swingerman/engineer/next

next

The session-start skill — the human's re-entry point. Surveys every source of DAE state and answers one question: what should I pick up now?

Read-only and advisory. It changes nothing, produces no artifact, and emits no handoff (the recommendation is the whole output; a handoff would just restate it). This is a deliberate exception to the agentic-summary contract — next is a query, not a task that changes state.

next is the read-side bookend to session-summary: session-summary writes "here's where I left off, next tasks" at session end; next consumes that — and everything else — at session start.

When to use

Start of a session, or any "what now?" moment. Project-scope; surveys the whole project — there is no narrower scope.

Not for: loading context on a feature you've already chosen (prime-context); checking artifact consistency (consistency-check); a per-feature next-step (each skill's handoff already carries recommended_next).

Workflow

Step 1 — Resolve and survey

Resolve the methodology root + manifest via ${CLAUDE_PLUGIN_ROOT}/scripts/dae_resolve.py (see references/resolving.md). Then read, read-only:

• In-flight features — every features/*/feature.md (status) + progress.md (current checkpoint; whether it's blocked or ready to advance). For each in-flight feature, run ${CLAUDE_PLUGIN_ROOT}/scripts/dae_reconcile.py <feature-dir> (read-only): if it reports needs_reconcile: true, the feature's PR is merged but its local state hasn't caught up — this is the authoritative merged-detection (via gh, so it catches squash-merge and `git.manual`, which the current-branch git-ancestry check below misses). Surface these under NEEDS YOUR DECISION and, at autonomy medium/high, offer /engineer.post-merge to reconcile (which flips the status and propagates). A flag: merged-unverified means it shipped without CP7 — call that out.
• Consolidation backlog — .engineer/consolidation.md if present (coverage backlog + triage order)
• Roadmap (strategic feature list) — the next unstarted roadmap item(s) via the driver (local = ${CLAUDE_PLUGIN_ROOT}/scripts/dae_roadmap.py next-unstarted; MCP/CLI/API-backed = the connected channel — see references/roadmap.md). This is the "what's next in the roadmap" altitude — candidate features not yet promoted to a feature folder, the forward-looking complement to the consolidation backlog. If manifest.roadmap.type is none, or its host is unreachable (MCP disconnected / CLI gone), skip it with a one-line note — degrade gracefully, never error.
• Parked ideas — features/*/ with status: parked; .engineer/discussions.log
• Tracker captures (triage queue) — tasks a human added directly to the onboarded tracker: rows with no Slug yet (the reconcile() orphan-as-intake case, see references/tracker.md), or .engineer/inbox.md lines in local mode. These are quick-captured bugs/ideas/tasks awaiting triage — not yet a feature or fix.
• Pending human actions — recent handoffs/*.md (per-feature and .engineer/handoffs/) flagged human_action_needed: yes
• Session-log next-tasks — the latest session-log.md entry per active feature ("Next tasks")
• Dispatched to cloud — features whose latest handoff carries cloud_session_url (running on a cloud agent; PR pending). Surface as DISPATCHED — not actionable until the PR lands.
• CHARTER.md + manifest — autonomy levels and path overrides, for execution-mode advice
• Open fixes — .engineer/fixes/*.md via ${CLAUDE_PLUGIN_ROOT}/scripts/dae_fix.py list_open_fixes (status != closed)
• Stale merged branch — if the current branch is not main/master, run git fetch origin --quiet then check git merge-base --is-ancestor HEAD origin/HEAD (fallback origin/main). If true, the branch is merged and lingering — surface it as a STALE BRANCH item. (This git check only catches the current branch and only non-squash merges; the per-feature dae_reconcile.py probe above is the broader, gh-based merged-detection.)

Step 1.5 — Offer branch cleanup (if stale)

If a stale merged branch was detected, defer to the post-merge skill — it owns the full cleanup flow (checkout, pull, branch -d, prune, tracker update). At autonomy high/medium, auto-invoke /engineer.post-merge before continuing to Step 2. At low, surface the finding and stop until the user invokes it themselves. Do not inline the cleanup commands here — keep next advisory and let the dedicated skill do the work, so the post-merge handoff lands in the audit trail.

