feat: promote lab workflow improvements for 0.3.0

.agents/skills/omp-flow-brainstorm/SKILL.md

@@ -16,10 +16,11 @@ description: Explore and clarify a consented omp-flow task Bundle when direction
 
 For a non-trivial Explore, use repository/Wiki evidence and human calibration to form or revise a
 provisional first-principles anchor（第一性锚定）before broad Research. Keep it in useful authored
-prose, not a required file or fixed checklist. It should be sufficient to expose the observable
-problem, current principal contradiction（主要矛盾）, irreducible outcome, human boundaries, a
-strong counter-hypothesis, and evidence that would revise the framing. Mechanical low-ambiguity
-work may proceed directly when this step would not improve a decision.
+prose, not a required file or fixed checklist. It should be sufficient to expose the concrete
+actors and action, observable intended outcome, current principal contradiction（主要矛盾）,
+irreducible outcome, human boundaries, a strong counter-hypothesis, and evidence that would
+revise the framing. Mechanical low-ambiguity work may proceed directly when this would not improve
+a decision.
 
 Concentrate interaction on the material decision frontier:
 
@@ -33,11 +34,13 @@ Concentrate interaction on the material decision frontier:
 ## Interview Contract
 
 1. Read the entry Concept and follow only useful links from the Bundle.
-2. Record the initial request and known repository facts in the assigned framing Concept.
+2. Record the initial request, concrete actors/action/outcome, current principal contradiction,
+   and known repository facts in the assigned framing Concept.
 3. Inspect code, tests, docs, prior tasks, relevant history, and applicable durable knowledge
    through the native `omp-flow-wiki` Skill before asking repository-answerable questions.
-4. Ask the user one material decision question at a time. Ask only about intent, preference,
-   scope, risk tolerance, or ambiguity that evidence cannot answer.
+4. Ask the user one material decision question at a time. Ask only about purpose, value, risk
+   tolerance, preference, or use that repository and other relevant practice evidence cannot
+   establish.
 5. Keep confirmed facts, assumptions, disagreements, alternatives, constraints, success criteria,
    and open questions distinct.
 6. Present materially different directions only when a real choice exists. Follow the human-first

.agents/skills/omp-flow-check/SKILL.md

@@ -18,6 +18,15 @@ Fail closed unless all are available:
 
 ## Review Order
 
+Only Main/coordinator may dispatch, correlate operations/receipts, obtain or record human
+calibration, and choose a workflow transition. Those coordinator actions are inapplicable to an
+already-dispatched Reviewer: it must not dispatch or self-redispatch, govern, calibrate,
+transition, or selectively reinterpret coordinator clauses. The Reviewer still owns the complete
+bounded independent Review and assigned Review Concept: compare actual code and verification with
+the intended work/Design consequence, and report evidence contradicting the Design or its
+principal-problem framing through that output rather than let a handoff or superficial green check
+substitute for the practice result.
+
 1. Resolve the completed predecessor operation and use its output path as the required handoff.
    Read the work Concept and handoff, verify they identify each other, then inspect the real diff
    and changed files.

.agents/skills/omp-flow-decompose/SKILL.md

@@ -12,6 +12,14 @@ description: Map an approved omp-flow design into descriptive, bounded work Conc
 
 ## Procedure
 
+Only Main/coordinator may dispatch, correlate operations/receipts, obtain or record human
+calibration, and choose a workflow transition. Those coordinator actions are inapplicable to an
+already-dispatched Architect: it must not dispatch or self-redispatch, govern, calibrate,
+transition, or selectively reinterpret coordinator clauses. After linked human QbD 1 approval, the
+Architect still owns the complete bounded Decompose responsibility and assigned work outputs: map
+the approved Design into bounded, independently reviewable work Concepts and an authored prose
+execution view, without inventing approval or encoding a machine DAG.
+
 1. Identify the smallest independently implementable and reviewable work items. Fold scaffolding and
    docs into the deliverable that needs them.
 2. Write one descriptive work Concept per item with objective, in/out scope, useful linked inputs,
@@ -26,8 +34,10 @@ description: Map an approved omp-flow design into descriptive, bounded work Conc
 
 ## Exit Gate
 
-Prepare QbD 2 only when every work item is bounded, testable, understandable from its entry path,
-and worth independent review. Load `omp-flow-qbd` next.
+In Main/coordinator context only, prepare QbD 2 and load `omp-flow-qbd` when every work item is
+bounded, testable, understandable from its entry path, and worth independent review. The
+already-dispatched Architect returns the assigned work map and does not choose or perform that
+transition.
 
