Deploy Agent Skill

Deploy

No gate skipped, no failure ignored. If any gate fails, halt and report. Never --no-verify. Never bypass.

Failure recovery (rule 14). Worker errors and Quality Gate failures follow the canonical policy in skills/hyperflow/failure-recovery.md. Gate failures are user-surfaced, never auto-fixed — print the failing command + full stderr and halt the push. Never --no-verify, never force-push to main.

Per-Step Agent Map

| Step | Sub-phase | Workers | Reviewer | Notes |

|---|---|---|---|---|

| 1a | Repo-state scan | Worker A (git status), Worker B (git log) | Reviewer | — |

| 1b | Tool detection | Worker A (profile.md + lockfiles), Worker B (testing.md + devDeps) | Reviewer | — |

| 2a | Lint + typecheck (parallel) | Worker A (linter), Worker B (formatter), Worker C (tsc) | Reviewer | Step 3 (Security Sweep) runs in parallel with Step 2 at orchestrator level; 2a halts chain on any failure before 2b |

| 2b | Build gate | Worker A (prod build), Worker B (dev build) | Reviewer | Depends on 2a PASS |

| 2c | Test gate | Worker A (unit), Worker B (integration/E2E) | Reviewer | Parallel (P1); depends on 2b PASS |

| 3a | Secrets scan | Worker A (diff pattern), Worker B (file pattern) | security-reviewer | Runs in parallel with Step 2 (pre-build; read-only) |

| 3b | Dependency audit | Worker A (CVE audit), Worker B (license check) | vulnerability-reviewer (web-research-first) | — |

| 4 | Commit | single Worker | Reviewer | atomic-exempt (DOCTRINE 12.2) |

| 5a | Release execution | single Worker | Reviewer | atomic-exempt (DOCTRINE 12.2) |

| 5b | Version sync | Worker A (manifests), Worker B (changelog) | Reviewer | — |

| 6 | Push gate | AskUserQuestion | — | structural gate; atomic-exempt |

| 7 | Output | single print | — | atomic-exempt (§12.1) |

Step 1 — Survey State

Sub-phases run in parallel (P1).

Step 1a — Repo-state scan

Two Workers in parallel:

• Worker A — git status --short — uncommitted changes, staged files
• Worker B — git log origin/<branch>..HEAD --oneline — commits ahead of remote; detect branch name

Reviewer — verdict on repo state (clean / has uncommitted / ahead by N). If detached HEAD or no remote configured → halt with reason.

Step 1b — Tool detection

Two Workers in parallel:

• Worker A — Read .hyperflow/profile.md for package manager and project type; fallback: inspect package.json, pyproject.toml, Cargo.toml, go.mod
• Worker B — Check .hyperflow/testing.md for test runner; fallback: detect from package.json devDependencies (vitest, jest, playwright, pytest, etc.)

Reviewer — produce a single tool manifest (package manager, test runner, typed-project flag, build script presence). Used by Step 2 gates.

Step 2 — Quality Gates

Step 2 runs in parallel with Step 3 (Security Sweep) at the orchestrator level — both are pre-build, read-only checks. Both must reach PASS before Step 4 (Commit) may proceed. Within Step 2, sub-phases 2a → 2b → 2c run sequentially (2b depends on 2a PASS; 2c depends on 2b PASS). Halt at the first NEEDS_REVISION verdict.

Wall-clock note: default flow runs 3 gates simultaneously (lint + security + typecheck in parallel), then build, then tests — roughly max(lint, security, typecheck) + build + max(unit, integration), versus the old 4× sequential gate duration. Typical saving: ~40% wall-clock reduction. Under --thorough, a standalone final-integration review pass is added after all gates complete, and intra-sub-phase Workers serialize (DOCTRINE §12.2/clarification), so the full saving collapses to 2c's unit + integration pair only.

Print Gate <letter> — <name> before each sub-phase.

Step 2a — Lint + typecheck (parallel; no build artifact required)

Three Workers in parallel (P1). None depend on build output — safe to run alongside Step 3.

• Worker A — Detect and run primary linter: npm run lint / pnpm lint / bun run lint / eslint .. On failure: auto-fix via --fix, re-run once; report final error count.
• Worker B — Detect and run formatter check: prettier --check . / biome check . / equivalent. Report diff count.
• Worker C — Root typecheck: tsc --noEmit / npm run typecheck. Skip if not a typed project (per Step 1b tool manifest). Also run per-package typecheck if workspace detected (pnpm/yarn workspaces): iterate packages with tsc --noEmit in each.

