humanize1
The Claude Code plugin the rest of humanfia grew out of, proposed at UCLA PolyArch, as three flows you run one after another: gen-idea opens a loose idea into a draft, gen-plan turns the draft into a plan two agents agreed on, and rlcr builds that plan under review until nothing is left to say. Each phase is a run of its own, and what passes from one to the next is a file you can read and edit before going on. Its loop, RLCR, is a Ralph loop with an independent code review every round: the reviewer arrives with no memory, reads the repository, and its findings are what the builder hears next.
In hmz, open /flow → Flowverses → official and install humanize1. Or run the release without installing anything:
hmz exec -f 'git+https://github.com/humanfia/humanize1-flow@v0.2.0#humanize1:rlcr' …- repository
- humanfia/humanize1-flow
- release
- v0.2.0 · commit
59e1ff7 - earlier
- v0.1.1 · v0.1.0
- licence
- Apache-2.0
Read from humanfia/flowverse when this site was built.
1 · gen-idea
One agent explores several directions for the idea and writes a draft, with no code.
The whole run, at rest. Step through it with the buttons, or drag the bar.
One turn, one file: a main direction and the rest as alternatives, and no code. The ticks are the six directions, each explored against the repository by a read-only helper inside the drafter’s own turn.
The run, turn by turn
- drafter — undo and redo: 6 directions; a session opened for this turn; hands the finish in the tree: ideas/undo-redo.md
It ends when the draft is written, or the budget runs out.
2 · gen-plan
A planner writes the plan and an analyst reviews it, fresh each time, until they agree.
The whole run, at rest. Step through it with the buttons, or drag the bar.
The side that writes remembers and the side that reads does not — the rule all of humanize1 is built on.
The run, turn by turn
- analyst — about this repo?; a session opened for this turn; hands analyst as words: yes: src/editor
- analyst — lists the risks; a session opened for this turn; hands planner as words: 3 risks
- planner — writes the plan; a session opened for this turn; hands analyst in the tree: docs/plan.md
- analyst — reviews the plan; a session opened for this turn; hands planner as words: 2 objections
- planner — revises it; another turn of the session it already had; hands analyst in the tree: the plan, revised
- analyst — reviews it fresh; a session opened for this turn; hands planner as words: agreed
- planner — finalizes it; another turn of the session it already had; hands the finish in the tree: docs/plan.md
Then round again: up to 3 rounds, until the two agree.
It ends when the plan is written, the draft is not about this repository, a decision is left PENDING, or the budget runs out.
3 · rlcr
A builder works through the plan; a fresh reviewer checks each round, then the whole change.
The whole run, at rest. Step through it with the buttons, or drag the bar.
The builder works until it believes the whole plan is done; the round’s checks run, and what the reviewer finds is what the builder hears next, in the same session. Once the reviewer finds nothing left, a code review of the whole change has to come back clean too.
The run, turn by turn
- you — 2 questions on the plan; no turn of a model; hands builder as words: answered, or skipped
- builder — builds undo/redo; a session opened for this turn; hands the gates in the tree: believes it is done
- the gates — 15 stop checks; no turn of a model; hands reviewer as words: passed
- reviewer — reviews round 1; a session opened for this turn; hands builder as words: [P1] redo after edit
- builder — fixes the [P1]; another turn of the session it already had; hands the gates in the tree: believes it is done
- the gates — 15 stop checks; no turn of a model; hands reviewer as words: passed
- reviewer — says COMPLETE; a session opened for this turn; hands reviewer as words: nothing left
- reviewer — code review; a session opened for this turn; hands the finish as words: clean
Then round again: a refused check or a finding is the builder’s next prompt.
It ends when complete: nothing left, code review clean, max rounds, or stop, or the budget runs out.
See also: rlar, the same loop without a plan, and recursive_lean_prover, built out of gen-plan and rlcr.