knowledge-capture

Shared logic for all /dex:* commands. Follow these procedures exactly — do what the calling command asks, nothing more.

Project Discovery

Discover the project's knowledge infrastructure fresh on every invocation. Scan the filesystem directly — no config files, no cached state.

Discovery Steps

1. Find the project root: Run git rev-parse --show-toplevel
2. Find the instructions file: Check in order, use the first found:
   • <root>/CLAUDE.md
   • <root>/.claude/CLAUDE.md
   • <root>/AGENTS.md — a direct AGENTS.md project (e.g. a Codex-only repo with no CLAUDE.md). Use it as the instructions file with ai_dir = .ai, then skip the CLAUDE.md indirection in step 3a.
3. Resolve instructions file and ai_dir: a. If a CLAUDE.md was found, check whether it redirects to AGENTS.md:
   • Symlink: Run readlink <claude_md_path>. If the target filename is AGENTS.md, use the symlink target.
   • Include directive: Read the file content. If the entire content is a single @AGENTS.md line (with optional whitespace/newlines), find the referenced AGENTS.md relative to CLAUDE.md's directory.
   • If no indirection is detected, continue using CLAUDE.md. b. Derive ai_dir from the resolved instructions file:
   • CLAUDE.md → ai_dir = .claude
   • AGENTS.md → ai_dir = .ai
4. Find knowledge directory: Check for <root>/<ai_dir>/docs/
   • Migration check: If ai_dir is .ai but <root>/.ai/docs/ does not exist and <root>/.claude/docs/ does exist, flag a migration mismatch. Use .claude/docs/ as the active knowledge directory for now, but record the mismatch for init and status to surface.
5. List existing subdirectories: Check for learnings/, patterns/, decisions/, research/ within the knowledge directory
6. Count lines: Run wc -l on the resolved instructions file

If a CLAUDE.md or .claude/docs/ does not exist, proceed with what you have. Missing infrastructure is a normal state — handle it per the calling command's instructions.

Discovery Output

Build this mental model before proceeding:

project_root:       /path/to/project
instructions_file:  /path/to/project/CLAUDE.md  (387 lines)
ai_dir:             .claude
knowledge_dir:      /path/to/project/.claude/docs/
  learnings:        exists (7 files)
  patterns:         exists (3 files)
  decisions:        exists (2 files)
  research:         exists (1 file)

If AGENTS.md indirection was resolved:

project_root:       /path/to/project
instructions_file:  /path/to/project/AGENTS.md  (387 lines, resolved from CLAUDE.md)
ai_dir:             .ai
knowledge_dir:      /path/to/project/.ai/docs/
  learnings:        exists (7 files)
  patterns:         exists (3 files)
  decisions:        exists (2 files)
  research:         exists (1 file)

If migration mismatch was detected (AGENTS.md + existing .claude/docs/):

project_root:       /path/to/project
instructions_file:  /path/to/project/AGENTS.md  (387 lines, resolved from CLAUDE.md)
ai_dir:             .ai
knowledge_dir:      /path/to/project/.claude/docs/  (migration available → .ai/docs/)
  learnings:        exists (7 files)
  patterns:         exists (3 files)
  decisions:        exists (2 files)
  research:         exists (1 file)

Variable Substitution

Throughout this skill and all /dex:* commands:

• "CLAUDE.md" refers to the resolved instructions file (CLAUDE.md or AGENTS.md)
• .claude/ and .claude/docs/ refer to the resolved ai_dir (<ai_dir>/ and <ai_dir>/docs/)

Substitute the resolved values in all paths and user-facing messages. For example, when ai_dir is .ai: "Promote to AGENTS.md?", "Create .ai/docs/?", scaffolding creates .ai/docs/learnings/ etc.

If .claude/docs/ doesn't exist, commands should ask the user whether to scaffold it using the current host's input mechanism before proceeding (except /dex:status, which reports the absence).

Scaffolding

When scaffolding is needed, create these directories:

• .claude/docs/learnings/
• .claude/docs/patterns/
• .claude/docs/decisions/
• .claude/docs/research/

Create empty directories only. No README files, no templates, no boilerplate.

Document Formats

All documents are agent-first: the first section contains the actionable directive — an AI agent reading only the Rule/Pattern/Decision section gets enough to act. Context and examples follow for depth.