 ## Red Flags
 

.agents/skills/omp-flow-design/SKILL.md

@@ -13,11 +13,20 @@ description: Turn a selected synthesis in an omp-flow task Bundle into linked re
 
 ## Procedure
 
-1. Dispatch the native Architect with Bundle root, selected synthesis entry, bounded objective,
-   relevant linked Concepts, explicit output ownership, actor ID, and receipt. Create the
-   operation first and forward its complete returned assignment unchanged with native item
-   `id = actor_id = actorId` and matching descriptor role. Preserve the strict v1
-   `ompFlowDispatch` descriptor as the first non-blank line.
+Before native dispatch, only Main/coordinator may dispatch, correlate operations/receipts, obtain
+or record human calibration, and choose a workflow transition. Those coordinator actions are
+inapplicable to an already-dispatched Architect: it must not dispatch or self-redispatch, govern,
+calibrate, transition, or selectively reinterpret coordinator clauses. The Architect still owns
+the complete bounded Design responsibility and assigned PRD/Design/decision/interface outputs:
+explain each proposed component's concrete causal purpose in plain language, without relying on its
+own terminology, and solve the evidenced principal contradiction rather than a secondary or
+invented mechanism.
+
+1. In Main/coordinator context only, dispatch the native Architect with Bundle root, selected
+   synthesis entry, bounded objective, relevant linked Concepts, explicit output ownership, actor
+   ID, and receipt. Create the operation first and forward its complete returned assignment
+   unchanged with native item `id = actor_id = actorId` and matching descriptor role. Preserve the
+   strict v1 `ompFlowDispatch` descriptor as the first non-blank line.
 2. Rewrite `prd.md` into observable requirements, non-goals, constraints, acceptance criteria, and unresolved product decisions. Remove temporary brainstorm duplication.
 3. Write `design.md` with components, interfaces, data flow, state ownership, error behavior, migration/compatibility, verification strategy, and rejected alternatives.
 4. Add linked decision, interface, brief, or finding Concepts only when they improve navigation or
@@ -34,8 +43,10 @@ description: Turn a selected synthesis in an omp-flow task Bundle into linked re
 - No unresolved question blocks decomposition.
 - Linked sources and accepted contracts agree with the Design.
 
-Load `omp-flow-qbd` for independent challenge and record the human decision as a linked Concept.
-Do not create implementation work before human QbD 1 approval.
+In Main/coordinator context only, load `omp-flow-qbd` for independent challenge and obtain and
+record the human decision as a linked Concept. The already-dispatched Architect returns its
+assigned outputs and does not choose or perform that transition. Do not create implementation work
+before human QbD 1 approval.
 
 ## Red Flags
 

.agents/skills/omp-flow-finish/SKILL.md

@@ -17,14 +17,34 @@ description: Complete integration verification, deliberate knowledge harvest, co
 2. Compare the integrated result against PRD acceptance criteria, cross-row interfaces, migrations, docs, and compatibility requirements.
 3. Resolve failures through the owning work/design Concept and repeat independent review or the
    applicable gate. Do not hide failures with warning-only fallbacks.
-4. Review accepted decisions, repeated patterns, and confirmed findings for durable value.
-5. Use the native `omp-flow-wiki` Skill to promote only evidenced, reusable knowledge. If nothing qualifies, state that no harvest was needed.
-6. Inspect final diff and repository status. Commit through the Harness-native Git workflow when requested.
-7. Archive the Bundle only after completion is real and no runtime operation is active.
+4. Review accepted decisions, repeated patterns, and confirmed findings for immediate durable value.
+   Use the native `omp-flow-wiki` Skill only when the evidence already supports promotion; do not
+   invent knowledge to satisfy Finish.
+5. Inspect final diff and repository status. Create the requested Git checkpoint before Archive so
+   the Task Bundle has a reproducible source commit/tree.
+6. Archive the Bundle only after completion is real and no runtime operation is active. Inspect the
+   returned `sleepSource`; never reconstruct one from the archive path.
+7. When `sleepSource.ready` is true, run `sleep start --source <receipt> --actor-id <native-id>`,
+   load the native `omp-flow-sleep` Skill, and forward its complete assignment unchanged as the sole
+   native assignment input. Normal `sleep start` delivery is primary. If that delivery is lost or
+   visibly truncated while its active receipt is known, recover only by capturing the complete
+   `sleep show <receipt>` subprocess/API stdout through successful process completion, parsing all
+   captured stdout as JSON into the show run mapping, requiring string `run.assignment`, and
+   forwarding that decoded string unchanged as the sole native assignment input, including its
+   trailing LF. Rendered Bash/terminal output is diagnostic only: never copy or parse it, add a
+   prefix or suffix, reconstruct or reserialize the assignment, fall back to `sleep list`, or issue
+   a duplicate `sleep start`. A nonzero exit, incomplete capture, invalid JSON, or missing or
+   non-string `run.assignment` is delivery failure and must stop before native dispatch. Complete
+   the Sleep run before final handoff; zero Candidates is valid.
+8. When the source is not ready, report its exact reason instead of asking the user to operate Sleep
+   manually or manufacturing a checkpoint. Sleep Candidate review and Wiki promotion remain
+   separate work after the run.
 
