clawstrap 1.3.0 → 1.4.1

package/README.md

@@ -8,14 +8,21 @@
    ╚═════╝╚══════╝╚═╝  ╚═╝ ╚══╝╚══╝ ╚══════╝   ╚═╝   ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝
 ```
 
-### Scaffold a production-ready AI agent workspace in under 2 minutes.
+### Scaffold a production-ready AI agent workspace — then keep it sharp automatically.
 
 ![npm](https://img.shields.io/npm/v/clawstrap)
 ![License](https://img.shields.io/badge/license-MIT-blue)
 ![Node](https://img.shields.io/badge/node-%3E%3D18-brightgreen)
-![Built for](https://img.shields.io/badge/built%20for-Claude%20Code-blueviolet)
+![AI Runtime](https://img.shields.io/badge/AI%20runtime-claude%20%7C%20ollama%20%7C%20codex-blueviolet)
+![Adaptive Memory](https://img.shields.io/badge/adaptive%20memory-watch%20daemon-orange)
 
-Most AI agent systems fail not because the models are bad, but because there's no foundation underneath them. Sessions die. Context vanishes between runs. Agents drift without guardrails. Quality degrades silently mid-batch. Clawstrap generates a complete, opinionated workspace that solves all of this before you write your first prompt.
+## What Is Clawstrap
+
+Clawstrap scaffolds a complete, opinionated AI agent workspace in under 2 minutes. Run `npx clawstrap init`, answer five questions, and you get a governance-first directory: a `CLAUDE.md` master rules file loaded every session, approval-first workflow rules, quality gates, cross-session memory, subagent definitions, and a spec template — all wired together before you write your first prompt. Sessions no longer die without context. Agents no longer drift without guardrails.
+
+The watch daemon takes it further. Run `clawstrap watch` once and it stays alive in the background, doing three things automatically: processing session summaries from `tmp/sessions/` through an LLM adapter to extract decisions, corrections, and deferred ideas — writing them to `MEMORY.md` and `gotcha-log.md`; mining your git history for co-changing file pairs, high-churn directories, and recurring commit patterns; and scanning your codebase on a configurable schedule to maintain `.claude/rules/conventions.md` with live naming, import, and error-handling patterns. Your workspace gets more accurate over time without any manual work.
+
+Most AI workspace tools generate files and stop. Clawstrap keeps going.
 
 ## Quick Start
 
@@ -84,6 +91,11 @@ Done. Open GETTING_STARTED.md to begin.
 
 That's it. Your workspace is ready. Open `GETTING_STARTED.md` and start your first session.
 
+```bash
+# Then start the adaptive memory daemon:
+clawstrap watch
+```
+
 ## What You Get
 
 The generated workspace adapts to your answers. Here's the full structure with everything enabled:
@@ -98,7 +110,8 @@ my-project/
 │   ├── rules/
 │   │   ├── context-discipline.md    # Flush cadence, thin orchestrator pattern
 │   │   ├── approval-first.md        # Plan → approve → execute
-│   │   └── quality-gates.md         # QC as structural gate (team/production only)
+│   │   ├── quality-gates.md         # QC as structural gate (team/production only)
+│   │   └── conventions.md           # Auto-generated by clawstrap analyze (watch only)
 │   ├── agents/                      # (multi-agent workspaces only)
 │   │   ├── primary-agent.md         # Main orchestrator definition
 │   │   └── _template.md             # Template for new subagents
@@ -118,12 +131,13 @@ my-project/
 ├── specs/
 │   └── _template.md                 # Spec template (SDD only)
 ├── tmp/                             # Gitignored session workspace
+│   └── sessions/                    # Session summaries processed by watch daemon
 ├── research/                        # Reference material
 ├── context/                         # Session checkpoints
 └── artifacts/                       # Durable output
 ```
 
-Solo single-agent workspaces get a leaner structure — no agent definitions, no quality-gates rule file, no memory directory. The workspace scales with your needs.
+Solo single-agent workspaces get a leaner structure — no agent definitions, no quality-gates rule file, no memory directory. The watch daemon works with any configuration.
 
