Sync ctx d07ce54bd2f1

README.md

@@ -1,5 +1,6 @@
 ---
 license: mit
+pretty_name: ctx
 tags:
 - agents
 - mcp
@@ -10,7 +11,6 @@ tags:
 - harness
 - codex
 - claude-code
-pretty_name: ctx
 ---
 
 # ctx — Skill, Agent, MCP & Harness Catalog
@@ -18,7 +18,7 @@ pretty_name: ctx
 [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE)
 [![Python 3.11+](https://img.shields.io/badge/Python-3.11+-green.svg)](https://python.org)
 [![PyPI](https://img.shields.io/pypi/v/claude-ctx.svg)](https://pypi.org/project/claude-ctx/)
-[![Tests](https://img.shields.io/badge/Tests-3447_collected-brightgreen.svg)](#)
+[![Tests](https://img.shields.io/badge/Tests-3448_collected-brightgreen.svg)](#)
 [![Graph](https://img.shields.io/badge/Graph-104%2C079_nodes_/_3.0M_edges-red.svg)](graph/)
 [![Docs](https://img.shields.io/badge/docs-MkDocs_Material-blue.svg)](https://stevesolun.github.io/ctx/)
 

mkdocs.yml

@@ -21,8 +21,8 @@ strict: true
 # in the rendered nav; the skill-router section summarizes it.
 not_in_nav: |
   SKILL.md
-
-theme:
+
+theme:
   name: material
   features:
     - navigation.instant

skills/find-skills/SKILL.md

@@ -1,65 +1,65 @@
----
-name: find-skills
-description: Discover installable agent skills from ctx's shipped Skills.sh catalog, the Skills.sh search API, and the npx skills CLI. Use when a user asks whether a skill exists, wants to add/update a skill, or needs a repeatable procedure for finding candidate skills safely.
----
-
-# Find Skills
-
-Use this when the user wants a skill recommendation or when ctx's curated graph
-does not already contain the capability they need.
-
-## Source Of Truth
-
-1. Query ctx first: use the shared recommendation surface so curated skills,
-   agents, MCP servers, harnesses, and the shipped Skills.sh external catalog are
-   ranked together.
-2. If ctx is stale or too narrow, search upstream:
-
-```bash
-npx skills find "<query>"
-```
-
-3. For the upstream discovery skill itself, the canonical install command is:
-
-```bash
-npx skills add https://github.com/vercel-labs/skills --skill find-skills
-```
-
-## Recommendation Checklist
-
-Before recommending or installing an upstream skill:
-
-- Prefer official or high-reputation sources.
-- Check install count, upstream repo activity, and whether a license is present.
-- Read the skill body before installing; never rely only on search rank.
-- Run the security review below for new or updated skills.
-- If a matching ctx entity already exists, use the update-review flow and compare
-  benefits, risks, lost tags/capabilities, and security findings before replacing
-  anything.
-
-## Security Review
-
-Flag the candidate for manual review if it asks the agent to:
-
-- Run network-fetched shell code, such as `curl ... | sh`, `wget ... | bash`, or
-  `Invoke-Expression`.
-- Exfiltrate secrets or environment variables.
-- Disable tests, lint, CI, permissions, sandboxing, auth, TLS, or audit logging.
-- Run destructive commands such as `rm -rf`, `git reset --hard`, or broad file
-  deletion without a scoped path.
-- Install packages or tools from unpinned, unknown, or typo-squatted sources.
-
-Do not install or import a flagged skill automatically. Present the user with
-the exact concern, the likely benefit, the safer alternative, and the explicit
-command they would need to approve.
-
-## Updating ctx
-
-When adding accepted skills to ctx itself:
-
-1. Add or update the source skill file.
-2. Run `ctx-skill-add --skill-path <path>/SKILL.md --name <slug>`.
-3. Run `ctx-wiki-graphify`.
-4. Repack `graph/wiki-graph.tar.gz` and refresh repo stats.
-5. Verify with `ctx-scan-repo --repo . --recommend` or
-   `ctx__recommend_bundle` for a query that should return the skill.
+---
+name: find-skills
+description: Discover installable agent skills from ctx's shipped Skills.sh catalog, the Skills.sh search API, and the npx skills CLI. Use when a user asks whether a skill exists, wants to add/update a skill, or needs a repeatable procedure for finding candidate skills safely.
+---
+
+# Find Skills
+
+Use this when the user wants a skill recommendation or when ctx's curated graph
+does not already contain the capability they need.
+
+## Source Of Truth
+
+1. Query ctx first: use the shared recommendation surface so curated skills,
+   agents, MCP servers, harnesses, and the shipped Skills.sh external catalog are
+   ranked together.
+2. If ctx is stale or too narrow, search upstream:
+
+```bash
+npx skills find "<query>"
+```
+
+3. For the upstream discovery skill itself, the canonical install command is:
+
+```bash
+npx skills add https://github.com/vercel-labs/skills --skill find-skills
+```
+
+## Recommendation Checklist
+
+Before recommending or installing an upstream skill:
+
+- Prefer official or high-reputation sources.
+- Check install count, upstream repo activity, and whether a license is present.
+- Read the skill body before installing; never rely only on search rank.
+- Run the security review below for new or updated skills.
+- If a matching ctx entity already exists, use the update-review flow and compare
+  benefits, risks, lost tags/capabilities, and security findings before replacing
+  anything.
+
+## Security Review
+
+Flag the candidate for manual review if it asks the agent to:
+
+- Run network-fetched shell code, such as `curl ... | sh`, `wget ... | bash`, or
+  `Invoke-Expression`.
+- Exfiltrate secrets or environment variables.
+- Disable tests, lint, CI, permissions, sandboxing, auth, TLS, or audit logging.
+- Run destructive commands such as `rm -rf`, `git reset --hard`, or broad file
+  deletion without a scoped path.
+- Install packages or tools from unpinned, unknown, or typo-squatted sources.
+
+Do not install or import a flagged skill automatically. Present the user with
+the exact concern, the likely benefit, the safer alternative, and the explicit
+command they would need to approve.
+
+## Updating ctx
+
+When adding accepted skills to ctx itself:
+
+1. Add or update the source skill file.
+2. Run `ctx-skill-add --skill-path <path>/SKILL.md --name <slug>`.
+3. Run `ctx-wiki-graphify`.
+4. Repack `graph/wiki-graph.tar.gz` and refresh repo stats.
+5. Verify with `ctx-scan-repo --repo . --recommend` or
+   `ctx__recommend_bundle` for a query that should return the skill.