Step 2 — Triage into eight buckets

NEEDS YOUR DECISION   — blocked features; handoffs flagged human_action_needed.
                        These stall progress — surface them first.
OPEN FIXES            — bug-fix artifacts not yet closed. Sub-priority within the bucket:
                         1. blocks_user: true AND workaround: none      ← top (user is hit right now)
                         2. blocks_user: true AND workaround: <text>    ← middle (user has a workaround)
                         3. blocks_user: false                          ← bottom
                         Within each tier, order by severity (critical, high, medium, low).
READY TO ADVANCE      — in-flight features sitting at a checkpoint that can proceed.
READY TO DISPATCH     — features / consolidation tasks that can go to a cloud agent
                        right now (bounded, automatable verification, no mid-stream
                        human input needed). At medium/high autonomy the dispatch
                        router auto-sends these (gated by dae_delegable.py).
DISPATCHED            — already running on a cloud agent (handoff carries
                        cloud_session_url). Show the session / PR link; review when
                        it lands. Not actionable until then.
TRIAGE                — tasks added directly to the tracker (a row with no Slug
                        yet) or to .engineer/inbox.md — bugs/ideas/tasks dropped
                        in for DAE to pick up, not yet a feature or fix. For each:
                        promote (→ /fix, /discuss, /feature-init) or defer. A
                        triage bug that blocks users ranks with OPEN FIXES top tier.
ON THE ROADMAP        — the next unstarted strategic feature(s) — roadmap items
                        (status: planned, no feature_slug yet), ordered by
                        horizon (now → next → later) then priority. This is the
                        "what's next in the roadmap" answer: the highest-altitude
                        candidate to START, distinct from finishing in-flight
                        work. Promote via /discuss or /feature-init (writes
                        roadmap_ref + marks the item in-progress — see
                        references/roadmap.md). Items already in-flight (they
                        carry a feature_slug) are hidden here — they appear under
                        READY TO ADVANCE instead.
COULD START           — the next consolidation-backlog item; parked ideas worth
                        promoting; fresh work.

Step 3 — Rank and recommend

Priority order:

1. Unblocking first — a blocked feature or a pending human-action stalls everything downstream; clearing it usually beats starting something new.

1b. User-blocking defects — an OPEN FIX with blocks_user: true and no workaround ranks above starting fresh feature work; a real defect hitting users now is more urgent than forward progress on new scope. A TRIAGE capture that reads as a user-blocking bug ranks here too — promote it to a /fix first.

1. Triage priority / dates — consolidation-backlog order, feature target dates.
2. Dispatchability — if the human has limited time, favour surfacing what can be dispatched (freeing the human) over what needs them.

Present the eight buckets, then a two-altitude answer — because "what's next" has two valid readings and the human asked for both:

• Next to ADVANCE/finish — the single best actionable next step across in-flight work, fixes, and dispatch (the existing top pick). Unblocking and user-blocking defects win here.
• Next to START — the top ON THE ROADMAP item: the next strategic feature to begin once there's capacity. Always shown so "what's next in the roadmap" is answered directly, even when the top actionable pick is finishing something else.

Give one top recommendation with a one-line rationale and a suggested execution mode (you, local subagent, agent team, coordinated, or cloud agent — for the last, the dispatch router gates via dae_delegable.py). Suggest coordinated (/engineer.run <slug>, see ${CLAUDE_PLUGIN_ROOT}/references/coordinator.md) for a non-express feature at autonomy medium/high that has two or more checkpoints left before CP8. It resumes from the latest complete checkpoint, so a half-done feature qualifies. name the roadmap's next-to-start alongside it. If in-flight work is clear (nothing to advance, no open fixes), the roadmap's next-to-start becomes the top pick.

Promoting a TRIAGE item is the human's call — next only recommends. Work it now → promote to the right unit and start: bug→/fix, idea→/discuss, task→/feature-init (each reads the tracker row as intake, reuses its tracker_ref, and writes the assigned Slug back — no duplicate row). Later → leave it, optionally raise a priority/Status. Drop → set Status: wontfix.

Step 4 — Stop

next recommends; it does not act. The human picks, then invokes the relevant skill themselves (prime-context, discuss, feature-init, a consolidation task, …). No handoff is emitted.

Example shape

NEEDS YOUR DECISION (1)
  • 042-customer-export — plan.md has a charter deviation; approve amendment ADR-012 or revise
 
OPEN FIXES (2)
  • 2026-05-20-login-crash [critical] blocks_user=true workaround=none — login fails on iOS 17
  • 2026-05-18-export-timeout [high] blocks_user=true workaround="retry after 30 s"
 
READY TO ADVANCE (1)
  • 015-image-formats — at Checkpoint 2; ACs done, ready for /atdd:atdd (or /engineer.run 015-image-formats to drive CP3→CP8)
 
READY TO DISPATCH (2)
  • consolidation #1 core-image-generation — full ATDD coverage; remote one-shot
  • consolidation #3 mcp-connector — remote one-shot
 
TRIAGE (2)
  • bug: CSV export drops the UTF-8 BOM | blocks_user=yes — added to tracker, no slug yet
  • idea: dark mode for the settings page
 
ON THE ROADMAP (2)
  • bulk-export [now/p1] — admins export all records as CSV; not yet started
  • dark-mode [next/p1] — settings dark theme
 
COULD START (1)
  • parked: bulk-admin-export — discussed 3 weeks ago; promote if priorities allow
 
→ TOP PICK (advance): fix 2026-05-20-login-crash. Users are hitting a critical crash
  on iOS 17 with no workaround — this outranks new feature work. Run /fix to continue.
  Then approve 042-customer-export's plan to unblock its pipeline.
→ NEXT TO START (roadmap): bulk-export — the top "now" item, unstarted. When the
  crash is cleared, promote it with /discuss or /feature-init.

When NOT to use this skill

• You already know the feature and just need to start → prime-context
• You want full project state validated, not a recommendation → consistency-check --project
• You want to wrap up a session → session-summary

References