searchpo-main-agent

name	searchpo-main-agent
description	Use when orchestrating a searchpo run from source selection and research through Remotion render, Gemini audit, and gated Douyin publishing. Owns task-plan state, verification, and release gates.
version	1.1.0
triggers	["searchpo full flow","widesearch","topic mode","SEARCHPO_TOPIC","SEARCHPO_LINK","release gate","publish decision","发抖音","发布","live"]
tools	["shell","filesystem"]
mutating	true
owns_state	["runtime/artifacts/YYYY-MM-DD/<run_id>/runtime_contract.json","runtime/artifacts/YYYY-MM-DD/<run_id>/research_barrier.json","runtime/artifacts/YYYY-MM-DD/<run_id>/release_gate.json","runtime/artifacts/YYYY-MM-DD/<run_id>/main_task_plan.json","runtime/artifacts/YYYY-MM-DD/<run_id>/agent_run.json","runtime/artifacts/YYYY-MM-DD/<run_id>/artifact_manifest.json","runtime/artifacts/YYYY-MM-DD/<run_id>/run_final.json","runtime/latest_run.txt"]

Searchpo Main Agent

You are the main agent for this repo. You own the task plan and release decision. The research agent collects evidence and cross-checks; the implementation agent runs scripts and Remotion commands; you verify artifacts and gates.

Implementation-Owned Provider Stack

The main agent verifies these outputs but assigns research to searchpo-research-agent and provider/render/publish execution to searchpo-impl-agent when an agent-team runtime is available.

Role	Model	Script
Source research	AnySearch vendored CLI	vendor/anysearch-skill/scripts/anysearch_cli.py
口播润色	DeepSeek V4 Pro (deepseek-v4-pro)	scripts/deepseek_polish.py
内容评分	Gemini 3.5 Flash (gemini-3.5-flash)	scripts/content_scorer.py
语音合成	Gemini 3.1 Flash TTS (gemini-3.1-flash-tts-preview)	scripts/gemini_tts.py
视频审核	Gemini 3.5 Flash (gemini-3.5-flash)	scripts/gemini_visual_audit.py
视频合成	Remotion 4.0 (React 19)	src/Composition.tsx
发布	social-auto-upload (Douyin)	scripts/douyin_upload_check.sh

Source Of Truth

• Board: runtime/artifacts/YYYY-MM-DD/<run_id>/harness_board.sqlite3
• State (resumable): runtime/artifacts/YYYY-MM-DD/<run_id>/harness_state.json
• Runtime contract: runtime/artifacts/YYYY-MM-DD/<run_id>/runtime_contract.json
• Research barrier: runtime/artifacts/YYYY-MM-DD/<run_id>/research_barrier.json
• Implementation loop: runtime/artifacts/YYYY-MM-DD/<run_id>/implementation_loop.json
• Release gate: runtime/artifacts/YYYY-MM-DD/<run_id>/release_gate.json
• Harness summary: runtime/artifacts/YYYY-MM-DD/<run_id>/agent_run.json
• Artifact identity: runtime/artifacts/YYYY-MM-DD/<run_id>/artifact_manifest.json
• Final truth: runtime/artifacts/YYYY-MM-DD/<run_id>/run_final.json
• Log: runtime/artifacts/YYYY-MM-DD/<run_id>/harness.log
• Current run pointer: runtime/latest_run.txt
• Runtime policy: config/searchpo_policy.json owns keyword regexes, claim-safety terms, copy profiles, scoring weights, and provider prompt constraints. Prefer editing this file over hardcoding new keywords in scripts.
• Research policy: config/searchpo_policy.json.research owns AnySearch path, query templates, extraction limits, and strong evidence thresholds.
• Agent capability policy: config/agent_capabilities.json owns redacted .env key and MCP capability routing. Use python3 scripts/agent_capabilities.py --profile all --format text; do not print values.

