agentleFS
fritzprixfritzprix/libr-agent/.agents/skills/release-head-self-e2e/SKILL.md

release-head-self-e2e

Skill12 starsChanged 6 days agoMITView on GitHub

Plan and run version-aware Session API self-E2E from the latest published release tag through current HEAD inside LibrAgent: pin baseline/head SHAs, map git diff to matrix cases, dedupe prior tracking tickets, optionally create/update a GitHub tracking issue when the user explicitly authorizes writes, then orchestrate parent→ child delegates for zero-human execution and independent verification using real builtin tools and the debug-only E2E control plane (LIBRAGENT_E2E_CONTROL). Triggers: "HEAD부터 릴리스까지 E2E", "last release..HEAD E2E", "self e2e matrix", "#2565 자동", "릴리스 이후 회귀 E2E", "tracking ticket then self E2E". Not Harbor-only orchestration (Harbor is optional CI path); not Cursor libr-delegate unless the user runs from Cursor. Never invent tools; use agent__* and workspace__*.

The file

We split this skill into the rules it gives an agent

7 rules. About 0 are common, 2 are distinctive, and 5 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.

  • 5

    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
  1. Distinctive:

    Read references/scopes.md before acting.

    line 22
  2. Very specific:

    If ambiguous, stay in read-only and ask one question.

    line 78
  3. Very specific:

    Search existing GitHub issues (title/body keywords: release-head-self-e2e, baseline tag, #2565).

    line 114
  4. Very specific:

    Only in github-write scope.

    line 122
  5. Very specific:

    Update progress.jsonl and per-case folder under .libragent/work/release-head-self-e2e/<run-id>/cases/<id>/.

    line 151
  6. Very specific:

    If github-write: comment on tracking issue with summary + artifact paths (no secret values).

    line 163
  7. Distinctive:

    Do not merge or change default settings.

    line 164

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 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

Release → HEAD self-E2E

Internal LibrAgent workflow to turn last release..HEAD into a deduped E2E plan, an optional GitHub tracking ticket, and delegated unattended runs with independent verifiers.

Read references/scopes.md before acting. Default is read-only / dry-run until the user names a scope.

Integrate — do not fork

NeedUse
Child spawn / isolationbundled delegate
Proof-before-done verifier childdelegation-eval-loop
Matrix + Harbor tasksbenchmarks/harbor/libragent-session-e2e-matrix/
Debug approvals / user prompts (ui__askUser, storeAs)LIBRAGENT_E2E_CONTROL=1 + /api/e2e-control/* (debug build only)
Optional external CI batchpnpm bench:session-e2e* (Harbor); secondary to this skill
Cursor HTTP harness.agents/skills/libr-delegate (out-of-app only)

Hard rules

  • Zero-human for every case the repo can run unattended; otherwise mark blocked / platform-unsupported — never pass and never “needs-input”.
  • No rerun-until-green; one honest attempt per case unless the user orders a rerun.
  • Failures get a diagnosis, not silent retry loops.
  • No fabricated tools — only names from the live tool list (agent__*, workspace__*, etc.).
  • Parent survives child stop — agent__stopSession on children only; parent keeps orchestrating. Outer app restart/crash tests may use a watchdog; nothing else.
  • Preserve protections: release has no E2E routes; dev needs explicit opt-in; loopback + capability header; isolated data dir — never production profile or real secrets.
  • Independent verifier accepts only objective artifacts (API messages JSON, workspace files, verifier scripts, Harbor session_e2e_evidence.json) — not child prose alone.
  • No product code fix / merge / global settings in this skill unless the user separately authorizes it.
  • Creating this skill or loading it does not grant standing permission to open GitHub issues.

Known product limits (honest)

These remain outside full #2565 unattended coverage (see PR #2632):

  • E2E control drives user prompts (ui__askUser questions and storeAs / requireUserInput secret inputs, via the shared pending-user-prompt) and tool approvals. There is no separate interactive-shell control surface.
  • Native GUI paths outside the app window are not drivable from the control plane.
  • Full matrix still needs Windows or Linux desktop debug beside the app; cloud/Linux without the app cannot claim native GUI E2E.

Mark dependent checklist rows blocked with reason; continue independent cases.

Phases

0. Confirm authorization scope

Parse the user message for one of:

ScopeAllowed
read-onlyPin SHAs, diff, read manifest, read GitHub (no writes), write local plan under .libragent/work/
planAbove + propose case list + ticket body draft (still no GitHub POST)
github-writeCreate/update tracking issue once when user explicitly says so (e.g. “이슈 올려”, “create tracking issue”)
execute-e2eRun delegated children + verifiers; requires debug app + LIBRAGENT_E2E_CONTROL=1 on isolated profile

If ambiguous, stay in read-only and ask one question.