skills/repo-stats-autoupdate/SKILL.md

@@ -1,56 +1,56 @@
----
-name: repo-stats-autoupdate
-description: Keeps README badge + inline counts in sync with the real number of skills, agents, graph nodes/edges, communities, converted pipelines, and pytest tests. Runs automatically on every commit via a git pre-commit hook. Use when the README drifts from reality or before publishing a release.
-type: maintenance
-priority: 30
-always_load: false
----
-
-# repo-stats-autoupdate
-
-## What this skill does
-
-Reads the **authoritative sources** for the ctx repo's key numbers and patches `README.md` in place so badges and inline counts never drift:
-
-| Number | Source of truth |
-|---|---|
-| Skills | `~/.claude/skill-wiki/graphify-out/graph.json` — count nodes where `type == "skill"` |
-| Agents | same file — count nodes where `type == "agent"` |
-| Graph nodes | `len(graph["nodes"])` |
-| Graph edges | `len(graph["edges"])`, formatted as `642K` / `1.2M` |
-| Communities | `graph/communities.json` → `total_communities` |
-| Converted pipelines | `~/.claude/skill-wiki/converted/` subdir count |
-| Tests | `pytest --collect-only -q` from `src/` |
-
-Fields that can't be resolved (e.g. wiki not deployed) are left untouched — the updater never blocks a commit.
-
-## How to use it
-
-**One-time install (per clone):**
-```bash
-git config core.hooksPath .githooks
-```
-
-After that, every `git commit` runs `.githooks/pre-commit` which calls the updater and re-stages `README.md` if any number drifted.
-
-**Manual run:**
-```bash
-python src/update_repo_stats.py          # patch README
-python src/update_repo_stats.py --check  # exit 1 if stale (for CI)
-```
-
-## When NOT to use it
-
-- When you've deliberately phrased a number descriptively (e.g. "over 1,700 skills") — the regex patterns only match exact digits, so prose phrasings are safe, but double-check after big imports.
-- When the wiki isn't deployed locally. The updater degrades gracefully: it prints a warning and skips fields it can't resolve. It does **not** invent numbers.
-
-## Known gaps (intentionally out of scope)
-
-- **Does not rebuild the graph or wiki.** Rebuilding `graph/wiki-graph.tar.gz` takes minutes and churns a 159 MB tree — too heavy for per-commit. Run `python src/wiki_graphify.py` and repack the tarball manually when the skill catalog changes materially.
-- **Does not verify graph integrity.** If `graphify-out/graph.json` is corrupt, you'll get junk numbers. Run the wiki health check (`python src/wiki_orchestrator.py --check`) separately.
-
-## Related files
-
-- [src/update_repo_stats.py](../../src/update_repo_stats.py) — the worker
-- [.githooks/pre-commit](../../.githooks/pre-commit) — the per-commit trigger
-- `README.md` — the target being kept in sync
+---
+name: repo-stats-autoupdate
+description: Keeps README badge + inline counts in sync with the real number of skills, agents, graph nodes/edges, communities, converted pipelines, and pytest tests. Runs automatically on every commit via a git pre-commit hook. Use when the README drifts from reality or before publishing a release.
+type: maintenance
+priority: 30
+always_load: false
+---
+
+# repo-stats-autoupdate
+
+## What this skill does
+
+Reads the **authoritative sources** for the ctx repo's key numbers and patches `README.md` in place so badges and inline counts never drift:
+
+| Number | Source of truth |
+|---|---|
+| Skills | `~/.claude/skill-wiki/graphify-out/graph.json` — count nodes where `type == "skill"` |
+| Agents | same file — count nodes where `type == "agent"` |
+| Graph nodes | `len(graph["nodes"])` |
+| Graph edges | `len(graph["edges"])`, formatted as `642K` / `1.2M` |
+| Communities | `graph/communities.json` → `total_communities` |
+| Converted pipelines | `~/.claude/skill-wiki/converted/` subdir count |
+| Tests | `pytest --collect-only -q` from `src/` |
+
+Fields that can't be resolved (e.g. wiki not deployed) are left untouched — the updater never blocks a commit.
+
+## How to use it
+
+**One-time install (per clone):**
+```bash
+git config core.hooksPath .githooks
+```
+
+After that, every `git commit` runs `.githooks/pre-commit` which calls the updater and re-stages `README.md` if any number drifted.
+
+**Manual run:**
+```bash
+python src/update_repo_stats.py          # patch README
+python src/update_repo_stats.py --check  # exit 1 if stale (for CI)
+```
+
+## When NOT to use it
+
+- When you've deliberately phrased a number descriptively (e.g. "over 1,700 skills") — the regex patterns only match exact digits, so prose phrasings are safe, but double-check after big imports.
+- When the wiki isn't deployed locally. The updater degrades gracefully: it prints a warning and skips fields it can't resolve. It does **not** invent numbers.
+
+## Known gaps (intentionally out of scope)
+
+- **Does not rebuild the graph or wiki.** Rebuilding `graph/wiki-graph.tar.gz` takes minutes and churns a 159 MB tree — too heavy for per-commit. Run `python src/wiki_graphify.py` and repack the tarball manually when the skill catalog changes materially.
+- **Does not verify graph integrity.** If `graphify-out/graph.json` is corrupt, you'll get junk numbers. Run the wiki health check (`python src/wiki_orchestrator.py --check`) separately.
+
+## Related files
+
+- [src/update_repo_stats.py](../../src/update_repo_stats.py) — the worker
+- [.githooks/pre-commit](../../.githooks/pre-commit) — the per-commit trigger
+- `README.md` — the target being kept in sync