 ## See the Difference
 
@@ -149,26 +163,31 @@ Claude: Reading progress tracker... Last session completed endpoints for
         tmp/api-migration/plan.md. Ready to proceed?
 ```
 
-### Quality during batch work
+### Living memory across sessions
 
-**Without Clawstrap** — process everything, discover problems at the end:
+Without Clawstrap watch — corrections evaporate:
 
 ```
-You: Extract data from these 50 supplier files.
-Claude: Done! Here are all 50 results.
-You: (reviews results, finds item #12 has wrong field mapping)
-You: The mapping was wrong. Redo all 50.
+You (session 2): Stop using var. We use const and let everywhere.
+Claude: Got it, fixing now...
+
+You (session 3): Why are you using var again?
+Claude: I apologize — I don't have context from last session.
+You: (explains again)
 ```
 
-**With Clawstrap** — quality gates catch it early:
+With Clawstrap watch — corrections survive automatically:
 
 ```
-You: Extract data from these 50 supplier files.
-Claude: Processing batch. QC checkpoint at item #5 — reviewing most
-        complex item... Item #3 looks correct. Continuing.
-        QC checkpoint at item #10 — reviewing item #8...
-        ⚠ Item #12: field mapping mismatch. Confidence below threshold.
-        Stopping to fix before continuing. Only 7 items to redo.
+You (session 2): Stop using var. We use const and let everywhere.
+Claude: Understood. Writing session summary to tmp/sessions/...
+
+[watch daemon detects new summary, extracts the correction,
+appends to .claude/gotcha-log.md]
+
+You (session 3): Continue the refactor.
+Claude: Reading gotcha-log.md... Noted: use const/let, avoid var.
+        Continuing with that convention.
 ```
 
 ### Subagent context control
@@ -196,6 +215,116 @@ Claude: Research complete. Reading the 3 low-confidence entries for
         review before proceeding to the report.
 ```
 
+## The Watch Daemon
+
+The watch daemon is a background process that keeps your workspace's memory current without any manual work. It runs as a detached subprocess after `clawstrap watch`, survives terminal closes, and orchestrates three subsystems continuously.
+
+**Transcript processor** — watches `tmp/sessions/` for new `.md` session summaries; when one appears, sends it to your configured LLM adapter, extracts decisions, corrections, and deferred ideas as structured JSON, and appends them to `MEMORY.md`, `gotcha-log.md`, and `future-considerations.md` respectively.
+
+**Git observer** — on first run, reads your full commit history; on subsequent runs, reads only commits since the last processed SHA; extracts co-changing file pairs, high-churn directories, and recurring themes from commit messages; writes findings to `MEMORY.md`.
+
+**Convention scanner** — walks the codebase on a configurable interval (default: weekly) and maintains `.claude/rules/conventions.md` with detected naming conventions, import style, test patterns, error handling approach, and comment density; preserves any sections you have manually edited.
+
+```bash
+clawstrap watch           # Start daemon (detached, persists after terminal closes)
+clawstrap watch --once    # Run all observers once and exit
+clawstrap watch --stop    # Stop the running daemon
+clawstrap analyze         # Run convention scan immediately
+```
+
+The daemon uses an LLM to process session transcripts. Four adapters are available:
+
+| Adapter | Requires | Notes |
+|---------|----------|-------|
+| `claude-local` (default) | Claude Code CLI | Uses your existing subscription, no API cost |
+| `claude-api` | `ANTHROPIC_API_KEY` | Uses claude-haiku by default, configurable |
+| `ollama` | Ollama running locally | Fully offline/air-gapped |
+| `codex-local` | OpenAI Codex CLI | Uses your existing Codex setup |
+
+Configure in `.clawstrap.json`:
+
+```json
+{
+  "watch": {
+    "adapter": "claude-local",
+    "scan": { "intervalDays": 7 },
+    "silent": false
+  }
+}
+```
+
+The daemon causes zero disruption to Claude Code sessions. It runs as a detached background process and only writes to files already managed by Clawstrap.
+
+## Spec-Driven Development
+
+Enable SDD mode to enforce a spec-first workflow. Before Claude implements anything, it writes a structured spec, saves it to `specs/{name}.md`, and gets your explicit approval.
+
+```bash
+npx clawstrap init --sdd
+```
+
+What you get:
+
+- **`specs/_template.md`** — reusable spec format with problem statement, acceptance criteria, technical constraints, out-of-scope, and implementation notes
+- **`.claude/rules/sdd.md`** — governance rule Claude loads every session: "Never implement from a vague prompt"
+- **`.claude/commands/spec.md`** — `/spec` slash command that walks Claude through writing, saving, and presenting a spec for approval
+
+In practice:
+
+```
+You: Build a CSV export for the orders table.
+
+Claude: Writing spec to specs/csv-export-orders.md...
+
+        # Spec: CSV Export — Orders Table
+        Status: draft | Created: 2026-04-04
+
+        ## Problem Statement
+        Users need to export order data for reporting in external tools.
+
+        ## Acceptance Criteria
+        - [ ] Export all orders or filter by date range
+        - [ ] CSV includes: order_id, customer, total, status, created_at
+        - [ ] Empty state handled gracefully
+        - [ ] File named orders-export-{date}.csv
+
+        ## Technical Constraints
+        - Must use existing DB connection pool
+        - Must stream for large datasets (>10k rows)
+
+        ## Out of Scope
+        - PDF export
+        - Scheduled exports
+
+        Does this spec look right? Any changes before I start?
+
+You: Looks good, go ahead.
+
+Claude: Implementing from spec...
+```
+
+## The Five Principles
+
+### 1. File-First Persistence
+
+If it's not on disk, it didn't happen. Every finding, decision, and correction gets written to a durable location immediately — not at session end. Sessions are disposable. The work isn't.
+
+### 2. Approval-First Workflow
+
+Plan. Approve. Execute. In that order, every time. No agent acts without human sign-off. If scope changes mid-task, the agent stops and re-confirms. You stay in control.
+
+### 3. Quality as a Structural Gate
+
+Validation isn't optional — it's built into the workflow. Checkpoint every 5 outputs. Grade below B means stop, don't continue. Quality failures caught late cost 100% rework. Catch them early.
+
+### 4. Governed Subagents
+
+Every subagent gets a full definition: tools it can use, output schema, step-by-step procedure, and governance rules. No agent runs ungoverned. No "figure it out" prompts. Predictable agents ship reliable work.
+
+### 5. Binary Decision Architecture
+
+Complex decisions decomposed into sequential binary choices. One question at a time. Yes or no. This or that. Ambiguity kills reliability — binary decisions eliminate it.
+
 ## Working with Your Workspace
 
