restarter backlog-workflow (Dynamic Workflow asset)
A Dynamic Workflow asset (see CLAUDE.md -> "Dynamic Workflow Assets"; review-workflow is the reference example). backlog.workflow.js is executed by the Workflow tool, invoked from /lets:backlog's Workflow Mode via:
Workflow({ scriptPath: "${CLAUDE_PLUGIN_ROOT}/skills/backlog-workflow/backlog.workflow.js", args })
${CLAUDE_PLUGIN_ROOT} is substituted at command-load time, so backlog.md carries the literal absolute path. Treat backlog.workflow.js as a template the command points at, not a script to reproduce inline.
Why this is a transparent performance lever (not autonomous)
/lets:backlog review --workflow produces the SAME result as the standard Task path - only the per-agent backlog-idea dumps stay off-context. The explorer scout (Phase 1) and the signal-driven agent selection (Phase 2) stay IN-CONTEXT in backlog.md; the ideate fan-out (Phase 3) and the aggregate/cluster (Phase 4) move into the workflow (off-context). The interactive dialog (Phase 5) + capture (Phase 6) stay in-context after the aggregate returns. There is no Web Research stage: the standard backlog Phase 3 ideates over the project's own profile + CLAUDE.md without web search, so adding web here would diverge the two paths - --workflow must stay a pure context-savings lever, same findings. The cluster is a SEMANTIC dedupe + impact-sort (an off-context merge), not a quality gate.
What it does (off-context)
- Ideate - fan out the selected
lets:<name>agents in parallel over the explorer's Project State Profile + CLAUDE.md (structuredIDEA_SCHEMA: title, category=missing|adjust|theme, impact, description).adjustideas embed the referenced task as Title (task-id) in description. - Cluster (off-context agent + deterministic attribution) - a neutral subagent (no
agentType) groups raw ideas into themes, merging ideas that make the same point even when worded differently (CLUSTER_SCHEMA: themes carrying member indices).buildThemescomputes each theme'sagents[]+ impact deterministically (high if any member is high) and guarantees no idea is dropped. Falls back to title-only dedup (clusterIdeas) if the cluster agent errors or returns nothing; skips the agent for <=1 idea.
args contract (built by /lets:backlog, passed as a JSON value)
| key | type | meaning |
|---|---|---|
profile |
string | the explorer's Project State Profile (in-context Phase 1 output, ~500 words) |
backlogState |
string | the orchestrator's capped BACKLOG STATE block (Phase 1) - the agents' only task data; they have no tracker access of their own |
agents |
[{name}] |
selected agent short-names (no lets: prefix; exclude actor) |
projectRoot |
string | absolute project root (agents must not read outside it) |
claudeMd |
string | CLAUDE.md context (first ~100 lines) |
Returns
{ ideas[], counts{agents, agents_responded, raw_ideas, ideas, clustering} }. ideas are semantically clustered + impact-sorted, each with agents[] attribution - agents[] length > 1 means multiple agents converged on that theme, a first-class signal the command should lead with. counts.clustering is semantic (cluster agent ran) | fallback_title (title-only fallback) | trivial (<=1 idea). The schema carries category (missing/adjust/theme) + impact only - Phase 4's "Gaps" / "Quick Wins" buckets are derived by judgment at render time, not labelled in the schema (same as the standard path). Anti-silent-fail: counts.agents_responded === 0 means every ideator errored - the command must surface "no ideas returned" and offer a re-run, never render a fabricated result.
Constraints (Dynamic Workflow runtime)
- No filesystem - the script returns data;
/lets:backlogrenders + records. - No sibling
import- all logic stays inline. - No
Date.now()/Math.random()/new Date(). - Top-level
await/returnare used, so the file is NOT Node-importable - no committed unit test; the pure logic (buildThemes,clusterIdeas) is category-agnostic and kept in sync withbacklog.mdPhase 4 prose by discipline + a live smoke test.