skills/skill-router/SKILL.md

@@ -1,37 +1,37 @@
----
-name: skill-router
-description: Alive skill router — reads the current project's stack and loads/unloads skills dynamically. Invoke at session start or when project context changes.
-type: meta
-priority: 99
-always_load: true
----
-
-# skill-router
-
-Reads `~/.claude/skill-manifest.json` and `~/.claude/pending-skills.json`.
-Executes the 5-stage pipeline. **No skipping stages.**
-
-## Pipeline
-
-```
-01-scope → 02-plan → 03-build → 04-check → 05-deliver
-```
-
-## Invoke When
-
-- Session starts (automatic via using-superpowers pattern)
-- User switches projects or opens a new repo
-- `pending-skills.json` exists (mid-session signals detected)
-- User asks "what skills are loaded?" or "update my skills"
-
-## Stage Files
-
-All stage files are in `references/`. Read and follow each in order.
-
-1. **[01-scope](references/01-scope.md)** — Is a scan needed?
-2. **[02-plan](references/02-plan.md)** — Run the scanner
-3. **[03-build](references/03-build.md)** — Resolve the manifest
-4. **[04-check](references/04-check.md)** — Validate before loading
-5. **[05-deliver](references/05-deliver.md)** — Present to user
-
-See [check-gates.md](check-gates.md) for validation questions.
+---
+name: skill-router
+description: Alive skill router — reads the current project's stack and loads/unloads skills dynamically. Invoke at session start or when project context changes.
+type: meta
+priority: 99
+always_load: true
+---
+
+# skill-router
+
+Reads `~/.claude/skill-manifest.json` and `~/.claude/pending-skills.json`.
+Executes the 5-stage pipeline. **No skipping stages.**
+
+## Pipeline
+
+```
+01-scope → 02-plan → 03-build → 04-check → 05-deliver
+```
+
+## Invoke When
+
+- Session starts (automatic via using-superpowers pattern)
+- User switches projects or opens a new repo
+- `pending-skills.json` exists (mid-session signals detected)
+- User asks "what skills are loaded?" or "update my skills"
+
+## Stage Files
+
+All stage files are in `references/`. Read and follow each in order.
+
+1. **[01-scope](references/01-scope.md)** — Is a scan needed?
+2. **[02-plan](references/02-plan.md)** — Run the scanner
+3. **[03-build](references/03-build.md)** — Resolve the manifest
+4. **[04-check](references/04-check.md)** — Validate before loading
+5. **[05-deliver](references/05-deliver.md)** — Present to user
+
+See [check-gates.md](check-gates.md) for validation questions.

