azcomputerguru/claudetools
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

refactor: optimize CLAUDE.md context footprint (-49%)

Browse Source 
Extract Ollama docs and PROJECT_STATE locking protocol to on-demand
reference files. Trim Work Mode to detection table only. Remove verbose
anti-pattern examples and credential encryption details.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Mike Swanson
2026-04-20 12:09:17 -07:00
parent 30af2d31c7
commit a50af7faf1
4 changed files with 395 additions and 493 deletions
Download Patch File Download Diff File Expand all files Collapse all files

397
.claude/CLAUDE.md
View File

@@ -18,22 +18,15 @@ This repo is shared across multiple team members. **At every session start, BEFO
"machine": "<HOSTNAME>" "machine": "<HOSTNAME>"
} }
``` ```
- Also set local git config for this repo: - Set local git config: `git config user.name "<full_name>"` and `git config user.email "<email>"`
```bash - Set git remote (read `gitea_username` from users.json): `git remote set-url origin https://<gitea_username>@git.azcomputerguru.com/azcomputerguru/claudetools.git`
git config user.name "<full_name>" - Add hostname to user's `known_machines` in users.json and commit.
git config user.email "<email>" - **Show the user `.claude/ONBOARDING.md`** — present section by section, explain the WHY, answer questions.
``` 3. **If hostname doesn't match any known machine** for the identified user, update their `known_machines` in users.json.
- Set git remote to use the user's own Gitea account (read `gitea_username` from users.json):
```bash
git remote set-url origin https://<gitea_username>@git.azcomputerguru.com/azcomputerguru/claudetools.git
```
- Add the machine hostname to the user's `known_machines` list in `.claude/users.json` and commit.
- **IMPORTANT: Show the user `.claude/ONBOARDING.md`** — present it section by section, explain what each part does and WHY, answer any questions. This is their orientation to the system.
3. **If the hostname doesn't match any known machine** for the identified user, update their `known_machines` in users.json.
### Session log attribution ### Session Log Attribution
Every session log MUST include a `## User` section at the top: Every session log MUST include a `## User` section:
```markdown ```markdown
## User ## User
- **User:** Mike Swanson (mike) - **User:** Mike Swanson (mike)
@@ -41,41 +34,30 @@ Every session log MUST include a `## User` section at the top:
- **Role:** admin - **Role:** admin
``` ```
### Git commit attribution Commits use local git config (user.name / user.email). Gitea push account is shared (azcomputerguru) but commit authorship tracks the actual person.
Commits use the local git config (user.name / user.email), which is set per-user during identity setup. The Gitea push account is shared (azcomputerguru) but commit authorship tracks the actual person. ### Current Team
### Current team | User | Role | Notes |
|---|---|---|
| **Mike Swanson** (mike) | admin | Owner, President of Arizona Computer Guru LLC |
| **Howard Enos** (howard) | tech | Employee, technician. Full trust — same access as admin. |
| User | Role | Access | Notes | ---
|---|---|---|---|
| **Mike Swanson** (mike) | admin | Full | Owner, President of Arizona Computer Guru LLC |
| **Howard Enos** (howard) | tech | Full | Employee, technician. Full trust — same access as admin for all MSP tracking and daily work. |
Both users have identical access. No permission gating between them. If a new team member is added later, their role and access scope should be defined in `.claude/users.json` before they sync. ## Work Mode
## Work Mode (auto-detect + color) Auto-detect on every user message (first match wins):
Claude operates in one of five modes. Mode determines terminal color and operational posture. **Auto-detect on every user message** using these priority rules (first match wins): | Mode | Triggers | Posture |
|------|----------|---------|
| **remediation** | "remediation tool", "365", "breach", "tenant sweep", M365 keywords | Graph API focus, compliance language, full audit trail |
| **client** | client name, `clients/` work, "for \<client\>" | Careful with data, session logs in `clients/`, name the client |
| **infra** | server names/IPs, SSH, firewall, DNS, deploy, service restart | Confirm before destructive ops, backup-first |
| **dev** | code, build, Rust/cargo, npm, GuruRMM dev, `projects/` work | Delegate freely, less confirmation friction |
| **general** | default | Lightweight |
1. **remediation** (purple) — "remediation tool", "365", "breach", "tenant sweep", M365 keywords On mode change: announce `[MODE -> infra]`, tell user to run `/color <color>`. Full details: `.claude/commands/mode.md`
2. **client** (orange) — client name mentioned (check `clients/` dirs), work under `clients/`, "for \<client\>"
3. **infra** (red) — server names/IPs (AD2, Jupiter, 172.16.x.x), SSH, firewall, DNS, deploy, service restart
4. **dev** (cyan) — code, build, compile, Rust/cargo, npm, GuruRMM dev, testing, work under `projects/`
5. **general** (blue) — default
**On mode change:** update `identity.json` "mode" field, change terminal color, announce briefly: `[MODE -> infra]`. Don't interrupt workflow.
