@a5c-ai/babysitter-github 5.0.1-staging.ddee3ed9 → 5.0.1-staging.e195a4980892

package/README.md

@@ -18,7 +18,8 @@ directly.
 - **Node.js 22+**
 - **GitHub Copilot CLI** (`copilot`) -- requires an active GitHub Copilot
   subscription
-- **Babysitter SDK CLI** (`@a5c-ai/babysitter-sdk`) -- installed globally
+- **Babysitter CLI** (`@a5c-ai/babysitter`) -- installed globally when using
+  the SDK helper
 
 ## Installation
 
@@ -28,7 +29,7 @@ Register the a5c.ai marketplace and install the plugin:
 
 ```bash
 # Register the marketplace
-copilot plugin marketplace add a5c-ai/babysitter
+copilot plugin marketplace add a5c-ai/babysitter-claude
 
 # Install the plugin
 copilot plugin install babysitter
@@ -36,42 +37,50 @@ copilot plugin install babysitter
 
 ### Direct GitHub install
 
-Install directly from the Git repository using Copilot CLI. Copilot CLI
-discovers the plugin via `.github/plugin/marketplace.json` at the repo root:
+Install directly from the generated external plugin repository:
 
 ```bash
-copilot plugin install a5c-ai/babysitter
+copilot plugin install a5c-ai/babysitter-github-copilot
 ```
 
-### Alternative Installation (npm / development)
+### Alternative Installation (SDK helper / development)
 
-For development or environments where the Copilot CLI plugin system is not
-available, install via npm:
+For development, automation, or environments where the Copilot CLI plugin system is not available, install through the Babysitter SDK helper. This is the canonical scriptable path used by the installer tests and resolves to `npx --yes @a5c-ai/babysitter-github install ...` under the hood:
 
-Install the SDK CLI first:
+Install the Babysitter CLI first:
 
 ```bash
-npm install -g @a5c-ai/babysitter-sdk
+npm install -g @a5c-ai/babysitter
 ```
 
-Then install the GitHub Copilot plugin globally:
+Then install the GitHub Copilot plugin globally or into a workspace:
 
 ```bash
-npm install -g @a5c-ai/babysitter-github
-babysitter-github install
+# Global install
+babysitter harness:install-plugin github-copilot
+
+# Workspace install
+babysitter harness:install-plugin github-copilot --workspace /path/to/repo
 ```
 
-Or install from source:
+You can also run the published package installer directly:
 
 ```bash
-cd plugins/babysitter-github
-node bin/install.js
+npx --yes @a5c-ai/babysitter-github install --global
+npx --yes @a5c-ai/babysitter-github install --workspace /path/to/repo
+```
+
+Or install from generated source:
+
+```bash
+npm run generate:plugins
+node artifacts/generated-plugins/github-copilot/bin/install.js
 ```
 
-Install into a specific workspace:
+Install into a specific workspace from generated source:
 
 ```bash
-babysitter-github install --workspace /path/to/repo
+node artifacts/generated-plugins/github-copilot/bin/install.js --workspace /path/to/repo
 ```
 
 ### GitHub Copilot cloud agent installation
@@ -253,8 +262,7 @@ Copilot CLI looks for the plugin manifest in these paths, checked in order:
 The first match wins. This plugin uses `plugin.json` at the package root.
 
 For marketplace discovery, Copilot CLI looks for `.github/plugin/marketplace.json`
-at the repository root. This file lists all available plugins in the repo and is
-used when installing via `copilot plugin install OWNER/REPO`.
+at the repository root of the generated plugin repository.
 
 ### plugin.json Schema
 
@@ -404,7 +412,7 @@ repository root in `.github/plugin/marketplace.json`:
       "name": "babysitter",
       "description": "Multi-step workflow orchestration with event-sourced state",
       "version": "0.1.0",
-      "source": "./plugins/babysitter-github"
+      "source": "./"
     }
   ]
 }
@@ -461,7 +469,7 @@ These registries are available without running `marketplace add`.
 ## Plugin Structure (Directory Layout)
 
