sam/pi-config

Fork 0

Files

Sam Rolfe 31b4110c87 Add 5 pi extensions: pi-subagents, pi-crew, rpiv-pi, pi-interactive-shell, pi-intercom

2026-05-08 15:59:25 +10:00

18 KiB

Raw Blame History

name, description, argument-hint

name

description

argument-hint

research

Answer structured research questions about a codebase using targeted parallel analysis agents, then synthesize findings into a research document in thoughts/shared/research/. Internally dispatches the scope-tracer agent to formulate trace-quality research questions, then answers them. Use when the user wants in-depth research on a codebase area, asks to "research X", or needs answers to architecture or behavior questions before designing changes.

free-text research prompt

Research

You are tasked with answering structured research questions by spawning targeted analysis agents and synthesizing their findings into a comprehensive research document. This skill internally dispatches the scope-tracer agent to formulate trace-quality research questions, then answers them.

Input: $ARGUMENTS

Step 1: Trace the Investigation Scope

Argument is empty:

Please provide a free-text research prompt.

Then wait for input.

Detect chained discover artifact: If $ARGUMENTS mentions a path matching thoughts/shared/discover/.*\.md, read it FULLY using the Read tool (no limit/offset) before scope-tracer dispatch:
- Translate each ### [Decision title] block in the FRD's ## Decisions section into a Developer Context entry: **Q (discover: <Decision title>): <Question text>** followed by A: <Chosen text>. Hold these entries in main context — they're recorded in the research artifact's Developer Context section in Step 4 (write document).
- Use the FRD's ## Recommended Approach text (1-2 sentences naming the architectural shape) as the topic body for the next sub-step's scope-tracer prompt. The full discover artifact path stays in $ARGUMENTS so scope-tracer's "read mentioned files first" rule picks up the file naturally for additional context.
- Carry the FRD's Open Questions forward verbatim into the research artifact's Open Questions section in Step 4.
- If $ARGUMENTS is plain free-text or mentions a non-discover path, skip this sub-step and proceed directly to scope-tracer dispatch with $ARGUMENTS as the topic.
Dispatch the scope-tracer agent to formulate trace-quality research questions for the user's topic:
```
Agent({
  subagent_type: "scope-tracer",
  description: "trace scope",
  prompt: "$ARGUMENTS"
})
```
The agent reads any mentioned files, sweeps anchor terms via grep/find/ls, reads 5-10 key files for depth, then emits a Discovery Summary + 5-10 dense numbered questions inline in its final message. Nothing is written to disk.
Parse the agent's final message as the questions artifact body. Extract: Discovery Summary (3-5 sentence file-landscape overview), Questions (numbered dense 3-6 sentence paragraphs).
Read key shared files referenced across multiple questions into main context — especially shared utilities, type definitions, and integration points that multiple questions mention.
Analyze question overlap for grouping:
- Parse all question paragraphs and extract file references from each
- Identify questions that share 2+ file references — these are candidates for grouping
- Group related questions together (2-3 questions per group max)
- Questions with no significant file overlap remain standalone
- Target: 3-6 agent dispatches total (grouped + standalone)

Report scoped status:

[Scoped]: ran scope-tracer. {N} questions in {G} groups, {M} shared files.

Step 2: Dispatch Analysis Agents

Spawn analysis agents using the Agent tool. All agents run in parallel.

Default agent: codebase-analyzer for all codebase questions. This agent has Read, Grep, Glob, LS — it can trace code paths, find patterns, and analyze integration points.

Exception: Questions that explicitly reference external documentation, web APIs, or third-party libraries → web-search-researcher.

Agent prompt — question-as-prompt:

Each agent receives the dense question paragraph(s) directly as its prompt. The question IS the instruction.

For standalone questions (no grouping):

Research topic: {topic from frontmatter}

Answer the following research question thoroughly with file:line references. Read the files mentioned, trace the code paths described, and provide a complete analysis.

{Full dense question paragraph}

Provide your analysis with exact file:line references. Focus on DEPTH — trace the actual code, don't just locate it.

