Orchestrate research programs under `kb/programs/`, including program state, open questions, evidence requests, decision logs, reporting events, and skill routing.
Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.
A direct command skips the review prompt. Inspect the source before running it.
Orchestrate research programs under `kb/programs/`, including program state, open questions, evidence requests, decision logs, reporting events, and skill routing.
Use this skill to anchor work to a concrete research program.
Workflow
Create or reopen a program under kb/programs/<program-id>/.
Keep state.yaml aligned with workflow counts and selected context.
Persist workflow/open-questions.yaml, workflow/evidence-requests.yaml, workflow/decision-log.md, and workflow/reporting-events.yaml.
Treat route keywords as factual hints only. For ambiguous, negated, or multi-step requests, the runtime Agent authors an ordered route over the complete formal owner catalog; the script validates formal membership and dependency order but never limits the Agent to keyword hits or decides the workflow semantically. The kb-cli adapter is not a routable owner; source-unit analysis routes to the unit-analyst facade while its internal implementation keeps the historical owner identity.
Keep user constraints and resource boundaries visible in the program state.
Program writes are serialized per program; attach-unit also backfills the unit-side program_ids.
Treat cross-program planning as an Agent judgement over a complete factual candidate snapshot. Scripts enumerate and validate; they never assign semantic value scores or choose a winner.
Shared Contract
Program coordination artifacts are durable inputs for later reopen, not chat-only summaries.
An Agent-authored RouteDecision is a closed record with the current task digest, positive and negated intents, ordered steps, rationale, formal owners, dependencies, and governance gates. The verifier checks shape/current binding/order only; it never derives intents from keywords or grades the rationale.
Any promised resumable deliverable, such as a batch survey or technical roadmap, must be written into a program next_actions entry before the conversation says kb next can resume it. Persisted program actions and loose maintenance suggestions are peer candidates for the Agent to compare; neither category has a fixed priority. Completed units do not generate work merely because a refresh ran.
Decision records require non-empty agent-authored canonical claims with verified evidence and stay pending_user_confirmation until the user explicitly confirms the current receipt. A call without claims only prepares workflow/decision-fill.yaml in awaiting_agent_fill; it does not append a decision, update last_decision, or emit a reportable decision event.
Public review routes an accepted decision to private confirm-decision with current-message authorization, or a rejected decision to private reject-decision without requiring a signature. Rejection changes the canonical decision and all of its claims to rejected, updates last_decision, and never creates a confirmed reporting event.
Pending and confirmed decision events both bind the canonical decision subject, claim ids, content digest, and verification receipt. Report consumers must resolve that binding; an event name or confirmation_status string cannot manufacture trust.
report-author should read from workflow/reporting-events.yaml, so important state changes must emit reporting events.
Script-generated timestamps are stored in UTC.
kb next is a pure read. If no current portfolio decision exists, or its candidate/state/preference binding is stale, request an Agent planning pass instead of falling back to a fixed priority rule.
A PortfolioDecision is Agent-filled and must bind decision_id, the current candidate_snapshot_digest, one or more selected_action_ids, non-empty rationale, expected_information_gain, , a current effective , and a timezone-aware . The script checks shape and current bindings only; it never grades the rationale.
When a program actually involves phased experiments, route experiment planning, run logs, diagnosis, and conclusion gates to experiment-workbench. The Agent may persist phase-specific inputs, outputs, resource budgets, evaluation criteria, failure handling, and artifact inventories as program actions or design notes. GPU allocation, training schedules, simulation environments, ablations, and executor-agent feedback formats are domain-specific options—not mandatory requirements for ordinary research programs. Any winner, causal conclusion, or phase-advance judgement still needs evidence and the existing user-confirmation gate.
启动澄清（Agent 用）
新建 program 还是续接已有？默认续接同题 program。
研究问题与本阶段目标一句话？默认从当前对话提炼后复述确认。
现在要挂接哪些 unit？默认稍后随分析逐步 attach。
cost_and_risk
research-orchestrator + plan
preference_selection_id
decided_at
Candidate snapshots contain every legal persisted next action plus open evidence requests, open questions, ready Agent work, every cross-owner pending judgement, resumable composite survey state, loose-unit maintenance, due monitor subscriptions, and unresolved completed-monitor outcomes. blocking, declared priority, and due time are facts for the Agent, not an automatic winner. Terminal programs produce no ordinary program-work candidates.
Planning-required JSON carries no compatibility winner list: items is empty until a current PortfolioDecision exists. A side judgement binds its exact review snapshot; a completed-monitor outcome binds run revision, run digest, and outcome digest until it receives a durable disposition.
Portfolio history is append-only. State, evidence, unit content, candidate membership, or effective-preference changes make the latest bound decision stale and require the Agent to plan again.
A human gate is never safe to continue automatically. A planning choice that itself asserts a research winner, baseline, idea, causal conclusion, or other judgement must reference a verified program decision and continue through the existing user-confirmation gate; portfolio planning cannot confirm it.