 ## Final Handoff
 
-Report behavior delivered, important decisions, verification results, archived task path, commit identity when created, and remaining non-blocking risks.
+Report behavior delivered, important decisions, verification results, archived task path, commit
+identity when created, Sleep run/Candidate outputs or the exact unavailable reason, and remaining
+non-blocking risks.
 
 ## Red Flags
 

.agents/skills/omp-flow-implement/SKILL.md

@@ -18,6 +18,14 @@ Fail closed unless all are available:
 
 ## Procedure
 
+Only Main/coordinator may dispatch, correlate operations/receipts, obtain or record human
+calibration, and choose a workflow transition. Those coordinator actions are inapplicable to an
+already-dispatched Implementer: it must not dispatch or self-redispatch, govern, calibrate,
+transition, or selectively reinterpret coordinator clauses. The Implementer still owns the
+complete bounded implementation and assigned handoff: return the approved Design to practice
+through code, execution, and real verification, and report contradictory evidence as a
+Design-return signal rather than special-case reality to preserve Design authority.
+
 1. Read the work Concept first. Follow only useful links to design, decisions, sources, and
    relevant existing code/tests.
 2. Restate the bounded objective, output boundary, done conditions, and verification before editing.
@@ -32,7 +40,8 @@ Fail closed unless all are available:
 Write or update the promised handoff Concept and link it back to the work it implements. Return
 `DONE`, `DONE_WITH_CONCERNS`, `NEEDS_CONTEXT`, or `BLOCKED`, plus changed files, commands, test
 counts, decisions, caveats, output path, actor ID, and receipt. Do not mutate runtime/session
-records. Implementation success goes to an independent review operation.
+records. The Implementer returns this assigned handoff; only Main/coordinator chooses and
+dispatches the independent Review operation.
 
