Workflow

Workflow Skill

You are the driver: you compile a milestone (or a ready plan.json) into a run, schedule its tasks into dependency waves, and fan them out as one parallel execution. The orchestrator's full-mode machinery does the per-task work (worktrees, validators, review, merge); you wrap it with a scrum→plan compile, wave scheduling, status mirror-back, and merge-conflict auto-rebound.

Source of truth stays in prove.db. The compiled plan.json is an ephemeral execution view — disposable, regenerated from the milestone.

prove never spawns Claude — you do. Every "dispatch" below is you launching a Claude Code subagent (the Agent tool, or Claude Code's dynamic-workflows fan-out), never an external process or a prove-rendered script. prove only emits artifacts (the plan, the wave schedule, per-task prompts) and the CLI commands those subagents run.

---

Input Resolution

Parse $ARGUMENTS. The first non-flag token is the target:

Target	Meaning	Path
Milestone id (e.g. auth-v1-mp8…)	Compile its tasks + dep-graph to a plan.	Phase 1 (compile), then 2-4
Path to a plan.json	Already executable — skip compile.	Phase 3 (execute) directly
(none)	If exactly one open milestone exists, offer it; else AskUserQuestion listing open milestones from scrum status.	—

A milestone mirrors status back to scrum (Phase 4). A raw plan.json does not — it has no scrum tasks to update.

---

Phase 1: Compile milestone → plan.json

Goal: produce a standard plan.json (the schema the orchestrator already runs) from the milestone's tasks and blocked_by edges. One command does it:

claude-prove scrum compile-plan --milestone <id> --out .prove/runs/<branch>/<slug>/plan.json

This writes the plan.json and a scrum-map.json sidecar ({ "<plan-task-id>": "<scrum-task-id>" }) that Phase 4 uses to resolve each plan task back to its scrum task. Compile rules (handled by the CLI): actionable tasks only (skips done/cancelled); deps[] = in-scope blocked_by predecessors; wave = longest-path depth + 1; mode = full at >= 4 tasks; one step per task. Dependency cycles error out.

The plan is regenerable — to change it, re-run compile rather than hand-editing. For richer per-task step trees, pass --decompose to follow up with /prove:plan per task (else one step/task). <slug> is the milestone slug.

---

Phase 2: Backend selection

Default --backend auto:

Backend	When	How
dynamic	Dynamic-workflows preview available (Opus 4.8, Max/Team).	Launch a Claude Code dynamic workflow that executes the wave-plan schedule, fanning per-batch task subagents out in the background. The session stays responsive; plan state lives in prove.db, not the context window.
native	No dynamic-workflows preview.	Run the Phase 3 loop in-session: dispatch each batch's task subagents with the Agent tool (run_in_background: true), as orchestrator full mode already does.
auto	Default.	Detect and pick dynamic, else native.

Both backends drive the same claude-prove commands and the same wave-plan schedule — only the fan-out mechanism differs (dynamic workflow vs. Agent tool). Don't add an abstraction layer over that shared vocabulary.

---

Phase 3: Execute (delegate to orchestrator full-mode)

Run the compiled plan through the orchestrator's full mode — do not reimplement worktrees, dispatch, validation, review, or merge.

claude-prove run-state init --branch <branch> --slug <slug> \
  --plan .prove/runs/<branch>/<slug>/plan.json

Compute the dispatch schedule once, up front:

claude-prove orchestrator wave-plan --run-dir .prove/runs/<branch>/<slug> --max-agents <n>

It returns the waves split into batches capped at --max-agents, plus dispatch_rounds and peak_concurrency. Dispatch each batch in order; fan out in parallel within a batch. This is the scheduler both backends share — no ad-hoc ordering.

Then drive the standard full-mode loop (skills/orchestrator/SKILL.md, "Full Mode") over those batches: create a worktree per task → launch one general-purpose subagent per task, each prompted via claude-prove orchestrator task-prompt --run-dir <dir> --task-id <id> → run validators → principal-architect review loop → sequential merge-back. As in orchestrator full mode, subagents record typed findings, commit, and exit; you own every step and scrum write — including the findings backstop: when a worker's handoff message reports findings missing from the reasoning log, transcribe them as typed hack/risk/decision/assumption entries before writing synthesis, so milestone-close curation can sweep them.

