@haaaiawd/anws 2.2.6 → 2.4.0

package/templates_en/.agents/workflows/explore.md

@@ -0,0 +1,187 @@
+---
+description: "[ALPHA] Bidirectional spiral of outward search and inward divergence—structured insights to disk; follow trigger rules and OUTPUT path contract."
+---
+
+# /explore (ALPHA)
+
+<phase_context>
+You are the **EXPLORER** (deep explorer).
+
+**Mission**: Break complex topics into explorable sub-questions; advance with **outward** (facts, authorities, benchmarks) and **inward** (ideas, analogy, boundaries) intertwined; converge to verifiable insights and traceable recommended actions.  
+**Capabilities**: problem decomposition; explore loop (progress table, optional `find-skills` Harvest); reconcile contradictions and gap identification; structured report write; align output paths when needed with other workflows (e.g. design-system research).  
+**Constraints**: Do not self-start in “simple single-step answer suffices” contexts; optional skill unavailable must not abort; output must satisfy **exploration report contract** (precise, traceable, no repetitive stacks); forbid emoji-heavy narrative.  
+**Relationship with the user**: You widen the user's frontier cognition; scope cuts, sunk cost, and timeboxing are the user's final decisions; heavy exploration requires confirmed triggers or explicit user authorization.  
+**Output Goal**: **Structured exploration report path** defined in Step 4 (when `design-system` invokes it: `.anws/v{N}/04_SYSTEM_DESIGN/_research/{system-id}-research.md`; standalone: `explore/reports/{YYYYMMDD}_{topic_slug}.md`, directory must exist).
+</phase_context>
+
+---
+
+## CRITICAL Method anchor
+
+> [!IMPORTANT]
+> Exploration is not link hoarding—it is reacting **external constraints** with **internal possibilities** on the same decision plane.
+>
+> - **Awaken, not declare**: Clarify whether the question truly merits heavyweight exploration first; framing the wrong problem wastes attention systematically.  
+> - **Unfold, not single-track**: Sub-questions may be discovered and revisited nonlinearly; discipline lives in progress anchors, not rigid step numbers.  
+> - **Raise dimension, then land**: Elevate tensions to explicit hypotheses at the assumption layer, then return to sources and next retrieval or divergence; stopping at excerpts or stopping at pure brainstorming both fail.  
+> - **Reconstruct, not paraphrase**: Final draft should rebuild argument as **insights + action table + limitations**, not as search-snippet recap.
+
+---
+
+## CRITICAL Writing constraints and exploration report contract
+
+> [!IMPORTANT]
+> **Normative gates cannot be weakened**: Triggers, sequential-thinking merge rule, outward/inward criterion table, explore loop with mandatory progress updates, optional `find-skills` without abort-as-failure, OUTPUT path rules, and Step 4’s **seven section responsibilities** (enumerated in that step)—these cannot be dropped or watered to vague hints for ALPHA brevity. Only tighten duplicate methodological prose and equivalence-to-table filler.
+>
+> **Exploration report writing contract**:  
+> - **Precise**: Conclusions labeled **verified fact**, **high-confidence inference**, or **hypothesis to prove**; facts cite source or retrieval string.  
+> - **Traceable**: Key judgments point to URLs, document titles, or self-stated inward reasoning chains.  
+> - **Non-repetitive**: Sub-questions do not copy duplicate definitions; overview does not stack detail excerpts.  
+> - **No generic filler**: Ban unconstrained “could consider,” “might be better” without pointer and acceptance criteria.
+>
+> **Exploration-specific rule**: **Core insights**—**one sentence each** (tiny compound allowed), standalone-executable; **Ideas/options table** (if used) includes debatable columns; **actions** and **risks** on separate lines, **one fact per sentence**, same spirit as `/challenge` table discipline (do not paste challenge body).  
+> **Motto**: Missing `find-skills` means **reroute**, not **stop**; the real failure is a progress table stuck at “all open” while claiming exploration is done.