**Manual override:** user can run `/mode <name>` to force a mode. `/mode auto` re-runs detection. `/mode` shows current.
**Posture by mode:**
- **client (orange):** careful with data, session logs go to `clients/`, always name the client
- **dev (cyan):** delegate freely to Coding/Testing agents, less confirmation friction
- **infra (red):** confirm before destructive ops, backup-first, double-check IPs
- **general (blue):** lightweight, default
- **remediation (purple):** Graph API focus, compliance language, full audit trail
Full details: `.claude/commands/mode.md`
--- ---
@@ -100,198 +82,64 @@ You are NOT an executor. You coordinate specialized agents and preserve your con
**Do yourself:** Simple responses, reading 1-2 files, presenting results, planning, decisions. **Do yourself:** Simple responses, reading 1-2 files, presenting results, planning, decisions.
**Rule:** >500 tokens of work = delegate. Code or database = ALWAYS delegate. **Rule:** >500 tokens of work = delegate. Code or database = ALWAYS delegate.
**DO NOT** query databases directly (no SSH/mysql/curl to API). **DO NOT** write production code. **DO NOT** run tests. **DO NOT** commit/push. Use the appropriate agent. **DO NOT** query databases directly. **DO NOT** write production code. **DO NOT** run tests. **DO NOT** commit/push.
### Model Routing (Complexity-Based) ### Model Routing (Complexity-Based)
Before spawning an agent, pick a tier:
| Tier | Model | When | | Tier | Model | When |
|------|-------|------| |------|-------|------|
| 0 | **Ollama** (local) | Low-stakes: summarize, classify, extract, draft, format — no code changes, output reviewed before use | | 0 | **Ollama** (local) | Low-stakes: summarize, classify, extract, draft — no code changes, output reviewed before use |
| 1 | `haiku` | Ollama unavailable, or task needs tool use / file access an agent provides | | 1 | `haiku` | Ollama unavailable, or task needs agent tool use / file access |
| 2 | (inherit) | Standard code, DB, tests, git — most work | | 2 | (inherit) | Standard code, DB, tests, git — most work |
| 3 | `opus` | Architecture, security, ambiguous failures, production risk | | 3 | `opus` | Architecture, security, ambiguous failures, production risk |
**Tier 0 rule:** Always try Ollama first for low-stakes work. It's free, fast, and private. Use `qwen3:14b` for general tasks; `codestral:22b` for code suggestions. Fall back to Haiku only if Ollama is unreachable or the task requires agent tool use. **Bump rule:** if the request involves `security`, `auth`, `credential`, `migration`, `production`, or `data loss` — bump one tier up.
**Bump rule:** if the request involves `security`, `auth`, `credential`, `migration`, `production`, or `data loss` — bump one tier up (Ollama → Haiku, Haiku → inherit, inherit → opus). Pass `model: "haiku"` or `model: "opus"` explicitly. Omit for Tier 2. Tier 0 is a direct Bash call — see `.claude/OLLAMA.md`.
Pass `model: "haiku"` or `model: "opus"` explicitly. Omit for Tier 2 (inherits session model). Tier 0 is a direct Bash call, not an agent spawn — see Ollama section below.
### Coordination Flow
```
User request -> Main Claude (coordinator) -> Launches agent(s) -> Agent returns summary -> Main Claude presents to user
```
- Independent operations run in parallel
- Skills (Skill tool) enhance/validate. Agents (Agent tool) execute/operate.
--- ---
## Automatic Context Loading (CRITICAL) ## Automatic Context Loading (CRITICAL)
**BEFORE responding to user's first message or when switching projects, AUTOMATICALLY load context:** **BEFORE responding to the first message or when switching projects, AUTOMATICALLY load context:**
### Trigger 1: Project Keywords Detected ### Trigger 1: Project Keywords Detected
If user mentions **GuruRMM**, **Dataforth**, **tunnel**, **VASLOG**, **AD2**, **testdatadb**, etc: If user mentions **GuruRMM**, **Dataforth**, **tunnel**, **VASLOG**, **AD2**, **testdatadb**, etc:
1. **Immediately read** the matching project CONTEXT.md: 1. Read the matching project CONTEXT.md:
- GuruRMM keywords → `projects/msp-tools/guru-rmm/CONTEXT.md` - GuruRMM keywords → `projects/msp-tools/guru-rmm/CONTEXT.md`
- Dataforth keywords → `projects/dataforth-dos/CONTEXT.md` - Dataforth keywords → `projects/dataforth-dos/CONTEXT.md`
- General → `CONTEXT.md` (root) - General → `CONTEXT.md` (root)
2. **If a `PROJECT_STATE.md` exists alongside the CONTEXT.md, read it immediately after.** This file tracks live inter-session state: active locks (what other Claude instances are touching), architecture decisions that are locked, recent changes since the last session, and pending work. Check it for blocking conflicts before starting any task. 2. If `PROJECT_STATE.md` exists alongside CONTEXT.md, read it **and** `.claude/PROJECT_STATE_PROTOCOL.md`.
