openclaw-skills
AGENTS.md3 starsChanged 10 days agoMITView on GitHub
This file provides shared guidance to coding agents (Claude Code, Codex, and any other assistant) working with code in this repository. CLAUDE.md is a compatibility entrypoint that points here; keep this file as the single source of repository guidance. A public collection of independent OpenClaw skills maintained by Garibong Labs. Each top-level directory is one self-contained skill that gets copied into an OpenClaw workspace. There is no root build system, package.json, or shared library — skills share nothing but conventions.
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 31,439 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository11 other files it gives its agents.
- CLAUDE.mdCLAUDE.md
- acp-discord-orchestratorSkill
- brave-api-setupSkill
- daum-trendsSkill
- ipo-alertSkill
- ku-portalSkill
- music-preference-recoSkill
- nano-banana-proSkill
- olympic-alertSkill
- seoul-subway-crowdSkill
- tistory-publishSkill
The file
We split this AGENTS.md into the rules it gives an agent
3 rules. About 0 are common, 0 are distinctive, and 3 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
3
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 9
Every skill has a SKILL.md with YAML frontmatter (name, description, optionally metadata.openclaw.requires/triggers, runtime, credentials).
- Very specific: line 15
No install step. Local versions in use: Python 3.13, Node 24.
- Very specific: line 100
Commit style: conventional prefixes scoped by skill, e.g.
≈ “Common” is an estimate: at least 1 in 5 of the 40,246 AGENTS.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
AGENTS.md
This file provides shared guidance to coding agents (Claude Code, Codex, and any other assistant) working with code in this repository. CLAUDE.md is a compatibility entrypoint that points here; keep this file as the single source of repository guidance.
What this repo is
A public collection of independent OpenClaw skills maintained by Garibong Labs. Each top-level directory is one self-contained skill that gets copied into an OpenClaw workspace. There is no root build system, package.json, or shared library — skills share nothing but conventions.
Every skill has a SKILL.md with YAML frontmatter (name, description, optionally metadata.openclaw.requires/triggers, runtime, credentials). The frontmatter is read by OpenClaw and by its security scanner; runtime/credentials declarations exist because the scanner flagged undeclared credential use (see tistory-publish/CHANGELOG.md v5.1.x).
Most skills are a single stdlib-only Python script (ipo-alert, olympic-alert, daum-trends, music-preference-reco, seoul-subway-crowd, nano-banana-pro) or Node script (brave-api-setup) with no tests. The two substantial, tested skills are tistory-publish and acp-discord-orchestrator. Docs in tistory-publish/ are written in Korean; keep that style when editing them.
Commands
No install step. Local versions in use: Python 3.13, Node 24. CI (.github/workflows/tistory-publish-ci.yml) only covers tistory-publish, path-scoped, on Python 3.9 and 3.13 and Node 24 — keep tistory-publish Python compatible with 3.9.
# tistory-publish — Python unit tests (CI command)
python3 -m unittest discover -s tistory-publish/tests -p "test_*.py" -v
# single file / single test (substring match on test name)
python3 -m unittest tistory-publish/tests/test_publish_post_upload.py
python3 -m unittest tistory-publish/tests/test_publish_post_upload.py -k file_chooser
# tistory-publish — JS helper tests (plain assert script, no runner)
node tistory-publish/tests/test_tistory_editor_helpers.js
# tistory-publish — syntax checks run in CI
python3 -m py_compile tistory-publish/scripts/*.py
node --check tistory-publish/scripts/tistory-editor-helpers.js
node --check tistory-publish/scripts/tistory-publish.js
bash -n tistory-publish/scripts/login.sh tistory-publish/scripts/publish-post.sh tistory-publish/scripts/publish.sh
# acp-discord-orchestrator — node:test suites (not in CI; run manually)
node --test acp-discord-orchestrator/scripts/test-acp-reporting-contract.mjs
node --test acp-discord-orchestrator/scripts/test-acp-host-transport.mjs
node --test acp-discord-orchestrator/scripts/test-acp-report-pump.mjs
node --test acp-discord-orchestrator/scripts/test-acp-report-controller-preparation.mjs
node --test acp-discord-orchestrator/scripts/test-acpx-foreground-supervisor.mjs
node --test acp-discord-orchestrator/scripts/test-acpx-runtime-preflight.mjs
node --test acp-discord-orchestrator/scripts/test-claude-acp-launcher.mjs
node --test tests/acp-discord-orchestrator-cli.test.mjs
node --test --test-name-pattern="distinct exit codes" acp-discord-orchestrator/scripts/test-acpx-foreground-supervisor.mjs
# acp-discord-orchestrator — installed-boundary evidence probe (not a node:test file)
node acp-discord-orchestrator/scripts/check-openclaw-automation-boundary.mjsThe launcher suite's execve end-to-end tests self-skip on runtimes without process.execve (needs 22.15+/23.11+ or any later line; POSIX only) — run it on a runtime that has it for full coverage.