Reviewer — aggregate verdict across all three Workers:

• PASS — all clean (or absent/untyped)
• NEEDS_REVISION — any gate fails → halt before 2b. Report which specific gate(s) failed and why. Do NOT proceed to build.
• ESCALATE — config errors preventing execution of any gate

Step 2b — Build gate (sequential; depends on 2a PASS)

Two Workers in parallel:

• Worker A — Production build: npm run build / pnpm build / bun run build. Capture output; report size or artifact path if printed.
• Worker B — Dev/preview build if a separate script exists (npm run build:dev, vite build --mode development, etc.). Skip if no separate dev-build script.

Reviewer — verdict:

• PASS — production build succeeds
• NEEDS_REVISION — production build fails → halt with output
• ESCALATE — build tool absent or script missing (skip silently, not failure)

Step 2c — Test gate (parallel; depends on 2b PASS)

Two Workers in parallel (P1):

• Worker A — Unit tests: run full unit suite per runner from Step 1b (vitest, jest, pytest, cargo test, etc.). Full suite — not just affected. Report count.
• Worker B — Integration / E2E tests if runner detected separately (playwright, cypress, etc.). Skip if no integration runner found.

Reviewer — verdict:

• PASS — all tests pass (or integration absent)
• NEEDS_REVISION — failing tests → halt with failing test names. Do NOT skip. Do NOT increase timeout.
• ESCALATE — runner misconfigured or no tests found and test runner is declared

See quality-gates.md for gate details.

Step 3 — Security Sweep

Runs in parallel with Step 2 at the orchestrator level (P3 — concurrent independent pre-conditions; DOCTRINE §12.2). Both Step 2 and Step 3 are pre-build, read-only checks with no shared state. Both must reach PASS before Step 4 (Commit) may proceed. Halt on SECURITY_VIOLATION immediately — no retry, no 2a must also complete first.

Sub-phases 3a and 3b run in parallel (P1).

Step 3a — Secrets and keys scan

Two Workers in parallel:

• Worker A — Pattern scan staged + recent diff for hardcoded secrets: API keys, private keys, connection strings, tokens. Use git diff HEAD~1..HEAD as scan surface.
• Worker B — File-level scan of files modified in this changeset for common secret patterns (SG., sk-, ghp_, AKIA, BEGIN RSA PRIVATE KEY, etc.).

Reviewer — dispatched as the security-reviewer specialist — aggregate findings from 3a Workers. If any secret found → halt immediately with SECURITY_VIOLATION: <file>:<line> — <pattern>. No auto-remediation — user must rotate + remove. (Add compliance-reviewer when the changeset touches PII / regulated data.)

Step 3b — Dependency audit

Two Workers in parallel:

• Worker A — npm audit --audit-level=high / pnpm audit / pip-audit / cargo audit. Report critical and high CVEs only.
• Worker B — License check: scan new dependencies added in this changeset for prohibited licenses (GPL in a proprietary project, etc.) if .hyperflow/profile.md declares a license policy.

Reviewer — dispatched as the vulnerability-reviewer specialist (deploy is a gated flow → web-research-first on current advisories, proving each CVE's applicability to the pinned versions) — verdict:

• PASS — no critical/high CVEs; no license violations
• NEEDS_REVISION — critical CVE found → halt and surface CVE IDs
• ESCALATE — audit tool absent → skip silently (not a failure); missing license policy → skip

Step 4 — Commit

Atomic — single Worker → Reviewer pair with no parallel angles. Exempt from sub-phase decomposition per DOCTRINE 12.2 atomic exemption.

• Worker-introduced fixes from Step 2 → commit automatically with a conventional commit message.
• Pre-existing user-owned uncommitted changes → use AskUserQuestion to confirm inclusion. Per DOCTRINE rule 8, this is a binary action gate — no recommendation marker:

  Include uncommitted user changes in this commit?
    Include — your local work + the pre-push fixes ship together
    Exclude — commit only the worker fixes; user changes stay local

If the popup UI is unavailable on a portable surface (Codex / OpenCode / Grok), print the same inclusion gate as a Hyperflow Question chat block and wait for the user's answer.

• Never add Co-Authored-By: Claude in commit messages — see git-workflow.md.

Step 5 — Release

Sub-phases run sequentially (5b depends on 5a output).

Step 5a — Release script execution

Single Worker (no parallel angle — single mechanical action):

• Worker — scripts/release.sh exists → run it. release-please / changesets detected → use it. "Nothing to release" or no releasable commits → skip and record Release: skipped.