3. Read ENTIRE file (infrastructure, current state, anti-patterns) 3. THEN respond with full context.
3. Note recent session logs mentioned in CONTEXT.md
4. THEN respond with full context
### Trigger 2: Continuation/Resume Words ### Trigger 2: Continuation/Resume Words
If user says "continue", "let's work on", "back to", "resume", "finish": If user says "continue", "let's work on", "back to", "resume", "finish":
1. Detect project from message 1. Detect project from message, read project CONTEXT.md.
2. Read project CONTEXT.md if found 2. If PROJECT_STATE.md exists, read it and `.claude/PROJECT_STATE_PROTOCOL.md`.
3. Check "Current State" and "Recent Session Logs" sections 3. Check "Current State" and "Recent Session Logs" sections, then proceed.
4. Proceed without asking for context
### Trigger 3: Infrastructure/Deployment Questions ### Trigger 3: Infrastructure/Deployment Questions
If user asks about **servers**, **databases**, **credentials**, **deploy**, **IP**, **password**: If user asks about **servers**, **databases**, **credentials**, **deploy**, **IP**, **password**:
1. Check current directory for CONTEXT.md 1. Check current directory for CONTEXT.md, then `projects/*/CONTEXT.md`.
2. If not found, check projects/*/CONTEXT.md 2. Answer from CONTEXT.md — never ask for info that's already there.
3. Answer from CONTEXT.md (never ask user for info that's in CONTEXT.md)
### Trigger 4: Uncertainty >5% ### Trigger 4: Uncertainty >5%
If you're <95% certain about infrastructure, recent work, or next steps: If you're <95% certain about infrastructure, recent work, or next steps: read CONTEXT.md before asking the user.
1. Search for CONTEXT.md in working directory
2. Search for CONTEXT.md in projects/*/
3. Read before asking user
### ANTI-PATTERN Examples (NEVER DO THIS): ### Anti-Pattern
❌ **Wrong:** Never ask "What did we do last time?" or "What's the server IP?" — read the CONTEXT.md first. If it's not there, then ask.
```
User: "Look at the Dataforth DFWDS folders"
You: "I don't recall what we've done with Dataforth. Let me search session logs..."
```
✅ **Correct:**
```
User: "Look at the Dataforth DFWDS folders"
You: [Detects "Dataforth" → reads projects/dataforth-dos/CONTEXT.md in <3 seconds]
"I see from CONTEXT.md that DFWDS is at C:\Shares\testdatadb\ on AD2 (192.168.0.6).
Recent work (2026-04-12) extended SCMVAS/SCMHVAS pipeline. Service is testdatadb on port 3000.
What would you like me to check?"
```
❌ **Wrong:**
```
User: "Continue working on GuruRMM tunnel"
You: "What phase are we on? Which server is this deployed to?"
```
✅ **Correct:**
```
User: "Continue working on GuruRMM tunnel"
You: [Reads projects/msp-tools/guru-rmm/CONTEXT.md]
"Tunnel Phase 1 is complete (v0.6.0, deployed to 172.16.3.30:3001).
Phase 2 is channel implementation (Terminal, File, Registry, Service).
2/6 agents online. Ready to proceed."
```
### Session Start Protocol
At session start:
1. Check for CONTEXT.md in current working directory
2. If found, read it silently (don't announce to user)
3. Be ready to answer questions about any project listed
4. When user specifies project, load that project's CONTEXT.md automatically
### Benefits
- ✅ Never ask "What's the server IP?" (it's in CONTEXT.md)
- ✅ Never ask "What did we do last time?" (recent logs in CONTEXT.md)
- ✅ Never ask "Where's the database?" (infrastructure table in CONTEXT.md)
- ✅ Start work immediately with full context
- ✅ Follow anti-patterns automatically (CONTEXT.md lists common mistakes)
**See:** `.claude/AUTO_CONTEXT_SYSTEM.md` for full implementation details
---
## PROJECT_STATE.md Action Protocol (MANDATORY for any project that has this file)
This protocol prevents conflicts between concurrent Claude sessions working on the same project. It is **not optional** — follow it for every significant action.
