Linked Issue Or Context

Closes #360 — feat(prompts): inject a "verify before assume" external-knowledge discipline into the planner so it knows WHEN to use the already-existing web_fetch capability.

Summary

The web_fetch action already exists and is mechanically described, but nothing tells the planner when to consult external docs. This adds that discipline:

• New exported constant EXTERNAL_KNOWLEDGE_PROTOCOL in packages/core/src/llm/prompts.ts, parallel in style/tone/format to PROGRESSIVE_EXPLORATION_PROTOCOL (Chinese protocol block, heading + numbered points). It tells the planner to consult authoritative sources (official docs / API reference) via web_fetch before planning or writing code when facing 不熟悉 / 版本特定 / 最近才出现 libraries, frameworks, or APIs — instead of guessing API shape or default behavior — and includes a reverse guard so 已经熟知且稳定 knowledge is not needlessly fetched.
• Wired ${EXTERNAL_KNOWLEDGE_PROTOCOL} into BOTH planner system prompts (generatePlanInTwoPhases outline system and generatePlanSinglePhase system) in packages/core/src/llm/plan-generation.ts, immediately after the existing ${PROGRESSIVE_EXPLORATION_PROTOCOL} interpolation; imported alongside it.

Impact Scope

Prompt-only. No executor, schemas, tool routing, or web_fetch runtime behavior changed. No new action enums. web_search intentionally out of scope.

GitNexus Impact Summary

• Risk level: LOW
• Critical skeleton changes: None. Additive prompt-string content only; no function signatures, return types, control flow, schemas, or tool routing changed.
• GitNexus impact: impact(generatePlanInTwoPhases, upstream) → risk LOW, 0 direct callers broken (impactedCount 0). impact(generatePlanSinglePhase, upstream) → risk LOW, 1 direct caller (generatePlan, CALLS), no signature change so nothing breaks. detect_changes (worktree) → changed symbols: generatePlanInTwoPhases, generatePlanSinglePhase (both prompt-string edits) plus the test const noopDeps; affected processes GeneratePlan → NormalizePlan and GeneratePlan → GeneratePlanInTwoPhases at step 5 only. detect_changes reports a heuristic medium because the two planner functions sit on execution flows, but the edits are additive prompt text with no behavioral/contract change, so the effective risk is LOW.
• Verification: pnpm --filter @frontagent/core test -- plan-generation (23 passed, incl. 4 new), pnpm typecheck (25 ok), pnpm biome check on the 3 touched files (clean), pnpm quality:precommit (green).

Verification

• pnpm --filter @frontagent/core test -- plan-generation → 23 passed (4 new: constant semantics + both planner system prompts inject the protocol, asserted via captured generateObject system args).
• pnpm typecheck → 25 tasks successful.
• pnpm biome check packages/core/src/llm/{prompts,plan-generation,plan-generation.test}.ts → clean.
• pnpm quality:precommit → green (ran via pre-commit hook).
• pnpm quality:local → green (ran via pre-push hook).

Checklist

• I have linked an issue or explained why this PR stands alone.
• I have kept the diff focused on the stated change.
• I have run pnpm quality:precommit, or explained why it could not run.
• I have run pnpm quality:local for critical skeleton changes, or explained why it could not run.
• I have updated docs or tests when behavior, public APIs, or Harness contracts changed.
• For critical skeleton changes, I have filled the GitNexus impact summary with concrete results.

🤖 Generated with Claude Code