Reviewer — capture output: new version string (if bumped) or skip reason. Feed version to Step 5b.

Step 5b — Version sync verification

Two Workers in parallel (only runs if 5a produced a new version):

• Worker A — Verify version appears consistently across all manifests: package.json, plugin.json, marketplace.json, any other version-bearing files identified in Step 1b.
• Worker B — Verify CHANGELOG was updated by the release script: check that the new version header exists in CHANGELOG.md (or equivalent). Skip if no changelog file.

Reviewer — verdict:

• PASS — all manifests in sync; changelog updated
• NEEDS_REVISION — version mismatch or changelog missing entry → halt
• (Skip entirely if Step 5a returned Release: skipped)

Step 6 — Push (honors push pre-election from Scope Step 2.6 · STRUCTURAL GATE when push=ask)

Read the push arg from chain args (propagated from Scope Step 2.6 when chain-mode=auto). Three paths:

push=auto — push immediately without asking. Print Push: pre-elected (auto) — pushing branch + tags…. Run git push, then git push --tags if release created tags. Skip the AskUserQuestion call. Per DOCTRINE rule 8, this is NOT an invented skip — the user already gave consent at Scope Step 2.6.

push=never — skip the push step entirely. Print Push: pre-elected (never) — branch held local. Run \git push\ manually when ready. Do not call git push.

push=ask (default; also fires when no operational pre-election was made — e.g. deploy invoked standalone) — fire the structural-gate AskUserQuestion. Per DOCTRINE rule 8, this is a binary action gate — no recommendation marker on either option.

Push to origin/<branch>?
  Push — all gates pass · safe to ship
  Hold — keep local; you can push later

• Never force-push to main or master, regardless of push value. push=auto is a plain git push; if the remote rejects it (non-fast-forward), surface the error and stop — do NOT add --force.
• On yes (or push=auto) — git push, then git push --tags if release created tags.
• If the popup UI is unavailable on a portable surface (Codex / OpenCode / Grok) for push=ask, print the push gate as a Hyperflow Question chat block and wait for the user's answer. If no interactive channel is available at all, hold the push and print Push: held — interactive confirmation required.

Step 7 — Output

── Ship Result ───────────────────
Branch: <name>
Gates: lint pass · typecheck pass · build pass · tests pass (<n> passed)
Security: pass
Commit: <sha> <message>
Release: v<x.y.z> (or skipped)
Push: confirmed (or held)
──────────────────────────────────

On gate failure:

── Ship Result ───────────────────
Branch: <name>
Gates: lint pass · typecheck fail · build skipped · tests skipped
  typecheck: 3 errors in src/auth/middleware.ts
Halted at Step 2a
──────────────────────────────────

Use pass / fail / skipped as plain words — no ✓ / ✗ / — symbols.

Anti-patterns

• --no-verify, --no-gpg-sign, bypassing hooks
• Ignoring failing tests
• Force-pushing to main
• Auto-pushing without explicit confirmation
• Committing Co-Authored-By: Claude

Memory

After successful ship, append to .hyperflow/memory/patterns.md if any new pattern was confirmed during gates. Skip if nothing new.

Doctrine

Full rules in DOCTRINE.md. Output style in output-style.md.

Overview

/hyperflow:deploy runs the pre-push gates (lint + typecheck + security sweep in parallel, then build, then tests), composes any worker-introduced fixes into a clean commit, runs the release script if present, and asks before pushing. Standalone — never auto-invoked from the chain. Push always requires an explicit AskUserQuestion confirmation. Never bypasses hooks, never force-pushes to main, never adds AI attribution to commits.

Prerequisites

• Git repository with a remote configured (for the push step).
• Lint / typecheck / build / test scripts detectable in package.json or via .hyperflow/testing.md. Missing scripts are skipped silently (not failed).
• scripts/release.sh (or release-please / changesets) optional — if present, runs at Step 5; otherwise release is user-managed.
• For security sweep: the security-reviewer and vulnerability-reviewer specialist agents are dispatched. Sweep is mandatory; failure to dispatch = halt.