skills/skill-router/references/01-scope.md

@@ -1,31 +1,31 @@
-# Stage 1: Scope — Is a Scan Needed?
-
-Determine which action to take before spending tokens on a full scan.
-
-## Check Order
-
-1. **Read** `~/.claude/pending-skills.json`
-   - If exists and `generated_at` is < 2 hours old → action = `apply_pending`
-   - Delete the file after reading (it is one-shot)
-
-2. **Read** `~/.claude/skill-manifest.json`
-   - Extract `repo_path` and `generated_at`
-   - If `repo_path` ≠ current working directory → action = `full_scan`
-   - If `generated_at` > 1 hour ago → action = `full_scan`
-   - Otherwise → action = `use_cached`
-
-3. **If no manifest exists** → action = `full_scan`
-
-## Output
-
-Emit one of:
-```
-ACTION: full_scan      — proceed to Stage 2
-ACTION: apply_pending  — skip to Stage 3 (use pending-skills.json as manifest delta)
-ACTION: use_cached     — skip to Stage 5 (just display current manifest)
-```
-
-## Fast Path
-
-If `use_cached`: read current manifest load list and jump to Stage 5 directly.
-If `apply_pending`: merge pending into current manifest, jump to Stage 4.
+# Stage 1: Scope — Is a Scan Needed?
+
+Determine which action to take before spending tokens on a full scan.
+
+## Check Order
+
+1. **Read** `~/.claude/pending-skills.json`
+   - If exists and `generated_at` is < 2 hours old → action = `apply_pending`
+   - Delete the file after reading (it is one-shot)
+
+2. **Read** `~/.claude/skill-manifest.json`
+   - Extract `repo_path` and `generated_at`
+   - If `repo_path` ≠ current working directory → action = `full_scan`
+   - If `generated_at` > 1 hour ago → action = `full_scan`
+   - Otherwise → action = `use_cached`
+
+3. **If no manifest exists** → action = `full_scan`
+
+## Output
+
+Emit one of:
+```
+ACTION: full_scan      — proceed to Stage 2
+ACTION: apply_pending  — skip to Stage 3 (use pending-skills.json as manifest delta)
+ACTION: use_cached     — skip to Stage 5 (just display current manifest)
+```
+
+## Fast Path
+
+If `use_cached`: read current manifest load list and jump to Stage 5 directly.
+If `apply_pending`: merge pending into current manifest, jump to Stage 4.

skills/skill-router/references/02-plan.md

@@ -1,34 +1,34 @@
-# Stage 2: Plan — Run the Scanner
-
-Only reached if Stage 1 determined `full_scan`.
-
-## Steps
-
-1. **Identify current repo root**
-   - Use current working directory from the session
-   - Confirm it looks like a project (has package.json, pyproject.toml, Cargo.toml, go.mod, etc.)
-   - If it's the home dir or system dir → warn and use limited scan
-
-2. **Run the scanner**
-
-```bash
-python3 ~/.claude/ctx/scan_repo.py \
-  --repo <cwd> \
-  --output /tmp/skill-stack-profile.json
-```
-
-(Replace `~/.claude/ctx` with the actual ctx_dir from `~/.claude/skill-registry.json`)
-
-3. **Read the output** (`/tmp/skill-stack-profile.json`)
-   - Report top 5 detected stacks with confidence scores
-   - Note any monorepo packages detected
-
-## Expected Duration
-
-< 10 seconds for repos up to 10K files. If it takes longer, something is wrong.
-
-## On Failure
-
-If scanner exits non-zero or output file is missing:
-- Fallback: use `use_cached` action if manifest exists
-- Otherwise: proceed to Stage 3 with an empty profile (will load only meta-skills)
+# Stage 2: Plan — Run the Scanner
+
+Only reached if Stage 1 determined `full_scan`.
+
+## Steps
+
+1. **Identify current repo root**
+   - Use current working directory from the session
+   - Confirm it looks like a project (has package.json, pyproject.toml, Cargo.toml, go.mod, etc.)
+   - If it's the home dir or system dir → warn and use limited scan
+
+2. **Run the scanner**
+
+```bash
+python3 ~/.claude/ctx/scan_repo.py \
+  --repo <cwd> \
+  --output /tmp/skill-stack-profile.json
+```
+
+(Replace `~/.claude/ctx` with the actual ctx_dir from `~/.claude/skill-registry.json`)
+
+3. **Read the output** (`/tmp/skill-stack-profile.json`)
+   - Report top 5 detected stacks with confidence scores
+   - Note any monorepo packages detected
+
+## Expected Duration
+
+< 10 seconds for repos up to 10K files. If it takes longer, something is wrong.
+
+## On Failure
+
+If scanner exits non-zero or output file is missing:
+- Fallback: use `use_cached` action if manifest exists
+- Otherwise: proceed to Stage 3 with an empty profile (will load only meta-skills)

skills/skill-router/references/03-build.md

@@ -1,38 +1,38 @@
-# Stage 3: Build — Resolve the Manifest
-
-Convert the stack profile into a concrete load/unload list.
-
-## Steps
-
-1. **Run the resolver**
-
-```bash
-python3 ~/.claude/ctx/resolve_skills.py \
-  --profile /tmp/skill-stack-profile.json \
-  --wiki ~/.claude/skill-wiki \
-  --output ~/.claude/skill-manifest.json \
-  --intent-log ~/.claude/intent-log.jsonl
-```
-
-2. **Read the manifest** (`~/.claude/skill-manifest.json`)
-   - Extract: `load[]`, `unload[]`, `warnings[]`, `suggestions[]`
-
-3. **Sync to wiki**
-
-```bash
-python3 ~/.claude/ctx/wiki_sync.py \
-  --profile /tmp/skill-stack-profile.json \
-  --manifest ~/.claude/skill-manifest.json \
-  --wiki ~/.claude/skill-wiki
-```
-
-## apply_pending Fast Path
-
-If Stage 1 returned `apply_pending`:
-- Read `pending-skills.json` suggestion list
-- Add each suggested skill to the manifest load list (if available on disk)
-- Skip re-running the full resolver
-
-## On Failure
-
-If resolver fails: use previous manifest if < 24 hours old. Report error as warning.
+# Stage 3: Build — Resolve the Manifest
+
+Convert the stack profile into a concrete load/unload list.
+
+## Steps
+
+1. **Run the resolver**
+
+```bash
+python3 ~/.claude/ctx/resolve_skills.py \
+  --profile /tmp/skill-stack-profile.json \
+  --wiki ~/.claude/skill-wiki \
+  --output ~/.claude/skill-manifest.json \
+  --intent-log ~/.claude/intent-log.jsonl
+```
+
+2. **Read the manifest** (`~/.claude/skill-manifest.json`)
+   - Extract: `load[]`, `unload[]`, `warnings[]`, `suggestions[]`
+
+3. **Sync to wiki**
+
+```bash
+python3 ~/.claude/ctx/wiki_sync.py \
+  --profile /tmp/skill-stack-profile.json \
+  --manifest ~/.claude/skill-manifest.json \
+  --wiki ~/.claude/skill-wiki
+```
+
+## apply_pending Fast Path
+
+If Stage 1 returned `apply_pending`:
+- Read `pending-skills.json` suggestion list
+- Add each suggested skill to the manifest load list (if available on disk)
+- Skip re-running the full resolver
+
+## On Failure
+
+If resolver fails: use previous manifest if < 24 hours old. Report error as warning.