Token Efficiency

These docs are consumed by AI agents, not read by humans. Every formatting choice must earn its tokens:

• Bare paths over markdown links — Details: path/to/file.md not [details](path/to/file.md). Agents don't click links.
• Plain metadata — Date: 2026-02-13 not **Date:** 2026-02-13. Bold markers serve human eyes only.
• Direct file:line references — src/gateway.php:45 not "see the gateway file". Include line numbers when relevant.
• No prose filler — omit transitions, summaries of what follows, and empty template sections.

Learning Format

Use for: discoveries, fixes, gotchas, debugging insights, non-obvious behaviors.

# Short directive title

Date: YYYY-MM-DD
Tags: tag1, tag2, tag3

## Rule

The actionable directive — what to do or not do. An agent reading
only this section should know enough to apply the knowledge.

## Context

Why this matters. What went wrong or what was non-obvious.
Technical explanation of the root cause.

## Examples

Code examples showing correct and incorrect approaches.
Use CORRECT / WRONG labels for clarity.

<example type="CORRECT"> # Always pass --user=1 for WP-CLI REST calls with auth

Date: 2026-02-13 Tags: wp-cli, rest-api, authentication

Rule

Always pass --user=1 when making WP-CLI REST API calls that have permission callbacks. Without it, the call runs as unauthenticated. </example>

<example type="INCORRECT"> # WP-CLI REST API Issue

Today I discovered that WP-CLI REST API calls need authentication. This was really confusing and took a while to debug. The error was a 403... </example>

Pattern Format

Use for: reusable approaches, conventions, anti-patterns, recurring solutions.

# Short pattern name

Date: YYYY-MM-DD
Tags: tag1, tag2, tag3

## Pattern

The reusable approach — what it is and how to apply it.

## When to apply

- Condition 1
- Condition 2

## Alternatives

- When [condition 1], prefer [alternative approach] instead
- When [condition 2], [simpler method] works better because [reason]

## Reference implementation

Direct file:line reference (e.g., `src/gateway.php:45-60`), or inline code example.

Decision Format

Use for: architectural choices, trade-off decisions, technology selections.

# Short decision statement

Date: YYYY-MM-DD
Tags: tag1, tag2, tag3

## Decision

What was chosen and the one-line rationale.

## Alternatives considered

- Option A: Description — why rejected
- Option B (chosen): Description — why chosen
- Option C: Description — why rejected

## Why this choice

Detailed reasoning, trade-offs, and constraints that led to this choice.

Research Format

Use for: extensive investigations, multi-hour debugging sessions, API explorations, trial-and-error findings across environments.

# Short title describing what was researched

Date: YYYY-MM-DD
Tags: tag1, tag2, tag3
Environment: key versions, OS, configs that matter
Status: current

## Summary
2-3 sentence overview of key findings for quick scanning.

## What Works
Proven approaches with evidence (commands, configs, code).

## What Doesn't Work
Failed approaches and WHY they failed.

## Key Findings
Detailed empirical observations, numbered for reference.

## Reproduction Steps
How to verify or reproduce the findings.

## Open Questions
Unresolved issues or areas needing further investigation.

Key differences from Learning:

• Environment field — versions, OS, configs for assessing relevance over time
• Status field — current / outdated / superseded for freshness tracking
• No line limit (but every section must earn its place — omit empty sections)
• No CLAUDE.md promotion (reference material, not rules)

<example type="CORRECT"> # PHP 8.3 readonly property behavior with WooCommerce hooks

Date: 2026-02-15 Tags: php-8.3, readonly, woocommerce, hooks Environment: PHP 8.3.4, WooCommerce 9.6.0, WordPress 6.7 Status: current

Summary

PHP 8.3 readonly properties cannot be re-initialized after clone. WooCommerce hook callbacks that clone objects hit fatal errors with readonly properties.

What Works

Use backed enums or private properties with getters instead of readonly.

