release-next

name	release-next
description	Recommend the next startable release work from the Marshall store — across every release in flight, or within one. Use when the user says /release-next, what is next, what should I work on, or I have a time budget.

Release Next (Marshall)

Recommend 1–3 startable components, reading the shared Marshall store through its MCP tools instead of a repo-local release-plan.json. This is read-first; it never claims work.

It is the proof of the lift — startability is computed server-side from the dependency graph and live claims, so two people (or an agent and a human) on different machines see the same truth.

Several releases in flight is normal, not an edge case: claims are held per component, so teammates on different releases never contend. So "what should I work on?" is answerable without naming a release first — scope the question to whatever the user actually gave you.

How it talks to the store

• mcp__plugin_marshall_marshall__release_next — startable work, ranked by what it unblocks
• mcp__plugin_marshall_marshall__release_status — who holds what + drift
• mcp__plugin_marshall_marshall__release_get — the full release document

If those tools aren't available (the MCP server isn't connected), fall back to the marshall CLI: marshall next --release <version> --json. Same data, secondary path — but it is per-release only, with no cross-release scan, so on that path you do have to ask the user which release.

Procedure

1. Scope it from what the user said. Never ask for a release they didn't name.
   • Named a release → { release: "<version-or-slug>" }. Add project only to disambiguate the same version across projects.
   • Named a project → { project: "<slug>" } — its unshipped releases.
   • Named nothing ("what should I work on?") → {} — every unshipped release in the workspace. This is the default for an open-ended ask.
2. Call mcp__plugin_marshall_marshall__release_next.
3. Render the result:
   • startable non-empty → recommend the top 1–3 (already ranked).
   • startable empty → report what's blocking. Read releases[] for the per-release breakdown — which release is stuck and why — not just the merged blocked_summary, which cannot say where the problem is.
   • Tool error → surface it verbatim. Do not paper over it.

Reading the response

• startable[] — ranked. Each carries its release ({version, slug, project}) and unblocks (how many components it frees).
• blocked_summary — counts by reason, merged across the scope (e.g. { "unverified": 1 }, { "blocked": 2 }).
• releases[] — one entry per release considered, each with its own startable count and blocked_summary.
• scope — release:<version> | project:<slug> | workspace.
• release — the version when you named one, else null.

Ranking across releases is a proxy, not gospel. unblocks is counted within each release's own graph, so unblocking 5 in a 40-component release is not the same currency as unblocking 3 in a 4-component one. When the list spans releases, weigh the ranking against what is actually near shipping, and say that you are — the server has no basis for that judgement, so it does not pretend to.

Errors

Surface them verbatim. Each is recoverable in a specific way:

• release_ambiguous — the key matched more than one release. candidates[] names each one's project + slug; retry with project set to a candidate's project and release set to its slug — that pair is unique. (project alone is enough when only one candidate is in it.) Never retry with the same bare key.
• release_not_found — nothing matched. Ask; don't guess another key.
• project_not_found — that project isn't in this workspace.
• An auth/connection problem — report it. Don't work around it.

Output

For each recommendation (already ranked by what it unblocks):

• tracker ref + title (e.g. #42 Audit trace CLI)
• its release — whenever the list spans more than one. A cross-release list is useless if the reader can't tell what belongs where.
• why it's startable now (no unmet blockers, graph verified, unheld)
• what it unblocks (its position on the critical path)

End with:

• start <ref> to begin work — claim it with mcp__plugin_marshall_marshall__claim_component
• to take more than one, claim each with /marshall:release-topic, or dispatch a whole wave concurrently with /marshall:release-parallel

Guardrails

• Do not claim, heartbeat, or release here — recommending is read-only. The claim lifecycle is a deliberate next step.
• Do not invent startability. If release_next returns nothing, the answer is "nothing is startable, here's why" — never relax the server's verdict.
• If a graph is unverified, say so plainly: that release needs /release-graph before work is safe to start (the store's fail-safe default).
• Scanning covers unshipped releases only — a shipped release has no work left to recommend. Naming one explicitly still reports on it.