skills/skill-router/references/04-check.md

@@ -1,28 +1,28 @@
-# Stage 4: Check — Validate Before Loading
-
-Binary YES/NO gates. Any NO = stop and report before proceeding.
-
-## Gates
-
-- [ ] **G1** Load list has ≥ 1 skill?
-- [ ] **G2** Load list has ≤ 15 skills?
-- [ ] **G3** No skill appears in both load AND unload?
-- [ ] **G4** All loaded skills have a valid path that exists on disk?
-- [ ] **G5** No two conflicting skills in load list? (e.g., flask + fastapi, jest + vitest)
-- [ ] **G6** Manifest `generated_at` is within the last 24 hours?
-
-## Failure Handling
-
-| Gate | Failure action |
-|------|---------------|
-| G1 | Load only meta-skills (skill-router), warn user |
-| G2 | Trim to top 15 by priority, warn about excluded skills |
-| G3 | Remove the skill from unload (keep loaded version) |
-| G4 | Remove missing-path skills from load, add to suggestions |
-| G5 | Keep highest-priority conflicting skill, warn about the other |
-| G6 | Warn that manifest may be stale, offer to re-scan |
-
-## Pass Condition
-
-All gates YES (or failures handled per table above).
-Proceed to Stage 5 only after resolving all failures.
+# Stage 4: Check — Validate Before Loading
+
+Binary YES/NO gates. Any NO = stop and report before proceeding.
+
+## Gates
+
+- [ ] **G1** Load list has ≥ 1 skill?
+- [ ] **G2** Load list has ≤ 15 skills?
+- [ ] **G3** No skill appears in both load AND unload?
+- [ ] **G4** All loaded skills have a valid path that exists on disk?
+- [ ] **G5** No two conflicting skills in load list? (e.g., flask + fastapi, jest + vitest)
+- [ ] **G6** Manifest `generated_at` is within the last 24 hours?
+
+## Failure Handling
+
+| Gate | Failure action |
+|------|---------------|
+| G1 | Load only meta-skills (skill-router), warn user |
+| G2 | Trim to top 15 by priority, warn about excluded skills |
+| G3 | Remove the skill from unload (keep loaded version) |
+| G4 | Remove missing-path skills from load, add to suggestions |
+| G5 | Keep highest-priority conflicting skill, warn about the other |
+| G6 | Warn that manifest may be stale, offer to re-scan |
+
+## Pass Condition
+
+All gates YES (or failures handled per table above).
+Proceed to Stage 5 only after resolving all failures.

skills/skill-router/references/05-deliver.md

@@ -1,34 +1,34 @@
-# Stage 5: Deliver — Present to User
-
-Show a concise summary. Keep it under 20 lines.
-
-## Format
-
-```
-─── Skill Router ──────────────────────────────────
-Project: <repo_name> (<project_type>)
-Detected: <top 3 stacks with confidence>
-
-Loading  [N]: <skill1>, <skill2>, <skill3> ...
-Unloading[M]: <skill_a>, <skill_b> ... (first 5)
-
-⚠ Warnings: <if any, one per line>
-💡 Suggestions: <missing but needed skills, with install hint>
-───────────────────────────────────────────────────
-```
-
-## Rules
-
-- List loaded skills in priority order (highest first)
-- If unload count > 5: show first 5 then "+ N more"
-- Warnings: only show if there are any (no empty section)
-- Suggestions: include install hint if skill is in a marketplace
-
-## After Delivery
-
-The session continues normally. The loaded skills are now available
-via the Skill tool. The PostToolUse hook watches for new intent signals.
-
-If the user asks to re-evaluate: re-run from Stage 1.
-If the user asks to force-load a skill: add `always_load: true` to its wiki page.
-If the user asks to block a skill: add `never_load: true` to its wiki page.
+# Stage 5: Deliver — Present to User
+
+Show a concise summary. Keep it under 20 lines.
+
+## Format
+
+```
+─── Skill Router ──────────────────────────────────
+Project: <repo_name> (<project_type>)
+Detected: <top 3 stacks with confidence>
+
+Loading  [N]: <skill1>, <skill2>, <skill3> ...
+Unloading[M]: <skill_a>, <skill_b> ... (first 5)
+
+⚠ Warnings: <if any, one per line>
+💡 Suggestions: <missing but needed skills, with install hint>
+───────────────────────────────────────────────────
+```
+
+## Rules
+
+- List loaded skills in priority order (highest first)
+- If unload count > 5: show first 5 then "+ N more"
+- Warnings: only show if there are any (no empty section)
+- Suggestions: include install hint if skill is in a marketplace
+
+## After Delivery
+
+The session continues normally. The loaded skills are now available
+via the Skill tool. The PostToolUse hook watches for new intent signals.
+
+If the user asks to re-evaluate: re-run from Stage 1.
+If the user asks to force-load a skill: add `always_load: true` to its wiki page.
+If the user asks to block a skill: add `never_load: true` to its wiki page.