Instructions

The 7 numbered steps live in Step 1 — Survey State through Step 7 — Output above. Summary:

• Survey state — two sub-phases in parallel: 1a repo-state scan (git status + ahead count), 1b tool detection (package manager, test runner, typed-project flag).
• Quality gates — three sequential sub-phases: 2a lint+typecheck (3-wide parallel Workers, no build artifact needed), 2b build (depends on 2a PASS), 2c tests (2-wide parallel, depends on 2b PASS). Runs in parallel with Step 3 at orchestrator level. Halt at first NEEDS_REVISION.
• Security sweep — runs in parallel with Step 2 (P3, pre-build read-only). Two sub-phases in parallel: 3a secrets/keys scan (security-reviewer specialist), 3b dependency audit. Halt on SECURITY_VIOLATION or critical CVE. Both Step 2 and Step 3 must PASS before Step 4.
• Commit — atomic. Worker fixes auto-committed; AskUserQuestion for pre-existing uncommitted user changes.
• Release — two sequential sub-phases: 5a run release script, 5b verify version sync across manifests.
• Push gate — atomic structural gate. Honors push pre-election (auto/never/ask). push=ask fires AskUserQuestion. Never force-push to main.
• Print structured ship result.

Output

See the ship result block in Step 7 — Output above. Two formats: success (all gates pass, listed inline) and failure (halt at first failing gate, listed in order). Always uses plain words (pass / fail / skipped) — no decorative symbols.

Error Handling

| Failure | Behavior |

|---|---|

| Step 2a — lint fails | Auto-retry once with --fix. Still failing → halt with error count. Do NOT proceed to 2b. |

| Step 2a — typecheck fails | Halt at 2a. No auto-fix — typecheck errors require human eyes. |

| Step 2b — build fails | Halt with build output. Pre-existing build issues likely pre-date the change set. |

| Step 2c — tests fail | Halt with failing test names. Do NOT skip failing tests. Do NOT increase timeout. |

| Security sweep finds secrets | Halt with SECURITY_VIOLATION: marker and the file:line. User decides remediation (revert the secret + rotate the credential). |

| scripts/release.sh says "nothing to release" | Skip release; print Release: skipped (nothing to release). Push step still fires for non-release commits. |

| Push rejected (non-fast-forward) | Refuse to force-push. Print: Push rejected — branch is behind origin. Pull/rebase first. |

| AskUserQuestion popup unavailable (Codex / OpenCode / Grok) | Print the push or commit-inclusion gate as a Hyperflow Question chat block and wait for the user's answer. |

| Headless / non-interactive | Refuse push step entirely. Print structured result with Push: held — interactive confirmation required. |

| Pre-existing uncommitted user changes | Use AskUserQuestion to ask whether to include or exclude from the commit. Default: include. |

Examples

Clean release path

/hyperflow:deploy

Step 2a — Lint + typecheck (parallel with Step 3 security sweep)
Worker A — running lint
Worker B — running formatter check
Worker C — running tsc
Step 3a/3b — security sweep (parallel)
Step 2a Reviewer — all clean
Step 3 Reviewer — no secrets found
Step 2b — Build
Step 2c — Tests (parallel)

? Push to origin/main?
   Push — all gates pass · safe to ship
   Hold — keep local; you can push later

[user picks Push]

── Ship Result ───────────────────
Branch: main
Gates: lint pass · typecheck pass · build pass · tests pass (147 passed)
Security: pass
Commit: dc38564 fix(skills): marketplace validator compliance
Release: v3.1.2
Push: confirmed
──────────────────────────────────

Gate failure halts the pipeline

/hyperflow:deploy

Step 2a — Lint + typecheck (parallel with Step 3 security sweep)
Worker A — running lint
Lint failed: 3 errors in src/auth/middleware.ts
Auto-fix attempted... still failing.
Step 2a Reviewer — NEEDS_REVISION: lint gate failed (3 errors in src/auth/middleware.ts)
Halted at Step 2a. Build and tests skipped.

── Ship Result ───────────────────
Branch: main
Gates: lint fail · typecheck skipped · build skipped · tests skipped
  lint: 3 errors in src/auth/middleware.ts (unused vars, missing return type)
Halted at Step 2a
──────────────────────────────────

Security violation

/hyperflow:deploy

Gates pass: lint · typecheck · build · tests
**Reviewer** — security sweep
SECURITY_VIOLATION: src/config/email.ts:12 — hardcoded SendGrid API key (SG.xxx...)

Halted before commit. Rotate the credential and remove the literal from source before retrying.

Resources

• DOCTRINE.md — orchestration rules (especially #8 push confirmation gate).
• quality-gates.md — full lint/typecheck/build/test policy.
• security.md — security sweep policy and blocklists.
• git-workflow.md — branch/commit conventions, no AI attribution rule.
• output-style.md — ship result formatting.