本文へ移動
cccskills
無料GitHub で公開

gz-flighttest

Use when the operator says "fly sortie S<N>" or "run a flight test"; flies one sortie with a human Go/No-Go.

インストール方法を見る

含まれるファイル(1)

  • SKILL.md9.1 KB

SKILL.md(原文)

インストールする前に、エージェントに与えられる指示の中身を確認できます。

gz-flighttest

Drive one sortie of the Flight-Test Program against a target substrate. A sortie is one flight flown in a single sitting — a chained run of test points that proves a gzkit workflow behaves as designed, with the target repo's ledger and receipts as the flight-data recorder.

This skill operates inside the target repo, where gzkit is the system-under-test. It is authored and versioned in gzkit and ships via distribution, but it runs on the ground in the target — flying gzkit's own workflows there. ("Operate in" ≠ "live in.")

The product is feedback that refines gzkit — not a built target. A sortie's yield is proving-outcomes about gzkit's design; the target is scaffolding. Because the skill runs in the target, that feedback reaches gzkit cross-repo: gzkit-directed squawks are filed via gz issue file (the gz-issue-file skill), which routes to the gzkit repo regardless of the target's git remote — never /ghi-author, which would file in the target (the wrong tracker).

The program (target-agnostic doctrine) lives in gzkit docs/flighttest/. The campaign instance and flight log live in the target repo (scaffolded from docs/flighttest/templates/). This skill is the propellant that flies a sortie and writes its evidence to the target's flight log.

Flight-test campaigns are a distinct work-stream: they govern flight-test engagements against a target, and never contend with the Build-to-1.0 Magna Carta, which rules gzkit's own build sessions.

Persona

Active driver: flight-test-engineer — read .gzkit/personas/flight-test-engineer.md and adopt its behavioral identity before flying. Falsifier-precommitment, black-box-evidence, and chase-deference are not rules you follow; they are who you are in the air.

Roles (who does what)

RoleWhoThis skill's relationship
Test directorHuman operatorAuthorizes each sortie (Go/No-Go); witnesses Gate-5; rules on squawk disposition. Never an agent. The skill pauses for these decisions — it does not make them.
Flight-test engineerThis skill (agent)Authors the card, flies the chain, collects evidence, drafts the debrief.
Chasespec-reviewer / quality-reviewer subagentConfirms the pass from the black box alone; never flew the sortie.

The Iron Law

A SORTIE IS NOT COMPLETE UNTIL ITS DEBRIEF IS RULED ON AND ITS SQUAWKS TRACKED.
A PASS IS NEVER SELF-DECLARED — THE CHASE VERDICTS; THE HUMAN AUTHORIZES.

"The commands ran" is not completion. "It worked when I flew it" is not a pass. The sortie runs through its debrief, and the pass/fail verdict is authored by an independent Chase reading evidence you did not narrate. Soliciting the human's Go before the card is frozen, or claiming a pass before the Chase has read the black box, is a gate bypass.

Rationalization Prevention

ThoughtReality
"The chain ran clean, let me mark it passed"You do not grade your own landing. Freeze evidence, dispatch the Chase.
"The hook blocked me — let me hand-write the marker to keep flying"The block IS the finding. Record it; never route around it (that is the vibing you were sent to catch).
"I'll write the expected observables from what I just saw"The falsifier is authored before the flight, from the design. Post-hoc expectations are worthless.
"This corner sortie is interesting, let me fly it first"Build up the envelope. A corner sortie with no spine baseline is unreadable data.
"I'll ask for Go/No-Go and Gate-5 together to save a round-trip"Go/No-Go gates entry; Gate-5 gates completion. Bundling them is the bundled-attestation anti-pattern.

Preconditions

  1. A target substrate satisfying docs/flighttest/README.md §7 (boring & bounded, real-buildable, greenfield/near-cold, separate repo).
  2. A campaign instance in the target repo (scaffold from docs/flighttest/templates/campaign-instance.md on first run) and a flight log (from flight-log.md).
  3. Green floor: the target's uv run gz check is not red at sortie open (a sortie does not fly into a broken baseline).

The sortie sequence (one run = one sortie)

Stage 1 — Resolve the topmost sortie

  • Read the target's campaign instance. Identify the topmost unflown sortie whose entry gate is met (build-up order; a sortie whose DEPENDS ON has not passed is not eligible). Do not reorder by interest.
  • Confirm the target's baseline is green (uv run gz check).