 ### Starting a session
@@ -231,8 +360,6 @@ You: Yes, go ahead.
 Claude: Starting implementation...
 ```
 
-Without Clawstrap, Claude would start writing code immediately. With it, the approval-first rule fires — you always see the plan before anything happens.
-
 ### Ending and resuming sessions
 
 When you're done for the day, tell Claude to wrap up. The session handoff checklist runs automatically:
@@ -250,16 +377,6 @@ Claude: Running session handoff checklist:
         Plan file: tmp/auth/plan.md
 ```
 
-Tomorrow, open Claude Code in the same directory. It reads `CLAUDE.md`, sees the plan file, and picks up:
-
-```
-You: Let's continue.
-
-Claude: Reading session plan from tmp/auth/plan.md...
-        Last session completed: login, register, JWT middleware.
-        Next: rate limiting middleware. Starting there.
-```
-
 ### Delegating to subagents
 
 When a task needs parallel work, tell Claude to delegate. The governance rules ensure subagents stay disciplined:
@@ -295,45 +412,22 @@ Claude: Writing correction to .claude/rules/context-discipline.md...
 
 The correction is saved to a durable rule file first, then applied. Next session, Claude follows the same rule without you repeating it.
 
-## The Five Principles
-
-### 1. File-First Persistence
-
-If it's not on disk, it didn't happen. Every finding, decision, and correction gets written to a durable location immediately — not at session end. Sessions are disposable. The work isn't.
-
-### 2. Approval-First Workflow
-
-Plan. Approve. Execute. In that order, every time. No agent acts without human sign-off. If scope changes mid-task, the agent stops and re-confirms. You stay in control.
-
-### 3. Quality as a Structural Gate
-
-Validation isn't optional — it's built into the workflow. Checkpoint every 5 outputs. Grade below B means stop, don't continue. Quality failures caught late cost 100% rework. Catch them early.
-
-### 4. Governed Subagents
-
-Every subagent gets a full definition: tools it can use, output schema, step-by-step procedure, and governance rules. No agent runs ungoverned. No "figure it out" prompts. Predictable agents ship reliable work.
-
-### 5. Binary Decision Architecture
-
-Complex decisions decomposed into sequential binary choices. One question at a time. Yes or no. This or that. Ambiguity kills reliability — binary decisions eliminate it.
-
-## Why Clawstrap Exists
-
-Built from real failures, not theory. After months of running AI agent workflows, the same problems kept killing projects: context lost between sessions with no recovery path, agents drifting into unsupervised work that had to be thrown away, and quality collapsing silently mid-batch because nobody was checking.
-
-Every rule in a Clawstrap workspace exists because something went wrong without it. The governance isn't theoretical — it's scar tissue encoded as structure.
-
 ## Commands
 
 ```
-clawstrap init [directory]                    Scaffold a new workspace (interactive)
-clawstrap init [directory] --yes              Use defaults, skip prompts
-clawstrap init [directory] --sdd              Enable Spec-Driven Development mode
-clawstrap add agent <name>                    Add a new agent definition
-clawstrap add skill <name>                    Add a new skill with SKILL.md
-clawstrap add project <name>                  Add a new project with README + process.md
-clawstrap status                              Show workspace configuration and structure
-clawstrap export --format paperclip           Export workspace as Paperclip company template
+clawstrap init [directory]                  Scaffold a new workspace (interactive)
+clawstrap init [directory] --yes            Use defaults, skip prompts
+clawstrap init [directory] --sdd            Enable Spec-Driven Development mode
+clawstrap add agent <name>                  Add a new agent definition
+clawstrap add skill <name>                  Add a new skill
+clawstrap add project <name>                Add a new project
+clawstrap status                            Show workspace configuration and structure
+clawstrap watch                             Start adaptive memory daemon
+clawstrap watch --once                      Run all observers once and exit
+clawstrap watch --stop                      Stop the running daemon
+clawstrap watch --silent                    Start daemon without output
+clawstrap analyze                           Run convention scan immediately
+clawstrap export --format paperclip         Export workspace as Paperclip company template
 ```
 
