Rise Pde Session Multi Agents V2

Hard Rules (Non-Negotiable)

These rules exist because real sessions failed when they were violated.

Read these references before proceeding:

• references/rise-framework.md — RISE methodology
• references/prompt-decomposition.md — PDE as EAST practice
• references/creative-orientation.md — creative vs reactive distinction
• references/structural-tension.md — tension charting methodology
• references/ceremonial-technology.md — when session involves ceremony/sacred work

For sub-agent prompt templates: read agent-templates.md. For orchestration patterns (incl. cascading parts): read orchestration-patterns.md.

Phase 0: Assess

1. Read the user's input carefully
2. Identify the session UUID — either from __.md, from the user's message, or generate one
3. Create workdir: mkdir -p .pde/<SESSION_UUID>/
4. Determine if this is a single mission or multi-part cascading mission
   • Multi-part: user says "Part 1... Part 2..." or numbered sections
   • See orchestration-patterns.md → Cascading Parts Pattern

Phase 1: Decompose (EAST 🌅)

1. Call pde_decompose(prompt) with the full user input
2. Process the decomposition — generate the JSON yourself
3. Call pde_parse_response(llm_response, original_prompt, workdir) — stores to .pde/
4. Note the PDE UUID — you'll need it throughout

For multi-part missions: decompose the ENTIRE input once, then identify part boundaries in the action stack.

Phase 2: Create STC Chart

1. Call import_pde_decomposition(pde_id, workdir) — creates .coaia/pde/<session>.jsonl
2. Present the STC to the human:
   • Desired Outcome: what the session creates
   • Current Reality: what exists now (including ambiguities)
   • Action Steps: the ordered stack with direction tags
3. Note the STC chart ID for progress tracking

Phase 3: Commit & SLEEP

1. git add only the .pde/<UUID>.md and .pde/<UUID>.json and .coaia/pde/<session>.jsonl
2. Commit: "[pde] Decompose: <summary> — STC chart_<id>"
3. Tell the human what was decomposed and what the STC shows

THEN YOU MUST RUN THIS EXACT COMMAND USING THE BASH TOOL:

sleep 1500

THIS IS NOT OPTIONAL. THIS IS NOT A SUGGESTION. YOU MUST RUN sleep 1500 IN BASH RIGHT NOW.

Do NOT:

• ❌ Skip the sleep
• ❌ Say "I'll wait for you" instead of running the command
• ❌ Say "sleep 1500" in text without actually calling the Bash tool
• ❌ Proceed to Phase 4 without the sleep completing
• ❌ Spawn any agents before the sleep finishes

Do:

• ✅ Call the Bash tool with command sleep 1500 and timeout 600000
• ✅ Wait for it to complete (it will take 1500 seconds = 25 minutes)
• ✅ Only AFTER the bash tool returns, proceed to Phase 4

The human + pde-review-companion use these 25 minutes to review and edit .pde/*.md files. If you skip this, you waste the human's tokens and break the review window.

Phase 4: Wake & Read Delta

After waking:

1. git diff .pde/ — see what the human changed
2. git diff issues/ — check plan edits
3. Read ALL .pde/*.md files (they are now the authoritative intent)
4. Read the STC chart file
5. Build the unified action stack from all PDE facets
6. Identify how many independent work streams exist
7. For multi-part missions: identify which part to execute first

Phase 5: SOUTH — Deep-Search Per Facet

Reminder: ALL agents you spawn MUST use model: "claude-opus-4.6" in the Agent tool call.

For EACH PDE facet, spawn a sub-agent (model: "claude-opus-4.6") that does deep-search. Results land in .pde/<SESSION_UUID>/deep-search/.

Use templates from agent-templates.md → Agent S1, S2, S3, S4.

Scale to facet complexity:

• Simple facet (single file change): 1 deep-search agent (model: "claude-opus-4.6")
• Multi-repo facet (rispecs + kinship): 1 deep-search + 1 rispecs-survey agent (both model: "claude-opus-4.6")
• Complex facet (implementation + ceremony): 1 deep-search + 1 academic/methodology agent (both model: "claude-opus-4.6")

All SOUTH agents write to .pde/<SESSION_UUID>/deep-search/ only — never source files.

Phase 6: Synthesize Deep-Search Results

After ALL deep-search agents return (batch — no anchoring):

1. Read all .pde/<SESSION_UUID>/deep-search/ findings
2. Read .pde/rispecs-survey.md if it exists
3. Update the plan with enriched understanding
4. Assign action stack items to executor agents (model: "claude-opus-4.6") based on:
   • File ownership (no two agents touch the same file)
   • Dependency ordering (respect the action stack)
   • Natural groupings (rispecs together, implementation together, diary separate)

Phase 7: NORTH — Parallel Execution

Reminder: ALL agents you spawn MUST use model: "claude-opus-4.6" in the Agent tool call.

Spawn N executor agents (each with model: "claude-opus-4.6"), each handling a work stream.

Use templates from agent-templates.md → Agent N1, N2, N3, N4, N5.

WEST — Validation Agent (runs parallel with executors)

Spawn 1 validation agent (model: "claude-opus-4.6") alongside NORTH executors. Use template from agent-templates.md → Agent W1.

Phase 8: Integration & Commit

1. Collect all executor outputs
2. Run validation checklist from WEST agent
3. Cross-reference: did rispecs reference each other properly?
4. Verify kinship mappings follow protocol
5. Run build/lint if applicable
6. git add only files actually created or modified
7. Commit: "[pde] Multi-agent execution: <summary>"

For Multi-Part Cascading Missions

After committing Part N:

1. Present Part N results to human
2. If more parts remain → return to Phase 5 for Part N+1
3. Each part gets its own commit: "[pde] Part N: <summary>"
4. See orchestration-patterns.md → Cascading Parts Pattern

Phase 9: Human Checkpoint

Present:

• What each executor created
• WEST validation results
• Any unresolved ambiguities
• git diff --stat
• STC progress update (which action steps completed)

Human adjusts or approves.

Phase 10: Session Close

After human approval:

1. Present final git diff --stat
2. Write a brief session summary to .pde/<SESSION_UUID>/session-summary.md
3. The session gist export is handled OUTSIDE this skill — the orchestrator launches copilot-cli with --share-gist and receives the gist_url when the session exits. That gist_url then gets routed to mino-mcp by the orchestrator, not by this session.

You do NOT need to export a gist. The --share-gist flag on copilot-cli handles that automatically.

Rispecs & RISE Framework Integration

Unchanged from v1 — see agent-templates.md for rispecs workflow.

Key files the skill reads automatically:

Rules

• bash sleep 1500 — actual command, non-negotiable, 25 minutes for human review
• All subagents: model: "claude-opus-4.6" — stated here, enforced everywhere
• Workdir: .pde/<SESSION_UUID>/ — all scratch, deep-search, summaries go here
• Never git add . or git add -A — only files you or your agents touched
• RISE creative orientation: structural tension → desired outcome → natural progression
• NO problem-solving language: no "gaps", "fixes", "bridge the gaps"
• Sub-agents write to .pde/<SESSION_UUID>/deep-search/ during SOUTH, never source files
• Only NORTH executor agents touch source files
• Batch ALL agent results before synthesizing (avoid anchoring bias)
• Scale agent count to task complexity — 2 agents for simple, 10+ for multi-repo
• Respect repo boundaries when spanning multiple repos
• MCP tool pattern: mcp__mcp-pde__<tool> for PDE, mcp__coaia-pde__<tool> for STC
• Session summary at close — gist export is handled by --share-gist at launch, not inside the session