check-openclaw-automation-boundary.mjs is deliberately outside the node:test suites: it executes the installed OpenClaw 2026.8.1 bundles to pin the three model-callable behaviours the portable mocks cannot see — that capCronJobToolsAllow re-emits a finite toolsAllow in creator tool-surface order rather than the requested order, that headless scripts run under agent:main:cron:<jobId>:trigger, and that a successful automations(action:"remove") reaches the model as { content, details: { removed: true } } with no top-level removed. It locates bundles by content signature (never by hashed filename or whole-bundle hash) and reports SKIP with exit 0 when no matching install is present, so it never fails a machine that has nothing to probe. Report its check count separately from the node:test totals.
tistory-publish architecture
Browser automation of the Tistory editor via Playwright attached to an existing Chrome over CDP (Tistory's Open API is dead). The pieces:
scripts/publish-post.sh— the current orchestrator (~2300 lines). It is a bash wrapper that parses args / applies template presets (mk-review,daum-trends,simple-post), then runs one large Python Playwright program embedded in a<< 'PYTHON_SCRIPT'heredoc. The Python program opens/manage/newpost, injects the JS helper withpage.add_script_tag(path=HELPER_JS), and drives the editor step by step (category/title → body → inline images → banner → OG cards → representative image → tags → duplicate preflight → publish → post-publish verification). Last stdout line is a JSON result ({"success":..., "postUrl":...}); non-zero exit on failure.scripts/tistory-editor-helpers.js— the current in-page helper (TinyMCE insertion, OG placeholder handling, tags, image alt/representative image). Functions are called from Python viapage.evaluate.scripts/publish.sh+scripts/tistory-publish.js— legacy pipeline kept for compatibility. Helper changes have historically been mirrored into both JS files (see CHANGELOG); when adding a helper function, check whether the legacy file needs the same change.scripts/login.sh— the only script allowed to touch Kakao credentials (--cred-file/TISTORY_CRED_FILE).publish-post.shnever reads credentials; it shells out tologin.sh(forwarding--blogand--cdp-port) when it lands on a login redirect.scripts/seo_check.py— standalone pre-publish static SEO check, wired in via--seo-check off|warn|strict.
Behavioral invariants worth knowing before changing it:
ALLOW_DIRECT_TISTORY_PUBLISH=1must be set orpublish-post.shrefuses to run — it is meant to be invoked from a private wrapper, not directly.- The code is deliberately fail-closed: OG-card retries/fallbacks only trigger on exactly-classified
/manage/scrapresponses (confirmed40002→ DCInside mobile/desktop pair; confirmed500+40009→ Daum next-article candidate fromdata-og-fallback-urls); unknown/unparsed responses abort.daum-trendsrepresentative image must be the first non-comic inline image or the publish aborts. Don't loosen these into "try something else" fallbacks. - A per-blog (or per-CDP-port)
fcntl.flocklock at/tmp/tistory-publish-{key}.lockserializes concurrent publishes; duplicate-title preflight queries RSS + manage list before clicking publish. - Body HTML contract:
<p data-ke-size="size16">, OG placeholders<p data-og-placeholder="URL">​</p>, image markers<p data-image-marker="...">, Tistory<hr data-ke-type="horizontalRule">.
How the tests work (this shapes how you write code):
- Python tests do not run the shell script. They regex the embedded heredoc out of
publish-post.sh(<< 'PYTHON_SCRIPT' ... PYTHON_SCRIPT),ast.parseit, pick named top-leveldefs, andexecthem into a namespace seeded with fakes (FakePage,FakeTime, etc.). So embedded helpers must be module-level functions whose collaborators (page,time,log,fail,re, …) are resolvable by name from the namespace — no hidden closures or module-import side effects. - JS tests load
tistory-editor-helpers.jsinto anode:vmsandbox with a minimal fakewindow/documentand then appendthis.__helpers = { ... }to export specific functions. A new helper is only testable if you add it to that export list intests/test_tistory_editor_helpers.js. tistory-publish/CHANGELOG.mdhas an## Unreleasedsection — add an entry for any behavior change.