 ## Export to Paperclip
@@ -354,71 +448,21 @@ Generates a directory with `paperclip.manifest.json`, agent definitions with fro
 | `--mission`, `-m` | Company mission statement |
 | `--adapter`, `-a` | Agent adapter type (default: `claude_local`) |
 
-## Spec-Driven Development
-
-Enable SDD mode to enforce a spec-first workflow. Before Claude implements anything, it writes a structured spec, saves it to `specs/{name}.md`, and gets your explicit approval.
-
-```bash
-npx clawstrap init --sdd
-```
-
-What you get:
-
-- **`specs/_template.md`** — reusable spec format with problem statement, acceptance criteria, technical constraints, out-of-scope, and implementation notes
-- **`.claude/rules/sdd.md`** — governance rule Claude loads every session: "Never implement from a vague prompt"
-- **`.claude/commands/spec.md`** — `/spec` slash command that walks Claude through writing, saving, and presenting a spec for approval
-
-In practice:
-
-```
-You: Build a CSV export for the orders table.
-
-Claude: Writing spec to specs/csv-export-orders.md...
-
-        # Spec: CSV Export — Orders Table
-        Status: draft | Created: 2026-04-04
-
-        ## Problem Statement
-        Users need to export order data for reporting in external tools.
-
-        ## Acceptance Criteria
-        - [ ] Export all orders or filter by date range
-        - [ ] CSV includes: order_id, customer, total, status, created_at
-        - [ ] Empty state handled gracefully
-        - [ ] File named orders-export-{date}.csv
-
-        ## Technical Constraints
-        - Must use existing DB connection pool
-        - Must stream for large datasets (>10k rows)
-
-        ## Out of Scope
-        - PDF export
-        - Scheduled exports
-
-        Does this spec look right? Any changes before I start?
-
-You: Looks good, go ahead.
-
-Claude: Implementing from spec...
-```
+## Why Clawstrap Exists
 
-## Flags
+Built from real failures, not theory. After months of running AI agent workflows, the same problems kept killing projects: context lost between sessions with no recovery path, agents drifting into unsupervised work that had to be thrown away, and quality collapsing silently mid-batch because nobody was checking.
 
-| Flag | Description |
-|------|-------------|
-| `--yes`, `-y` | Skip prompts, use sensible defaults (init only) |
-| `--sdd` | Enable Spec-Driven Development mode (init only) |
-| `--version` | Show version |
-| `--help` | Show help |
+Every rule in a Clawstrap workspace exists because something went wrong without it. The governance isn't theoretical — it's scar tissue encoded as structure.
 
 ## Roadmap
 
 | Version | Status | What |
 |---------|--------|------|
-| **v1.0** | Done | `init` command, 4 workspace profiles, full governance templates, `--yes` mode |
-| **v1.1** | Done | `add agent`, `add skill`, `add project`, `status` commands |
-| **v1.2** | Done | `export --format paperclip` — Paperclip company template export |
-| **v1.3** | **Now** | `--sdd` flag — Spec-Driven Development mode with `/spec` slash command |
+| **v1.0** | Done | `init` command, 4 workspace profiles, governance templates, `--yes` mode |
+| **v1.1** | Done | `add agent/skill/project`, `status` command |
+| **v1.2** | Done | `export --format paperclip` |
+| **v1.3** | Done | `--sdd` flag, Spec-Driven Development mode |
+| **v1.4** | Now  | `clawstrap watch` — adaptive memory daemon, `clawstrap analyze` |
 | **v2.0** | Planned | Multi-model support, `upgrade` command, ClipMart publishing |
 
 ## Contributing