Ctx

Read the linked workflow file and follow its instructions.

Report Discovery

When any subcommand other than brief is invoked, locate the active report:

1. Search: ls ~/.agent-memory/*/ctx-report.md
2. If exactly one report found → use it. The appendix file (ctx-appendix.md) is always in the same directory
3. If multiple found → list each report's title line (# Ctx: ...) and whether it has unfilled placeholders (in-progress). Prioritize in-progress reports. Ask the user to choose
4. If none found → tell the user to run /ctx brief first

When passing file paths to subagents, always pass both the main report path and the appendix path.

Prerequisite Check

On subcommand start, verify that the prerequisite placeholders are all replaced (no {{ remaining for those placeholders). If unmet, inform the user which prior subcommand to run.

Progress Check (default)

When invoked with no arguments:

1. Run Report Discovery
2. If no report found → display: "分析がまだ開始されていません。/ctx brief で開始してください"
3. If report found → run grep '{{' <report-path> <appendix-path> to find unfilled placeholders across both files
4. Map unfilled placeholders to subcommands and display progress:

## 進捗: <レポートタイトル>

✅ brief — リサーチブリーフ
✅ facts — 事実抽出・整理
⬜ phases — フェーズ定義
⬜ context — 共通パターン抽出
⬜ synthesis — 統合

→ 次のステップ: `/ctx phases`

Next-step suggestion logic:

• brief incomplete → brief
• facts incomplete → facts
• phases incomplete → phases
• context incomplete → context
• synthesis incomplete → synthesis

Placeholder → subcommand mapping:

Data Flow

The brief subcommand creates two files from the report template (report-template.md):

• ctx-report.md (main report): Analysis results — phase definitions, common patterns, and synthesis
• ctx-appendix.md (appendix): Raw data — fact tables, case stories, per-case phase narratives

Each subsequent subcommand replaces its placeholders ({{XXX}}) in the appropriate file after user confirmation. When all subcommands complete, the two files together form the finished deliverable.

• Read from appendix, write to main report: context reads fact data from the appendix, then writes common patterns to the main report and per-case narratives to the appendix. synthesis reads from the main report and writes to the main report.
• Replace after confirm: Each subcommand's output replaces its placeholder(s) after user approval at the confirmation gate. Content is wrapped in HTML comment boundary markers (<!-- BEGIN XXX -->...<!-- END XXX -->) for re-replacement if the user requests revisions.
• Atomic replacement: When a subcommand has multiple placeholders, all are replaced together after the confirmation gate. No partial replacement state.
• Re-replacement: If the user requests a revision after confirmation, locate the content between <!-- BEGIN XXX --> and <!-- END XXX --> markers and replace it with the revised content.
• Temporary files: context uses temporary files (_narrator_tmp.md, _analyst_tmp.md) for intermediate subagent outputs. These are deleted after the confirmation gate.

Section Dependency Map

Cross-section dependencies. Consult this when manually editing or re-running subcommands to identify downstream impact.

Re-run for Case Addition

To add cases to an existing analysis, re-run subcommands in this order:

1. facts: Append new case's fact table and Background/Events
2. phases: Re-evaluate phase definitions with the new case's data (user confirms adjustments)
3. context: Run Narrator for new case only (append to temporary file) → Re-run Analyst-Critic and Integration with all cases (re-replacement)
4. synthesis: Re-run all three steps (re-replacement)

Post-hoc Correction

When a missed pattern or error is discovered after a confirmation gate:

Small correction (no pattern ID additions/deletions/renames):

• Edit the affected section directly (re-replace)
• Consult the dependency map above and verify consistency in only the downstream sections that reference the changed content

Large correction (pattern IDs added/deleted/renamed):

• Re-run the affected subcommand (re-replacement) rather than manual patching
• Manual patching across multiple sections with ID changes is error-prone

Re-replacement Consistency Rule

• ALWAYS: After any re-replacement or manual edit, consult the dependency map and verify consistency in only the downstream sections that reference the modified content
• Verification items: (a) any P*-S*/P*-St* added or changed in the edit exists and is correctly referenced downstream, (b) quantitative phrases like "N cases" match the actual case count
• If inconsistencies are found, re-replace the downstream sections as well

Report paths: ~/.agent-memory/<scope>/<date>_<topic>/ctx-report.md and ctx-appendix.md

Output Language Rules

• All user-facing output (table headers, section titles, labels) must be written in Japanese
• English terminology may be used as-is in subagent prompts for analysis accuracy
• Established loanwords whose original meaning would be lost in Japanese translation (e.g., エビデンス) should remain in English

Inference Permission Hierarchy

Each subcommand has explicit bounds on what kind of inference is permitted:

Strict Rules

• Do not interpret "why" people acted as they did — describe what happened and how patterns connect, not the underlying motives or dynamics
• Do not propose or suggest solutions — this skill is observational analysis only
• Do not skip from facts to pattern extraction without data
• Do not proceed past a confirmation gate (→) without user approval
• Do not treat customer opinions or stated preferences as behavioral facts
• Do not score, rank, or prioritize
• NEVER ask the user to verify completeness — always perform self-review against the source material before presenting results at any confirmation gate
• RQ is the starting point for analysis, not a filter. Fact extraction is RQ-independent — every observable behavior and verbatim quote must be recorded regardless of apparent relevance to the RQ
• Output written to the report or appendix (subagent output and integration results) must NOT contain file paths, plan references, or any external file references — the report must be a self-contained deliverable

Completion

This skill is complete when all conditions are met:

• All subcommands' placeholders are filled in both the main report and appendix files
• The user has approved each confirmation gate
• The user confirms analysis is complete, or directs additional investigation