acp-discord-orchestrator architecture
Core ESM modules under scripts/:
acpx-foreground-supervisor.mjs— runs one ACPX turn in the foreground on behalf of an agent and emits newline-delimited JSON events. It owns the centralized closed-set agent gate (assertCanonicalSupportedAgent:invalid_agent_unsupported/invalid_agent_not_canonical, enforced first both in the config loader — before any unrelated filesystem I/O — and in the in-memoryrunSupervisorpath, symmetric between claude and codex), the agent-neutral process-integrity env baseline (ACP_INJECTION_ENV: Node module/preload, dynamic-linker, and proxy selectors) that every supported agent gets automatically before runtime import, and the local-commit-only permission boundary that rejects opaque execute requests plus direct remote Git/repository-hosting actions before approval;CLAUDE_INJECTION_ENVlayers the Anthropic-specific selectors (CLAUDE_PROVIDER_INJECTION_ENV) on top for Claude only.acp-reporting-contract.mjs— the pure reporting-contract validator and canonical start/intermediate/terminal message builder. Current schemaacp-reporting-v3, whose enabledreportPumpbinds the exact job id plus the non-secret deterministic controller-script structure (version/digest, 60-second timeout, five-call allowlist, delivery none) without carrying the substituted script, lease token, or a static report.acp-reporting-v2(watchdog shape) stays accepted for already-prepared configs, and legacy v1 is Claude-only. It owns the closed agent presentation mapping and model/report identity rules.acpx-runtime-preflight.mjsandacpx-runtime-preflight-cli.mjs— the canonical machine-enforced runtime-module preflight. Theattestaction consumes the raw stdout JSON ofopenclaw plugins info acpx --jsondirectly (saved unmodified to an owner-private file, never reshaped by the caller), uniquely selects theplugin.dependencyStatus.dependenciesentry named exactlyacpx(rejecting missing/duplicate exact matches, relative or malformed paths, the active plugin package root bound from the rawplugin.rootDir— plusinstall.installPathwhen present — and any raw-schema drift including the legacy synthetic top-leveldependenciesshape; the@openclaw/acpxmanifest-name gate stays as defense in depth), validates the package (name, ACPX ≥ 0.11.2 release version, realdist/runtime.js, and the authoritativecreateAcpRuntime/createRuntimeStore/createAgentRegistrycapability exports), and writes an owner-private digest-bound attestation (acpx-runtime-attestation.v1). Theassembleaction re-verifies that attestation fail-closed (absent/stale/mismatched/invalid) and writes the final supervisor config with the attestedruntimeModule, replacing the template sentinelRUNTIME_MODULE_FROM_PREFLIGHT— callers never hand-copy or chooseruntimeModule; already-prepared configs with a validated absolute path stay valid. Bounded stable codes only; no path or plugin-info payload is ever echoed.acp-private-json-input.mjs,acp-start-message-cli.mjs, andacp-report-message-cli.mjs— the shared owner-private JSON reader plus operator message builders. The reader stats and reads one normalized path, distinguishes missing/unreadable/empty/oversized/invalid JSON, and the CLIs emit only canonical public content or one bounded error event.acp-host-transport.mjs,acp-host-transport-cli.mjs, andacp-host-transport-runner.mjs— the POSIX/tmux production host transport (schemaacp-host-transport.v2; v1 records/inputs fail closed). A privatepreparecall creates one detached-but-owned tmux session and returns its exact handle before activation; a separateactivatecall writes the activation record into that exact PTY.status,claim-report,begin-delivery,ack-report,cancel, andreconcileuse only the retained transport record and handle. Every record mutation runs under an exclusive sibling lease file with bounded wait and no pathname-based stale takeover (atomic sibling-temp rename prevents torn files but is never the serialization), and a monotonic publication fence orders claims across calls.statusis observation-only;claim-reportis the single obligation-minting path (bound to the exact pump job identity and destination, with bounded attempts, explicit claim_acquired/delivery_pending/acknowledged/uncertain/missing delivery states, and a permanenttracking_lostpublication halt for a dead session without evidence); acknowledgements require the exact fenced attempt identity, re-derive canonical report content from stored reporting identity, and verify its SHA-256 attestation before publication can close.acp-lifecycle-ledger.mjs— the host-activation and owner-private lifecycle ledger contract. It parses the one exact-handle stdin activation record, keeps ACP runtime mutation behind that gate, stores bounded last-event/terminal-intent evidence, and reconciles either a mapped process exit or stabletracking_lostwithout exposing the handle. Repeated reconciliation re-verifies the current handle/outcome/exit evidence. End-to-end completion requires normalized terminal evidence, mapped exit, canonical terminal receipt acknowledgement, and private-ledger reconciliation; the transport never closes the ledger before that acknowledgement.acp-report-pump.mjs— the content-attested in-process pump imported by the installed controller plugin. Each invocation binds the exact transport/handle/destination/job identity, performsclaim-report→ fresh canonical derivation →begin-delivery, and returns the fenced