Tenax

Tenax Skill

⚠️ MANDATORY PROTOCOL

You MUST follow these steps. This is not optional.

At Session Start

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/get-summary.ts"

Do this FIRST, before responding to any user request.

Before Any Action - MANDATORY CHECKS

⛔ DO NOT skip these steps. ALWAYS check before acting.

Step 1: Check Conversation Context (Current Session)

Before executing ANY task, check your conversation history:

• Did you just complete this exact task in the last few messages?
• If yes: "I just did X moments ago - is there an issue, or should I redo it?"

Step 2: Search Tenax (Previous Sessions)

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/search.ts" "<relevant-topic>"

Check search results for:

1. Recency - Was this task done recently?

   • "Uploaded frontend - 3hr ago" → "This was done 3 hours ago. Redo it?"

2. Prior Decisions - Does your choice conflict with history?

   • User says "use SQLite" but search shows "[D] database: PostgreSQL chosen" → Surface the conflict before proceeding

This applies to:

• Adding new features or functionality
• Changing existing code or configuration
• Choosing libraries, tools, or approaches
• Making architectural or design decisions
• Any task that involves writing or modifying code
• Repeating tasks that may have been done recently

WRONG: User says "upload frontend" → you immediately upload RIGHT: User says "upload frontend" → check context (did I just do this?) → search "frontend upload" → note recency → then act

WRONG: User says "let's use SQLite" → you implement SQLite RIGHT: User says "let's use SQLite" → search "database" → find prior PostgreSQL decision → "Session 005 decided PostgreSQL. Switch to SQLite, or keep PostgreSQL?"

If search returns no results, briefly note: "Searched for X - no prior decisions found."

After Deciding - Use ALL FOUR Markers

DO NOT default to only [D]. Each response should include the appropriate markers:

Why this matters: Without search, you may contradict prior decisions. Without ALL markers, patterns and insights are lost.

Before Sending Your Response - VERIFY

⛔ Do not send your response until you've confirmed:

• Checked conversation context for recent duplicates
• Searched Tenax for prior decisions on this topic
• Noted any conflicts or recency issues to the user
• Used appropriate markers ([D], [P], [T], [I]) for any choices made

If you skipped any step, go back and complete it now.

Quick Start Example

When user asks: "Should we use SQLite or PostgreSQL?"

Step 1: Search first

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/search.ts" "database"

Step 2: Respond with marker INLINE

[D] database: I recommend SQLite for your embedded storage needs - zero config, single file, no server required. The main advantages are portability and simplicity for your use case.

CRITICAL: The marker starts the sentence. Don't write your response then add a marker afterward. The marker labels the prose that follows it.

Session Start

At conversation start, load project context:

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/get-summary.ts"

Then acknowledge: "Loaded project memory: X sessions, Y decisions."

Markers

Use these markers to label decisions, patterns, tasks, and insights:

Multi-Line Format

For detailed decisions, use multi-line blocks (ends at blank line):

[D] frontend-redesign: Implementing new design system with JetBrains Mono + Outfit typography, bento grid layout with asymmetric cards, and a timeline-style "How It Works" section. These changes improve visual hierarchy and modernize the aesthetic.

For lists, keep the marker at the start:

[D] api-endpoints: Creating three new endpoints - /users for account management, /projects for CRUD operations, and /webhooks for event subscriptions.

Search Flow Details

When search returns results:

1. Results found → Reference them: "Based on session 003, we decided to use X..."
2. Conflict found → Surface it: "This differs from session 005 which chose Y. Should I proceed or align with the prior decision?"
3. Recent task found → Note it: "This was done 2hr ago in session 017. Redo it?"
4. No results → Search raw transcripts:

   grep -r "<keyword>" "${PROJECT_DIR}/.claude/tenax/sessions/"
   

5. Still nothing → Note "Searched for X - no prior decisions" → Proceed with [D] marker

Recording Knowledge

Single items:

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/record-decision.ts" -t "<topic>" -r "<rationale>" "<decision>"
npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/scripts/record-pattern.ts" -n "<name>" -u "<usage>" "<description>"

Batch recording (single permission):

npx tsx "${CLAUDE_PLUGIN_ROOT}/skills/tenax/cli.ts" batch --json '<json>'

Batch JSON format:

{
  "decisions": [{"topic": "api", "decision": "Use REST", "rationale": "Simpler"}],
  "patterns": [{"name": "error-handler", "description": "Try-catch wrapper", "usage": "API routes"}],
  "tasks": [{"title": "Add tests", "priority": "high"}],
  "insights": [{"content": "Insight text"}]
}

When to Use Markers

CRITICAL: Mark ALL Decisions

Every response that involves judgment should include a marker. This includes:

The Rule

If you made a choice, verified something, or concluded anything - mark it inline.

WRONG (marker as afterthought):

The gitignore looks correct.
[D] gitignore: Verified

RIGHT (marker labels the explanation):

[D] gitignore: Verified correct - properly excludes .claude/, backups, and test/ directories.
[D] commit-scope: This release (v2.1.0) includes install scripts and removes deprecated commands.
[P] retry-logic: Using exponential backoff with max 3 attempts for all API calls.
[I] The rate limiter resets per-user, not globally - discovered while debugging timeouts.

DON'T Mark

• Temporary workarounds you'll remove
• Debugging console.logs
• Exploratory code that won't persist
• Explicit "just for now, will change later"

Common Mistake

❌ Treating operational tasks as "just execution" - commits, verifications, and process choices ARE decisions that should be marked.

User Commands

Users can invoke these commands:

Session Capture

Sessions are captured automatically at:

1. PreCompact - Before auto-compaction (may happen mid-session)
2. SessionEnd - When session ends normally

Because PreCompact can trigger anytime, use markers throughout the session - not just at the end.

Referencing Past Sessions

When session data is loaded, cite it:

Based on the decision from session 003, we're using SQLite.
Following the error-handling pattern from session 002...
Session 004 noted important subprocess handling considerations.

This creates continuity across sessions.