agentleFS
protocolprotocol/perun/.claude/skills/agentic-delivery/SKILL.md

agentic-delivery

Skill2 starsChanged 28 days agoMITView on GitHub

Use when implementing a feature, migration or end-to-end change: build, add, ship, deliver, fix a ticket. Invoke FIRST, before inspecting files, even if the dir looks empty. Gated delivery: independent QA and security, one writer per worktree, human approval on push/merge/deploy. Not a typo fix.

1 other project uses a version of this.

Adopted by

Most recently taken up by

The file

We split this skill 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
  1. Very specific:

    Use when a change spans build plus QA or security, writers would collide, or the owner asked for a feature, migration or release.

    line 42
  2. Very specific:

    Every packet names deep-code-review, its scope (FULL/DIFF <base>/FILE <paths>) and the base or reviewed SHA.

    line 68
  3. Very specific:

    Ask as Issue: one line, A: (recommended), B:, ≤12 words each.

    line 80

≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill 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.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

Agentic delivery

Gated-delivery pattern, not a runtime or standing swarm. Does not replace deep-code-review (review is the bar; this is how work reaches it); one delivery OS per repo. Persisted artifacts are normal English. No files/tools? Say so once, never claim a reference loaded or a command run: write would run: <cmd>, mark unverified-depth. Full gate table, roles, worktree, human-gate, failure text: references/full-map.md.

Task to entry — a match skips INDEX.md; else read it and open only matching rows.

TaskOpen (bare *.md = references/)
Land PRs as a train; red gate or unionscripts/merge_train.py, scripts/train_land.sh, merge-queue-worktrees.md; gate-epistemology.md, red-gate-triage.md; flow-guards.md (dedupe hook, fitness check, flow alarms, freeze)
Start a lane; isolate its worktreetemplates/lane-preamble.md in every brief; worktree-isolation.md, roles.md; templates/loops/
Verify or relay a result; slow pipelineverification-handback.md; pipeline-optimizer.md
Long session, compaction, new session, memory or docs bloatreferences/focus-kit.md: update .claude/handoff.md (scripts/focus_kit.py handoff set) at each gate, before a context-heavy step, before handback
New machine or fleetoperating-discipline.md
Audit a repo's GitHub settings (branch protection, Actions token, deploy keys, webhooks)run python3 scripts/repo_settings_audit.py OWNER/NAME... [--repos-file FILE] [--json]: read-only, a 403 or 404 is not checked; admin token for the full set
PR state; hand off; changelogscripts/pr_state.py <n>; scripts/land_ready.py (open PRs ready to land, or why not); scripts/review_status.py --sha S --red N (post perun/review); scripts/premise_check.py <n>...; scripts/ci-gates.sh fragments
Cost cut, lane cap; "enforced" claim, CI costcost-quality-guardrails.md; host-enforcement.md
Fan-out size, host probefanout-host-sizing.md, fast-agentic-delivery.md
Dev server, local stackdev-env-ownership.md
Resume, durable state; peer sessionsproject-state.md; multi-session-coordination.md
Tracker, auto-close, multi-hour loopwork-tracking.md, unattended-trackers.md, unattended-operating-mode.md
Product surface; ADR; briefproduction-grade-product-playbook.md, template-adr.md, template-onepager.md
Incident, support; G10, postmortem; installincident-response.md, support-ops.md; retrospective.md, template-postmortem.md; install.md
scriptsscripts/review_gate.py, flake_ledger.py, focus_kit.py, handback_cap.py, heavy_gate.py, operating_selfcheck.py, outcome_claim_guard.py, pipeline_profile.py, queue_guard.py, serial_gate.py, session_brief.py, wired_check.py; templates/review-tiers.tsv

Open references/full-map.md for multi-lane work, a red gate, host sizing, an unattended loop or a human-gate edge case. Sections cited as "SKILL.md <section>" (Conductor operating rhythm, Worktrees and occupancy, Local environment, Environment probe, Gate epistemology, Output contract) now live there.

When to Use

Use when a change spans build plus QA or security, writers would collide, or the owner asked for a feature, migration or release; not a typo. Load idea-critic before any agent-originated "we should" reaches the owner.

Hard rules

  • Smallest sufficient team, hats from risk. The Conductor never self-approves. A builder never reviews builder and never wears QA, security or release-approval. Physical or social-engineering security is human-led under written ROE.
  • Conductor reacts to a blocked gate, receipt, budget breach or collision; never does a lane's work. After two equivalent failures change approach.
  • Low-blast work may collapse adjacent gates; it may not remove independent verification or an applicable human approval.
GateHard condition
G0 IntakeGoals, non-goals, appetite; idea-critic first. Shape references/template-onepager.md
G1 SpecAcceptance criteria, deep-code-review scope, pinned base SHA
G2 PlanOne writer per worktree; every paid lane names a token/dollar budget before G4, none is blocked
G3 DesignADRs / contracts: interfaces, NFR budgets, data/security decisions explicit. Shape references/template-adr.md
G4 ImplementTests before or with the change
G5 VerifyLocal = affected tests; CI on the exact SHA is the receipt; UI needs headed-browser evidence
G6 ReviewIndependent of builder; Blocker/Critical block, High needs a named owner
G7 IntegrateOne owner; CI on the exact final SHA
G8 ReleaseRollback proven; owner approves outward/production action
G9 Prod verifyServed SLOs; rollback on breach
G10 LearnEscaped gap becomes a regression test; strip third-party identifiers

Missing evidence is UNVERIFIED, never pass. Missing price is UNPRICED, never zero. Missing spend cap is BLOCKED, never unlimited. A lane's completion is G7; G8 is separate and owner-gated.

Exact revision

Every packet names deep-code-review, its scope (FULL/DIFF <base>/FILE <paths>) and the base or reviewed SHA. QA and security get the final SHA; a stale SHA fails closed.

Worktrees

One writer per worktree; assume a shared tree until proven isolated. Preflight: list worktrees and open PRs, claim the work, never duplicate a running lane. Use the repo's own script for repeatable steps. A child lane branches off the parent's local ref, never polls the remote. Spawn a heavy lane only while free RAM >15% and swap is flat (scripts/host_probe.py).

Human gates

Agents prepare; humans approve push, open/merge PR, publish, deploy, external replies or refunds, scopes/secrets/IAM, destructive migrations, force-push, flag flips, waiving a Blocker/High finding.

Standing grant. A grant counts only from an owner-authored artifact: the owner's own commit (author = owner) to CLAUDE.md or the state record, or their message quoted verbatim with date. The agent never creates, widens, extends, or re-dates a grant. Absent or ambiguous means gated. A merge that triggers deploy or publish is a deploy, never covered. Force-push, history rewrite, deploy, secrets, spend, external sends and destructive data stay gated. Park a gated item with its next step and continue.

Ask as Issue: one line, A: (recommended), B:, ≤12 words each; never a list of questions.

Output and failure

Workers return role, exact SHA, artifacts, commands run with exit status, findings with severity and location, remaining risk, cost or UNPRICED. A work item needs Where, Done-when, Verify, Why; missing Done-when or Verify is sent back.

A start or auth failure is no work done. A running lane is not a finished one. Done means verified — a green gate at the exact SHA. A model outage fails the lane closed. No uncommitted writer work without a recovery record.

Download now (ZIP)

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
What did you use it for? e.g. Reviewing pull requests in a Rails app

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.