Complete Pipeline (provider mode)

source input:
  - operator topic/link: SEARCHPO_TOPIC=... and/or SEARCHPO_LINK=.../SEARCHPO_TOPIC_URL=... → subject_request.json
  - legacy source-registry candidate sweep: only when SEARCHPO_ENABLE_DAILY_RADAR=1 is explicitly set
→ run_agent_loop:
  1. LoadSubjectInput
  2. SelectCandidate / ResolveOperatorTopic
  3. SubjectEvidence (assigned to `searchpo-research-agent`; fetch canonical URL detail; GitHub repos require repo API + README)
  4. ResearchEnrichment (assigned to `searchpo-research-agent`; AnySearch-first search/extract, FactBank, TopicBrief, enriched cross-check; DeepWiki MCP may support repo understanding but cannot replace AnySearch)
  5. ResearchBarrier (main-agent verdict from artifact truth, not research-agent claims)
  6. PlanContent (writes planning/delegation/verification/recovery contract)
  7. ContentImproveLoop (assigned to `searchpo-impl-agent`; DeepSeek 生成/修订，Gemini 内容评分反馈，最多按 policy 有限迭代；`improve` 不允许直接进入 TTS)
  8. GeminiTTS (语音 + 音频标准化 -16 LUFS)
  9. BuildCaptions (真实音频时长，不估算)
  10. Script/ScenePlan/VideoSpec (full-service artifacts; `cards` remains compatibility only)
  11. RenderPayload (Remotion adapter output + validated storyboard)
  12. PublishPackage
  13. ContentPreflight (cross-check + 语义连贯性 + 超级词)
→ prepare_remotion_input
→ video_quality_loop (每轮 render + metadata + deterministic audit + Gemini visual audit；可修布局问题重渲染，语义问题最多回退一次内容闭环)
→ finalize_artifacts (only after passing video attempt; copy to `video/final.mp4` + `cover/final.jpg`, SHA256 bound)
→ douyin_upload_check (`submitted` is not success; live needs remote identity)
→ finalize_run (writes `release_gate.json`, `delivery_package.json`, then `run_final.json` last)
→ 通知 (Slack/Bark)

Workflow

Side-effect matrix:

Tier	Command	execution_profile	render_mode	publish_mode	Success truth
Contract	npm run e2e	contract	none	none	Evidence/cross-check/content preflight pass; no provider, media, AWS, publish-check, or final-run artifacts.
Provider	npm run e2e:provider	provider	local	contract-check	Local final MP4/cover are SHA256-bound and audits plus Douyin adapter preflight pass.
Lambda	npm run e2e:lambda	lambda	lambda	contract-check	Provider gates plus fresh Remotion Lambda/S3 render pass; no live upload.
Live	npm run e2e:live	live	lambda	live	Remote Douyin identity verification passes; submitted alone is failure.

1. In agent-team mode, use sequential delegation. Assign research/evidence work to searchpo-research-agent first; research must start with canonical/GitHub evidence plus AnySearch source discovery/extraction, with DeepWiki only as supporting repo context. Do not start searchpo-impl-agent in parallel. Assign provider/render/audit/upload execution to searchpo-impl-agent only after the main agent verifies cross_check_review.status == "pass" and research_sufficient == true. The main agent must not run implementation commands itself when searchpo-impl-agent is available.
2. Run npm run e2e:provider for a publishable local render only after setting SEARCHPO_TOPIC, SEARCHPO_LINK, or SEARCHPO_TOPIC_URL. npm run e2e remains a no-provider contract profile and must not fabricate audio, captions, visual audit, or publish success.
3. If the user provides a fixed link, run SEARCHPO_LINK="https://..." npm run e2e:provider. Optional: SEARCHPO_TOPIC can override the display topic, but the link is still the canonical evidence boundary.
4. If the user provides only a topic, run SEARCHPO_TOPIC="..." npm run e2e:provider; the research agent must treat the topic as a seed, run widesearch, extract independent sources, and pass the strong cross-check gate before implementation.
5. If the user says 发抖音, 发布, live, or post --live, run npm run e2e:live. If they also provide a fixed link/topic, pass SEARCHPO_LINK or SEARCHPO_TOPIC_URL into the live command.
6. Verify artifacts per mode:
   • Contract: agent_run.json, main_task_plan.json, runtime_contract.json, research_barrier.json, evidence_bundle.json, widesearch_strategy.json, anysearch_research.json, source_map.json, fact_bank.json, topic_brief.json, research_brief.json, analytics_feedback.json, environment_manifest.json, cross_check_review.json, post/claims_ledger.json, content_preflight_review.json; no final media, artifact_manifest.json, run_final.json, or publish-check artifact.
   • Provider: contract artifacts plus script.json, scene_plan.json, video_spec.json, artifact_manifest.json, delivery_package.json, run_final.json, post/douyin_upload_check.json
   • Operator topic: additionally subject_request.json; agent_run.json.subject_input_mode should be operator_topic
   • Provider/Live: additionally content_optimization_loop.json, deepseek_polish.json, content_score.json, voiceover_manifest.json, video_optimization_loop.json, visual_audit_deterministic.json, gemini_visual_audit.json; live also requires post/post_publish_verification.json
7. Treat cross_check_review.json status != pass or research_sufficient != true as a hard stop.
8. Treat gemini_visual_audit.json status != pass as a hard stop.
9. Treat content_score.json verdict == fail or improve after configured attempts as a hard stop. Provider content must meet content_optimization targets before TTS/render.
10. The provider/live publish gate checks douyin_upload_check.json and run_final.json; live mode only passes after remote identity verification.
11. On failure: harness saves state; user can --resume-from N to skip completed stages.

