Questions before a task
A dwarf can't stop and ask you anything mid-run. Three opt-in features close that gap: brief mode asks first, task summaries keep you informed, and retro notes tell you what got in the way.
Brief mode
Before each task starts, a cheap read-only model reads the plan and writes the questions only you can answer, each with why it matters, a few options, and the default forge will assume if you don't answer. Your answers go into that task's prompt as decisions you made.
/forge "add auth: tokens, middleware, login" --decompose-level medium \ --profile balanced --brief gated --brief-model haiku:medium
| Level | What waits |
|---|---|
advisory | Nothing. Unanswered questions reach the prompt with their defaults. |
gated | A task with an unanswered must-answer question is held until you answer. The rest run. |
strict | The whole run waits while any question of a task that hasn't started is open. |
forge brief show "$PLAN" # open questions, flagged first
forge brief answer "$PLAN" login-route q1 "one message for both cases"
You can also answer in forge-ui's inspector. Answer while earlier waves run: an answer that lands after its task started is kept as late, and forge prints the command to queue a follow-up. When a merge changes files a waiting task depends on, that task's open questions are flagged as possibly stale. forge never answers for you and never picks the brief model.
Questions for a wave or the whole plan
Some questions cover more than one task. With --brief-scope wave or plan, those are answered once, and the answer becomes a proposal. A wave answer may only change that wave's tasks that haven't started; a plan answer can also add, remove or split tasks. Nothing changes until you approve the exact diff by its hash, and only the tasks it affects get their questions asked again.
forge brief proposals "$PLAN" # what each would change forge brief apply "$PLAN" p1 --approve 3f9c2a # that exact diff forge brief reject "$PLAN" p1
A wave is a group of tasks with a title and a goal. The planner writes them; for groups forge derives from the waves, forge group set gives them yours.
Task summaries
A few plain sentences per task, plus one line for the task table, written by a cheap model from the task's facts only: its status, attempts, verdicts and findings. It never sees your code. A summary is rewritten only when those facts change.
/forge "..." --decompose-level medium --summaries on --summaries-model haiku:low forge control set "$PLAN" summaries on # switch it during or after a run forge control set "$PLAN" summaries off --task docs # or per task
forge-ui shows each summary in the task inspector with its model, time and sources, and marks one out of date once the task has moved on.
Retro notes
With --retro (or forge config set retro on), builders and reviewers note what got in their way: a missing tool, a flaky test, a sandbox that blocked a command, a wrong instruction. Each note carries its command and error and is checked against the run's own logs.
forge retro report "$PLAN" # retro.md: forge's problems, your project's, the environment's forge retro issues "$PLAN" # drafts redacted issues for forge's own repository
Nothing leaves your machine unless you approve a draft by its hash. The notes are about your setup as much as forge's: a missing dependency or a slow test lane shows up there too.
Full reference Every flag and edge case: references/decompose.md