### What counts as an action
Any of the following requires the full read → lock → update → release cycle:
- Editing or creating source code files
- Git commit or push
- SSH command that modifies a server (deploy, install, config change, service restart)
- Database schema change or data migration
- Build pipeline modification
- Any operation listed in the project's Component Ownership Map
Reading files, planning, and answering questions do NOT require a lock.
### The protocol
**Step 1 — Read before acting**
Before starting any action, re-read PROJECT_STATE.md:
- Check Active Session Locks: is anything locked that you need to touch?
- If there is a conflicting lock that is NOT stale (< 2 hours old): stop, report the conflict to the user, ask how to proceed.
- If locks are stale (> 2 hours, last updated timestamp): note it to the user, clear the stale row, proceed.
**Step 2 — Claim your lock**
Immediately before performing the action, add a row to the Active Session Locks table:
| Session | Working On | Status | Blocks | Started |
|---------|-----------|--------|--------|---------|
| DESKTOP-0O8A1RL/Claude | Brief description | IN_PROGRESS | What others must avoid | HH:MM UTC |
Use `{machine}/{Claude or agent description}` as Session identifier.
Use the machine name from identity.json.
**Step 3 — Perform the action**
**Step 4 — Update on completion OR failure**
Immediately after the action finishes (regardless of outcome):
1. Remove your lock row from Active Session Locks
2. Add an entry to Recent Changes:
- Status: `COMPLETE`, `FAILED`, `PARTIAL`, or `ROLLED_BACK`
- If FAILED or PARTIAL: describe what state things were left in
3. Update the Current Project State table if any component status changed
4. Check off completed items in Pending / Next Up if applicable
### Anti-patterns (never do these)
❌ Starting work without reading PROJECT_STATE.md first
❌ Forgetting to claim a lock before touching a component
❌ Leaving a lock row in place after finishing (even on failure)
❌ Skipping the Recent Changes entry when work completes
❌ Letting a lock go stale — update the timestamp if a task takes longer than expected
### Stale lock rule
A lock older than 2 hours with no timestamp update is considered abandoned. Any session may clear it and claim the component, but must note in Recent Changes: `[Cleared stale lock from {session}]` before proceeding.
### Example — correct flow
```
[Reading PROJECT_STATE.md — no active locks on server/src/]
[Updating PROJECT_STATE.md — adding lock: "DESKTOP-0O8A1RL/Claude | POST /api/enroll endpoint | IN_PROGRESS | server/src/, migrations/ | 14:32 UTC"]
[... writing code, running migration ...]
[Updating PROJECT_STATE.md — removing lock, adding Recent Changes: "POST /api/enroll implemented | COMPLETE", updating component status: "Enrollment endpoint | IMPLEMENTED"]
```
--- ---
## Projects ## Projects
**ClaudeTools** -- MSP Work Tracking System (Production-Ready) **ClaudeTools** — MSP Work Tracking System (Production-Ready)
- Database: MariaDB 10.6.22 @ 172.16.3.30:3306 | API: http://172.16.3.30:8001 - Database: MariaDB 10.6.22 @ 172.16.3.30:3306 | API: http://172.16.3.30:8001
- 95+ endpoints, 38 tables, JWT auth, AES-256-GCM encryption - 95+ endpoints, 38 tables, JWT auth, AES-256-GCM encryption
- DB creds in vault: `bash D:/vault/scripts/vault.sh get-field projects/claudetools/database.sops.yaml credentials.password` - DB creds: `bash D:/vault/scripts/vault.sh get-field projects/claudetools/database.sops.yaml credentials.password`
**GuruRMM** -- Remote Monitoring & Management (Active Development) **GuruRMM** — Remote Monitoring & Management (Active Development)
- Server: Rust/Axum @ 172.16.3.30:3001 | Dashboard: https://rmm.azcomputerguru.com - Server: Rust/Axum @ 172.16.3.30:3001 | Dashboard: https://rmm.azcomputerguru.com
- Repo: `azcomputerguru/gururmm` on Gitea (active), `guru-rmm` is a stale copy - Repo: `azcomputerguru/gururmm` on Gitea (active), `guru-rmm` is a stale copy
- Roadmap: `projects/msp-tools/guru-rmm/ROADMAP.md` - Roadmap: `projects/msp-tools/guru-rmm/ROADMAP.md`
@@ -300,51 +148,41 @@ A lock older than 2 hours with no timestamp update is considered abandoned. Any
## Key Rules ## Key Rules
- **NO EMOJIS** - Use ASCII markers: `[OK]`, `[ERROR]`, `[WARNING]`, `[SUCCESS]`, `[INFO]` - **NO EMOJIS** — Use ASCII markers: `[OK]`, `[ERROR]`, `[WARNING]`, `[SUCCESS]`, `[INFO]`
- **No hardcoded credentials** - Use SOPS vault (`vault get-field <path> <field>`) or 1Password as fallback - **No hardcoded credentials** — Use SOPS vault (`vault get-field <path> <field>`) or 1Password as fallback
- **SSH:** Use system OpenSSH (on Windows: `C:\Windows\System32\OpenSSH\ssh.exe`, never Git for Windows SSH) - **SSH:** Use system OpenSSH (`C:\Windows\System32\OpenSSH\ssh.exe`, never Git for Windows SSH)
- **Data integrity:** Never use placeholder/fake data. Check SOPS vault, credentials.md, or ask user. - **Data integrity:** Never use placeholder/fake data. Check SOPS vault, credentials.md, or ask user.