For grouped questions:

Research topic: {topic from frontmatter}

Answer the following related research questions thoroughly with file:line references. These questions share overlapping code paths — use your cross-question context to provide deeper, more connected analysis.

Question 1: {Full dense question paragraph}

Question 2: {Full dense question paragraph}

For each question, provide your analysis with exact file:line references. Note connections between the questions where the same code serves multiple roles. Focus on DEPTH — trace the actual code, don't just locate it.

Precedent sweep (git-gated):

When commit is available (not no-commit), spawn one precedent-locator agent alongside the question agents with prompt: "Find similar past changes involving {list key files from Discovery Summary}. Search git log for commits that touched these files, similar commit messages, and follow-up fixes. Research topic: {original query}."

Findings go into Precedents & Lessons. Otherwise skip and note "git history unavailable" there.

Wait for ALL agents to complete before proceeding.

Step 3: Synthesize and Checkpoint

Compile findings:
- Match each agent's response to the question(s) it answered
- Cross-reference findings across questions — look for patterns, conflicts, and connections
- Prioritize live codebase findings as primary source of truth
- Use thoughts/ findings as supplementary historical context
- Include specific file paths and line numbers
- Build Code References as jump-table entries for the planner, not narrative (file:startLine-endLine format)
- No multi-line code blocks (>3 lines) — use file:line refs + prose. No implementation recipes — facts only.
- No artifact summaries — link plans/designs in Historical Context, don't summarize their contents. Research describes current codebase state.
Developer checkpoint — grounded questions one at a time:

Start with grounded questions referencing real findings with file:line evidence. Ask ONE question at a time, waiting for the answer before the next. Use a ❓ Question: prefix. Each question must pull NEW information from the developer — not confirm what you already found:

Every question MUST embed at least one file:line reference in the question text — not just in surrounding context. Examples:
- "❓ Question: src/events/orders.ts:45-67 has 3 event hooks but no error recovery path. Is there a retry mechanism elsewhere I'm not seeing?"
- "❓ Question: Pattern-finder found manual mapping at src/services/OrderService.ts:45 (8 uses) vs AutoMapper at src/services/UserService.ts:12 (2 uses). Which should new code follow?"
- "❓ Question: Precedent commit abc123 required a follow-up fix at src/handlers/key.ts:158 for connection leak. Should we account for that pattern in this design?"
Anti-patterns — NEVER ask these:
- "Is this research to understand X or prepare for Y?" — confirmatory, pulls zero new information
- "Does this look correct?" / "Should I continue?" — asks developer to validate YOUR work instead of providing NEW context
- Questions without file:line — if you can't ground it in code, it's not a research question
Question patterns by finding type:
- Pattern conflict: "Found 2 implementations of {X} — which is canonical?" with options citing file:line + occurrence count
- Scope boundary: "Question {N} references files {A,B,C} but analysis shows {D} is the real integration point. Extend scope?" with yes/no + "describe what I missed"
- Priority override: "Questions Q1 and Q2 have competing implications for {area}. Which is load-bearing?" with options
- Integration ambiguity: "Found no connection between {X} and {Y}. Is there an indirect path?" (free-text — can't predict the answer)
Choosing question format:
- ask_user_question tool — when your question has 2-4 concrete options from code analysis (pattern conflicts, integration choices, scope boundaries, priority overrides). The user can always pick "Other" for free-text. Example:
  
  Use the ask_user_question tool with the following question: "Found 2 patterns for retry logic — which is canonical?". Header: "Pattern". Options: "Event-sourced retry (Recommended)" (src/events/orders.ts:45-67 — 3 hooks, matches precedent commit abc123); "Direct retry loop" (src/services/OrderService.ts:112 — single use, no event traceability).
- Free-text with ❓ Question: prefix — when the question is open-ended and options can't be predicted (discovery, "what am I missing?", corrections). Example: "❓ Question: src/events/orders.ts:45-67 has 3 event hooks but no error recovery path. Is there a retry mechanism elsewhere I'm not seeing?"
Anti-pattern — do NOT dump a verbose paragraph mixing analysis with a trailing question:

❌ "The premise inversion is load-bearing for prioritization — it means Site A is a no-op, and the real bloat only hits general-purpose dispatches. Given this, where is the bloat actually landing? Do your skills dispatch named bundled agents — in which case append-mode is irrelevant — or general-purpose — in which case it IS the dominant source?"