One delta this skill applies beyond the schedule:

• --verify <tag>: tasks carrying <tag> always run the adversarial principal-architect review (refute-until-approved), even if global review is off.

---

Durable execution directives (execution block)

A plan task carries an optional execution block of declarative directives the driver honors and the run record persists. The engine records them; the driver executes them — they are control flow, not judgment.

Directive	Shape	Driver behavior
retry	{ max: N }	On a task's terminal failure, re-dispatch it up to N times before halt-and-drain. Rebuild on the current integration HEAD (the same reset path a merge-conflict rebound uses).
loop	{ max_iterations: N }	Repeat the task body until its own exit condition or N iterations — N is the runaway floor, NOT a target; the body decides early exit.
fanout	{ batch_size: N }	Fan the task's sub-work out N-wide; split larger sets into sequential batches at the cap (the same batching wave-plan applies per wave).
on_fail	<task-id>	On terminal failure, branch to the named task instead of halting the branch. Absent = halt-and-drain (the default).
concurrency	parallel | singleton	singleton = at most one in-flight instance of this task across the run; a second dispatch waits for the first to reach a terminal state. parallel = no limit. (A story-close task runs singleton.)

Absent block = run-once, no retry, no loop, fan-out 1, halt-on-fail, parallel — the pre-directive behavior. The directives compose: a retried task that still fails takes its on_fail branch; a singleton task's retries never overlap. Because they live in the durable run record, a re-dispatch after a session break reads the same directives — the run's retry/loop/fanout/branch policy survives the handoff rather than resetting.

---

Cross-team step: kind:<team-slug>

A plan step whose kind names a team slug (rather than a normal implementation step) is a request to another team — it delegates the work to that team and waits for the team's published output. This is sugar over the cross-team ask protocol: it files an ask, lets the responding team triage it, and collects the team's exposed outputs once the delegated work is done. You compose three CLI primitives — you do not reimplement the ask flow.

When you hit a kind:<team-slug> step, drive this loop. The judgment (the triage verdict) belongs to the responding team's driver; the polling is mechanical.

1. File the ask. The blocking artifact is the step's own task — it stays blocked until the other team delivers.

   claude-prove scrum ask file \
     --from-team <this-team> --to-team <team-slug> \
     --ask-type <type> --blocking-artifact <this-step-task-id>
   

   The final stdout line is the new ask id. Capture it.

2. Let the responding team triage. The to-team's driver decides accept / reject / counter and applies the verdict with claude-prove scrum ask respond <ask-id> --verdict accept|reject|counter [--comment …]. On accept the store creates one child task under the to-team and wires the dependency automatically.

3. Poll for resolution. Re-run the mechanical primitive until it reports a terminal phase (terminal: true in its JSON). It spawns no model and never mutates — it just derives the current phase:

   claude-prove scrum ask await <ask-id>
   

   phase	Meaning	terminal	What the step does
   pending	Filed, not yet triaged.	false	Poll again later.
   waiting	Accepted; the delegated child task is not done yet.	false	Poll again later.
   ready	Accepted and the child reached done.	true	Read outputs (the to-team's published outputs) and continue the plan.
   rejected	The team declined; reason carries why.	true	Surface and stop — do not loop.
   countered	The team proposed an alternative; reason carries it.	true	Surface and stop — re-plan or re-file against the counter.

4. Collect the outputs. On ready, the report's outputs array is the responding team's published outputs (its active exposed interface) — that is the value the step returns to the rest of the plan.

Reject and counter never hang. They are terminal phases the poll surfaces explicitly, so a kind:<team-slug> step that the other team declines or counters resolves the step with a visible result the calling plan can act on — it never spins waiting for a delivery that will not come. Treat a rejected/countered step like a halted task: surface the reason, stop polling, and re-plan (drop the step, narrow scope, or re-file against the counter) rather than blocking the wave.

