plan

name	plan
description	Create graph-informed implementation plans before writing code. Use before any non-trivial task. Skip for small changes (1-2 files, <10 lines).
mcp	codebase-memory-mcp

/supergraph:plan

Scan codebase, map blast radius, create machine-readable plan.

Announce: "📐 /supergraph:plan — scanning codebase, creating plan..."

Quick Gate

< 10 lines, 1 file, no hub/bridge nodes → skip to /supergraph:tdd.

Steps

0. Read CONTEXT.md (if exists):

cat CONTEXT.md 2>/dev/null | head -60

Use domain vocabulary from CONTEXT.md in all plan task descriptions — never use raw file/class names where a domain term exists.

1. Read the codebase (MANDATORY before planning):

• Read config file → language, framework, versions
• Read 2-3 source files near target area → naming, imports, error handling
• Read 1-2 test files → test structure, assertion style

2. Ensure graph: Reuse graph context from /supergraph:scan. If scan not done → run /supergraph:scan first. Require CBM_PROJECT and healthy index_status. If stale/degraded, call index_repository with the absolute repository path before analysis.

3. Graph analysis: Call detect_changes(project=CBM_PROJECT), search_graph for target symbols, trace_path inbound/outbound/data-flow, and get_architecture for overview, clusters, boundaries, and hotspots. After get_graph_schema, run named contract recipes hubs, bridges, test-gaps, and cross-boundary. Derive risk from impact plus recipe evidence; never invent missing results. Preserve escalation: more than 20 files stops for discussion; hub/bridge changes require approval; cross-boundary findings require justification.

3b. Serena symbol analysis (optional — deepens blast radius): If /supergraph:scan was not run this session, call mcp__serena__initial_instructions() first. For key symbols in target files:

mcp__serena__find_referencing_symbols(symbol=<key_symbol>)
mcp__serena__find_implementations(symbol=<interface_or_abstract>)

Cross-reference with graph blast radius — add any missed callers to task Blast radius fields. Optional — for large blast radius (> 10 files) or hub/bridge nodes: persist caller list for future sessions:

mcp__serena__write_memory(title="<plan-slug>-blast-callers", content="[caller list from find_referencing_symbols]")

Skip gracefully if Serena unavailable — log "Serena unavailable, skipping symbol analysis".

4. Discuss approach with user (MANDATORY, use user's language): Before creating tasks, present findings from steps 1-3 to the user:

• What was found in codebase (naming, conventions, patterns)
• Graph risk (blast radius, hubs, bridges affected)
• Proposed task breakdown (just summaries, not full tasks yet)

Ask for approval in the user's language. If user disagrees with approach → revise. If user wants changes → incorporate, then re-present. Do not proceed to step 5 until user approves the approach.

5. Create plan tasks — each task 2-5 min. Use exact machine-readable format:

## Task N: [Short description]
Status: pending
Risk: low|medium|high
Dependencies: none | Task 1, Task 2

Files:
- Create: path/to/new-file.ext
- Modify: path/to/existing-file.ext
- Test: path/to/test-file.ext

Blast radius:
- path/to/affected-file.ext

Acceptance:
- [observable behavior/result]
- [test/assertion that proves completion]

TDD:
- Behavior: [single externally visible behavior]
- Test file: [exact test path]
- Test name: [behavior-focused test name]
- RED command: `$FOCUSED_TEST_CMD`
- Expected RED failure: [missing behavior, not setup/import/syntax error]
- Minimal GREEN change: [smallest implementation idea]
- Refactor candidates: [optional, only after GREEN]
- Mocking: none | [why unavoidable]

Steps:
1. RED: [write exact failing test]
   Command: `$TEST_CMD`
   Expected: FAIL
2. GREEN: [write minimal implementation]
   Command: `$TEST_CMD`
   Expected: PASS
3. REFACTOR: [safe cleanup or "none"]
4. VERIFY:
   - `$TEST_CMD`
   - `$LINT_CMD` (skip if none)

Checkpoint:
- Files: `path/to/test-file.ext path/to/source-file.ext`
- Commit: `type: short description`

Task status values: pending, in_progress, completed, stuck (managed by executor)

6. Validate:

• Every task uses ## Task N: heading exactly
• Every task has all 8 fields: Status, Risk, Dependencies, Files, Acceptance, TDD, Steps, Checkpoint
• No placeholders (TBD, TODO, "add validation", "similar to Task N")
• Test commands real (from .supergraph-env)
• Hub nodes have review steps
• Each behavior task has expected RED failure reason
• NO indentation under field lines — Status: pending starts at column 0, not spaces
• No extra blank lines between fields within a task section

7. Save plan: docs/supergraph/plans/YYYY-MM-DD-<slug>.md

8. Analysis Review Gate (if /supergraph:analyze was used): If analyze step was completed — verify plan aligns with documented analysis decisions:

## Analysis Decisions
- Approach: [chosen approach from analyze step]
- Alternatives rejected: [from analyze step with reasons]

If analysis was skipped for ambiguous task → WARN user: "No analyze step — proceeding with plan as-is."

9. Environment Context (MANDATORY at plan end):

## Environment Context
- **Language:** [X] v[Y]
- **Test command:** [from .supergraph-env]
- **Linter command:** [from .supergraph-env]
- **Formatter command:** [from .supergraph-env]
- **Build command:** [from .supergraph-env]
- **Branch:** [current]
- **Conventional commit style:** [e.g., "feat: / fix:"]

**Codebase conventions:** [naming, imports, error handling, test structure]

**Graph Context:**
- Blast radius: M files | Hub nodes: [list]
- Bridge nodes: [list] | Communities crossed: [list]

10. Auto-review: Dispatch supergraph:plan-reviewer subagent. Fix issues. Do not hand off to execute until Approved.

11. User Review Gate (MANDATORY): Present plan summary to user:

Plan: [plan path]
Tasks: N ([list summaries])
Blast radius: M files | Hub nodes affected: [list/none]
Review: Approved (by plan-reviewer)

Ask user for approval in their language: "[yes / modify / reject]" If modify → incorporate feedback, re-run auto-review. If rejected → ask for direction, return to design.

12. Report:

✅ /supergraph:plan complete
- Plan: docs/supergraph/plans/YYYY-MM-DD-<slug>.md
- Tasks: N | Blast radius: M files | Review: Approved
- User: yes | modify | rejected
- Next: /supergraph:execute plan <slug> (multi-task) or /supergraph:tdd (single-task)

Rules

• Codebase first, plan second — never plan blindly
• Environment Context mandatory — executor depends on it
• Exact file paths, commands, code — no vagueness
• Task headings stay ## Task N: for executor parsing
• No placeholders, no "TBD", no "similar to Task X"
• Never execute code — only create plans