- **Coding standards:** `.claude/CODING_GUIDELINES.md` (agents read on-demand, not every session) - **Coding standards:** `.claude/CODING_GUIDELINES.md` (agents read on-demand)
--- ---
## Automatic Behaviors ## Automatic Behaviors
- **Frontend Design:** Auto-invoke `/frontend-design` skill after ANY UI change (HTML/CSS/JSX/styling) - **Frontend Design:** Auto-invoke `/frontend-design` skill after ANY UI change (HTML/CSS/JSX/styling)
- **Sequential Thinking:** Use for genuine complexity - rejection loops, 3+ critical issues, architectural decisions, multi-step debugging - **Sequential Thinking:** Use for genuine complexity — rejection loops, 3+ critical issues, architectural decisions
- **Task Management:** Complex work (>3 steps) -> TaskCreate. Persist to `.claude/active-tasks.json`. - **Task Management:** Complex work (>3 steps) → TaskCreate. Persist to `.claude/active-tasks.json`.
--- ---
## Context Recovery ## Context Recovery
When user references previous work, use `/context` command. Never ask user for info in: When user references previous work, use `/context` command. Never ask for info in:
- `credentials.md` - Infrastructure reference (being migrated to SOPS vault at D:\vault) - `credentials.md` — Infrastructure reference (being migrated to SOPS vault at D:\vault)
- `session-logs/` - Daily work logs (also in `projects/*/session-logs/` and `clients/*/session-logs/`) - `session-logs/` — Daily work logs (also in `projects/*/session-logs/` and `clients/*/session-logs/`)
- `SESSION_STATE.md` - Project history - `SESSION_STATE.md` — Project history
### Credential Access (SOPS Vault - Primary) ### Credential Access (SOPS Vault)
Credentials are stored in SOPS+age encrypted YAML files in a dedicated Gitea repo.
**Vault repo:** `D:\vault` (git.azcomputerguru.com/azcomputerguru/vault, private)
**Structure:** infrastructure/, clients/, services/, projects/, msp-tools/
**To read credentials:**
```bash ```bash
bash D:/vault/scripts/vault.sh search "keyword" # Search (no decryption needed) bash D:/vault/scripts/vault.sh search "keyword" # Search without decrypting
bash D:/vault/scripts/vault.sh get-field <path> <field> # Get specific field bash D:/vault/scripts/vault.sh get-field <path> <field> # Get specific field
bash D:/vault/scripts/vault.sh get <path> # Decrypt full entry bash D:/vault/scripts/vault.sh get <path> # Decrypt full entry
bash D:/vault/scripts/vault.sh list # List all entries bash D:/vault/scripts/vault.sh list # List all entries
``` ```
**Encryption:** AES-256 via age. Metadata stays plaintext for searchability. Vault repo: `D:\vault` — structure: `infrastructure/`, `clients/`, `services/`, `projects/`, `msp-tools/`
**age key location:** `%APPDATA%\sops\age\keys.txt` (Windows) / `~/.config/sops/age/keys.txt` (Linux/Mac) **1Password fallback:** service account token in `infrastructure/1password-service-account.sops.yaml`
### 1Password (Fallback)
Service account token in vault: `infrastructure/1password-service-account.sops.yaml`
--- ---
@@ -353,115 +191,34 @@ Service account token in vault: `infrastructure/1password-service-account.sops.y
| Command | Purpose | | Command | Purpose |
|---------|---------| |---------|---------|
| `/checkpoint` | Dual checkpoint: git commit + database context | | `/checkpoint` | Dual checkpoint: git commit + database context |
| `/save` | Comprehensive session log (credentials, decisions, changes) | | `/save` | Comprehensive session log |
| `/context` | Search session logs, credentials.md, and 1Password | | `/context` | Search session logs, credentials.md, and 1Password |
| `/1password` | 1Password secrets management integration | | `/1password` | 1Password secrets management |
| `/sync` | Sync config from Gitea repository | | `/sync` | Sync config from Gitea repository |
| `/create-spec` | Create app specification for AutoCoder | | `/create-spec` | Create app specification for AutoCoder |
| `/frontend-design` | Modern frontend design patterns (auto-invoke after UI changes) | | `/frontend-design` | Modern frontend design (auto-invoke after UI changes) |
| `/remediation-tool` | M365 breach checks, tenant sweeps, gated remediation via Claude-MSP-Access Graph API app | | `/remediation-tool` | M365 breach checks, tenant sweeps, gated remediation |
