fixit is the only variant that can recommend peer variants and write a
fresh advisory report. Every other variant expects an active plan in the
live store already.
Entry point¶
radioactive_ralph run --variant fixit --advise \
--topic m3-completion \
--description "Close the remaining M3 docs and supervisor gaps"
--advise puts fixit into advisor mode. It scans the repo, assesses
what’s known vs. unknown, writes a recommendation report to
.radioactive-ralph/plans/<topic>-advisor.md, syncs that proposal into
the durable plan DAG on first creation for the topic slug, and can
optionally hand off to the recommended variant. Without --advise,
fixit behaves as a bounded ROI worker.
The six stages¶
1. Intent capture¶
fixit records the operator ask from --description, TOPIC.md, or
its fallback prompt context. The goal is to preserve the original ask
before any ranking or decomposition starts.
2. Analysis¶
fixit walks the repo, reads the docs and project metadata, and builds an analysis record of the codebase’s current shape. This is separate from the operator intent so later passes can detect drift.
3. Initial task set¶
Using the intent plus analysis, fixit drafts a first-pass task list and variant recommendation.
4. Confidence threshold + refinement¶
If confidence lands below the threshold, fixit refines the proposal up
to --max-iterations. The point is to get from vague advice to a
report that is actually actionable.
5. Acceptance criteria synthesis¶
fixit synthesizes acceptance criteria for the recommendation so the operator or downstream variant has a concrete definition of done.
6. Report emission, DAG sync, and optional handoff¶
Finally, fixit writes .radioactive-ralph/plans/<topic>-advisor.md.
If this is the first durable plan for that repo/topic slug, fixit also
creates the executable DAG plan. If --auto-handoff is enabled and the
recommendation is unambiguous, fixit launches the recommended variant as
a follow-up run.
Handoff¶
Once the report is written, fixit either:
Prints the recommendation and exits so the operator can inspect it.
Auto-handoffs (with
--auto-handoff), launching the recommended variant immediately. This is the “leave me alone for the weekend” mode.
Re-advising¶
fixit can re-run the same topic after the repo changes:
radioactive_ralph run --variant fixit --advise \
--topic m3-completion \
--description "Re-score the remaining M3 work after the Sphinx pivot"
That re-runs the analysis and refinement loop and overwrites the advisor report with current context. Today it does not rewrite an existing durable DAG plan with the same slug; it refreshes the human report and leaves the stored plan in place.