Malformed workflow-declared agent files are silently dropped here, so the fail-closed path never runs. listMarkdownAgentFiles logs the parse/read error and returns no entry; applyClaudeSubagentModels then reports success and Claude starts with the unresolved staged agent, reproducing the registration/model failure this change is intended to prevent. Treat parse failures under the workflow staging directory as fatal (while retaining lenient handling for repository agents), and add a regression test.

Fixed in 1f97c55. A staged workflow sub-agent file that can't be read or parsed (or a non-ENOENT read error on the staging directory) now stops Claude before it starts. listMarkdownAgentFiles takes an onInvalid callback, and applyClaudeSubagentModels returns ok: false. Repository agent files are still only warned about. Regressions are in engine_subagent_models.test.cjs (staged: fails closed; repository: lenient) and claude_harness.test.cjs (exit 1, no spawn).

The /reflect refresh promise can reject, and this await refreshOnce() is outside the surrounding resolution try/catch. Because the promise is memoized, one transient fetch failure aborts the entire batch; the Claude harness does not catch applyClaudeSubagentModels, so the run exits via an uncaught rejection without recording the remaining agents or emitting the intended catalog-unavailable diagnostic. Catch refresh failures inside refreshOnce and return null so the existing failure classification remains reliable.

Fixed in 1f97c55. refreshOnce now catches refresh failures, including synchronous throws from refetchReflectData. It logs a warning and returns null, so the batch still resolves and records every agent, and an incomplete catalogue is classified as catalog_unavailable. A regression test checks that a rejected refresh is attempted once and both agents are recorded as catalog_unavailable.

This makes an incomplete /reflect catalog a hard startup failure for every fatal Claude workflow agent, even when the provider may still serve the configured model. catalog_unavailable covers discovery being incomplete or absent, not evidence that the model is invalid; exiting before the workflow starts introduces a new availability dependency and can turn transient discovery failures into total run outages. Keep catalog-unavailable non-fatal (or make this policy explicit), and reserve fail-closed behavior for a complete catalog proving model_unavailable.

Keeping this as is, on purpose. For a workflow-declared Claude sub-agent, catalog_unavailable fails closed. That's the same policy #70389 uses for Copilot sub-agents and #70320 uses for the Copilot SDK main model: gh-aw doesn't start an agent whose declared model it can't confirm. A transient failure gets one shared refresh first. Repository agents and Codex agents only record and warn. The policy is now documented in the PR body under "catalog_unavailable policy (intentional)".

[/codebase-design] classifyUnlistedModel is computed before refreshOnce() as initialReason, then the post-refresh branch re-derives it with a terse ternary: initialReason === "model_unavailable" || !refreshed ? initialReason : classifyUnlistedModel(refreshed, providers). This is hard to verify at a glance without tracing both call sites.

const reason = !refreshed ? initialReason : classifyUnlistedModel(refreshed, providers);

@copilot please address this.

Fixed in 1f97c55. The line is now initialReason === "catalog_unavailable" && refreshed ? classifyUnlistedModel(refreshed, providers) : initialReason. Behaviour is unchanged. A comment states the invariant: a complete catalogue without the model already proves model_unavailable, so only an incomplete one is reclassified from refreshed data.

[/diagnosing-bugs] applyClaudeSubagentModels runs after fetchAWFReflect() and the pre-flight noop/safeInitialArgs checks, but there's no unit/integration test asserting the exact ordering relative to removeClaudePlugin(pluginDir) cleanup on a mid-stream failure of resolveClaudePromptFileArgs (the earlier try/catch at line ~517). If applyClaudeSubagentModels throws instead of returning { ok: false } (e.g. an unexpected error inside resolveEngineSubagentModels's outer try/catch), this call site has no surrounding try/catch, so pluginDir cleanup would be skipped on an unhandled rejection.

@copilot please address this.

Fixed in 1f97c55. The call is now wrapped in try/catch. An unexpected throw logs fatal: unexpected error while resolving Claude sub-agent models: …, then goes through the existing ok: false path: removeClaudePlugin(pluginDir), then process.exit(1). A harness test injects a throw and checks the exit code, that nothing was spawned, the message, and that cleanup ran.

