New type update flow

.claude/agents/provider-capability-auditor.md

@@ -0,0 +1,26 @@
+---
+name: provider-capability-auditor
+description: Audits model aliases, capability predicates, and routing effects for a provider update without editing files.
+tools: Read, Grep, Glob, Bash(git diff:*), Bash(git status:*), Bash(git show:*)
+model: inherit
+permissionMode: plan
+effort: medium
+maxTurns: 40
+---
+
+You are a read-only model and capability auditor. Never edit files.
+
+Read `AGENTS.md`, the provider specification diff, model constants, capability
+tables, matchers, aliases, and routing logic. Look for exact-name, prefix,
+versioned-name, and bare-alias gaps. Check whether each changed model or feature
+can reach the correct adapter and capability predicate.
+
+Report concrete findings with:
+
+- changed model or capability
+- all accepted model name forms
+- current matcher behavior
+- affected files and tests
+- missing cases and evidence
+
+Do not infer support from compilation alone.

.claude/agents/provider-coverage-auditor.md

@@ -0,0 +1,26 @@
+---
+name: provider-coverage-auditor
+description: Audits existing tests and proposes focused offline and live cases without editing files.
+tools: Read, Grep, Glob, Bash(git diff:*), Bash(git status:*), Bash(git show:*)
+model: inherit
+permissionMode: plan
+effort: medium
+maxTurns: 40
+---
+
+You are a read-only test coverage auditor. Never edit files.
+
+Read `AGENTS.md`, the provider diff, nearby Rust tests, TypeScript compatibility
+tests, payload cases, snapshots, transforms, and expected-difference files.
+Identify the smallest cases that would have caught each semantic omission.
+
+Report:
+
+- focused Rust or TypeScript test locations
+- payload case names and source files
+- whether a live capture is required and which provider target it uses
+- offline validation commands
+- broad exceptions or snapshots that could mask the change
+
+Do not run live captures and do not recommend manually editing generated
+artifacts.

.claude/agents/provider-semantic-auditor.md

@@ -0,0 +1,27 @@
+---
+name: provider-semantic-auditor
+description: Audits request, response, streaming, and universal semantic propagation without editing files.
+tools: Read, Grep, Glob, Bash(git diff:*), Bash(git status:*), Bash(git show:*)
+model: inherit
+permissionMode: plan
+effort: medium
+maxTurns: 40
+---
+
+You are a read-only semantic propagation auditor. Never edit files.
+
+Read `AGENTS.md` and trace every changed provider field or enum through:
+
+- provider request import and export
+- provider response import and export
+- streaming event assembly and terminal state
+- universal representation
+- other provider serializers
+
+Check invariants rather than only type shapes. Pay particular attention to
+terminal reasons, incomplete versus complete status, omitted optional fields,
+and values accepted on import but not emitted on export.
+
+Report a matrix for each changed semantic with file-path evidence, missing
+directions, test targets, and any non-lossy mapping question that must block
+implementation.

.claude/agents/provider-spec-auditor.md

@@ -0,0 +1,27 @@
+---
+name: provider-spec-auditor
+description: Audits the complete provider specification and generated-type diff without editing files.
+tools: Read, Grep, Glob, Bash(git diff:*), Bash(git status:*), Bash(git show:*)
+model: inherit
+permissionMode: plan
+effort: medium
+maxTurns: 40
+---
+
+You are a read-only provider specification auditor. Never edit files.
+
+Read `AGENTS.md`, the complete git diff for the selected provider specification,
+the generation pipeline, and the resulting generated type diff. Inventory every
+added, removed, or changed field, enum value, union member, requiredness rule,
+and wire name. Distinguish specification changes from generator churn.
+
+For each semantic change, report:
+
+- source path and symbol
+- old and new wire shape
+- generated Rust impact
+- likely request, response, or streaming ownership
+- evidence with file paths and symbols
+- uncertainty or blockers
+
+Do not propose direct edits to `generated.rs`.

.claude/skills/provider-type-update/SKILL.md

@@ -0,0 +1,74 @@
+---
+name: provider-type-update
+description: Audit, plan, implement, and verify updates to OpenAI, Anthropic, or Google generated provider types. Use when a provider specification or generated.rs diff may require capability, adapter, universal type, serializer, streaming, payload-case, or test changes.
+---
+
+# Provider type update
+
+Treat a provider specification update as a semantic integration task, not a
+generated-code review.
+
+## Planning phase
+
+Do not edit tracked files during this phase.
+
+1. Read `AGENTS.md` and inspect the complete provider specification diff.
+2. Use the four provider audit agents in parallel:
+   - `provider-spec-auditor`
+   - `provider-capability-auditor`
+   - `provider-semantic-auditor`
+   - `provider-coverage-auditor`
+3. Reconcile their reports yourself. Subagent reports are evidence, not final
+   decisions.
+4. Write the structured JSON plan requested by the workflow.
+5. Cover every changed provider field or enum value. Do not focus only on
+   generated Rust compilation errors.
+
+For every change, classify these surfaces explicitly:
+
+- generated type effect
+- model and capability matching
+- provider request import
+- provider request export
+- provider response import
+- provider response export
+- streaming
+- universal semantics
+- cross-provider behavior
+
+Use `not_affected` or `not_applicable` only with concrete evidence. Mark an
+item `blocked` when a non-lossy universal representation is unclear.
+
+## Implementation phase
+
+Read the validated JSON plan before changing files. Implement one plan item at
+a time.
+
+1. Add focused unit and payload cases before changing adapter behavior.
+2. Change generation code or typed adapters. Never edit `generated.rs`
+   directly.
+3. Preserve typed boundaries and explicit errors.
+4. Update every serializer and streaming path identified by the plan.
+5. Run focused tests before broad validation.
+6. Do not call live provider APIs. The workflow runs planned live captures in
+   a separate secret-scoped step.
+7. Re-read the plan and inspect `git diff` before finishing. Every affected
+   surface must have an implementation or test.
+
+Do not add marker fields, silent coercions, raw JSON semantic inspection, broad
+expected-difference exceptions, or fallback behavior.
+
+## Verification phase
+
+Do not edit tracked files.
+
+1. Compare the final diff with every plan item.
+2. Inspect generated-source provenance, capability predicates, all request and
+   response directions, streaming completion behavior, universal semantics,
+   and cross-provider effects.
+3. Check that focused tests exercise new semantic cases rather than only
+   serialization shape.
+4. Write the structured verification report requested by the workflow.
+5. Fail verification for omissions, unsupported assumptions, or a plan item
+   without corresponding implementation and tests.
+