✅ Extract the 2 concrete options and call ask_user_question: "Where is the prompt bloat landing?". Header: "Bloat source". Options: "Named bundled agents (Recommended)" (Skills dispatch codebase-analyzer etc. — prompt_mode: "replace", no parent inheritance); "General-purpose agent" (default-agents.ts:11-28 — promptMode: "append", inherits full parent prompt).

Batching: When you have 2-4 independent questions (answers don't depend on each other), you MAY batch them in a single ask_user_question call. Keep dependent questions sequential.

CRITICAL: Ask ONE question at a time. Wait for the answer before asking the next. Lead with your most significant finding.

Present compiled scan (under 30 lines):

Task: {one-line summary}
Scope: {N files across M layers, K integration points}

{Layer name} — {key files and what they do}
{Layer name} — {key files and what they do}
Integration — {N inbound, M outbound, K wiring. Top concern if any}
History — {N relevant docs. Key insight if any}

Best template: {implementation to model after}
Precedents — {N similar changes found. Top lesson if any}
Inconsistencies: {count} found ({short names})

Wait for the developer's response before proceeding.

Incorporate developer input:

Classify each response:

Corrections (e.g., "skip the job scheduler", "use CreateProduct not GetUser"):
- Incorporate directly into synthesis. Record in Developer Context.
New areas (e.g., "you missed the events module"):
- Spawn targeted rescan: codebase-locator + codebase-analyzer on the new area (max 2 agents).
- Merge results into synthesis. Record in Developer Context.
Decisions (e.g., "yes, hook into that event chain"):
- Record in Developer Context. Remove corresponding item from Open Questions.
Scope/focus (e.g., "focus on API layer, UI is out of scope"):
- Record in Developer Context.
After incorporating all input, proceed to Step 4.

Step 4: Write Research Document

Determine metadata:
- Filename: thoughts/shared/research/YYYY-MM-DD_HH-MM-SS_{topic}.md
  - YYYY-MM-DD_HH-MM-SS: Current date and time
  - topic: Brief kebab-case description
- Repository name: from git root basename, or current directory basename if not a git repo
- Use the git branch and commit from the git context injected at the start of the session (or run git branch --show-current / git rev-parse --short HEAD directly)
- Timestamp: run date +"%Y-%m-%dT%H:%M:%S%z" — raw for date: and last_updated:, first 19 chars (T→_, :→-) for filename slug.
- Author: use the User from the git context injected at the start of the session (fallback: "unknown")
- If metadata unavailable: use "unknown" for commit/branch

Write the research document — this document is compressed context for a new session. Include everything the planner needs to make architectural decisions without re-researching:

---
date: {Current date and time with timezone in ISO format}
author: {User from injected git context}
commit: {Current commit hash}
branch: {Current branch name}
repository: {Repository name}
topic: "{User's Research Topic}"
tags: [research, codebase, relevant-component-names]
status: complete
last_updated: {Same ISO timestamp as `date:` above}
last_updated_by: {User from injected git context}
---

# Research: {User's Research Topic}

## Research Question
{Original user query from questions artifact}

## Summary
{High-level findings answering the user's question}

## Detailed Findings

### {Component/Area 1}
- Finding with reference (`file.ext:line`)
- Connection to other components
- Implementation details

### {Component/Area 2}
...

## Code References
- `path/to/file.py:123` — Description of what's there
- `another/file.ts:45-67` — Description of the code block

## Integration Points
{All connections to the researched area. Enumerate each consumer, dependency, and wiring point with file:line. Source from the questions artifact's Discovery Summary + new connections found by analysis agents.}

