sanctioned exception is a detached per-worker worktree that lands its diff back onmainviafak worktree worker prepare|land|reap(build isolation, #1334 / epic #3165): because it is detached (never a branch) and commits only through the serialized land under the worker's lane lease, it is not off-trunk and never tripsOFF_TRUNK. Feature branches and any other off-trunk commit stay forbidden. (Details inAGENTS.md.) - Commit via
fak commit --pathorfak sweepby explicit path — all commits must be made viafak commit --path <p> -m "<subject> (fak <leaf>)"orfak sweep --apply --lane <lane> -m "<subject>".fak commitlocks the lane, stages exactly your paths, provides automatic DCO sign-off (with-saccepted for compatibility), and verifies that no peer files were raced in (PATHSPEC_RACE). End every ship commit's Conventional-Commits subject with a bindable(fak <leaf>)trailer so thedos verifyreferee can bind it — e.g.fix(gateway): treat same-tick ready as positive (fak gateway). A bare un-stamped subject stays NOT_SHIPPED. The/commit-cleanskill mechanizes this rule end to end. WARNING: Raw git commits are strictly an emergency-only fallback, permitted ONLY when thefakbinary is unbuilt (git commit -s -m "<subject> (fak <leaf>)" -- <paths>); never usegit add -Aor uncoordinated raw git commits on the shared multi-session tree. (Full convention inAGENTS.md.) - Default is to ship green work and sync safely unprompted — do more work by default:
pre-flight sync with trunk (
fak sync check,fak sync reconcile --apply, orfak sync apply), verify on-device (fak validate --mine <paths>,go test ./internal/<pkg>/...), commit by explicit path (fak commit --path <p> -m "<subject> (fak <leaf>)"orfak sweep --apply --lane <lane> -m "<subject>"), and push unprompted viafak sync pushorfak commit --push. "Green" requires shift-left proof: for changes touching executable CLI verbs, gateway adapters, runtime logic, or hardware/compute paths, execute real paths in dogfood, integration tests, or live hardware sub-component runs rather than relying on mock-only or shallow tests. Bias heavily toward testing sub-components on live physical hardware (divide and conquer) with high volume and frequency (e.g.fak validate --strix --subkernels=...,make mac-perf,make cuda-test). Safe merge discipline: verify no in-flightMERGE_HEADexists before staging (if active, unstage and wait; never clobber, abort, or finish a peer's merge); usefak sync apply(--ff-only) for clean trunk convergence; route divergence viafak sync reconcile(disjoint integration, superset merge, or dirty parking viafak wip park); never force-push, use--autostash, or perform raw 3-way merges. Dual-repo synchronization invariant & Companion Dev Harness: when working across both repos or touching shared interfaces (pkg/*), keep bothfakandfak-privatesynchronized (fetch both remotes andrefs/fak/locks/*, fast-forward both, syncgo.workto prevent module skew, and audit with public leak scrubtools/scrub_public_copy.py --audit-stagedorfak-dev audit-leak --staged). Local development infakis first-class and companion-governed: pre-commit hooks infakautomatically run 5-gate boundary checks (cmd/fak-boundary) and secret leak audits whenfak-privateis present. In-tree CLI commandsfak-dev gate,fak-dev provenance,fak-dev audit-leak, andfak-dev ticketbridge directly to companion tools without requiring directory switches. Cross-repo issue quoting rule: bare#<num>strictly denotes a publicfakissue; quoting afak-privateissue in publicfakMUST be explicitly qualified asfak-private#<num>(oranthony-chaudhary/fak-private#<num>) to ensure provenance is unambiguous. Full default + verify command inAGENTS.md. - Divide and conquer: Delegate substantive work and keep this coordinator context clean; enforce capability-aware scoping and persistence —
structurally drive 10x subagent adoption by launching 4–8 (up to 16 on multi-core hosts) specialized subagents concurrently across pairwise tree-disjoint lanes.
Decompose substantive or multi-part requests into atomic single-concern units and deploy multi-agent triads per leaf (researcher/explore -> worker -> cross-validator/issue-auditor).
Divide and conquer applies equally to hardware verification: isolate compute workloads into discrete sub-components (sub-kernels, GEMV/GEMM microbenchmarks, Vulkan primitives) to test on live physical devices early and often, rather than bottlenecking on monolithic full-model runs.
Top-level coordinator (depth 0) aggressively fans out parallel subagents; leaf workers (depth 1) execute directly within assigned package boundaries and must NOT invoke nested
taskcalls (preventing recursion depth exhaustion, #12028). Use guarded headless agents or equivalent isolated workers for investigation, implementation, tests, and review. Constrain smaller models and workers to atomic S0/S1 leaf units (1–3 files, single package, exactly one witness). Scope abstention strictly to bounded high-difficulty aspects (concurrency, frozen ABI, kernel memory layout, security gates): emit a structuredABSTAINfor the escalated boundary while executing all independent, safe, solvable sub-components (such as reproduction tests or diagnostics). Treat guard refusals as actionable feedback rather than session stops: queryfak recover <TOKEN>, adapt execution or wait out transient locks, and maintain momentum on the objective without repeating failing calls. Keep only decisions and compact witnessed evidence here; independently verify worker effects before landing or reporting them. Reserve direct work for lightweight coordination and truly trivial tasks. Full contract inAGENTS.md. - The Go module is the repository root — run
gocommands from the clone root;go install github.com/anthony-chaudhary/fak/cmd/fak@latestresolves directly. - Strict Ban on PowerShell and Loose Scripts — Mandatory Modular Go Programs —
adding new PowerShell scripts (
.ps1), Bash/POSIX shell scripts (.sh), batch files (.bat,.cmd), or loose scripts is STRICTLY BANNED across this repo andfak-private. An exception requires an extraordinary, super heavily justified reason (basically ~1 in the whole repo, such as initial host bootstrapping before Go is installed). All new automation, test runners, harnesses, session orchestrators, and background tooling MUST be native Go programs (cmd/*, Go leaves registered asfakCLI verbs, or quarantined sub-modules). Think modular, integrated, long-term value.
Move forward over conclusions
Frame all diagnostics, benchmark reports, and investigations around forward momentum rather than terminal conclusions. When an approach underperforms, a benchmark trails baseline, or an experiment fails, focus on forward action rather than passive or defeatist editorializing (e.g. rather than "X didn't work ... therefore we suck...", state "the next step to get better performance is X"). Treat unmet targets as empirical data that eliminates a variable and always formulate the concrete next checkable step.
Doc map
Doc map for humans and agents: llms.txt. Full contributor contract:
CONTRIBUTING.md.
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.

