# Medium Reasoning Rollout

## Current Design

- `src/runner.ts` is the canonical source for the full support-analysis
  reasoning effort.
- Dry-run and live analysis share the same runner and currently request `high`.
- Classification, reply normalization, customer-reply validation, and report
  generation are separate calls with their own reasoning settings.
- Deploying a source change requires a new service image and the normal
  app-delivery image-bump flow.

## Risks

- Changing only one of dry-run or live would break their shared analysis
  contract.
- Changing the thread setting without the trace setting would make runtime
  evidence inaccurate.
- Broadening this change to classifier, normalization, validation, or report
  generation would mix independent quality and latency contracts.
- Existing uncommitted attachment-input work must not be included in this
  rollout.

## Target Contract

- Full dry-run and live support analysis use `medium` reasoning effort.
- The analysis thread and its context trace read the same canonical constant.
- Model IDs, prompts, gates, Jira write behavior, classifier, normalization,
  validator, reports, deployment configuration, and automation roles do not
  change.
- No compatibility fallback to `high` is introduced.

## Implementation Steps

1. Define one canonical full-analysis reasoning constant in `src/runner.ts`.
2. Use it for both thread creation and trace creation in dry-run and live paths.
3. Add a local contract test and update the configuration reference.
4. Run typecheck, build, and tests.
5. Publish through the normal staging image workflow and merge the generated
   app-delivery image-bump PR.
6. Verify the live pod image and a new analysis context trace.

## Verification Checklist

- [x] Source contains no remaining `high` setting in the full-analysis path.
- [x] Contract test asserts `medium`.
- [x] `npm run typecheck`, `npm run build`, and `npm test` pass.
- [x] The source PR contains no unrelated attachment-input changes.
- [ ] The staging image-bump PR is merged and the Deployment is ready.
- [ ] A new live or read-only live-style analysis trace reports `medium`.
- [ ] Classifier and normalization remain `low`; report generation remains
      `high`.