drive-check-health

name	drive-check-health
description	Run a project-health rollup at one of three cadence points: session-bookend (start / end), per-slice-merge (after each slice's DoD), or trigger-fired (operator-requested or surfaced by drift signals). Surfaces progress, drift signals, dispatch throughput, calibration story, and recommended next pick. Atomic skill — primary output is the rollup. Downstream skill recommendations are surfaced; auto-invoke only in unattended mode with operator policy.
metadata	{"version":"2026.5.28"}

Execution mode: orchestrator-direct. Atomic skill invoked by the Orchestrator directly. Outputs land in projects/<current-project>/rollups/ or the conversation surface. If the body would require running builds/tests or writing files outside projects/<current-project>/ — STOP. Dispatch. See drive/roles/README.md.

Drive: Check Health

A rollup answers five questions:

• Where is the project? Slices delivered / in flight / not started; direct changes done.
• Where is it drifting? Dispatches that failed dispatch-INVEST in flight; slices that failed slice-INVEST (coherence broke down); scope-shift candidates.
• How fast is it going? Dispatch throughput, failure rate, review-round distribution.
• What's the calibration story? Predictions vs actuals; retro-trigger frequency; spike outcomes that re-shaped the plan.
• What should we pick next? Recommended slice / direct change, given dependencies, parallelisation, and operator availability.

The rollup template lives at ./templates/rollup.template.md. Fill it; don't author from scratch.

Cadences

Three canonical cadences (team overlays in drive/health/README.md):

Cadence	When	Mode
Session-bookend	Start + end of a Drive session on the project	Most common in interactive
Per-slice-merge	After each slice merges, before picking the next	Surfaces drift between slices
Trigger-fired	Operator request OR a drive-build-workflow stop-condition OR cron-fired in unattended	On-demand

Skip the rollup if no decisions hang on it. Health checks have a purpose — cadence without trigger value is theatre.

Workflow

Step 1 — Load context

Read drive/health/README.md if it exists. Look for rollup template overlays, drift-signal thresholds specific to the team, pick-next heuristics, common false-positives (signals the team has decided to ignore in this repo).

Step 2 — Gather state

Read:

• projects/<project>/spec.md § Project-DoD (for DoD coverage).
• projects/<project>/plan.md (for slice composition + sequencing).
• Linear Project state (issue statuses for slices + direct changes; use Linear MCP).
• Recent slice retros (projects/<project>/retros.md if present) for context on what's been learned.
• Recent code-review.md round entries for in-flight slices (drift / throughput signals).
• Unattended-mode rollups: also read wip/unattended-decisions.md for decisions the orchestrator made without the operator that the operator should be aware of.

Step 3 — Compute the rollup

Fill the sections of ./templates/rollup.template.md.

Drift signals. Look for:

• Dispatches that failed dispatch-INVEST in flight — outcome fuzzier than the brief named, or scope expanded beyond the brief. Likely sizing miscalibration or under-specified outcome.
• Slices that failed slice-INVEST in flight — coherence broke down (a reviewer can't hold the PR's concerns together; the dispatch sequence drifted across unrelated outcomes). Candidate for promote.
• Slices whose dispatch count exceeded the plan — if actual > 2× planned, flag.
• Failed-dispatch rate — > 30% of recent dispatches required ANOTHER ROUND NEEDED (or worse) → calibration issue.
• Long-running in-flight slice — > N days (threshold in drive/health/README.md) → surface as warning or scope-shift-candidate.

Assign severity per signal:

• Informational — surface but no action needed.
• Warning — pattern emerging; consider retro at next merge.
• Scope-shift-candidate — recommend drive-start-workflow mid-flight for re-triage.

Step 4 — Render the rollup

Interactive: display in chat. Unattended: write to projects/<project>/rollups/<timestamp>.md and emit a notification surface for the operator to read on return.

Emit health-check-fired: Fields: cadence ("opening-rollup" | "per-slice-merge" | "closing-rollup" | "session-bookend" | "trigger-fired" — read from the invoking context; whichever cadence point called this health check), drift_signal_count (integer ≥ 0 — total drift signals surfaced in this rollup), max_drift_severity ("none" | "low" | "medium" | "high" — the highest-severity signal in the rollup), recommended_next (string describing the recommended next pick, or null if none), plus envelope fields (event_id, schema_version: "1", ts, project_run_id, orchestrator_agent_id). See the drive-record-traces skill — events.md § health-check-fired for the payload schema and emission.md § Append protocol for the file-append mechanics.

Step 5 — Downstream skill recommendations (policy-gated)

The rollup is complete after Step 4. This step recommends only unless unattended policy explicitly allows auto-invoke:

Signal	Recommend	Unattended
Scope-shift candidate	drive-start-workflow for mid-flight re-triage	Halt and surface unless policy permits auto-invoke
Retro trigger (not already retroed)	drive-run-retro	Auto-invoke OK
Project-DoD coverage gap	drive-plan-project to amend the plan	Surface
Recommended next pick	Present for operator confirmation	Log and proceed to drive-build-workflow only if operator policy allows

Pitfalls

1. Rollup without action. A rollup that doesn't change what happens next is theatre. Always include "recommended next pick" + any triggers; if neither fires anything, the rollup wasn't worth running.
2. Drift signals reported without severity. Without severity, every signal looks equally urgent; the operator can't prioritise.
3. Cadence followed without trigger value. Per-slice-merge rollups are cheap; session-bookend rollups are cheap. Mid-day cron rollups without a drift signal are expensive — only fire them if the project has a high failure rate that benefits from continuous oversight.
4. Throughput shown without context. "5 dispatches/day" means nothing without the team's normal range. Anchor in drive/health/README.md's baseline.
5. Pick-next ignoring dependencies. If the recommended next pick depends on an in-flight slice, the recommendation is wrong. Always trace dependencies.
6. Long-running in-flight slice silently accepted. If a slice has been in flight for > N days, that's a scope-shift candidate, not just an informational signal.

References

• ./templates/rollup.template.md — the fillable rollup template.
• drive/health/README.md — project-health cadence overlays, drift thresholds, pick-next heuristics.
• docs/drive/principles/sizing.md — INVEST at three altitudes (dispatch / slice / project), referenced by the drift signals.
• skills-contrib/drive-deliver-workflow/, drive-run-retro/, drive-start-workflow/, drive-plan-project/, drive-close-project/ — the skills this rollup recommends downstream.