--- ---
## File Placement (Quick Rules) ## File Placement
- **Dataforth DOS work** -> `projects/dataforth-dos/` - **Dataforth DOS work** → `projects/dataforth-dos/`
- **ClaudeTools API code** -> `api/`, `migrations/` (existing structure) - **ClaudeTools API code** → `api/`, `migrations/`
- **GuruRMM work** -> `projects/msp-tools/guru-rmm/` - **GuruRMM work** → `projects/msp-tools/guru-rmm/`
- **Client work** -> `clients/[client-name]/` - **Client work** → `clients/[client-name]/`
- **Session logs** -> project or client `session-logs/` subfolder; general -> root `session-logs/` - **Session logs** → project or client `session-logs/` subfolder; general → root `session-logs/`
- **Full guide:** `.claude/FILE_PLACEMENT_GUIDE.md` (read when saving files, not every session) - **Full guide:** `.claude/FILE_PLACEMENT_GUIDE.md`
--- ---
## Local AI (Ollama) ## Local AI (Ollama)
Ollama runs on Mike's workstation (DESKTOP-0O8A1RL) with GPU acceleration. Available to all team members via Tailscale. Tier 0 — use before spawning Haiku agents for low-stakes tasks (free, fast, private):
- **DESKTOP-0O8A1RL:** `http://localhost:11434`
| Model | Size | Use For | - **Other machines:** `http://100.92.127.64:11434` (Tailscale required)
|-------|------|---------| - **Models:** `qwen3:14b` (summarize/classify/draft), `codestral:22b` (code suggestions — always review)
| `qwen3:14b` | 9.3 GB | Summarization, classification, data extraction, drafting | - **Full reference:** `.claude/OLLAMA.md` (connection examples, model selection, review policy)
| `codestral:22b` | 12 GB | Code generation, refactoring suggestions, docstrings |
| `nomic-embed-text` | 274 MB | Embeddings only (used by GrepAI) |
### How to connect
**On Mike's workstation (local):**
```bash
curl -s http://localhost:11434/api/generate -d '{"model":"qwen3:14b","prompt":"...","stream":false}' | jq -r '.response'
```
**On any other machine via Tailscale:**
```bash
curl -s http://100.92.127.64:11434/api/generate -d '{"model":"qwen3:14b","prompt":"...","stream":false}' | jq -r '.response'
```
### Per-machine setup
Read `.claude/identity.json` to determine which machine you're on:
- **DESKTOP-0O8A1RL** (Mike's workstation): Ollama runs locally. Use `localhost:11434`.
- **Any other machine** (Howard's laptop, other workstations): Ollama is remote via Tailscale. Use `100.92.127.64:11434`. Requires Tailscale to be connected.
**To check if Ollama is reachable:**
```bash
curl -s http://100.92.127.64:11434/api/tags | python -c "import sys,json; [print(m['name']) for m in json.load(sys.stdin).get('models',[])]"
```
If it fails: verify Tailscale is connected (`tailscale status`), and that Mike's workstation is online.
### Access control
- Firewall rule on Mike's workstation allows port 11434 ONLY from Tailscale subnet (100.0.0.0/8)
- NOT exposed to LAN, VPN, or internet
- Binding: `OLLAMA_HOST=0.0.0.0:11434` (all interfaces, firewall restricts)
### Delegation pattern (Tier 0 — use instead of spawning a Haiku agent)
Determine the endpoint from identity.json, then call directly with the Bash tool:
```bash
# Resolve endpoint once per session
OLLAMA=$([ "$(jq -r .machine ~/.claude/identity.json 2>/dev/null)" = "DESKTOP-0O8A1RL" ] \
&& echo "http://localhost:11434" || echo "http://100.92.127.64:11434")
# General task (summarize, classify, extract, draft)
curl -s "$OLLAMA/api/generate" \
-d "{\"model\":\"qwen3:14b\",\"prompt\":\"$(echo "$PROMPT" | python3 -c 'import sys,json; print(json.dumps(sys.stdin.read()))'| tr -d '\"')\",\"stream\":false}" \
| python3 -c "import sys,json; print(json.load(sys.stdin).get('response',''))"
# Code suggestion (refactor ideas, docstrings — NOT production code)
# Same call, model: "codestral:22b"
```
**Practical shorthand** — for one-off inline prompts, use python3 to avoid escaping issues:
```bash
python3 -c "
import urllib.request, json, sys
url = 'http://localhost:11434/api/generate' # or 100.92.127.64
body = json.dumps({'model':'qwen3:14b','prompt': sys.argv[1],'stream':False}).encode()
res = json.loads(urllib.request.urlopen(urllib.request.Request(url, body)).read())
print(res['response'])
" "Summarize these changes in one sentence: ..."