1. Pin baseline and HEAD

From the repo root (absolute workspacePath / workspaceOverride="current"):

python3 .agents/skills/release-head-self-e2e/scripts/pin_release_range.py \
  --repo . --json > .libragent/work/release-head-self-e2e/<run-id>/pins.json

Script behavior:

  • Picks latest annotated/lightweight git tag reachable as ancestor of HEAD (default: skips -*alpha*, -*beta*, -*rc*, -*dev* prerelease names).
  • Sets baselineSha = tag peel, headSha = HEAD, verifies ancestry; if no tag or tag not ancestor → explicit error object (no silent fallback).

2. Map diff → E2E cases (version-aware)

  1. Load benchmarks/harbor/libragent-session-e2e-matrix/coverage-manifest.json.
  2. Run hints (optional):
python3 .agents/skills/release-head-self-e2e/scripts/map_diff_to_cases.py \
  --repo . --baseline "$(jq -r .baselineTag pins.json)" \
  --manifest benchmarks/harbor/libragent-session-e2e-matrix/coverage-manifest.json \
  --json > case-hints.json
  1. Human/agent merge: hints are suggestions only — ground each selected case in git log / git diff paths and the issue #2565 expectation text. Write plan.md with: case id, why in range, risk, runner (harbor-task | delegate-brief | blocked).

3. Dedupe prior tickets

  • Search existing GitHub issues (title/body keywords: release-head-self-e2e, baseline tag, #2565).
  • Merge with completed evidence paths cited in comments; do not re-run pass cases unless user requests.
  • Record dedupe.md with kept/skipped cases and links.

Requires github-write to post; otherwise output draft only.

4. Tracking ticket (optional)

Only in github-write scope. Single tracking issue per pinned baselineSha..headSha pair. Body must include: pins, plan table, artifact dir, explicit “no secrets” policy, blocked list. Link confirmed product failures only within user-granted write scope.

5. Execute self-E2E (execute-e2e)

Preconditions

  • Debug LibrAgent (pnpm tauri dev or debug build), disposable data dir.
  • LIBRAGENT_E2E_CONTROL=1. The capability token is harness-side only: the external harness reads it from the disposable instance data dir. Agents and child briefs never receive token material.
  • LIBRAGENT_E2E_SYNTHETIC_SECRET for masked-value fixtures only.
  • Serialize: one case at a time on one app instance (queue file queue.json).
  • Prompt resolution scope: the only auto-resolver today is Harbor's trial poll loop, which watches its own session. Executor children spawned via agent__spawnSession have nothing answering their requireUserInput prompts — route those cases (e.g. B-3 interactive echo) to the Harbor runner, or mark them blocked with reason no in-app prompt resolver.

Parent (orchestrator)

  1. Load @skill:delegate mechanics; keep workspace shared when children need repo + manifest.
  2. For each runnable case:
    • Spawn executor child (waitForResult=false first if parallel planning; serialize runs).
    • Brief: self-contained task, exact matrix id, synthetic secret policy, deliverable = evidence paths + API excerpt pointers.
    • Poll agent__checkSession with backoff; on paused, one agent__messageToSession resume attempt.
  3. Spawn verifier child with no executor transcript as sole proof — pass references/verifier-brief.md + paths to raw messages.json, workspace outputs, Harbor evidence.
  4. Update progress.jsonl and per-case folder under .libragent/work/release-head-self-e2e/<run-id>/cases/<id>/.
  5. On unsupported driver → blocked with reason; continue queue.
  6. Cleanup: executor terminate/delete after evidence harvest; verifier separate session; parent remains.

Negative controls (delegate brief or Harbor test_outputs.py patterns):

  • E2E routes 404 without opt-in; wrong capability header rejected; stale execution id rejected; synthetic secret absent from logs in verifier grep rules.

6. Closeout

  • summary.md: pass / fail / blocked counts, diagnosis per fail, limits section.
  • If github-write: comment on tracking issue with summary + artifact paths (no secret values).
  • Do not merge or change default settings.

Artifact layout

.libragent/work/release-head-self-e2e/<run-id>/
  pins.json
  case-hints.json
  plan.md
  dedupe.md
  queue.json
  progress.jsonl
  summary.md
  cases/<case-id>/
    executor-messages.json
    verifier-verdict.md
    evidence/...

Usage examples

Full command recipes: references/usage.md.

References

  • scopes.md — authorization matrix
  • usage.md — release→HEAD ticket + full self-E2E examples (bash + PowerShell)
  • verifier-brief.md — independent child template
  • executor-brief.md — unattended runner child template

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.