[callumreid]

• Add the six immutable Python metric draft/preview/publication commands and the persona interruption-instructions flag, with source-file input and agent next actions.
• Preserve the newly published request fields on existing commands: review batch mode/size, uploaded-conversation stereo channel and set-once agent assignment, and atomic test-set creation with inline cases. Preserve omitted fields, explicit batch resets, channel zero/null, and empty inline-case arrays; reject invalid null case lists before sending.
• Reconcile the November 2 catalog: 125 of 199 operations have first-class commands, with 74 explicitly reviewed route gaps. All 472 top-level fields on covered request operations are modeled. The request-field audit also rejects body calls it cannot parse. New exploratory audio-analysis/publication and transcript-comment resources are explicitly deferred with contract-specific reasons in the manifest.
• Keep this implementation PR at 0.9.2. The 0.10.0 bump will be a separate manifest-only PR after merge, followed by package/Homebrew verification and the prepared user-doc updates.

Validation: five new captured-wire regressions failed against 97eed40 before the November correction; all six final new tests pass. Python automation (65 tests), fmt, clippy, Python lint/format, release-manifest checks and the live 472/472 audit pass. The full local Rust run passed 226/227 tests, with the unchanged 500ms advisory update-notification test timing out once under host load; that isolated test then passed. The updater source is byte-identical to main. Fresh final-head CI is required before merge.

JSON numeric bounds and nested inline-case validation remain enforced by the API; CLI flags constrain channel 0-1 and batch size 1-50. No live conversation, review-project, test-set, draft, preview or publication writes were performed during validation.

Terminal preview follow-up (COVAL-7661)

Preserves the external COMPLETED/SUCCEEDED and queued-admission-warning corrections. An unavailable Python metric evaluation terminates as SKIPPED in the backend; it now retains its diagnostic without suggesting either polling or publishing. The captured-wire regression fails on ba2b63a before this correction and passes after it. All five Python metric CLI tests pass.

Local validation at 112ad1f: 9 unit tests passed; 219 of 220 integration tests passed in the full run. The unchanged 500ms update-notification advisory timed out under the full run and passed in isolation; no timeout or assertion was weakened. Clippy with warnings denied, formatting, and whitespace checks pass. Exact-head GitHub CI remains the required final source gate. This changes no database queries.

Final validation at 1104eb3 (additive merge of main's approved permission/Gate workflows; no source rewrite): all 11 GitHub checks passed, including the real Switchboard Gate from GitHub Actions app15368, all five platform builds, and CodeQL. CI run https://github.com/coval-ai/cli/actions/runs/41542332213 passed 67 automation tests, 9 Rust unit tests and all 220 integration tests. Local advisory timeout limitation above is retained for transparency; exact-head CI passed it. Existing human changes-requested review remains an independent merge gate.

[borgesius]

Requesting changes for the successful-preview status mismatch, which leaves the agent workflow polling after completion. The separate note about polling during uncertain admission is nonblocking. The existing Rust suite passes; correcting the success fixture to the API response status reproduces the mismatch.

[github-actions]

The weekly refresh left this branch unchanged because it has commits not made by the bot.
Merge this PR, or rebase it onto main and regenerate api-coverage-report.md, so the next run can pick up new API changes.

Run: https://github.com/coval-ai/cli/actions/runs/41541116967

[borgesius]

The previous preview-state findings are fixed on this head: COMPLETED offers publication, queued admission warnings retain polling, and SKIPPED is terminal with its diagnostic preserved. All 229 Rust tests and 67 automation tests pass locally. Looks good.