Orchestrate research programs under `kb/programs/`, including program state, open questions, evidence requests, decision logs, reporting events, and skill routing.
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.