 ## Red Flags
 

.agents/skills/omp-flow-learn/SKILL.md

@@ -0,0 +1,127 @@
+---
+name: omp-flow-learn
+description: Maintain and use repository-root .omp-flow/learn for deliberate, reciprocal, resumable Human–Agent co-learning when a human is present and pursuing understanding rather than production execution.
+---
+
+# OMP-Flow Learn
+
+Use repository-root `.omp-flow/learn/` when a human is actively pursuing mutual understanding,
+explanation, challenge, reconsideration, or reflection. Learn is a human-present dialogue in which
+both participants may correct their models; autonomous harvesting, passive one-way tutoring, or
+summary generation is not a substitute.
+
+Production work normally pauses during a Learn conversation. Source and code inspection, worked
+examples, counterexamples, and discussion remain available, but do not execute production work
+until the human returns the conversation to production.
+
+## Consult and bootstrap
+
+Open a known Learn Concept directly when its path is available. Otherwise start at
+`.omp-flow/learn/index.md` and follow only ordinary links useful to the current discussion. Indexes
+support navigation; they are not exhaustive manifests.
+
+When Learn is genuinely needed and the root is absent, bootstrap only the root index and the
+`daily/`, `threads/`, and `retrospectives/` indexes described by the canonical minimal structure.
+Do not pre-populate speculative encounters, a taxonomy, or companion records.
+
+## Preserve a resumable checkpoint
+
+Use judgment to select or create a bundle immediately below the relevant stable directory. Name a
+new bundle `YYYY-MM-DD-<topic>/` and maintain its authored `index.md` so a later human and Agent can
+recover the useful current understanding, disagreement, correction, or open point. Add only
+concrete descriptive Markdown Concepts that make the learning easier to resume or challenge.
+
+A Daily records a useful dated encounter and may stand alone. A Thread is warranted only when a
+topic needs continuity across encounters. A Retrospective is warranted only for a substantive
+lesson about understanding or collaboration. Do not manufacture a Thread or Retrospective to
+complete a set.
+
+Treat the human's corrections, rejected premises, values, business context, and objections as
+learning material. Make corrected Agent assumptions visible when that helps resumption. Accuracy
+comes from reciprocal correction and evidence, not from fluent prose or Agent confidence.
+
+## Authority and continuity boundaries
+
+Learn is provisional authored context, not proof of understanding or completion.
+
+- A Learn artifact cannot advance, pause-state, approve, dispatch, or otherwise alter an omp-flow
+  Task. A Task link is contextual; commitments remain authored in the applicable Task Bundle.
+- A Learn artifact is not Wiki authority and is never promoted automatically. Durable knowledge
+  reaches `.omp-flow/wiki/` only through ordinary evidence-based Wiki maintenance.
+- Cross-session continuity is limited to visible authored artifacts. Never claim a hidden user
+  profile, persistent personal memory, or model-weight change.
+
+## Select the smallest useful lens
+
+The lenses below are a selectable conversational repertoire, not stages, a checklist, a catechism,
+an assessment, or an Agent-led teaching script. Choose the smallest lens that addresses the live
+misunderstanding. Combine lenses only when one supplies a relevant corrective to another. Concrete
+project facts, code, sources, worked cases, and the human's correction outrank philosophical labels.
+Stop when the participants have a useful explanation, disagreement, or open point; do not march
+through the remaining lenses.
+
+### Feynman-style self-explanation
+
+Invite a plain account of one concrete causal link. Find the first hidden term, unsupported leap, or
+failed application, then repair it from a source or worked case. Fluency is only a possible gap
+diagnostic and never proof of mastery. Do not attribute the modern packaged “Feynman Technique” to
+Feynman.
+
+### Cargo Cult Science integrity
+
+For a consequential explanation, volunteer the material contrary fact, plausible alternative,
+uncertainty, and observation that could change the conclusion so the other participant can judge
+it. This is judgment about material disclosure, not four mandatory headings, a fixed response form,
+or a ritual inventory of every caveat.
+
+### Grounded first principles
+
+Distinguish observed fact, definition, hard constraint, value choice, assumption, and convention.
+Trace only premises capable of changing the present conclusion, retain the conditions under which
+it should be revised, and stop when further reduction would not help. Existing knowledge and
+history are evidence to examine, not impurities to discard.
+
+### Practice-led principal-contradiction analysis
+
+Offer the principal contradiction, dominant aspect, and material secondary coupling only as a
+revisable prioritization hypothesis tied to current effects. Name the changed condition or practical
+observation that would reorder it. Returning to practice may mean sources, code inspection,
+examples, counterexamples, a diagram, or later observable verification; it does not authorize
+production execution while Learn is paused. Do not import a universal single-cause claim or
+political doctrine.
+
+### Material history or genealogy
+
+Use a concise temporal account only when it clarifies the present. Connect a concrete problem and
+conditions to a prior arrangement, emergence or adoption, limitation, and later change; include a
+discontinuity where needed and label interpretive causality. Reject inevitable or present-purpose
+origin stories, mechanically economic single-causality, and political doctrine; examine interacting
+causes and source support.
+
+### Reciprocal Socratic questioning
+
+Use a few genuine reasons, counterexamples, and consequence questions to test either participant's
+claim. Either participant may reject a premise, ask the Agent to answer directly, or leave an aporia
+open. Questions yield to direct explanation or admitted uncertainty when more useful and never
+steer assent toward a conclusion the Agent selected in advance.
+
+### Minimal structural box-drawing
+
+Draw labeled ASCII, Unicode, or Mermaid boxes and arrows only when a few elements at one useful
+abstraction level make a particular structural inference cheaper than prose. Check the picture
+against sources, code, examples, or the participants' model; revise or discard it when it does not
+help. A diagram is not inherently clarifying.
+
+These lenses create optional corrective friction rather than a composite routine: explanation
+answers to integrity and evidence; alleged hard constraints to history; principal ranking to
+secondary couplings and changed conditions; historical narrative to discontinuity and interacting
+causes; questions to reciprocal answerability; and diagrams to prose, source, or practice
+comparison. They do not adopt a complete philosophical school or establish mastery.
+
+## Keep Learn conversational
+
+Do not introduce an entry command, pause or save signal, fixed encounter stages, required
+transcript or summary, mandatory headings, score, metric, completion gate, approval ceremony, or
+per-edit human approval rule. There is no parsed body schema, lifecycle record, second task system,
+or automatic learning judgment. Preserve only what is genuinely useful for reciprocal resumption,
+and return to production only when the human changes the conversation back to production work.