Context

RemoteClaw is middleware that delegates LLM execution to CLI agents (Claude Code, Gemini, Codex, OpenCode). The pre-fork embedded execution path was gutted during early milestones. Scattered test files and documentation pages still reference obsolete LLM-level concepts (provider selection, model catalogs, thinking-level config, mocks of the removed embedded runner).

Problem

Orphan tests

runEmbeddedPiAgent is the primary function of the gutted embedded execution engine. Verification:

# Production imports of the function:
rg "import.*runEmbeddedPiAgent|from.*pi-embedded" src/ --type ts | grep -v "\.test\." | grep -v "\.cases\."
# No hits. Zero production imports.

# Tests that mock or reference it:
rg "runEmbeddedPiAgent" src/ --type ts

Known files to audit:

• src/commands/agent.test.ts — thinking merge tests (expectDefaultThinkLevel, "prefers per-model thinking over global thinkingDefault", "defaults thinking to adaptive for Anthropic Claude 4.6 models"). Orphan. (May be removed in a concurrent thinking-gut PR — coordinate.)
• src/auto-reply/reply.block-streaming.test.ts — entire file may be orphan (runs against runEmbeddedPiAgent mock). Audit and confirm.
• src/auto-reply/reply/agent-runner.runreplyagent.e2e.test.ts — mixed: contains both real agent-runner exercises AND runEmbeddedPiAgentMock assertions. Orphan sections only; keep the live parts.
• src/auto-reply/reply.triggers.trigger-handling.filters-usage-summary-current-model-provider.cases.ts — contains expect(runEmbeddedPiAgentMock).not.toHaveBeenCalled() tautologies. Delete tautological assertions, keep whatever real coverage exists.

Onboarding and help-text

The onboarding flow and help pages must not prompt the user for LLM provider selection, model choice, or LLM API keys — those are CLI-agent concerns. The user only configures: channel tokens (Telegram, Slack, WhatsApp, etc.), CLI auth env vars (provider-level env vars consumed by the CLI subprocess), and workspace/session paths.

Scope to audit:

• docs/ directory
• README.md at repo root
• src/**/* help strings, onboarding prompts, wizard flows, CLI --help output
• Translations (if any): search for the same concepts in docs/**/*-fr.md, -de.md, etc.

Grep patterns (case-insensitive):

rg -i "provider.*select|model.*picker|api.?key.*(anthropic|openai|google)" docs/ src/
rg -i "choose.*(provider|model)" docs/ src/
rg -i "thinking.?level|thinking.?default" docs/ src/

Tasks

Phase 1: Audit + clean tests

• For each file in the "Known files to audit" list above, classify every test/assertion: (a) exercises live production code, or (b) orphan (asserts against runEmbeddedPiAgent mock or similar gutted paths)
• Delete confirmed orphan test files entirely (where all tests are orphan)
• For mixed files (agent-runner.runreplyagent.e2e.test.ts), extract the live parts and delete the runEmbeddedPiAgent-mock sections
• Delete tautological assertions (expect(runEmbeddedPiAgentMock).not.toHaveBeenCalled() is vacuously true since the function is imported nowhere)
• Verify: rg "runEmbeddedPiAgent" src/ --type ts returns zero hits (or only intentional stub-detection references)

Phase 2: Audit onboarding + docs

• Grep docs/ and src/ for each pattern in the "Grep patterns" section above
• For each match, classify:
  • LEGITIMATE — references channel-bound config (Telegram tokens, Slack OAuth, etc.) or CLI-auth env-var setup (e.g., "export ANTHROPIC_API_KEY before running Claude CLI")
  • OBSOLETE — references gutted embedded-execution features (e.g., "pick a model provider during setup")
  • STALE — references product strategy or features no longer present
• Update or remove OBSOLETE and STALE matches
• Preserve LEGITIMATE matches unchanged

Phase 3: Verification

• Run full test suite; confirm test count drops (orphans removed) with no new skips
• Spot-check onboarding end-to-end: a user running the setup wizard for a fresh install should see NO prompts for LLM provider, LLM model selection, or LLM API keys. Channel tokens and CLI auth env-var instructions are expected and correct.

Acceptance Criteria

• rg "runEmbeddedPiAgent" src/ returns zero hits, or hits only in a single intentional stub-detection file with clear commentary
• Grep of docs/ and onboarding sources for obsolete LLM concepts returns no OBSOLETE or STALE matches (LEGITIMATE channel/auth matches are fine)
• tsc --noEmit clean
• Test suite passes with a smaller test count (orphans removed) and no new skipped tests