build can now author the plan it implements
/plan-agent:build used to stop and tell you to go make a plan first.
Now, when you run the command without naming one, it walks you through making
that plan — proposal, plan, review — and then implements what comes back.
The second half of the change is a safety fix: when you ran build bare,
it used to quietly pick up whatever unfinished plan it happened to find. It now
shows you the candidates and asks. Nothing about the assistant's automatic
behaviour changed — the new chain is reachable only when you type the slash
command yourself.
build with no plan now makes oneThe command no longer dead-ends. It asks whether to start from a proposal or go straight to plan authoring, hands off to the skill that owns each stage, and implements the plan that comes back.
The command accepts free text. The first word decides how the argument is
read — a path if it ends in .md/.html or contains a
/, otherwise an objective.
build asks instead of assumingA single unfinished plan used to be adopted with no prompt. It is now presented as an option, alongside None of these — author a new plan. The offer shows at most three candidates and states how many were hidden, because the question widget renders at most four choices.
Flags are stripped before anything is classified, so --dir tmp/plans
on its own is still a bare build, not an objective named
--dir.
The check used to sit just before implementation. On a chained run that meant crossing a whole proposal loop and plan interview before being asked about files you already had open. It now runs first — and it ignores the plan's own spec, its rendered HTML, and any proposal the chain just wrote, since those are not pre-existing work.
A skill's model override lasts the rest of the turn and does not unwind when the skill ends, so without the pin a chained run would write source code on whatever model the last planning skill happened to declare.
When the question tool is unavailable, every gate stops and reports the choice it would have offered. It never picks for you.
Nine new checks covering the chain, the discovery offer, guard ordering, and the argument grammar. Each was proven load-bearing by reverting the step it guards and confirming it fails.
build back, and that nested run is the
one that writes the code. The outer chain then reports its result and stops.| Situation | Before | After |
|---|---|---|
| No plan named | Stopped, told you to run implementation-plan | Enters the chain: proposal → plan → review → implement |
| Exactly one unfinished plan | Adopted it silently | Offers it, alongside “author a new plan” |
| Many unfinished plans | Asked, unbounded list | Offers the newest three, states how many were hidden |
| A free-text objective | Read as a plan path, then failed | Skips discovery, goes straight to authoring |
| Objective whose first word has a slash | Bare list of paths tried | Names the misparse and says to reword |
| A mistyped plan path | Fell through to discovery, could implement a different plan | Stops and names both paths it tried |
| HTML-only legacy plan | Stopped | Still stops — needs its spec reconstructed, not a new plan on top |
| Uncommitted files present | Asked just before implementing | Asked first, before any authoring; plan artifacts excluded |
| Model during implementation | Inherited from the last planning skill | Pinned to Opus |
| Question tool unavailable | Undefined — resolved inconsistently | Every gate stops and reports the options |
| Ambient activation | Requires an existing plan | Unchanged, byte-identical description |
Step 1b makes real Skill() calls to build-proposal and
implementation-plan.
Rejected: copying their logic into build,
which would have produced a second plan-authoring implementation to keep in sync
with the first.
The plan skill's Step 8 already calls back into build, so a chained
run enters build twice.
Rejected: a re-entry flag or guard. Resolving the produced spec by its path lets the existing completed/resume preconditions do the deciding, with no new state to get wrong.
Skill() is synchronous, so by the time control returns the nested run
has already finished.
Rejected: continuing the outer run — it would ask whether to redo work that just completed, or restart a run the user deliberately stopped. This overrode two rows in the original proposal's appendix.
The description frontmatter is byte-identical to main, so
automatic activation keeps its old contract exactly.
Rejected: widening the trigger to catch “build a todo app” — the same widening also catches “build fails on CI” and “build the docker image”. The word is too overloaded.
A named-but-missing path, and an HTML-only legacy plan.
Rejected: chaining on both. Authoring an entire plan because of a typo is worse than stopping, and a legacy plan needs its spec reconstructed rather than a new plan written over it.
--dir is forwarded to plan authoring but not to proposal writing.
It names where the plan goes; build-proposal resolves its own
proposals directory.
Corrected during review: the first pass withheld it from both, which would have written the spec to the default directory and then failed to find it on return.
AskUserQuestion renders at most four options, and one slot is reserved
for None of these.
Rejected: an unbounded list, which would not render at all.
A pull request carries no record of what was tried and abandoned, so this section is thin by construction. Two items survived into the PR description because they came out of live runs rather than review:
add A/B testing support parses cleanly as an objective; only a slash
in the first token misreads the argument as a path.docs/plans/add-plan-authoring-to-build.md.implementation-plan appeared before
build-proposal, and check 13 accepts status: todo and
Implement now anywhere in the section rather than scoped to the Exit/workflow
clauses. Check 11 already does the offset comparison; 10 and 13 do not. Non-blocking, not
addressed.<dt>/<dd>
boundary, leaving mismatched backticks in
docs/plans/add-plan-authoring-to-build.html around line 1808. Cosmetic.plugin.json carries only name and description by
repo convention.model: opus pin, new
argument-hint, and Skill added to allowed-tools.
+163 lines.plan-agent.completed, with the Completion Report and
verification notes.build holding the no-plan authoring chain.implementation-plan: Implement now, Exit, or Run as workflow.build scanning the plans directory for specs whose status is todo or in-progress.build-proposal judges small enough to answer directly, producing no proposal document.