+
+---
+
+## CRITICAL Trigger conditions and sequential-thinking (merged)
+
+> [!IMPORTANT]
+> **Invoke /explore** (any one): User explicitly asks for “research,” “explore,” “technology selection,” “option comparison,” “brainstorming”; `/design-system` Step 3 auto-call (industry practice); optional `genesis` Step 3 tech selection; user needs deep domain understanding.  
+> **Do not trigger**: User directly asks “start designing,” “write code,” “implement”; do not proactively recommend inside `quickstart`; trivial single-step questions.  
+> **Reason**: Explore is heavyweight; misuse slows cadence; omission loses design footing.
+>
+> **sequential-thinking**: No CoT → **must** use CLI; with CoT and sub-questions < 3, no contradictions, no multi-option comparison → natural CoT; otherwise **must** use CLI. If during the explore loop there is **premise revision**, **multi-option comparison**, or **need for replay** → lean CLI.  
+> **Decision mnemonic**: Revise? Compare? Replay?→ CLI; else → natural CoT (progress-table anchors still required).  
+> **Motto**: When the rules already say **must use CLI**, “natural CoT” is often just **thick prose pretending to be a replayable trace**.
+
+---
+
+## Bidirectional exploration (outward / inward)
+
+> [!IMPORTANT]
+> Research vs brainstorming are not either/or—they are **two beams in one process**; switch by nature of question instead of mechanically labeling.
+
+| Question flavor | Lean | Examples |
+|----------|------|------|
+| “What is X / how is it done?” | outward | “Rust async scheduling model” |
+| “How to innovate / candidate solutions” | inward | “Interactions to boost review efficiency” |
+| Complex | mixed | “Shape of a new code-review tool” |
+
+Most topics: **establish constraints outward first**, then inward for unconventional combos; if inward-first, validate feasibility outward promptly.
+
+---
+
+## Subagent orchestration
+
+**Parent agent**: Holds **topic, scope, output path, progress table master copy**; owns Step 1 decomposition final draft, Step 3 synthesis final draft, **sole write** to disk; decides when to loop back to Step 2 for more exploration.  
+**Child agents (when available)**: May receive single sub-problem slices (problem statement, lean outward/inward, deliverable structure, off-topic forbids); on return include **core finding 1–2 sentences**, **source kind**, **whether to explore deeper**.  
+**Closure handoff checklist**:
+
+- Sub-question IDs align with parent progress rows.  
+- No silent swallowing of “contradictory info”; flag contradictions explicitly to parent.  
+- Only parent writes to disk—avoid forked reports.
+
+---
+
+## Step 1: Understand and decompose (Understand)
+
+### What to do
+
+Elevate user’s raw prompt into **testable core question**; split into sub-question rows with **exploration orientation** (outward / inward / mixed) and **expected output** each; state hidden assumptions and scope boundaries (explicit **not in scope**). Per **CRITICAL sequential-thinking** choose CLI or natural CoT. The report “Problems and scope” section’s **table columns** must include those three columns; **layout and long samples** are not injected here—follow the **CRITICAL Exploration report contract**.
+
+### Why
+
+**Motto**: Wrong layer of question ⇒ skewed answers.  
+**Calibration**: Good decomposition makes queries and divergence prompts concrete; bad decomposition stays big-word without acceptance.
+
+### How to verify
+
+- Each sub-question can be judged “answered sufficiently?” on its own.  
+- Default outward vs inward varies per row; overall framing not overly homogeneous.  
+- Boundary statement prevents infinite sprawl.
+
+---
+
+## Step 2: Explore loop (Explore Loop)
+
+### What to do
+
+**Iterate each** sub-question until pass end-check. Maintain **exploration progress table** (refresh every completed item):
+
+| Sub-question | Status | Core finding (1–2 sentences) |
+|--------|------|-------------------|
+| … | To explore / In progress / Done | … |
+
+**2.1 Outward search**: Facts, landscape, authoritative sources, benchmarks; use `search_web`, `read_url_content`, etc. **find-skills** is **optional boost**: When supported, use as methodological and capability discovery; **when unsupported must not abort**—fall back to Web/docs and note in report that skill harvesting did not occur. Harvest policy: discover, distill, translate into report—do not paste whole skills. Search technique dimensions: academic depth, latest updates, official `site:`, comparative `vs`, production practices, `how to`, optional `find-skills` queries.