### Inbound References
- `path/to/consumer.ext:line` — {What references the component and how}

### Outbound Dependencies
- `path/to/dependency.ext:line` — {What the component depends on}

### Infrastructure Wiring
- `path/to/config.ext:line` — {DI, routes, events, jobs, middleware}

## Architecture Insights
{Patterns, conventions, and design decisions discovered}

## Precedents & Lessons
{N} similar past changes analyzed.

### Precedent: {what was added/changed}
**Commit(s)**: `hash` — "message" (YYYY-MM-DD)
**Blast radius**: N files across M layers
  layer/ — what changed

**Follow-up fixes**:
- `hash` — "message" (date) — what went wrong

**Lessons from docs**:
- thoughts/path/to/doc.md — key lesson extracted

**Takeaway**: {one sentence — what to watch out for}

### Composite Lessons
- {Composite lesson 1 — most recurring pattern first, with relevant `commit hash` inline}
- {Composite lesson 2}

## Historical Context (from thoughts/)
{Links only — one line per doc, no summaries of their contents}
- `thoughts/shared/something.md` — {one-line description of what this doc covers}
## Developer Context
**Q (`file.ext:line`): {Question grounded in specific code reference}**
A: {Developer's answer}

## Related Research
- {Links to other research documents}

## Open Questions
{Only questions NOT resolved during checkpoint}

Step 5: Present and Chain

Research document written to:
`thoughts/shared/research/{filename}.md`

{N} questions answered, {M} findings across {K} files.

Please review and let me know if you have follow-up questions.

---

💬 Follow-up: describe the change in chat to append a timestamped Follow-up section to this artifact. Re-run `/skill:research` for a fresh artifact.

**Next step (choose one):**
- `/skill:design thoughts/shared/research/{filename}.md` — iterative design with vertical-slice decomposition (produces design artifact for plan)
- `/skill:blueprint thoughts/shared/research/{filename}.md` — lightweight fast path for smaller tasks; combined design + phased plan in one pass (produces implement-ready plan directly)

> 🆕 Tip: start a fresh session with `/new` first — chained skills work best with a clean context window.

Step 6: Handle Follow-ups

Append, never rewrite. Edit the artifact to add a ## Follow-up Research {ISO 8601 timestamp} section. Prior content stays immutable.
Bump frontmatter. Update last_updated + last_updated_by; set last_updated_note: "Added follow-up research for <brief description>".
Re-dispatch narrowly. Spawn ≤1–2 fresh analysis agents scoped to the new question. Do NOT re-run the full skill.
When to re-invoke instead. If scope changed materially (different feature surface, different research target), re-run /skill:research for a fresh artifact. The previous block's Next step: stays valid for the existing artifact.

Important Notes

Analysis only: This skill answers questions. Question formulation is delegated to the scope-tracer subagent at Step 1.
Single entry point: Free-text research prompt. Argument substitution is handled by rpiv-args; scope-tracer runs in-band before analysis dispatch.
Chained from discover: when $ARGUMENTS mentions a thoughts/shared/discover/*.md artifact, read it FULLY in Step 1 and translate each Decision into a Q (discover: <title>) / A: <Chosen> Developer Context entry. Pass the FRD's Recommended Approach text as the scope-tracer topic. Open Questions carry forward verbatim. The argument-hint stays free-text-only — discover artifact recognition is by path-mention, not by argument-hint widening.
Grouped dispatch: Related questions are batched per agent based on file overlap. Default agent: codebase-analyzer. This reduces token waste from redundant file reads and lets agents build cross-question context.
Downstream compatible: Research documents feed directly into design and plan — the same Code References / Integration Points / Architecture Insights sections they expect.
Agent-message parsing: scope-tracer emits Discovery Summary + numbered Questions inline in its final assistant message; parse the agent's final-message text (no file write).
Critical ordering: Follow the numbered steps exactly
- ALWAYS dispatch scope-tracer first (Step 1)
- ALWAYS analyze question overlap for grouping (Step 1)
- ALWAYS wait for all agents to complete (Step 2)
- ALWAYS run developer checkpoint before writing (Step 3)
- ALWAYS gather metadata before writing (Step 4)
- NEVER write the document with placeholder values
Frontmatter consistency: Always include frontmatter, use snake_case fields

18 KiB Raw Blame History Unescape Escape