Unblock

Objective

Turn .agenda tasks and plans that are blocked only on decisions into ready work. Settle each open technical decision with /decide and lock in its recommendation, ask the user only about decisions a person must own, and update the entries so /pending lists them as ready. This skill decides and edits agenda entries; it does not implement them.

Invoking this skill is the user's delegation of technical decisions to the agent. It overrides /decide's default of only proposing, and /agenda's rule that not-ready entries wait for the user, for the decisions this skill may settle on its own.

Follow /agenda for the workspace layout, filename statuses, the not-ready disclaimer, and plan sections.

Scope

1. Named entries. When the user names tasks or plans, by ID, title, or path, take exactly those. Process a named entry even if it looks exploratory, since naming it is the go-ahead; still apply the human-decision rule below.
2. Otherwise, all candidates. Classify every task with todo or in_progress status and every plan with planned or active status with /agenda-readiness, and take the ones it rates decision-blocked.
3. Skip, without asking, every other class: entries another agent is working on, exploratory or deferred entries, entries waiting on something other than a decision, and underspecified, manual, or already ready ones.
4. Limit to 20. When more than 20 candidates remain, take the 20 most recently created or edited. Order by the last commit touching each file (git log -1 --format=%ct -- <path>), treating uncommitted or untracked entries as newest and falling back to file modification time outside Git. Mention how many were left out.

Stale Claims

Before collecting candidates, release abandoned claims so their entries can be picked up again. This runs on every invocation, whatever the Scope, and covers every task, plan, and roadmap entry that carries an Assignee claim (see /agenda):

1. Check each claimant with /agenda-readiness's Activity rules. Release the claim only when the claimant is stale: no matching session or worktree activity in the last 30 minutes, or no claimant found and the entry file unchanged for 30 minutes. Never release a claim whose agent is still active, and keep the current session's own claims.
2. Remove both claim lines and append a note naming the previous assignee and branch, and recording any stale worktree or session so the next agent resumes it. Keep the entry's ID, slug, and status, so an unfinished in_progress or active entry stays that way.
3. Release the claim in the main checkout's copy, and in the claimant's worktree copy only when that worktree is stale and untouched; otherwise leave the worktree alone and report it.

Decision Types

Classify every open decision in an in-scope entry as technical or human, by /decide's "Human versus technical decisions":

• Technical: decide it with /decide and lock in its recommendation without asking.
• Human-required: still run /decide on it, then confirm with the user before locking it in (see Workflow).

Workflow

1. Collect the candidates per Scope, apply the skips and the limit, and list each in-scope entry's open decisions. Group decisions that one coherent solution settles together, including across entries.
2. Decide with /decide. Run it for each decision or group. Scale its research to the decision: local research is enough for small choices; use web research when the choice turns on external facts, such as an upstream API, a release, or a known issue. Keep this phase read-only apart from the agenda edits below.
3. Lock technical decisions. Update each entry right away, so the results survive an interruption:
   • Tasks: remove the answered items from ## Open Decisions and add each answer to a ## Decisions section with the choice and a one-line rationale. Fold any facts the research established into the description.
   • Plans: record each answer under ## Locked Decisions with its rationale and the rejected alternatives, and update the items it affects, following /agenda-plan-refine's rules for a ready plan.
   • When no open decision remains, remove both the > **Not ready:** disclaimer and the empty ## Open Decisions section.
4. Confirm human-required decisions. Present them together, each marked with ⚠️ and carrying /decide's recommendation, rationale, and main risk, and ask the user with the available question tool, listing the recommendation first. Lock each answer as in step 3. Leave any decision the user does not answer in ## Open Decisions with the /decide findings attached, so the next pass does not redo the research.
5. Do not drop work on your own. If /decide concludes an entry should not be done at all, treat that as a human-required decision instead of closing the entry.
6. Review each edited entry: /agenda-readiness must now rate it ready, manual when its remaining work acts on live external state, or decision-blocked when only human-required decisions remain. It must keep its ID and status and have no unrelated changes.

Report

List every in-scope entry with a status marker, then the skipped entries:

🔓 **Unblocked**

- 🟢 Task 12: Title — ready; decided: <choice>, <choice>
- ✋ Task 14: Title — decided: <choice>; manual: <live step needing you>
- 🟡 Plan 3: Title — waiting on your answer to ⚠️ <decision>
- 🔴 Task 9: Title — still blocked: <reason>

⏭️ Skipped: Task 7 (exploratory), Task 8 (waits on upstream release), 4 older entries over the limit
🔓 Released stale claims: Task 5 (`feature/x`, idle 3 days), Plan 2 (`main`, idle 2 hours)

• 🟢 now ready to implement, ✋ decided but manual because its remaining work acts on live external state, 🟡 waiting only on a human-required decision you were asked about, 🔴 still blocked for another reason found while deciding.
• Then show the entries that became ready in /pending's output format, so the user sees where they now rank among the ready work.
• Show the released-claims line only when a claim was released.
• If nothing was in scope, say so and list the skipped entries.

Boundaries

• Do not implement tasks or plan items; use /agenda-implement for that once the user asks.
• Do not change entry IDs, rename statuses beyond what /agenda requires, or delete entries.
• Do not run destructive, costly, outward-facing, or live system-state operations while researching or deciding.
• Do not commit unless the user asks; then use /git-commit.