Azure TypeSpec Assessment

Assess only the user-selected TypeSpec scope, normally before a PR exists. Never modify the user's assessed source, branch, index, or working tree; isolated temporary sparse worktrees and artifacts are allowed only under the chosen work directory.

Before starting a local assessment

Check whether the user explicitly supplied or confirmed the baseline for this assessment. If not, your first assessment action must be to ask: "Compare against origin/main, or a different branch or commit?" Use the host's user-question tool when available and wait for the answer. Do not start preflight, compute a merge base, create work artifacts, or run analysis before confirmation. Permission to assess the code is not permission to choose its baseline; merely announcing origin/main is not confirmation. Use a supplied baseline without asking again. Existing-report replay does not need this question; for PR assessment, use the PR's actual target baseline.

For every fresh assessment, the coordinator is the first operational command. For a PR, run run-assessment-analysis.ts --pr <url-or-number> --repo <repo> --output <work> immediately. For local code, run it immediately with --base, --specification, --repo, and --output once the baseline is known. Do not run separate Git status/diff/fetch/worktree, PR metadata, dependency, or project-discovery commands first. Diagnose only a concrete blocker returned by the coordinator.

Follow the complete workflow. Apply the classification rules, including the detailed downstream cases, perform the Azure Guidelines search against the official document catalog, check documentation completeness, and produce exactly the required outputs.

Required completion

After guarded finalization succeeds, the assessment is not complete until the rendered report is served and linked:

1. Start scripts/serve-assessment.ts --file <work-directory>/assessment.html through the host's attached background or long-lived process mechanism.
2. Immediately read startup output from the running process through the host's process-output reader. Do not wait for process completion or a completion notification: the server is intentionally long-lived.
3. Capture the printed http://127.0.0.1:<port>/assessment.html URL and confirm the process remains running.
4. Return that URL as the clickable Assessment report link and include the absolute assessment.json path for structured results.

Do not substitute a file: URL or only return filesystem paths.

Boundaries

• V1 is standalone and opt-in: run only when the user explicitly requests an assessment or review. Do not invoke this skill from azure-typespec-author, or automatically before or after its authoring and validation workflow. Integration is deferred to a future version.
• Run complete mode only: merge-base through HEAD, staged, unstaged, and relevant untracked changes.
• Derive semantic intents from changed TypeSpec source. Use AutoRest only to map those intents to REST operations and assess REST compatibility; use TCGC only for downstream SDK analysis.
• Check deterministic hunk coverage in model-input.json. Resolve only its declared artifactReferences and evidenceSetId entries when full evidence is needed. Skip inference when all hunks have candidates or explicit deterministic classifications. For unknown hunks only, write and validate inference.json before final judgment.
• When the coordinator returns awaiting-agent-judgment, continue immediately in the same task. Do not stop for a progress summary or require the user to say "continue." Read agent-workspace/agent-index.json once, then read its single bounded model-input.json exactly once. Read the compact decision schema listed in the index exactly once, resolving it relative to this skill directory; do not search for schemas. Do not recursively list the work directory, scan canonical artifacts broadly, or reread schemas during normal execution.
• api-version-publication and api-version-wide-change intents are deterministic informational items. A version-wide change is identified from a Versions declaration with no directly owned operation and only version-transition/governance operation mappings; it does not use an operation-count threshold. Their affected operations remain in the final Semantic intents report, but they and their exclusively associated candidates are excluded from model input, inference, Azure Guidelines decisions, breaking-change relationships, and findings.
• In the final bounded Agent phase, summarize each assessed Semantic intent once, judge every deterministic and inferred REST/downstream candidate, rank the catalog once across all assessed Semantic intents, fetch four official documents once with ranked fallback, record provenance and extracted guidance once, then assess each included intent against that shared document set. Guidance excerpts do not carry declaration IDs; use only the prefilled intent-scoped qualified declarationNames in the citing compliance judgment. Write only agent-workspace/agent-decisions.json; run the indexed materializer to produce inference, search-evidence, and judgment artifacts. The materializer never fetches or authors judgment. Do not assess each operation or invent URLs, operations, symbols, sources, or guidance.
• Treat a completed search with no governing guidance as no-applicable-guidance; count it as assessed and do not create a blocker. Reserve not-assessed for an incomplete or blocked Azure Guidelines assessment.
• Check documentation deterministically from compiler results only for newly added operation, model, enum, and interface declarations. A declaration is complete when the compiler returns a nonempty effective document, including inherited documentation. Missing or empty documentation creates a finding with the exact TypeSpec declaration. Never compare documentation text with code or send documentation to the Agent.
• Report Azure Guidelines and Documentation Completeness independently with explicit coverage. Documentation with no eligible newly added declarations is not-applicable; unavailable compiler evidence remains not-assessed.
• Overall safety covers REST and downstream SDK impact only.
• After successful materialization, run one guarded invocation of finalize-assessment.ts --work <work-directory>. Do not separately assemble, validate, and render. Finalization requires validated assessment.json and assessment.html; user-facing completion additionally requires the served report link described above. model-input.json, the Agent index, filesystem paths without the link, or a partial blocker is not completion.
• Retain blockers as Potential limits in the report appendix and stop after assessment. Do not author fixes or remediate TypeSpec.