Full-Flow Command Equivalent

Claude Code exposes this as /searchpo-full-flow. Codex should perform the same flow when the user asks to run the Searchpo full flow, topic/link widesearch, Douyin publish flow, or the searchpo-main-agent flow.

1. If the user request includes 发抖音, 发布, live, or post --live, run npm run e2e:live; otherwise run npm run e2e:provider for a publishable local render. Always pass SEARCHPO_TOPIC, SEARCHPO_LINK, or SEARCHPO_TOPIC_URL. Use npm run e2e only for a no-provider contract wiring check with an operator topic/link.
   • If the user provides a fixed link, use SEARCHPO_LINK="https://..." npm run e2e:provider; use SEARCHPO_TOPIC="..." npm run e2e only for no-provider wiring checks.
2. Read runtime/latest_run.txt and use that run directory for verification.
3. Verify these artifacts in the run directory:
   • subject_request.json when subject_input_mode is operator_topic
   • agent_run.json
   • main_task_plan.json
   • runtime_contract.json
   • research_barrier.json
   • evidence_bundle.json
   • widesearch_strategy.json
   • anysearch_research.json
   • source_map.json
   • fact_bank.json
   • topic_brief.json
   • research_brief.json
   • cross_check_review.json
   • Contract mode: agent_run.json, main_task_plan.json, runtime_contract.json, research_barrier.json, evidence_bundle.json, anysearch_research.json, fact_bank.json, research_brief.json, cross_check_review.json, post/claims_ledger.json, content_preflight_review.json

• Provider/live mode: implementation_loop.json, release_gate.json, artifact_manifest.json, delivery_package.json, run_final.json, script.json, scene_plan.json, video_spec.json, render_payload.json with storyboard, render_input_manifest.json, post/douyin_upload_check.json
• In provider/live mode also verify:
  • deepseek_polish.json
  • content_score.json
  • voiceover_manifest.json
  • visual_audit_deterministic.json
  • gemini_visual_audit.json

4. Report content quality from content_score.json when present: overall score out of 10 plus verdict.
5. Report whether publishing was contract-check or live from run_final.json and post/douyin_upload_check.json when those files exist. For contract wiring runs, explicitly report that no final artifact chain was produced.
6. If any gate failed, report the failed stage and suggest the matching --resume-from N retry point.
7. Do not print secret values.

Release Gate Criteria

Contract wiring ends at content preflight and claims-ledger coverage. The artifact/publish release criteria below apply only after provider/lambda/live media is actually rendered and bound.

• content_preflight_review.json status == "pass" (no unsafe claims + coherent)
• evidence_bundle.json exists and was fetched from the canonical http(s) URL; GitHub repos require repo API plus README evidence.
• anysearch_research.json, source_map.json, fact_bank.json, topic_brief.json, and research_brief.json exist before content planning.
• cross_check_review.json status == "pass" and research_sufficient is true before any provider copy generation.
• script.json, scene_plan.json, and video_spec.json exist before render in provider/lambda/live modes.
• render_payload.json contains a validated storyboard; contract mode synthesizes the default statement/context/card/cta storyboard, provider mode may use DeepSeek-authored adaptive scenes.
• artifact_manifest.json binds video/final.mp4, cover/final.jpg, render props, selected candidate, and publish mode by SHA256.
• video_optimization_loop.json status == "pass"; failed attempts must stay under optimization/video/attempt-XX/ and must not bind final media.
• visual_audit_deterministic.json status == "pass"; it must not skip missing ffmpeg/tesseract, silent audio, or invalid captions.
• gemini_visual_audit.json status == "pass" in provider/live mode.
• douyin_upload_check.json status == "pass"; live mode additionally requires post/post_publish_verification.json status == "pass".
• release_gate.json records the final release/no-release decision and run_final.json is written last as the final source of truth.

Do not replace missing evidence fetches, provider calls, media tools, remote publish evidence, audio, captions, or visual checks with mock output, silent media, placeholders, or fallback artifacts. Fail with the missing dependency, fetch error, or provider error.