wsr

Next inspection iteration: manual dispatch input definitions

Status: owner-authorized; implemented locally, integration pending. Baseline: PR #25 merged as 91fd3d5; literal runner-label lists are delivered. Owner authorized implementation; schedule settings remain unchanged.

Outcome

Inspect the manual-trigger mapping and input declarations used by both real .github/workflows/example-*.yml files. Report declared inputs, without selecting an event, binding values, applying defaults or evaluating expressions. Preserve the existing scalar on: workflow_dispatch behavior and output, and continue accepting fragments without on.

A representative excerpt:

on:
  workflow_dispatch:
    inputs:
      mode:
        description: Choose the reference outcome
        type: choice
        options: [pass, fail]
        default: pass
        required: true

Before: exit 1 because on expects a scalar string. Proposed successful output includes:

Trigger: workflow_dispatch (not evaluated)
  Input: mode
    Type: choice
    Description: Choose the reference outcome
    Required: true
    Default: "pass" (not applied)
    Options: ["pass", "fail"]

Keep the existing workflow header/disclaimer and job/runner/step/needs output unchanged. Order input identifiers deterministically; preserve option order and repetitions. Quote/escape user-controlled values. Render only explicitly declared metadata: omitted required/default/type values must not be invented or applied. State absence of a declared type if needed for clarity.

Bounded syntax and model

Use the existing trigger modeling guidance. At implementation start, verify supported shapes and validation rules against the pinned official workflow schema/converter already referenced there, and record the exact reference used. Do not infer GitHub rules solely from the examples or this output sketch.

Keep private GHA source decoding/validation in wsr-gha, provider-neutral inspection summaries in wsr-types, and file IO/rendering in the CLI per D-031. Extend the current trigger summary minimally rather than adopt legacy executable IR, add a generic event engine or introduce crates. No event binding, expression evaluation, required-input enforcement at invocation, environment lookup, secret access, action fetching or workflow execution. No runtime TypeScript sidecar.

Acceptance and delivery

  1. Refresh live Git/PR/checkpoint state; baseline focused CLI/inspection tests. Preserve unrelated work; one feature PR at a time in the existing checkout, based on current origin/main.
  2. Add one real CLI test using a dispatch mapping with the choice input from the examples. Observe the current string-shape failure, then minimally implement parsing, summary and output.
  3. Add empty/null dispatch bodies and missing/explicit metadata cases. Add invalid shapes, duplicates, unsupported events/fields and dynamic-value diagnostics incrementally; observe meaningful red before each new accepted behavior or validation rule. Existing rejections count as regressions, not behavioral red. Verify pinned type/default/choice rules through real CLI tests.
  4. Cover all five declared types and omitted type, deterministic input order, original scalar types, escaped text, option order, no partial output on failure, and unchanged scalar-trigger/no-trigger behavior. Add public parser-summary assertions for declaration presence and typed defaults.
  5. Retain runner/needs/step-ID regressions, malicious scripts and source/config sentinels. No command runs or repository mutation. Use temporary files and ordinary Cargo integration tests.
  6. Update CLI/status/testing evidence only for working behavior. Run focused tests, cargo xtask ci, Markdown target checks and git diff --check; review actual diff, then commit/push/create PR.
  7. Merge only the verified head after required checks pass and no unresolved review findings. Report captured old/new CLI and parser behavior from the merged implementation.

Use extracted representative workflow fragments for this slice; do not edit the canonical examples or claim they now inspect end-to-end. They still contain unsupported run-name, permissions, timeout, shell/environment, action inputs and working-directory fields. Full example integration and the finite milestone audit remain pending. Do not resume the Java worker or expand to execution.

Cost and checkpoint

Rough estimate: 14k–24k additional tokens including implementation, verification, PR and merge follow-up. Typed defaults and cross-field validation make this larger than runner lists. This is not measured billing or a conversion of remaining weekly allowance; unexpected findings may exceed it. No reset or model change is authorized by this plan. Use supported T3 controls for account management; do not automate redemption or patch tooling. Checkpoint revision, owned files, exact red/green evidence, PR head/checks and next action before interruption. No full milestone completion is claimed.

Implemented reference boundary and evidence

Pinned official reference: actions/languageservices 4043eda158e16579cc5fb1b0b07a4bce2a76f0b5: workflow schema and dispatch converter. Schema descriptions supply input-identifier syntax; no step-ID length/reserved-prefix rules were imported. Choice needs nonempty options; nonchoice rejects options. Boolean versus string family default checks follow the converter. The converter does not check choice-default membership. For number inputs it does not copy/check defaults; WSR retains schema-permitted scalar defaults, without asserting stronger numeric validation. Omitted type stays absent instead of copying the converter’s injected string type. YAML coercions are not adopted: WSR keeps strict source kinds. No new architecture or runtime compatibility is implied by these source-preservation differences.

Observed behavioral red/green and 25 CLI inspection tests plus two declaration-summary tests are recorded in the checkpoint. The old fixture failed with expected a string, exit 1 and empty stdout. The new fixture reports the choice declaration and ordered steps, exit 0. No reset consumed; owner reported a manual account reset before implementation. No live usage amount was inferred.