What Doesn't Work

• clone $order with readonly properties → Fatal error
• Reflection-based workaround → works but fragile across PHP versions </example>

<example type="INCORRECT"> # PHP research

Today I spent a few hours looking into PHP 8.3. I tried a bunch of things and some worked and some didn't. Here's what I found... </example>

Filename Convention

All files follow: YYYY-MM-DD-slug.md

Slug rules:

• Lowercase
• Replace spaces with hyphens
• Remove special characters except hyphens
• Truncate to keep full path under 100 characters

CLAUDE.md Promotion

When to Suggest Promotion

digraph promotion_decision {
    "Document captured" [shape=doublecircle];
    "What type?" [shape=diamond];
    "Always offer promotion" [shape=box];
    "Meets promotion criteria?" [shape=diamond];
    "Offer promotion" [shape=box];
    "Skip silently" [shape=doublecircle];
    "No promotion" [shape=doublecircle];
    "Count CLAUDE.md lines" [shape=box];
    "Lines < 500?" [shape=diamond];
    "Promote freely" [shape=doublecircle];
    "Lines 500-550?" [shape=diamond];
    "Warn: near budget" [shape=box];
    "STOP: Over budget — extract first" [shape=octagon, style=filled, fillcolor=red, fontcolor=white];

    "Document captured" -> "What type?";
    "What type?" -> "Always offer promotion" [label="Pattern"];
    "What type?" -> "Meets promotion criteria?" [label="Learning"];
    "What type?" -> "No promotion" [label="Decision/Research"];
    "Meets promotion criteria?" -> "Offer promotion" [label="yes"];
    "Meets promotion criteria?" -> "Skip silently" [label="no"];
    "Always offer promotion" -> "Count CLAUDE.md lines";
    "Offer promotion" -> "Count CLAUDE.md lines";
    "Count CLAUDE.md lines" -> "Lines < 500?";
    "Lines < 500?" -> "Promote freely" [label="yes"];
    "Lines < 500?" -> "Lines 500-550?" [label="no"];
    "Lines 500-550?" -> "Warn: near budget" [label="yes"];
    "Lines 500-550?" -> "STOP: Over budget — extract first" [label="no"];
}

Graduation flow: /dex:grok may offer to upgrade a learning to a pattern when reusability signals are detected. Delegates to /dex:pattern for Alternatives and When to apply sections.

Learning promotion criteria — at least one:

• Contains a do/don't directive that corrects a common mistake
• Addresses a recurring issue mentioned multiple times in conversation
• Is a project-wide constraint that applies broadly, not to one file

Skip promotion silently for informational learnings, one-off debugging insights, decisions, and research.

Budget 500-550: Warn using the current host's input mechanism: "CLAUDE.md is at X/500 lines." Options: "Add anyway" / "Extract a section first"

Budget 550+: Hard block. Tell user to extract sections first. Show sections ranked by line count, offer to extract largest. Proceed only after extraction brings count below 550.

Promoted Rule Format

One-liner + bare path to the source document:

- Always pass `--user=1` for WP-CLI REST calls with auth. Details: .claude/docs/learnings/2026-02-13-wp-cli-rest-auth.md

<example type="CORRECT" label="bare path saves tokens"> - Always pass `--user=1` for WP-CLI REST calls with auth. Details: .claude/docs/learnings/2026-02-13-wp-cli-rest-auth.md </example> <example type="INCORRECT" label="redundant markdown link wastes tokens"> - Always pass `--user=1` for WP-CLI REST calls with auth. See [details](.claude/docs/learnings/2026-02-13-wp-cli-rest-auth.md). </example>

Auto-Placement

When promoting a rule to CLAUDE.md:

1. Read CLAUDE.md and identify all ## section headings
2. Match the rule's tags and topic against section headings and content to find the most relevant section
3. Append the one-liner at the end of that section (before the next ## heading)
4. If no section is a clear match, append under the last section

Extraction Flow

When extracting a section from CLAUDE.md:

1. List all ## sections with their line counts
2. Ask the user "Which section to extract?" using the current host's input mechanism, and show sections ranked by size
3. Move the section content to .claude/docs/ as a standalone doc
4. Replace the section in CLAUDE.md with a 1-2 line summary + bare path:

   ## Section Name
   Full details: .claude/docs/section-name.md
   

5. Report the new line count

Knowledge Extraction from Conversation

How to Extract

Before drafting, analyze the relevant conversation exchange:

<pre_extraction_analysis>

• What type of insight is this? (rule, approach, choice, or investigation)
• What triggered the discovery? (error, surprise, repeated friction)
• What behavioral change should result? (do X instead of Y, always check Z) </pre_extraction_analysis>

Then draft:

1. Title — short directive statement (imperative or declarative)
2. Key section — Rule for learnings, Pattern for patterns, Decision for decisions, Summary for research
3. Tags — 3-5 from the technical domain (lowercase, hyphen-separated)
4. Present using the current host's input mechanism for confirmation

Focus on what an agent needs to act differently next time, not on narrating what happened during debugging.

