Phased Task Loop

A disciplined workflow for implementing multi-phase coding tasks. tasks.md tracks progress in real time; before each phase, a short code check finds where the plan is wrong about the code, so each phase builds from current facts.

Arguments: $ARGUMENTS — the plan (a .plans/<task>/ path or slug). Always run the whole plan.

Picked by pick-tier when the tier is L and the change touches a sensitive path or has more than 3 phases.

Read task-loop-rules.md in this skill's folder first. It holds the project commands, constraints, concepts, task-list discipline, and finishing steps all three task loops share. This file adds only what is specific to this loop.

Orchestrated mode: read ORCHESTRATED.md before step 1. It assigns each loop step to you or a subagent; the steps below still define what each must achieve. A subagent whose prompt says to run solo does every step itself instead, skipping ORCHESTRATED.md. Without that instruction, a subagent running this loop (like ship-plan's build helper) is the orchestrator.

Constraints (on top of the shared ones)

• Sync before the next phase. Run sync at the end of every phase.

Resolve every judgement call yourself. The rules file says when one needs a divergence note.

Concepts (on top of the shared ones)

Code check — The file check-N.md: where this phase's part of the plan is wrong about the code, one line each, with what the plan assumes, what the code does (file:line), and what you'll do instead. Not a how-to: no code snippets, no file-by-file plan.

- Plan says the picker reloads on save; it reloads only on focus (src/picker.rs:212). Added a task to reload on save.

The Loop

Repeat until every task is complete. Work per phase, top to bottom.

1. Check status. Read tasks.md. Every task [x] → go to When you are done. Otherwise start at the next task.
2. Check the plan against the code (skip if this phase already has a check-N.md). Re-read this phase in plan.md, then read the code it touches and the library source it calls into (the installed copy, not the docs). Write .plans/<task>/check-N.md, one line per place the plan is wrong or unclear. If the plan holds, write one line saying so and what you read. Then act on each line:
   • Fix tasks.md to match: add, cancel or move lines.
   • A fix that changes something the user approved → divergence note in plan.md.
   • Something reading can't settle (how another program behaves, say) → a line in risks.md.

3. Implement, one task at a time:

   a. Restate the current task in one sentence.

   b. Do the work for that task. Run check after each logical group of changes; if it fails, fix and re-run.

   c. Tick it in tasks.md.

   d. Pick the next task under the current phase header and repeat from (a). When the phase has no - [ ] left, go to 3.5.

3.5. Audit. Re-read tasks.md. Any task in this phase not [x] means you drifted: return to step 3 with it. The list must be clean before you test or commit.

4. Test. Run the full test suite; all tests pass. Fix any failure before going on.
5. Commit and sync. commit with a Conventional Commit message, <type>(<scope>): <summary>, one per phase, with the phase named in the summary or body, then sync (resolve conflicts, re-test, commit the merge). Commit only green code.
6. Go to step 1.

When you are done

Follow Finishing in the rules file.