From Pipeline Conductor
The sequencing contract for the pipeline CONDUCTOR. Drives one change end-to-end: recall/refresh the explorer codebase memory, run the spec-driven builder flow (or BUG-FIX MODE) honoring its own gates, then verify release-readiness with the deterministic release gate (verify-release.sh) and STOP if it fails. Everything is STATUS-driven (.claude/<module>/STATUS.json) so each phase reads the previous phase's machine-readable verdict instead of re-deriving it. Use when conducting /pipeline:run or /pipeline:fix; the conductor coordinates and delegates — it never does the heavy reading itself.
How this skill is triggered — by the user, by Claude, or both
Slash command
/pipeline:orchestrateThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
The `pipeline` plugin is a **conductor**: it does not explore, plan, or implement itself.
The pipeline plugin is a conductor: it does not explore, plan, or implement itself.
It sequences the existing explorer and builder plugins and then renders a release
verdict. Each stage communicates through two durable channels — the STATUS contract
(.claude/<module>/STATUS.json, written by bd_status_write, read by bd_status_read)
and the durable memory / artifacts under .claude/ — so the conductor reads short,
machine-readable state rather than re-deriving anything.
Prime directive — protect your own context window. You coordinate; the sub-flows do
the heavy reading. Delegate via the existing /explorer:start and /builder:start
orchestrators (or their sub-agents), have them write detail to .claude/, and read back
only their STATUS + a ≤12-line summary. Never paste a sub-agent transcript into your
context. Same discipline as builder's start.md.
scripts/pipeline-status.sh / /pipeline:status): read
explorer, builder, and pipeline STATUS and explorer freshness. This tells you
what (if anything) is already done so you can resume rather than restart..claude/explorer/MEMORY.md is missing or
STALE (its explored_commit ≠ current git HEAD), run the explorer flow first
(/explorer:start). If memory is present and fresh, do not re-explore — recall is
cheap, re-exploration is the expensive step the explorer already paid./builder:start for a feature spec; BUG-FIX MODE via
/pipeline:fix → builder's reproduce-first flow for a bug). Do not bypass any
builder gate: clarity ≥ threshold, plan rating ≥ threshold + validate-plan.sh, the
scope guard, per-edit feedback loop, hybrid QA, and (for bugs) the regression gate.
The builder writes its own STATUS, PLAN.md, CHANGELOG.md (with the per-task
edge-case coverage map), QA.md, and — for bugs — BUG.md + bugfix/results.txt.bd_status_write builder <phase> done. Treat a
non-done builder STATUS as not finished — do not advance to the release gate."${CLAUDE_PLUGIN_ROOT}"/scripts/verify-release.sh. It is
pure shell/awk (no python dependency) and checks, against existing artifacts/STATUS:
explored_commit == git HEAD);PLAN.md exists — every ## Tasks item has a
coverage-map line in CHANGELOG.md;BUG.md exists, the bug-fix net is green (repro red→green + characterization/linked
green, from bugfix/results.txt);CHANGELOG.md present and non-empty;.claude/pipeline/RELEASE.md + bd_status_write pipeline release <done|failed>.PIPELINE_ENFORCE=1 or settings.enforce_release=true it exits 2 on any REQUIRED
failure. Regardless of mode, treat a failed release STATUS as a hard stop: do not
claim the change is releasable.RELEASE.md). If it fails: STOP and
report the exact failing checks from RELEASE.md and the single next action to clear
each — never overclaim. Never state "100%".bd_status_read <module> <key> and the artifacts; don't
re-derive what a previous phase already recorded.npx claudepluginhub hafizmirhamza276-lab/backend-agentic-marketplace --plugin pipelineGuides completion of development work by verifying tests, detecting environment, and presenting structured options for merge, PR, or cleanup.
Guides creation and editing of skills using test-driven development with pressure scenarios and subagents to verify agent compliance.
Dispatches multiple subagents concurrently for independent tasks without shared state. Use when facing 2+ unrelated failures or subsystems that can be investigated in parallel.