lnwjud
AGENTS.md209 starsChanged 15 days agoMITView on GitHub
Do routine development directly on dev. Do not create feature, hotfix, or codex/* branches, or a new worktree for branch isolation, unless the user explicitly requests one. Keep main for verified dev -> main merges through the existing pull-request and branch-protection process; do not develop on main. After earlier temporary branches are merged and no open pull request or worktree uses them, remove them so the repository retains only dev and main. Check for uncommitted work and active ownership before…
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 32,671 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 repository7 other files it gives its agents.
- lnwjud-scheduled-continuationSkill
- ponytail-auditSkill
- ponytail-debtSkill
- ponytail-gainSkill
- ponytail-helpSkill
- ponytail-reviewSkill
- ponytailSkill
The file
Show its rules and text
We split this AGENTS.md into the rules it gives an agent
13 rules. About 0 are common, 5 are distinctive, and 8 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈5
Distinctive · estimate
This author's own choices. Worth a look.
8
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- could be automatic: a setting could enforce it
- Very specific: line 5
Do routine development directly on dev.
- Very specific: line 15
When a user-authorized repository task is still incomplete and the user has authorized scheduled continuation, load and follow …
- Distinctive: line 33
When a GitHub Actions workflow must be monitored until completion, use one authoritative long-running background/durable watcher for the …
- Very specific: line 36
Before starting a watcher, inspect existing authoritative shell/process tasks.
- Very specific: line 38
Treat the watcher task as the authoritative task for that CI run and keep it attached to the durable goal as a blocking_job when goal …
- Distinctive: line 39
Do not report CI success until the authoritative watcher is terminal and the GitHub workflow conclusion is confirmed.
- Distinctive: line 41
Never merge, tag, publish, or release from a SHA whose required CI run failed or is still non-terminal.
- Distinctive: line 69
Every application version prepared for release must also have a non-empty, exact-version entry in the in-app What's New registry, with …
- Distinctive: line 76
Before pushing any version-changing commit:
- Very specific: line 79
Run corepack pnpm@10.15.0 test:version immediately after set-version.
- Very specific: line 80
Run powershell -NoProfile -NonInteractive -ExecutionPolicy Bypass -File scripts/verify-release.ps1 -SkipWindowsPackaging and git diff …
- Very specific: line 93
Never weaken official release/CI signing gates.
- Very specific: line 94
When investigating a local Windows packaging failure, first distinguish among: package/build failure, SHA/provenance/cosign verification …
≈ “Common” is an estimate: at least 1 in 5 of the 41,581 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
Repository agent instructions
Branch workflow
Do routine development directly on dev. Do not create feature, hotfix, or
codex/* branches, or a new worktree for branch isolation, unless the user
explicitly requests one. Keep main for verified dev -> main merges through
the existing pull-request and branch-protection process; do not develop on
main. After earlier temporary branches are merged and no open pull request or
worktree uses them, remove them so the repository retains only dev and
main. Check for uncommitted work and active ownership before any cleanup.
Automatic scheduled continuation
When a user-authorized repository task is still incomplete and the user has authorized scheduled continuation, load and follow .agents/skills/lnwjud-scheduled-continuation/SKILL.md as the source of truth. Never create a schedule when the user has explicitly opted out for the current task.
For rolling continuation, the workspace skill is authoritative. Current mainline uses exactly one hourly recurring Native ChatGPT Scheduled Task for an active goal, in the current chat with cloud execution requested. Ordinary checkpoints and ordinary recurring wakes reuse the same native task; they do not create or retime a per-turn successor. Historical one-time watchdog rows are compatibility-only. Never use lnwjud scheduler, Windows Task Scheduler, schtasks.exe, cron, shell timers, browser/DOM automation, or undocumented scheduler APIs as a fallback. Never create or re-enable scheduled continuation when the user has explicitly disabled it for the current task.
Scheduled and manual durable workers are work-conserving. Discovering a fixable CI/test/code-scanning/packaging/release/tool failure is the next unit of work, not a status boundary: inspect the exact failure, diagnose it, repair it, and run the relevant validation in the same host turn while safe useful actions remain. Do not spend an hourly wake merely summarizing an intermediate problem that can still be acted on; user-visible status prose never substitutes for available execution.
