--- updated: "2026-10-02T23:16:00Z" source_commit: "0d098c9e06d1" update_event: "review_refresh" context: "changes=XL files=30" description: "Preserve complete caller-owned scope across answerless audit failures" --- # Small workflow starters Begin with the sequential implementation/review module in either [authoring lesson](../../../../../docs/workflows/create.md#choose-an-authoring-route). These modules deepen reusable interactions. After the dependency lesson, choose from task need using the [semantic approach guide](../../../../../skills/locus-pi-workflow-create/references/agentic-approaches.md), then adapt the prompts, criteria and finite allowances. They are starting points, not compulsory templates or runnable names in the Package catalog. | Source | Task need and mechanism | Terminal result | | ----------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------- | | [Evaluator-Optimizer](evaluator-optimizer.workflow.mjs) | Known implementation scope; reviewer-owned findings and choice, one correction followed by recheck | Accepted short status/path after artifact review, or non-success with the preserved implementation and findings files | | [Project tour](project-tour.workflow.mjs) | Optional independent file writers, then one merge owner | Exact explanatory guide in `guide.md`, without independent product acceptance | | [Caller audit](caller-audit.workflow.mjs) | Caller-owned per-source pipeline, then complete synthesis and coverage review | Exact reviewed `audit.md`, or incomplete candidate/findings; empty required items are non-success | | [Plan/replan](plan-replan.workflow.mjs) | Observed results change remaining work; planner replaces named files and executor reads the next step | Evidenced delivery account in `delivery.md`, or non-success with plan and latest execution | | [Reflection](reflection.workflow.mjs) | Improve a low-risk explanation through draft, critique and revision | Editorial revision in `document.md`; no independent acceptance claim | | [Parallel investigation + Reflection](parallel-reflection.workflow.mjs) | Independent evidence and reader investigation before synthesis, critique and revision | Editorial revision in `document.md`, preserving both investigations in its handoffs | The primary loop uses the [shared working context](../../../../../skills/locus-pi-workflow-create/references/design-and-build.md#folder-level-context): original Task, verified checkout/branch, delegated product root, and an explicit orchestration/evidence folder. Only the developer changes product files. It writes `implementation.md`; the reviewer reads it and writes only `findings.md`, then returns the actual DSL choice. Corrections reopen the files; source passes short statuses, never full review transcripts. Each correction requires fresh review, including the final allowed one. Project tour is secondary. It uses author-known README/package sources, separate assigned `purpose.md` and `commands.md` artifacts, and a single composer assigned `guide.md`. The composer waits for both writers and reopens their files. Parallel workers never share an output path. Other semantic input is one string: supply the request, required evidence and source locations for agents to inspect. Caller audit additionally requires structured tool `items`; slash syntax cannot supply them. Its callbacks receive previous value and composite index, not an original-item argument. Caller audit also passes all complete original units to verification, synthesis and coverage review. A provider failure may have no answer or source identity; retain that unassigned failure and compare required coverage with the original units rather than guessing a source from a slot. File handoffs such as `findings.md`, `plan.md` and `next-step.md` use exact caller-assigned paths in the whole input. The final guide/audit/delivery files are also written directly by agents; optional text snapshots do not attest their existence. Workflow source never reads or parses them. Choose a fresh workspace for an independent run; a mentioned path does not change the child project directory. All calls inherit configured model routing. Independent sessions may use the same model; the examples make no claim of measured model quality. Use [Design → review → Build](../../../../../skills/locus-pi-workflow-create/references/design-and-build.md) to adapt a starter into `.locus-pi/workflows//.workflow.mjs` with a matching `meta.name` and reviewed design. Check its exact bytes before running: ```text node --check npm run check:workflow-source -- --mode orchestration-only ``` The second command is available from a checkout. Installed Pi provides `workflow_check_source` with `mode: "orchestration-only"`. These checks establish syntax and supported source shape; they do not run children or certify judgment, file contents or task completion. Follow the [create guide](../../../../../docs/workflows/create.md) and [run guide](../../../../../docs/workflows/running.md) for the full authoring/launch route. Keep the deeper references when their mechanisms are needed: [fixed](../fixed.workflow.mjs), [refinement](../refinement.workflow.mjs), [adaptive design](../adaptive-design.workflow.mjs), [adaptive slices](../adaptive-slices.workflow.mjs), [caller-supplied decomposition](../decomposition.workflow.mjs), and [human continuation](../human-continuation.workflow.mjs). The last is a reviewed compatibility example; the six starters use orchestration-only source.