[/tdd] ensureClaudeSubAgentName is only applied when engineId.toLowerCase() === "claude", string-compared against the raw engineId parameter. If an upstream caller ever passes a differently-cased or aliased Claude engine id (e.g. "Claude-Code"), this silently skips the name injection and reintroduces the exact Agent type '<name>' not found bug this PR fixes, with no test coverage for that mismatch.

it('still ensures name for an unexpected engineId casing or alias', () => {
  // document whatever ids this function is actually called with across the codebase
});

@copilot please address this.

Fixed in 1f97c55. writeInlineSubAgents now normalises the engine id once (trimmed and lower-cased) and uses it for both the target directory and the name: injection, so every id that stages into .claude/agents gets the name. The only production caller is interpolate_prompt.cjs, which passes GH_AW_ENGINE_ID; the compiler emits that as the canonical claude. A test covers Claude and CLAUDE.

[/codebase-design] claudeCodexSubagentProvider collapses LLMProviderGitHub and the literal "copilot" string into "github-copilot", duplicating the normalization already expressed via llmProviderAliases in llm_provider.go (which maps copilot/github/github-copilot/github_models all to LLMProviderGitHub). Since resolveEngineLLMProviderFromModel already returns a normalized LLMProvider, the provider == "copilot" branch here looks unreachable (it would always be the already-normalized LLMProviderGitHub string "github"), which is subtly different from the alias map's source strings.

@copilot please address this.

Removed in 1f97c55. resolveEngineLLMProviderFromModel always returns a normalised LLMProvider, because normalizeLLMProvider and fallbackModelProvider both map copilot through llmProviderAliases. So the literal "copilot" branch was unreachable. A comment now explains the github → github-copilot mapping. The existing "claude on the copilot provider" declaration test still passes.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and this smoke workflow safe-outputs change isn't one of them. It's left for the maintainer to schedule.

Fixed in f55f661. The smoke workflow is now read-only:

• Disabled: the default create-issue, failure and failed-jobs issue reporting, issue creation for missing-tool, missing-data and report-incomplete, and threat detection.
• It finishes with a noop that doesn't publish (report-as-issue: false).

The regenerated lock grants no issues: write and has no create_issue. smoke_claude_sub_agents.test.cjs checks this.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and reserving the main role for sub-agent records isn't one of them. It's left for the maintainer to schedule.

Fixed in f55f661. main is now a reserved sub-agent name for Claude and Codex:

• A workflow-declared main stops the run with a message asking to rename the agent.
• A repository agent named main logs a warning and is skipped.

New tests check that both the main and worker model.resolution records survive, and that the harness stops before spawning when a declared main is present.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and the bare-mode plugin copy isn't one of them. The PR also lists bare-mode behaviour as out of scope. It's left for the maintainer to schedule.

Fixed in f55f661. With bare: true, the plugin's agents link is replaced by a run-local copy of the staged definitions, with resolved models written in. The staged and repository files are not changed.

claude_harness.test.cjs reads the definition through the real --plugin-dir/agents path, and engine_subagent_models.test.cjs uses the real claudeBareCapabilities.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and parser-based name: detection and flow-mapping insertion aren't among them. If such a file did come out unparseable, it now fails closed before Claude starts (1f97c55) rather than being skipped. It's left for the maintainer to schedule.

Fixed in f55f661. ensureClaudeSubAgentName now uses the YAML parser:

• An existing name key is detected, including a quoted "name":.
• For flow-mapping frontmatter, the name goes inside the {...}, so the YAML stays valid.

Staging tests cover a quoted key, {description: ..., model: small} and { }.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and expanding provider-qualified aliases in declarations isn't one of them. It's left for the maintainer to schedule.

Fixed in f55f661. Claude and Codex declaration patterns now strip a provider prefix that matches the engine's provider before alias expansion. For example, anthropic/fast with fast: [anthropic/*haiku*] expands to the Haiku pattern.

Prefixes for a different provider are left as they are. Declaration tests cover anthropic/fast, anthropic/small, copilot/small, openai/small, and a mismatched prefix.

Not changed in this round. The maintainer limited the review to items 1–4 in #70671 (comment), and matching inherited declarations in the audit consumer isn't one of them. It's left for the maintainer to schedule.

Fixed in f55f661. declaredModelMatches now matches inherit and inherited declarations against the parent model. That model comes from the main agent's observed models (requested and served); if none were observed, it falls back to the model in aw_info.

As a result, a successful inherited invocation is no longer reported as "Declared Sub-agent Model Not Observed". TestDeclaredInheritedSubagentModelMatchesParentModel covers this.