Track background work with goal-relative trackedTasks: blocking_job entries participate in liveness, while supporting_service entries do not; explicit provider routing and cancelWithGoal ownership prevent a shared service from being probed or stopped accidentally. Legacy activeTaskIds rows decode conservatively as goal-owned blocking jobs. Live or unknown liveness fails closed; stale-owner takeover must follow the bounded recovery rules in the scheduled-continuation skill. For recurring interval wakes, an existing runKey is idempotency history, not worker-liveness proof: only a true concurrent/stale duplicate may return already_claimed, while a same-interval retry observing the current lease must re-evaluate liveness and recover after the stale-heartbeat grace when safe. A request to disable scheduling stops only scheduled continuation and never abandons the durable goal. When the goal finishes, make the exact Native ChatGPT task non-runnable using the strongest operation actually exposed by the host (prefer true delete, otherwise host-confirmed disable), record truthful cleanup evidence, finish the goal, verify get_goal is terminal, and stop. Never report completion while the goal is active.
Durable checkpoint fidelity
A milestone checkpoint is durable reconstruction state, not a status blurb. For meaningful milestones and every handoff boundary, populate resumeContext with enough concrete state for a new worker to continue without guessing or repeating settled work: changed files, exact commands/results, decisions, failed attempts, pending validation, resume prerequisites, state facts, and artifacts. Keep blockers, tracked tasks, step status and next action truthful and current. summary is only a headline; never rely on summary text alone when detailed recovery facts exist. session_handoff must prefer durable goal + checkpoint resume context before Git diff or legacy trackers.
Goal freshness is also a user-visible Watcher contract. While a durable Goal is active, call checkpoint_goal immediately after any plan step changes state, a blocking task becomes terminal, a blocker is added/resolved, or a commit, push, hosted-CI conclusion, package, deploy, release artifact, or other meaningful milestone becomes known. During sustained active work with no natural milestone, refresh currentPhase, nextAction, blockers, trackedTasks, and truthful step state at least every 10 minutes. Do not batch several already-finished steps into one late end-of-task checkpoint. Before any user-visible progress/final status, first make the durable Goal reflect all work already completed.
Do not leave duplicate, abandoned, superseded, or completed work as stale active Goal cards. When authoritative liveness/evidence shows that one Goal has been incorporated into another, reconcile it promptly; when a Goal is genuinely complete, satisfy its acceptance criteria and finish it before reporting completion.
Authoritative CI Watcher Policy
When a GitHub Actions workflow must be monitored until completion, use one authoritative long-running background/durable watcher for the exact workflow run instead of repeated ad-hoc polling.
- Resolve and record the exact GitHub Actions run ID first. Never monitor only "the latest run on a branch" after the watcher starts.
- Before starting a watcher, inspect existing authoritative shell/process tasks. If a live watcher already targets the same exact run ID, reuse its task ID and inspect/wait/result it; do not start a duplicate watcher.
- Preferred watcher command:
gh run watch <RUN_ID> -i 20 --exit-status. - Treat the watcher task as the authoritative task for that CI run and keep it attached to the durable goal as a
blocking_jobwhen goal tracking is active. - Do not report CI success until the authoritative watcher is terminal and the GitHub workflow conclusion is confirmed.
- On failure, inspect the actual GitHub Actions job logs for that exact run before changing code. Do not guess from branch state or from a different run.
- Never merge, tag, publish, or release from a SHA whose required CI run failed or is still non-terminal.
- Multiple exact workflow run IDs may be chained inside one durable watcher process when appropriate, but preserve each workflow's exit code and fail the watcher if any required workflow fails.
- A CI watcher is process monitoring, not a ChatGPT Scheduled Task. Do not create or re-enable scheduled continuation merely to watch CI when the user has scheduling disabled.
Windows example only:
powershell -NoProfile -NonInteractive -Command " Write-Host 'Watching CI run 123456...'; gh run watch 123456 -i 20 --exit-status; $ciExit = $LASTEXITCODE; exit $ciExit "macOS/Linux must use a shell available on that platform (for example
sh/bash) rather than assuming PowerShell is installed:gh run watch 123456 -i 20 --exit-status
This policy is cross-platform because the monitoring contract is shared while the wrapper shell is platform-native. Do not label a PowerShell-only implementation as cross-platform.