delivery result plus the exact structured report input used for rendering, so acknowledgement never re-derives minute-sensitive fields.acp-report-controller-preparation.mjsowns private token/declaration-key generation, the two-stage model-callable job construction (a disabled inertautomationsadd with no reservedid— the publiccron.addschema has none — then one update on the scheduler-returned exact id that substitutes the token plus that id and enables the job atomically, with a bounded declaration-key replay for a lost add response), structural attestation, registration shape, ordered prepare/register/activate/commit flow, bounded exact registration replay/recovery, preactivation abort rollback, and commit-only recovery. It attests the persisted job'stoolsAllowas an exact duplicate-free set rather than an ordered list, because the model-callable boundary normalizes the order away, and it reads removal success through one strict bounded parser covering the wrappeddetails.removedenvelope that boundary actually returns. Its public automation template must stay byte-for-byte aligned with the controller plugin template. The production loader enforces the template and placeholder-script SHA-256 constants exported fromacp-reporting-contract.mjs; the normative runtime contract carries the exact plugin implementation pin, which must be replaced with the companion plugin's merge commit before a compatibility branch merges. Prepared controller ticks are inert so registration recovery remains owner-driven; every non-throwing controller-script path returns the bounded plain object{}required by OpenClaw 2026.8.1 scheduler result parsing.acp-lifecycle-reconcile-cli.mjs— the operator CLI that reads one private JSON reconciliation input and updates the private lifecycle ledger after the exact tracked host handle exits or is confirmed lost. It emits only a bounded reconciliation result/error and never echoes paths or handles.claude-acp-launcher.mjs— the canonical Claude-specific launcher that validates Claude setup-token auth and re-execs into the supervisor viaprocess.execve. Codex (and any future non-Claude agent) runs the supervisor directly and declares credentials through the genericrequiredEnv/forbiddenEnvcontract; genuinely Claude-provider-specific names (setup-token launcher,CLAUDE_CODE_OAUTH_TOKEN, theclaude-setup-token-env-fileauth profile,claude_*failure codes) stay Claude-named on purpose.
references/runtime-contract.md is the normative spec (config fields, exit codes 0/20/21/22/64, permission rejection rules, start-receipt/reporting/env/Claude-credential preflight gates, evidence boundary); SKILL.md is the operator-facing summary; templates/supervisor-config.json is the agent-neutral config template and templates/claude-auth-profile.json the provider-specific Claude auth profile. Keep all of them (code, contract, SKILL.md, templates) in sync when changing behavior.
- The supervisor module exports its internals (
loadSupervisorConfig,runSupervisor,buildPermissionHandler,normalizeRuntimeEvent,runStartReceiptPreflight,runReportingPreflight,EXIT_CODES,main, …) specifically so the sibling test file can unit-test them with an in-memory mock runtime module.tests/acp-discord-orchestrator-cli.test.mjsat the repo root spawns the real CLI against a mock runtime file to assert the process actually exits.scripts/acp-reporting-test-fixture.mjsis the shared integration fixture (parameterized by agent and schema version); the standalone contract suite deliberately keeps its templates literal instead. - Everything is fail-closed and evidence-minimal: no raw prompt/model/tool text in events, no env values or conversation IDs leaked, distinct
completed/cancelled/failednever collapsed, and the public agent label never caller-chosen. Preserve these properties when adding features — tests assert them. - Requires Node ≥ 22.13 and ACPX ≥ 0.11.2 (capability-detected, not version-string gated). The production host transport is POSIX-only and requires tmux plus the executable system
/usr/bin/envused to clear tmux-server environment residue. The Claude launcher additionally needsprocess.execve— Node 22.15+ in the 22.x line, 23.11+ in the 23.x line, or any later line. - A separate local Skill Workshop proposal named
acp-agent-authis planned for provider-specific auth of additional agents (e.g. Codex); until it lands, do not add non-Claude auth profiles to this skill.
Repo conventions
- This is a public repo. Per
docs/tistory-cdp-architecture.md, never commit production account names, credential paths, cron identifiers, browser profile locations, or account→blog maps; those belong in private wrappers. Deployment targets (--blog,--category,--cdp-port) are passed in by the caller, not hardcoded (themk-reviewpreset defaults are the one intentional exception). - Published skills get scanned by OpenClaw/VirusTotal; past flags (CHANGELOG v5.0.x–v5.1.x) were
child_process.execSync, a base64/atobcanvas-drop fallback, credential handling inside the publish script, and undeclaredruntime/credentialsfrontmatter. Keep credential handling confined tologin.shand declare new runtimes/credentials in SKILL.md frontmatter. - Commit style: conventional prefixes scoped by skill, e.g.
fix(tistory-publish): …,feat(acp-discord-orchestrator): …,docs(...),ci(...). Branches arefix/…,feat/…,chore/…,ci/…merged tomainvia PR.
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.