Stage 2 — Author the flight card (freeze the falsifier)

  • Draw the card from docs/flighttest/manifest.md for this sortie and the template in docs/flighttest/flight-card-template.md.
  • Author expected observables from the workflow's design, not from a trial run: the exact ledger event kinds/counts, receipt name prefixes, and gz state/gz status assertions that constitute a pass, plus the pre-registered falsifiers (including the "passed only by leaving the governed path" tripwire).
  • Freeze the card into the flight log as a CARD entry.

Stage 3 — Go/No-Go (human gate)

  • Present the frozen card to the test director and pause for explicit Go/No-Go. Record the ruling (verbatim) to the flight log.
  • No Go → the sortie does not open. This pause is not optional.

Stage 4 — Fly the chain

  • Execute each test point in order, through its governed path (the matching skill or gz verb named in the manifest). Feed each point's output into the next — the chain is the point.
  • A blocking hook or failing gate is a squawk, recorded, never defeated. If the governed path cannot carry a test point, halt per the card's abort criteria; do not hand-craft the artifact.
  • Collect the black box: the target's ledger slice for the sortie window and all emitted receipts.

Stage 5 — Chase (independent verdict)

  • Dispatch a spec-reviewer (evidence-to-card tracing) and/or quality-reviewer (integrity of the run) subagent. Hand them the ledger slice and receipts — not your narrative. Include a Why in the subagent prompt (filter signal from noise).
  • The Chase authors PASS/FAIL against the card's pre-registered falsifiers, reachable from evidence alone. If the Chase cannot reach PASS from the black box, the sortie has not passed.

Stage 6 — Debrief & route squawks

  • Draft the debrief (dispatch narrator for operator-value framing if useful): black box vs. card, each squawk, and the design signal (sound / awkward-but-correct / wrong — with the why).
  • Route every squawk — untracked squawk = nonexistent squawk. Because the sortie flies in the target repo, gzkit-directed feedback goes cross-repo:
    • gzkit defect (the product) → file against gzkit via gz issue file (the gz-issue-file skill) — routes to the gzkit repo regardless of the target's remote. Never /ghi-author, which files in the target. This is the engagement's yield.
    • gzkit correct-but-wrong-design → also gz issue file against gzkit; a gap between shipped surface and declared intent is a correction to the owning gzkit ADR, not an enhancement (operator doctrine).
    • Target/substrate-local defect → fix in the target or track it there; it gates the sortie only if it blocks the workflow under test.
    • Bad card → amend the card in the target's flight log; re-fly.
  • Present the debrief to the test director for a ruling.

Stage 7 — Record & advance

  • Append the sortie's full record (card, Go/No-Go, black box, Chase verdict, debrief, squawk dispositions) to the target's flight log — the append-only Layer-2 of the engagement.
  • Tick the campaign checkbox only with the Chase verdict + ledger evidence cited (the checkbox is Layer-3; the flight log is truth).
  • Surface the next eligible sortie; do not auto-open it (each sortie is human-authorized).

Validation

  • The flight log carries a complete record for the sortie (card frozen before fly; Go/No-Go verbatim; black-box slice; Chase verdict; debrief; squawk routes).
  • Every squawk has a tracked home — gzkit-directed ones as a cross-repo gz issue file against gzkit (never a target-local /ghi-author); a target-local defect or a card amendment otherwise.
  • The campaign checkbox change cites Chase verdict + ledger evidence, never narrative.
  • No test point passed by leaving the governed path.

Anti-patterns

  • Flying a sortie before its dependency sortie has passed (envelope violation).
  • Authoring expected observables after seeing the output (falsifier inversion).
  • Marking a pass without an independent Chase verdict (self-attestation).
  • Hand-writing a marker / using --no-verify to get past a block (path-defeating — the block was the finding).
  • Bundling Go/No-Go with Gate-5, or ticking a checkbox on prose (Layer-3 as truth).

レビュー

まだレビューはありません。使ってみた感想をお寄せください。

同じリポジトリのスキル

概要と使いどころ

Use to mine ../airlineops for governance habits gzkit lacks; produces a parity scan report (gzkit leads, AirlineOps adopts).

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

Use to file a GitHub issue when a finding needs its own work order or the operator asks; OBPI corrections stay in its change log.

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

ghi-close

無料

Use to do the work a GHI describes, verify it and close the issue with evidence in one sweep.

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

Use to triage the open GHI queue before planning or choosing what to pull; produces a ranked deliverable.

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

git-sync

無料

Run the guarded repository sync ritual; commit-time hooks enforce the cheap gates and the pre-push hook runs gz check.

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

git-sync

無料

Run the guarded repository sync ritual; the hooks your pre-commit config declares enforce the quality gates.

日本語の概要は準備中です。原文の説明を表示しています。

tvproductions/gzkit92026年10月18日 更新

tvproductions のスキルをすべて見る

このスキルの問題を報告する