Version Bump Policy
For any repository/application version change, use the canonical root version script first: corepack pnpm@10.15.0 run set-version <version>. Do not start by hand-editing package versions or current-version documentation.
After the script runs, inspect the diff and current-version references for drift. Manual per-file edits are fallback-only for genuinely uncovered references. If an uncovered reference belongs to the canonical current-version surface, update scripts/set-version.mjs in the same change so the next bump is automated. Preserve historical release notes, plans, and dated evidence unless the task explicitly requires changing history.
Every application version prepared for release must also have a non-empty, exact-version entry in the in-app What's New registry, with meaningful Thai and English text. README and GitHub Release notes are separate outputs and do not populate the in-app modal. The Desktop release-notes test must verify the current package version and pass before tagging; never publish a version whose in-app modal would show the empty state.
Before pushing any version-changing commit:
- Finish the functional change and its deterministic regression test before changing the version. Keep the version-sync diff mechanically isolated from unrelated behavior where practical.
- Run
corepack pnpm@10.15.0 test:versionimmediately afterset-version; a failure is version drift and must be fixed in the script or canonical references before continuing. - Run
powershell -NoProfile -NonInteractive -ExecutionPolicy Bypass -File scripts/verify-release.ps1 -SkipWindowsPackagingandgit diff --check. Do not push merely because a narrower targeted test passed. - Inspect the final diff and confirm that the development target changed while the current published version stayed unchanged unless publication was explicitly requested.
A GitHub Actions failure on a version-named commit is not automatically a version failure. Resolve the exact run ID and inspect the exact failed job/log before editing. Treat only version-contract, lockfile, artifact-name, tag/version, or provenance mismatches as version failures. Timing, process-lifecycle, OS-specific, and external MCP failures require a root-cause fix and deterministic regression test at their real seam. Never make a blind rerun the fix, and never hide nondeterminism by only increasing a timeout. A rerun that passes on the same SHA is evidence of a flaky test or race, not proof that the defect is resolved. Do not report the Action fixed until a new exact-SHA run completes successfully.
Local Windows Packaging / Signing Policy
Local Windows development builds normally run without a paid/commercial Windows code-signing certificate. Therefore Get-AuthenticodeSignature may legitimately report NotSigned/unsigned for locally built Setup or Portable artifacts. Do not treat that status by itself as a local build failure.
- Separate Windows Authenticode publisher signing from artifact integrity/provenance verification. They are different mechanisms.
- For a local dev package with no Windows signing credentials configured, success is based on the package command succeeding plus the required SHA-256/provenance/release-evidence checks passing. Record the unsigned Authenticode state truthfully; do not invent or require a certificate that is not configured.
- The canonical Sigstore verifier for this repository's Windows local/CI workflow is cosign v3.1.3. Prefer setting
LNWJUD_COSIGN_PATHto a trusted cosign v3.1.3 binary; on the current Windows workstation the usual cached binary isC:\Users\ABCz\AppData\Local\lnwjud-build-tools\cosign-v3.1.3\cosign-windows-amd64.exewhen present. cosignverifies pinned Sigstore provenance/integrity (for example the bundled tunnel-client). It does not Authenticode-sign lnwjud and is not a substitute for a Windows publisher certificate.- Never weaken official release/CI signing gates. If Windows signing secrets/certificates are configured for an official workflow, that workflow may require Authenticode
Valid; missing/invalid signing in that configured context is a real failure. - When investigating a local Windows packaging failure, first distinguish among: package/build failure, SHA/provenance/cosign verification failure, and merely
NotSignedAuthenticode status. Only the first two are automatically failures for an unsigned local dev build.
Full Bypass is an explicit runtime exception to application approval and scope enforcement: while a Full profile has the relevant Desktop or STDIO Full Bypass toggle enabled, permission prompts, Active Project scope, and other lnwjud approval gates may be skipped. Durable-goal ownership is different: if a workspace has a live rolling scheduled-goal mutation fence, ToolRegistry must still require and validate the current goalLease before mutation, even under Full Bypass. This prevents a stale worker from mutating after handoff. Ordinary unscheduled Full Bypass calls remain lease-free when no live rolling-goal fence exists.
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.