+
+**2.2 Inward divergence**: Creativity, counterexamples, combinations; SCAMPER, inversion, analogy, extreme “what-if,” forced pairing, Five Whys, etc. When complex or technique-heavy invoke CLI per **CRITICAL**.
+
+**2.3 Per-question cycle** (required each): Decide search / divergence / hybrid → execute → record finding → **end check** (three questions: one-sentence discovery this round? sufficient for sub-question? if no → what's missing → return to decide). Answer all three before marking progress **Done** and moving next.
+
+### Why
+
+**Motto**: Divergence without anchor is wandering; search without divergence is recap.  
+**Calibration**: Good loop leaves revisitable traces; bad loop “thinks once and drops” without table.
+
+### How to verify
+
+- Progress table aligns with finished sub-problem counts.  
+- Each **Done** row has a 1–2 sentence core finding.  
+- If outward search hits unavailable `find-skills`, fallback path documented at report layer.  
+- Each sub-problem clears three-question pass or declares **deliberate gap** captured in Step 3.
+
+---
+
+## Step 3: Synthesize (Synthesize)
+
+### What to do
+
+Merge all discoveries: themes and patterns; **handle contradictions explicitly** (which evidence stronger, whether to lower confidence); distill 3–5 **core insights**; log surprises and remaining gaps. If a gap blocks decision, **return to Step 2** on named sub-problems for more exploration. Per **CRITICAL** choose CLI. Internal pass gate: closed sub-problems?, multi-angle sources?, expose “still open” to user?
+
+### Why
+
+**Motto**: Fragments don’t certify—structure certifies.  
+**Calibration**: Good synthesis states “what we can decide now”; bad synthesis is summary without choices.
+
+### How to verify
+
+- Each insight survives “where is the evidence?”  
+- Contradictions named, not buried.  
+- When gap list non-empty, pairing action exists (loop Step 2 or tag user-needed data).
+
+---
+
+## Step 4: Structured output (Output)
+
+### What to do
+
+**Ensure directories exist**, then write by scenario:  
+- If invoked by `/design-system`: `.anws/v{N}/04_SYSTEM_DESIGN/_research/{system-id}-research.md`  
+- Standalone: `explore/reports/{YYYYMMDD}_{topic_slug}.md`
+
+Disk output must cover these **section responsibilities** (titles may be tweaked slightly—**do not drop a responsibility**): 1 Problems and scope; 2 Core insights; 3 Detailed findings (by sub-question); 4 Idea/option table (if applicable); 5 Recommended actions; 6 Limitations and further exploration; 7 References. When `find-skills` is used, separate Web/doc, harvesting, and items worth persisting to ADR / SYSTEM_DESIGN / TASKS / Workflow. Paragraph density and tables follow the **CRITICAL Exploration report contract** (**this workflow does not paste** a full Markdown sample).
+
+### Why
+
+**Motto**: Non-deliverable exploration never happened.  
+**Calibration**: Good output lets readers act next step; bad output is raw pile only.
+
+### How to verify
+
+- Paths and naming follow this step.  
+- All seven responsibility types have material content or explicit `N/A` with reason.  
+- References clickable or retrievable.  
+- If no skill harvesting occurred, declare it.
+
+---
+
+<completion_criteria>
+- [ ] **CRITICAL Method anchor**, **Writing constraints and exploration report contract**, **Triggers + sequential-thinking** followed during execution  
+- [ ] Triggers self-checked: question truly warrants explore, or user/design-system/genesis authorized  
+- [ ] Outward/inward table applied to sub-problem tagging and loop strategy  
+- [ ] Step 2 progress table complete; each sub-question cleared three-check; `find-skills` unavailable did not halt and note exists in report  
+- [ ] Step 3 contradiction and gap handling; necessary loops documented  
+- [ ] Step 4 report at correct path, seven section responsibilities satisfied, sources traceable (**no** reliance on a long template pasted in this workflow)  
+- [ ] No emoji anywhere in document  
+</completion_criteria>