---
name: humanize-workflow
description: Use adaptive Python workflows when a task benefits from staged work, evidence-driven branching, bounded iteration, independent verification or parallel agents. Select proactively; handle simple tasks directly.
---

# Choose the right amount of workflow

Use Humanize when coordination provides a concrete benefit. The user need not name this
plugin. Handle a simple edit, answer or single check directly. For larger work, select the
control structure from the task: sequence, bounded loop, conditional branch, parallel fanout/
join, nested flows, or combinations. Do not force every task into an actor/reviewer loop.

1. Inspect the objective, applicable AGENTS.md, existing changes and completion evidence.
   Check runtime readiness when unknown. Keep model, permission and budget choices explicit.
2. Search `humanize_library` for applicable curated flows. Read use_when/avoid_when and code;
   success count alone is not a ranking. Candidates are hidden by default. Inspect them with
   include_candidates=true when assessing a potentially useful recipe.
3. Reuse a suitable curated flow, or author one through `humanize-author-flow`. Call
   `humanize_run` with flow_file, entrypoint, agents keyed by the flow's own declared roles,
   and flow_params for its inputs. The two built-in dynamic/review modes remain conveniences.
   They are templates, not the definition of dynamic workflow. For a simple candidate trial,
   allow_candidate=true is an explicit choice, not a claim that it has been curated.
4. The manager adds independent verification and a separate status observer to generic flows.
   The authored Python decides its own internal sequence, loops, branches and concurrency.
   A verifier finding unfinished work produces needs_work; the manager does not secretly
   replace the authored structure with a repair loop. Decide the next bounded action from
   the evidence. No unapproved scope expansion or overlapping workspace writers.
5. Open/share `humanize_dashboard` once. The observer owns narration; runtime events show all
   active role sessions, including parallel work. The execution agents do not maintain the
   page or poll progress. In the parent, wait up to 60 seconds per call; report required
   decisions and final outcomes instead of repeating every stage.
6. Steer a generic flow by naming its task role. Untargeted guidance is stored for the flow;
   arbitrary Python must explicitly read it if needed. Check delivery receipts. Cancellation
   is complete only at terminal state; do not automatically resume a user-cancelled task.
7. Complete plus independent review evidence admits a flow to the private candidate pool.
   For useful recipes, assess reusability and limitations, test relevant control paths using
   humanize_validate_flow, then humanize_curate_flow. Promotion requires at least two passing
   scenarios for the exact code/entry and an applicability assessment. Reject brittle,
   task-bound or redundant candidates; more successful runs alone do not make a good flow.

Keep the original goal, user changes, protected processes and authorization boundaries.
Respect task-specific budgets. Sub-workflows cannot recursively launch this manager.