skills/toolbox/SKILL.md

@@ -1,61 +1,61 @@
----
-name: toolbox
-description: Pre/post dev toolbox — named bundles of skills/agents loaded before development work and councils of experts invoked after. Run /toolbox or the toolbox.py CLI to list, activate, initialize, export, import, and validate toolboxes. Invoke at the start or end of a dev session, when setting up a new repo, or when sharing toolboxes with a team.
-type: feature
-priority: 50
-always_load: false
----
-
-# toolbox
-
-Named bundles of skills/agents that run **before** (`pre`) or **after** (`post`)
-a development task. Learn patterns from past work and propose new bundles.
-Invoke slash, pre-commit, session-end, or file-save.
-
-## Quick reference
-
-| Task                                       | Command                                 |
-|--------------------------------------------|-----------------------------------------|
-| Seed 5 starter toolboxes                   | `python src/toolbox.py init`            |
-| List available toolboxes                   | `python src/toolbox.py list`            |
-| Inspect one                                | `python src/toolbox.py show ship-it`    |
-| Activate (global)                          | `python src/toolbox.py activate ship-it`|
-| Export for sharing                         | `python src/toolbox.py export ship-it`  |
-| Import a shared toolbox                    | `python src/toolbox.py import file.yaml`|
-| Validate config                            | `python src/toolbox.py validate`        |
-
-## Starter templates
-
-| Name              | When to use                                                     |
-|-------------------|-----------------------------------------------------------------|
-| ship-it           | End-of-feature: 7-expert council on the change set              |
-| security-sweep    | Security audit with guardrail blocking HIGH findings            |
-| refactor-safety   | Refactors with graph-informed blast radius                       |
-| docs-review       | When touching Markdown or API docs                              |
-| fresh-repo-init   | Blank repo: run intent interview and scaffold                    |
-
-## Data model
-
-Global: `~/.claude/toolboxes.json`.
-Per-repo (overrides global): `<repo_root>/.toolbox.yaml`.
-
-Each toolbox declares:
-- `pre` — skills/agents to load before work starts
-- `post` — agents that form the council after work ends
-- `scope.analysis` — `diff` | `full` | `graph-blast` | `dynamic`
-- `trigger` — `slash`, `pre_commit`, `session_end`, or `file_save` glob
-- `budget.max_tokens` / `budget.max_seconds` — hard stop on runaway cost
-- `dedup.policy` — `fresh` (always re-run) or `cached` (reuse within window)
-- `guardrail` — if true, block commit on HIGH findings
-
-## Invoke when
-
-- User says "I'm done with this feature" or commits a feature branch.
-- User opens a fresh repo (triggers `fresh-repo-init` suggestion).
-- User touches security-sensitive paths (triggers `security-sweep` file_save).
-- User asks "what toolboxes do I have?" or "run the council".
-
-## Reasoning
-
-See `docs/roadmap/toolbox.md` for the full design rationale, open decisions,
-and phase rollout.
+---
+name: toolbox
+description: Pre/post dev toolbox — named bundles of skills/agents loaded before development work and councils of experts invoked after. Run /toolbox or the toolbox.py CLI to list, activate, initialize, export, import, and validate toolboxes. Invoke at the start or end of a dev session, when setting up a new repo, or when sharing toolboxes with a team.
+type: feature
+priority: 50
+always_load: false
+---
+
+# toolbox
+
+Named bundles of skills/agents that run **before** (`pre`) or **after** (`post`)
+a development task. Learn patterns from past work and propose new bundles.
+Invoke slash, pre-commit, session-end, or file-save.
+
+## Quick reference
+
+| Task                                       | Command                                 |
+|--------------------------------------------|-----------------------------------------|
+| Seed 5 starter toolboxes                   | `python src/toolbox.py init`            |
+| List available toolboxes                   | `python src/toolbox.py list`            |
+| Inspect one                                | `python src/toolbox.py show ship-it`    |
+| Activate (global)                          | `python src/toolbox.py activate ship-it`|
+| Export for sharing                         | `python src/toolbox.py export ship-it`  |
+| Import a shared toolbox                    | `python src/toolbox.py import file.yaml`|
+| Validate config                            | `python src/toolbox.py validate`        |
+
+## Starter templates
+
+| Name              | When to use                                                     |
+|-------------------|-----------------------------------------------------------------|
+| ship-it           | End-of-feature: 7-expert council on the change set              |
+| security-sweep    | Security audit with guardrail blocking HIGH findings            |
+| refactor-safety   | Refactors with graph-informed blast radius                       |
+| docs-review       | When touching Markdown or API docs                              |
+| fresh-repo-init   | Blank repo: run intent interview and scaffold                    |
+
+## Data model
+
+Global: `~/.claude/toolboxes.json`.
+Per-repo (overrides global): `<repo_root>/.toolbox.yaml`.
+
+Each toolbox declares:
+- `pre` — skills/agents to load before work starts
+- `post` — agents that form the council after work ends
+- `scope.analysis` — `diff` | `full` | `graph-blast` | `dynamic`
+- `trigger` — `slash`, `pre_commit`, `session_end`, or `file_save` glob
+- `budget.max_tokens` / `budget.max_seconds` — hard stop on runaway cost
+- `dedup.policy` — `fresh` (always re-run) or `cached` (reuse within window)
+- `guardrail` — if true, block commit on HIGH findings
+
+## Invoke when
+
+- User says "I'm done with this feature" or commits a feature branch.
+- User opens a fresh repo (triggers `fresh-repo-init` suggestion).
+- User touches security-sensitive paths (triggers `security-sweep` file_save).
+- User asks "what toolboxes do I have?" or "run the council".
+
+## Reasoning
+
+See `docs/roadmap/toolbox.md` for the full design rationale, open decisions,
+and phase rollout.