```
**When to use which model:**
| Task | Model |
|------|-------|
| Summarize logs, diffs, session notes | qwen3:14b |
| Classify bug type, severity, category | qwen3:14b |
| Extract structured data from text output | qwen3:14b |
| Draft commit message from diff | qwen3:14b |
| Suggest refactor for a function (review output) | codestral:22b |
| Docstring / comment generation | codestral:22b |
**Review policy:**
- Low-stakes output (summary, label, draft) — use directly, no review needed
- Code suggestions from codestral — always review before applying
- Never use Ollama for: auth decisions, credential handling, production migrations, security review
**Review policy:** Always review Critical/High impact Ollama outputs (auth, security, migrations, production). Trust Low impact (classification, formatting). Flag uncertainty to user.
### GrepAI (Semantic Code Search) ### GrepAI (Semantic Code Search)
@@ -474,7 +231,7 @@ Use for intent-based search ("how does auth work"), exploring unfamiliar code, c
## Memory (Shared Across Machines) ## Memory (Shared Across Machines)
Stored in-repo at `.claude/memory/` -- syncs via Gitea to all workstations. Stored in-repo at `.claude/memory/` — syncs via Gitea to all workstations.
Index: `.claude/memory/MEMORY.md` Index: `.claude/memory/MEMORY.md`
**IMPORTANT:** Always write to `.claude/memory/` (repo-relative), NOT `~/.claude/projects/*/memory/`. **IMPORTANT:** Always write to `.claude/memory/` (repo-relative), NOT `~/.claude/projects/*/memory/`.
@@ -487,7 +244,9 @@ Index: `.claude/memory/MEMORY.md`
- **Agent definitions:** `.claude/agents/*.md` - **Agent definitions:** `.claude/agents/*.md`
- **MCP servers:** `MCP_SERVERS.md` - **MCP servers:** `MCP_SERVERS.md`
- **Coding standards:** `.claude/CODING_GUIDELINES.md` - **Coding standards:** `.claude/CODING_GUIDELINES.md`
- **Ollama connection + examples:** `.claude/OLLAMA.md`
- **PROJECT_STATE locking protocol:** `.claude/PROJECT_STATE_PROTOCOL.md`
--- ---
**Last Updated:** 2026-04-02 **Last Updated:** 2026-04-20

67
.claude/OLLAMA.md Normal file
View File

@@ -0,0 +1,67 @@
# Ollama — Local AI Reference
Ollama runs on Mike's workstation (DESKTOP-0O8A1RL) with GPU acceleration. Available to all team members via Tailscale.
## Models
| Model | Size | Use For |
|-------|------|---------|
| `qwen3:14b` | 9.3 GB | Summarization, classification, data extraction, drafting |
| `codestral:22b` | 12 GB | Code generation, refactoring suggestions, docstrings |
| `nomic-embed-text` | 274 MB | Embeddings only (used by GrepAI) |
## Endpoints
- **DESKTOP-0O8A1RL** (local): `http://localhost:11434`
- **Any other machine** (Tailscale required): `http://100.92.127.64:11434`
Check reachability:
```bash
curl -s http://100.92.127.64:11434/api/tags | python -c "import sys,json; [print(m['name']) for m in json.load(sys.stdin).get('models',[])]"
```
If it fails: verify Tailscale is connected (`tailscale status`) and Mike's workstation is online.
## Access Control
- Port 11434 allowed ONLY from Tailscale subnet (100.0.0.0/8)
- NOT exposed to LAN, VPN, or internet
- Binding: `OLLAMA_HOST=0.0.0.0:11434` (firewall restricts)
## Calling Ollama
Resolve endpoint from identity.json first:
```bash
OLLAMA=$([ "$(jq -r .machine .claude/identity.json 2>/dev/null)" = "DESKTOP-0O8A1RL" ] \
&& echo "http://localhost:11434" || echo "http://100.92.127.64:11434")
```
Preferred one-liner (avoids shell escaping):
```bash
python3 -c "
import urllib.request, json, sys
url = 'http://localhost:11434/api/generate'
body = json.dumps({'model':'qwen3:14b','prompt': sys.argv[1],'stream':False}).encode()
res = json.loads(urllib.request.urlopen(urllib.request.Request(url, body)).read())
print(res['response'])
" "Your prompt here"
```
For code suggestions, swap `qwen3:14b` for `codestral:22b`.