Extraction Quality

Before saving any document, verify it passes all four checks. If any check fails, revise before saving.

<extraction_quality_checklist>

• Self-contained: an agent reading this in isolation knows what to do
• Actionable: tells the agent what to DO, not what happened
• Specific: includes concrete examples, file paths, or code
• Concise: under 50 lines (learnings), under 80 (patterns/decisions). Research: no limit but omit empty sections </extraction_quality_checklist>

Write Failure Recovery

If writing a document file fails (permissions, disk full, invalid path), report the error and display the intended file path and full document content so the user can save it manually. Do not retry — the user may need to fix permissions or path issues first.

Agent Behavior Analysis

Shared logic for analyzing agent behavior in a conversation to find inefficiencies and capture fixes as project knowledge. Used by /dex:sharpen.

Inefficiency Categories

Scan the conversation for these categories of wasted effort:

| Category | What to look for | |----------|-----------------| | Wrong tool usage | Used Bash when Grep/Glob/Read was better; used find instead of Glob; used cat instead of Read | | Inefficient discovery | Took 5+ searches to find something findable in 1-2; didn't use IDE index tools when available | | Missed shortcuts | Didn't leverage existing project scripts, aliases, or conventions documented in CLAUDE.md | | Rediscovering known info | Debugged something already documented in CLAUDE.md or .claude/docs/ | | Incorrect assumptions | Assumed wrong file structure, API shape, or convention — then had to backtrack | | Unnecessary confirmation loops | Asked the user things that could be inferred from context or existing docs | | Over-broad scope | Read entire files when only a section was needed; searched too widely before narrowing |

Flag only inefficiencies that meet both criteria:

1. Cost significant time or tokens (multiple wasted tool calls, 3+ minutes of backtracking)
2. Would likely recur in future sessions without a fix

Skip: one-time wrong guesses, normal exploration overhead, reasonable first attempts that didn't pan out.

For each flagged inefficiency, note three things:

1. What happened — the specific moment of waste
2. What should have happened — the efficient alternative
3. Why — the missing knowledge that caused the inefficiency

Root Cause Classification

Map each identified inefficiency to its fix destination:

| Root cause | Output destination | |-----------|-------------------| | Missing rule (do/don't directive) | CLAUDE.md promotion candidate | | Missing knowledge (debugging insight, gotcha) | .claude/docs/learnings/ | | Missing approach (reusable workflow) | .claude/docs/patterns/ | | Skill gap (agent needs better instructions) | .claude/docs/learnings/ tagged skill-improvement |

Sharpen Extraction Quality

Apply the standard extraction quality checklist (self-contained, actionable, specific, concise) PLUS these three additional checks:

1. Agent-operational — focuses on how the agent should work, not domain knowledge about the codebase
2. Preventive — tells the agent what to do before hitting the inefficiency, not how to recover after
3. Non-obvious — captures project-specific knowledge an agent wouldn't discover without experience (e.g., "this project's test runner requires --user=1" qualifies; generic tool preferences do not)

<example type="CORRECT" label="agent-operational"> Rule: "In this project, always use `pnpm wp --user=1` for REST API calls in WP-CLI"

Agent-operational: tells the agent how to run commands in this project. Preventive: specifies what to do before making the call. Non-obvious: project-specific pnpm wrapper isn't discoverable without experience. </example>

<example type="INCORRECT" label="domain knowledge — use /dex:learn instead"> Rule: "The REST API returns 403 when permissions are missing"

This is domain knowledge (how WordPress works), not operational knowledge (how the agent should work). Capture via /dex:learn instead. </example>

Sharpen Audit Log

Running log at .claude/docs/.sharpen-log.md that tracks efficiency fixes across sessions. Used by /dex:sharpen to avoid duplicating previously captured knowledge.

Audit Log Format

# Sharpen Audit Log

## YYYY-MM-DD

- [category]: [1-sentence summary] → [file path]
- [category]: [1-sentence summary] → [file path]

Reading the Audit Log

Before analyzing, read .claude/docs/.sharpen-log.md if it exists. Compare previous findings against current conversation to skip inefficiencies already captured. If the log does not exist, proceed normally — it will be created after capture.

Writing to the Audit Log

After capturing fixes, append a dated section to .claude/docs/.sharpen-log.md. Create the file if it does not exist. Each entry is one line per captured fix.