src/tests/test_update_repo_stats.py

@@ -44,22 +44,24 @@ def test_tarball_stats_only_trust_safe_regular_members(
         [
             ("./graphify-out/graph.json", {"nodes": [{}, {}], "edges": [{}, {}, {}]}),
             ("./graphify-out/communities.json", {"total_communities": 4}),
-            ("./entities/skills/good.md", b"# skill"),
-            ("entities/agents/good.md", b"# agent"),
-            ("entities/mcp-servers/a/good.md", b"# mcp"),
-            ("shadow/entities/skills/ignored.md", b"# ignored"),
-            ("entities/skills/../ignored.md", b"# ignored"),
-        ],
+            ("./entities/skills/good.md", b"# skill"),
+            ("entities/agents/good.md", b"# agent"),
+            ("entities/mcp-servers/a/good.md", b"# mcp"),
+            ("entities/harnesses/good.md", b"# harness"),
+            ("shadow/entities/skills/ignored.md", b"# ignored"),
+            ("entities/skills/../ignored.md", b"# ignored"),
+        ],
     )
 
     assert urs._read_graph_from_tarball() == {
         "nodes": 2,
         "edges": 3,
-        "skills": 1,
-        "agents": 1,
-        "mcps": 1,
-        "communities": 4,
-    }
+        "skills": 1,
+        "agents": 1,
+        "mcps": 1,
+        "harnesses": 1,
+        "communities": 4,
+    }
 
 
 def test_tarball_stats_reject_suffix_impersonation(
@@ -128,6 +130,7 @@ def test_tarball_stats_uses_report_when_graph_json_is_large(
             ("entities/skills/good.md", b"# skill"),
             ("entities/agents/good.md", b"# agent"),
             ("entities/mcp-servers/a/good.md", b"# mcp"),
+            ("entities/harnesses/good.md", b"# harness"),
         ],
     )
 
@@ -137,6 +140,7 @@ def test_tarball_stats_uses_report_when_graph_json_is_large(
         "skills": 1,
         "agents": 1,
         "mcps": 1,
+        "harnesses": 1,
         "communities": 50,
     }
 
@@ -149,6 +153,7 @@ def test_test_badge_is_labeled_collected_not_passing() -> None:
         "skills": None,
         "agents": None,
         "mcps": None,
+        "harnesses": None,
         "communities": None,
     }
     patched = text
@@ -159,6 +164,27 @@ def test_test_badge_is_labeled_collected_not_passing() -> None:
     assert "_passing" not in patched
 
 
+def test_harness_aware_readme_prose_is_updated() -> None:
+    text = (
+        "walks a **1,000 skills, 20 agents, 30 MCP servers, "
+        "and 4 cataloged harnesses** graph"
+    )
+    stats = {
+        "nodes": None,
+        "edges": None,
+        "skills": 92815,
+        "agents": 464,
+        "mcps": 10787,
+        "harnesses": 13,
+        "communities": None,
+    }
+    patched = text
+    for pattern, replacement in urs.build_replacements(stats, tests=None, converted=None):
+        patched = pattern.sub(replacement, patched)
+
+    assert "**92,815 skills, 464 agents, 10,787 MCP servers, and 13 cataloged harnesses**" in patched
+
+
 def test_read_test_count_prefers_project_python(
     monkeypatch: pytest.MonkeyPatch,
 ) -> None:

src/update_repo_stats.py

@@ -128,14 +128,15 @@ def _read_graph_from_tarball_legacy() -> dict[str, int | None] | None:
         return None
     stats: dict[str, int | None] = {
         "nodes": None, "edges": None,
-        "skills": None, "agents": None, "mcps": None, "communities": None,
+        "skills": None, "agents": None, "mcps": None, "harnesses": None,
+        "communities": None,
     }
     try:
         with tarfile.open(tarball, "r:gz") as tf:
             # Count entity pages directly from the archive index.
             # MCP entities are sharded by first char (entities/mcp-servers/<shard>/)
             # so we match the whole subtree, not just one level.
-            s = a = m = 0
+            s = a = m = h = 0
             for member in tf.getmembers():
                 name = _safe_tar_name(member.name)
                 if name is None or not member.isfile() or not name.endswith(".md"):
@@ -144,9 +145,11 @@ def _read_graph_from_tarball_legacy() -> dict[str, int | None] | None:
                     s += 1
                 elif name.startswith("entities/agents/"):
                     a += 1
-                elif name.startswith("entities/mcp-servers/"):
-                    m += 1
-            stats["skills"], stats["agents"], stats["mcps"] = s, a, m
+                elif name.startswith("entities/mcp-servers/"):
+                    m += 1
+                elif name.startswith("entities/harnesses/"):
+                    h += 1
+            stats["skills"], stats["agents"], stats["mcps"], stats["harnesses"] = s, a, m, h
             # Graph + communities are smaller files — extract to read.
             for path in (_GRAPH_JSON_MEMBER, _COMMUNITIES_JSON_MEMBER):
                 body = _read_json_member(tf, path)
@@ -183,11 +186,12 @@ def _read_graph_from_tarball() -> dict[str, int | None] | None:
         return None
     stats: dict[str, int | None] = {
         "nodes": None, "edges": None,
-        "skills": None, "agents": None, "mcps": None, "communities": None,
+        "skills": None, "agents": None, "mcps": None, "harnesses": None,
+        "communities": None,
     }
     try:
         with tarfile.open(tarball, "r:gz") as tf:
-            s = a = m = 0
+            s = a = m = h = 0
             for member in tf.getmembers():
                 name = _safe_tar_name(member.name)
                 if name is None or not member.isfile() or not name.endswith(".md"):
@@ -198,7 +202,9 @@ def _read_graph_from_tarball() -> dict[str, int | None] | None:
                     a += 1
                 elif name.startswith("entities/mcp-servers/"):
                     m += 1
-            stats["skills"], stats["agents"], stats["mcps"] = s, a, m
+                elif name.startswith("entities/harnesses/"):
+                    h += 1
+            stats["skills"], stats["agents"], stats["mcps"], stats["harnesses"] = s, a, m, h
 
             report = _read_text_member(tf, _GRAPH_REPORT_MEMBER)
             if report is not None:
@@ -258,11 +264,12 @@ def read_graph_stats() -> dict:
     stats: dict[str, int | None] = {
         "nodes": None,
         "edges": None,
-        "skills": None,
-        "agents": None,
-        "mcps": None,
-        "communities": None,
-    }
+        "skills": None,
+        "agents": None,
+        "mcps": None,
+        "harnesses": None,
+        "communities": None,
+    }
 
     if graph_json.exists():
         g = json.loads(graph_json.read_text(encoding="utf-8"))
@@ -274,9 +281,10 @@ def read_graph_stats() -> dict:
         for n in g.get("nodes", []):
             t = n.get("type", "?")
             type_counts[t] = type_counts.get(t, 0) + 1
-        stats["skills"] = type_counts.get("skill")
-        stats["agents"] = type_counts.get("agent")
-        stats["mcps"] = type_counts.get("mcp-server")
+        stats["skills"] = type_counts.get("skill")
+        stats["agents"] = type_counts.get("agent")
+        stats["mcps"] = type_counts.get("mcp-server")
+        stats["harnesses"] = type_counts.get("harness")
 
     if communities_repo.exists():
         c = json.loads(communities_repo.read_text(encoding="utf-8"))
@@ -401,7 +409,21 @@ def build_replacements(stats: dict, tests: int | None, converted: int | None) ->
     if stats["skills"]:
         s = stats["skills"]
         reps.append((re.compile(r"badge/Skills-[0-9%,]+-"), f"badge/Skills-{s:,}-".replace(",", "%2C")))
-        # 3-type pattern: "1,789 skills, 464 agents, and 10,786 MCP servers"
+        # 4-type pattern: "92,815 skills, 464 agents, 10,787 MCP servers,
+        # and 13 cataloged harnesses". Keep this before the 3-type fallback
+        # so the README's harness-aware lead sentence stays machine-owned.
+        if stats["agents"] and stats["mcps"] and stats["harnesses"]:
+            reps.append((
+                re.compile(
+                    r"\*\*[\d,]+\s+skills,\s+[\d,]+\s+agents,\s+"
+                    r"[\d,]+\s+MCP\s+servers,\s+and\s+[\d,]+\s+"
+                    r"cataloged\s+harnesses\*\*"
+                ),
+                f"**{s:,} skills, {stats['agents']:,} agents, "
+                f"{stats['mcps']:,} MCP servers, and "
+                f"{stats['harnesses']:,} cataloged harnesses**",
+            ))
+        # 3-type pattern: "1,789 skills, 464 agents, and 10,786 MCP servers"
         # Order matters — this regex is more specific than the 2-type one
         # below, so match it first. Handles the MCP-aware tagline that
         # lands in the README after the Phase 7 MCP-first rewrite.