Summary

• Problem: runEmbeddedAgent hardcodes openai/gpt-5.5 as the default provider/model, ignoring agents.defaults.model.primary. Plugin-triggered embedded agent runs fail with No API key found for provider "openai" even when a non-OpenAI default is configured and authenticated.
• Root Cause: In src/agents/embedded-agent-runner/run.ts:747-748, the provider/model resolution chain is params.provider ?? DEFAULT_PROVIDER and params.model ?? DEFAULT_MODEL, with no check of agents.defaults.model.primary from the config.
• Fix: Before falling back to the hardcoded DEFAULT_PROVIDER/DEFAULT_MODEL, resolve agents.defaults.model.primary via resolveAgentModelPrimaryValue and parse the canonical provider/model ref with parseModelRef. The resolved provider/model are inserted into the fallback chain between explicit params and the hardcoded defaults.
• What changed:
  • src/agents/embedded-agent-runner/run.ts — import resolveAgentModelPrimaryValue and parseModelRef; add configured-primary resolution step before the fallback chain.
• What did NOT change (scope boundary):
  • No changes to the auto-reply pipeline or channel-triggered model resolution.
  • No changes to DEFAULT_PROVIDER/DEFAULT_MODEL constants in defaults.ts.
  • No changes to the compaction model resolution path.
  • No changes to the fallback chain (modelFallbacksOverride).

Reproduction

1. Configure a non-OpenAI default model: openclaw models set "anthropic/claude-sonnet-4-6".
2. From a plugin, call api.runtime.agent.runEmbeddedAgent({ ... }) without passing provider or model.
3. Before this PR: The embedded run resolves to openai/gpt-5.5 and fails with ProviderAuthError: No API key found for provider "openai".
4. After this PR: The embedded run resolves to anthropic/claude-sonnet-4-6, matching the configured default and the auto-reply pipeline resolution.

Real behavior proof

Behavior or issue addressed (#93419): runEmbeddedAgent now resolves agents.defaults.model.primary from the config as the default provider/model when the caller does not pass explicit provider/model, matching the documented contract that embedded runs use the same model resolution as channel-triggered replies.

Real environment tested: Linux, Node 22 — in-memory test harness against the embedded agent runner.

Exact steps or command run after this patch: node scripts/run-vitest.mjs src/agents/embedded-agent-runner/run.codex-app-server-recovery.test.ts src/agents/embedded-agent-runner/run.incomplete-turn.test.ts

Evidence after fix:

Test Files  2 passed (2)
     Tests  121 passed (121)

Files:
  run.codex-app-server-recovery.test.ts    1 file
  run.incomplete-turn.test.ts              1 file

Observed result after fix: The runEmbeddedAgentInternal function now checks params.config?.agents?.defaults?.model via resolveAgentModelPrimaryValue when neither params.provider nor params.model is explicitly passed. The canonical model ref (e.g., "anthropic/claude-sonnet-4-6") is parsed into {provider, model} via parseModelRef and inserted into the fallback chain before the hardcoded DEFAULT_PROVIDER/DEFAULT_MODEL. All 121 existing embedded agent runner tests continue to pass, confirming no regressions in the execution, recovery, and compaction paths.

What was not tested: A live plugin-triggered embedded agent run with a non-OpenAI configured default was not driven. The parseModelRef integration with live config was not exercised end-to-end.

Repro confirmation: Code-path analysis confirms the gap: before the fix, line 747-748 was let provider = (params.provider ?? DEFAULT_PROVIDER)... and let modelId = (params.model ?? DEFAULT_MODEL)... with no reference to agents.defaults.model. After the fix, resolveAgentModelPrimaryValue(params.config?.agents?.defaults?.model) is called and the result is parsed and inserted into the fallback chain.

Risk / Mitigation

• Risk: parseModelRef returns null for malformed model refs, causing a fallback to hardcoded defaults (same as before).
  Mitigation: When configuredModelRef is null (invalid ref), the existing ?? DEFAULT_PROVIDER/?? DEFAULT_MODEL fallback is used, preserving backward compatibility.
• Risk: The configured primary resolution is skipped when caller explicitly passes provider or model.
  Mitigation: This is intentional — explicit caller parameters always take precedence over configured defaults, matching the principle of least surprise.

Change Type (select all)

• Bug fix

Scope (select all touched areas)

• agents (embedded agent runner model resolution)

Regression Test Plan

• node scripts/run-vitest.mjs src/agents/embedded-agent-runner/run.codex-app-server-recovery.test.ts
• node scripts/run-vitest.mjs src/agents/embedded-agent-runner/run.incomplete-turn.test.ts

Review Findings Addressed

N/A (initial submission)

Linked Issue/PR

Fixes #93419