---

Phase 4: Mirror status back to scrum (milestone target only)

Resolve <scrum-id> from scrum-map.json (Phase 1). After each task reaches a terminal state, write its outcome:

Task outcome	Scrum write
Approved + merged	claude-prove scrum task status <scrum-id> done, then link the run: claude-prove scrum link-run <scrum-id> .prove/runs/<branch>/<slug> --branch <branch> --slug <slug>
Halted / failed after retry	claude-prove scrum task status <scrum-id> blocked — do not mark done. Its dependents stay blocked.

(link-run takes the task id and run path as positionals — both required.)

When no ready tasks remain, emit a milestone summary: completed / blocked / skipped counts and the blocked subtree, if any.

---

Flags

Flag	Default	Effect
--backend auto|dynamic|native	auto	Execution substrate (Phase 2).
--max-agents <n>	16 dynamic / 4 native	Per-batch fan-out ceiling; wave-plan splits oversized waves into sequential batches at this cap.
--verify <tag>	off	Force adversarial review on tagged tasks.
--decompose	off	Run /prove:plan per task for multi-step trees (else one step/task).
--max-rebounds <n>	2	Merge-conflict rebound attempts per task before halt-and-drain (Guards).
--dry-run	off	Compile + print the DAG, wave plan, and agent-count/cost estimate. Write nothing, dispatch nothing.

---

Guards & failure handling

• --dry-run before any large run. A milestone can spawn up to the dynamic-workflows ceiling (1000 agents/run) — print the projected wave plan first and dispatch nothing: claude-prove orchestrator wave-plan --run-dir <dir> --max-agents <n> --format md.

• Halt-and-drain. A failed task halts its branch only: independent branches keep running, dependents stay blocked, and the run reports partial completion. It does not wedge the whole milestone.

• Merge conflict → bounded rebound. Rebuild the task on the updated integration HEAD and retry, instead of wedging the run. git merge --abort, then, up to --max-rebounds (default 2):

  1. claude-prove worktree reset <slug> <task-id> — resets the task worktree to integration HEAD, discarding its commits and picking up already-merged work.
  2. Re-dispatch the task (task-prompt subagent → validators → review), then retry the merge. Rebuilt on the merged base, the retry fast-forwards instead of re-conflicting.

  When the rebound budget is spent, fall back to halt-and-drain: keep merging independent branches, and for a milestone target mark the task blocked via Phase 4. The rebound count is tracked per task, reset each run.

• Plan-only target. Skip Phases 1 and 4 entirely; just init + execute.

• Stopping a wave → the Layer-1 interrupt floor. To abort in-flight tasks, do not wait on cooperation: scrum task cancel <id> --cascade → run-state init --overwrite → re-dispatch, with the /workflows token budget + subagent timeout as the hard stop. See skills/orchestrator/SKILL.md → "Interrupting a run — the Layer-1 floor".

• Stopping a wave gracefully → the Layer-2 cooperative checkpoint-interrupt. When you want in-flight work preserved rather than discarded, raise a cancel-flag instead of a hard abort: write a CANCEL file under the run dir (.prove/runs/<branch>/<slug>/CANCEL, resolved from the main worktree). Task subagents poll it at natural checkpoints — when set, each writes a synthesis graceful-handoff entry (claude-prove acb log append), commits work-in-progress, and self-exits, so a re-dispatch RESUMES from the handoff. Clear the flag (rm -f .../CANCEL) before re-dispatching. This is best-effort and layers ON TOP of the Layer-1 floor — it never replaces it: a non-polling or stuck task only stops at the token budget / subagent timeout, so Layer 1 stays the backstop. The worker-side protocol ships in the prompt emitted by claude-prove orchestrator task-prompt.

---

Notes

• Phase 1 compile = scrum compile-plan; Phase 3 scheduling = orchestrator wave-plan; rebound reset = claude-prove worktree reset.
• Cross-team kind:<team-slug> step = scrum ask file → (responder scrum ask respond) → poll scrum ask await until terminal. await is the mechanical primitive; the triage verdict is the responding team's judgment.
• For large milestones, run the session at high effort (xhigh/ultracode) — the research recommends it for extended async fan-out.