## When to Use Which Model
| Task | Model |
|------|-------|
| Summarize logs, diffs, session notes | qwen3:14b |
| Classify bug type, severity, category | qwen3:14b |
| Extract structured data from text | qwen3:14b |
| Draft commit message from diff | qwen3:14b |
| Suggest refactor for a function | codestral:22b |
| Docstring / comment generation | codestral:22b |
## Review Policy
- Low-stakes output (summary, classification, draft) — use directly
- Code suggestions from codestral — always review before applying
- Never use Ollama for: auth decisions, credential handling, production migrations, security review

42
.claude/PROJECT_STATE_PROTOCOL.md Normal file
View File

@@ -0,0 +1,42 @@
# PROJECT_STATE.md Locking Protocol
This protocol prevents conflicts between concurrent Claude sessions. Follow it for every significant action on any project that has a PROJECT_STATE.md.
## What Requires a Lock
- Editing or creating source code files
- Git commit or push
- SSH command that modifies a server (deploy, install, config change, service restart)
- Database schema change or data migration
- Build pipeline modification
Reading files, planning, and answering questions do NOT require a lock.
## The Protocol
**Step 1 — Read before acting**
Re-read PROJECT_STATE.md before starting:
- Check Active Session Locks: is anything locked that you need to touch?
- Conflicting lock < 2 hours old: stop, report to user, ask how to proceed.
- Lock > 2 hours old (stale): note it to user, clear the row, proceed.
**Step 2 — Claim your lock**
Add a row to Active Session Locks before performing the action:
| Session | Working On | Status | Blocks | Started |
|---------|-----------|--------|--------|---------|
| DESKTOP-0O8A1RL/Claude | Brief description | IN_PROGRESS | What others must avoid | HH:MM UTC |
Use `{machine}/{Claude or agent description}` as the Session identifier.
**Step 3 — Perform the action**
**Step 4 — Update on completion OR failure**
1. Remove your lock row
2. Add a Recent Changes entry with status: `COMPLETE`, `FAILED`, `PARTIAL`, or `ROLLED_BACK`
3. Update Current Project State if any component status changed
4. Check off completed Pending items
## Stale Lock Rule
A lock older than 2 hours with no timestamp update is abandoned. Clear it, note `[Cleared stale lock from {session}]` in Recent Changes, then proceed.

34
session-logs/2026-04-20-session.md
View File

@@ -223,6 +223,40 @@ Committed: db157e3
--- ---
## Update: 19:08 UTC — CLAUDE.md Optimization
### Summary
Refactored `.claude/CLAUDE.md` to reduce context window pressure. Extracted verbose sections to on-demand reference files. No functional changes — same rules, smaller footprint.
### Changes Made
**`.claude/CLAUDE.md`** — 494 lines → 252 lines (~49% reduction)
- Work Mode section: replaced 22-line narrative with a 5-row detection table; dropped color references (color changes don't work programmatically anyway — mode.md handles that)
- PROJECT_STATE.md Action Protocol: removed entirely (~65 lines) — moved to dedicated file
- Ollama section: trimmed from ~90 lines to 5-line summary + pointer
- Auto Context Loading: removed Benefits list, condensed anti-pattern examples to one sentence
- Credential Access: kept the 4 vault commands, removed encryption/key location details
**`.claude/OLLAMA.md`** — new file (67 lines)
- Full Ollama reference: endpoints, models, connection examples, review policy
- Only loaded when Claude needs to call Ollama (Tier 0 tasks)
**`.claude/PROJECT_STATE_PROTOCOL.md`** — new file (42 lines)
- Full locking protocol: what requires a lock, how to claim/release, stale lock rule
- Only loaded when Auto Context Loading finds a PROJECT_STATE.md in a project
### Decisions
- PROJECT_STATE files already have brief locking instructions embedded in their headers — the full protocol in CLAUDE.md was redundant
- Work mode colors were dropped from CLAUDE.md because `/color` is a CLI built-in Claude can't invoke programmatically; mode.md already has the authoritative detail
- Ollama extraction was the biggest single win (~90 lines) — it's only needed for Tier 0 delegation, not every session
### Pending
- None. This was a self-contained optimization pass.
---
## Key Infrastructure Reference ## Key Infrastructure Reference
| Resource | Details | | Resource | Details |