 ```
-plugins/babysitter-github/
+artifacts/generated-plugins/github-copilot/
   plugin.json              # Plugin manifest (skills, hooks, metadata)
   .github/plugin.json      # Plugin manifest (alternate discovery path)
   hooks.json               # Hook configuration (sessionStart, sessionEnd, userPromptSubmitted)
@@ -579,14 +587,14 @@ compatibility where PowerShell execution is available.
 git clone https://github.com/a5c-ai/babysitter.git
 cd babysitter
 npm install
-cd plugins/babysitter-github
+npm run generate:plugins
 node bin/install.js
 ```
 
 ### Publishing
 
 ```bash
-cd plugins/babysitter-github
+cd artifacts/generated-plugins/github-copilot
 npm run deploy            # Publish to npm (public)
 npm run deploy:staging    # Publish to npm with staging tag
 ```
@@ -594,7 +602,7 @@ npm run deploy:staging    # Publish to npm with staging tag
 ### Team installation
 
 ```bash
-cd plugins/babysitter-github
+cd artifacts/generated-plugins/github-copilot
 npm run team:install
 ```
 

package/bin/install-shared.js

@@ -29,7 +29,7 @@ function writeFileIfChanged(filePath, contents) {
   try {
     const existing = fs.readFileSync(filePath, 'utf8');
     if (existing === contents) return false;
-  } catch {}
+  } catch (e) { process.stderr.write('[extension-mux] file read failed for ' + filePath + ', overwriting: ' + (e instanceof Error ? e.message : String(e)) + '\n'); }
   fs.mkdirSync(path.dirname(filePath), { recursive: true });
   fs.writeFileSync(filePath, contents);
   return true;
@@ -82,7 +82,7 @@ function writeJson(filePath, value) {
 function ensureExecutable(filePath) {
   try {
     fs.chmodSync(filePath, 0o755);
-  } catch {}
+  } catch (e) { process.stderr.write('[extension-mux] chmod failed for ' + filePath + ': ' + (e instanceof Error ? e.message : String(e)) + '\n'); }
 }
 
 function normalizeMarketplaceSourcePath(source, marketplacePath) {
@@ -104,7 +104,7 @@ function ensureMarketplaceEntry(marketplacePath, pluginRoot) {
     name: PLUGIN_NAME,
     source: relSource,
     description: "Orchestrate complex, multi-step workflows with event-sourced state management, hook-based extensibility, and human-in-the-loop approval",
-    version: "5.0.0",
+    version: "5.0.1-staging.e195a4980892",
     author: { name: "a5c.ai" },
   };
   if (idx >= 0) marketplace.plugins[idx] = entry;
@@ -147,7 +147,7 @@ function resolveCliCommand(packageRoot) {
   const versionsPath = path.join(packageRoot, 'versions.json');
   const versions = readJson(versionsPath) || {};
   const ver = versions.sdkVersion || 'latest';
-  return `npx -y @a5c-ai/babysitter-sdk@${ver}`;
+  return `npm exec --yes --package @a5c-ai/babysitter-sdk@${ver} -- babysitter`;
 }
 
 function runCli(packageRoot, cliArgs, options = {}) {

package/commands/call.md

@@ -4,4 +4,8 @@ argument-hint: Specific instructions for the run.
 allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSearch, Search, AskUserQuestion, TodoWrite, TodoRead, Skill, BashOutput, KillShell, MultiEdit, LS
 ---
 
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md).
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Then continue executing the returned instructions in this same turn. Do not stop after the Skill tool returns; carry the requested run through to completion proof.
+
+User arguments for this command:
+
+$ARGUMENTS

package/commands/check-forbidden-markers.md

@@ -0,0 +1,68 @@
+---
+description: Pre-deploy gate that scans built JS chunks for forbidden substring markers (saga-era / obsolete code paths) listed in a project-local forbidden-markers.txt
+argument-hint: "[--markers-file <path>] [--chunks-dir <path>] [--json] Optional overrides; defaults are project-relative."
+allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSearch, Search, AskUserQuestion, TodoWrite, TodoRead, Skill, BashOutput, KillShell, MultiEdit, LS
+---
+
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Compose the gate from the shared helper at `library/processes/shared/forbidden-markers-scanner.js` (issue #477).
+
+## What this gate does
+
+Reads a list of literal substring markers from `scripts/forbidden-markers.txt` (blank lines and `#`-prefixed comments stripped) and greps every `.js` chunk under `.vercel/output/static/_next/static/chunks/` (Next.js / Vercel default; configurable) for any occurrence. Reports structured hits per `(marker, chunk)` pair with occurrence counts. Designed to chain between `vercel build --prod` and `vercel deploy --prod`.
+
+Use this gate when a refactor or restart-from-baseline replaced load-bearing code paths and you need a structural guarantee the obsolete symbols never re-ship. Burned-in evidence: cookbook VI-9 / VI-12 near-miss revivals during the 2026-05 iOS-Safari saga; the prototype lives at `cookbook/scripts/check-no-forbidden.mjs` and shipped two upstream contributions before being generalized as this gate.
+
+## When to use
+
+- **Pre-deploy.** Insert after build, before deploy. Block the deploy when `ok: false`.
+- **Post-restart.** After a baseline rollback + step-by-step re-add, snapshot the saga-era markers in `forbidden-markers.txt` and let CI hold the line.
+- **Post-refactor.** When old helper / handler / module names must not coexist with the new ones in the same bundle.
+
+## Expected config locations
+
+- `scripts/forbidden-markers.txt` — one marker per line, `#` for comments. The list is the contract; the gate is mechanical. Commit this file to source control.
+- `.vercel/output/static/_next/static/chunks/` — default scan target. Override for non-Vercel frameworks via the `--chunks-dir` flag or the `chunksDir` task input.
+
+A missing markers file is a no-op (`ok: true`, `reason: 'missing-markers-file'`) — misconfiguration is never a deploy block. A missing chunks directory is likewise a no-op (`reason: 'missing-chunks-dir'`) so the gate is safe to chain into `check:all` before the build runs.
+
+## Exit semantics
+
+| Reason                  | `ok`   | Deploy decision                |
+|-------------------------|--------|--------------------------------|
+| `missing-markers-file`  | true   | Pass (no gate active)          |
+| `missing-chunks-dir`    | true   | Pass (run before build)        |
+| `empty-markers`         | true   | Pass (list is empty)           |
+| `no-chunks`             | true   | Pass (nothing to scan)         |
+| `clean`                 | true   | Pass — proceed to deploy       |
+| `hits`                  | false  | **BLOCK** — surface hits, ask for triage |
+
+For each hit, the gate emits `{ marker, chunk, count }` so the operator sees the exact marker string, the absolute chunk path, and the number of occurrences in that chunk. Multiple hits across chunks for the same marker are reported separately.
+
+## Programmatic surface
+
+```js
+import { scanForbiddenMarkers, checkForbiddenMarkersTask } from '@a5c-ai/babysitter-library/processes/shared';
+
+// Direct call:
+const result = await scanForbiddenMarkers({
+  markersFile: 'scripts/forbidden-markers.txt',
+  chunksDir:   '.vercel/output/static/_next/static/chunks',
+});
+if (!result.ok) {
+  // result.hits: Array<{ marker, chunk, count }>
+  // result.reason === 'hits'
+  process.exit(1);
+}
+
+// Or dispatched as a babysitter task:
+const gate = await ctx.task(checkForbiddenMarkersTask, {
+  projectDir: '.',
+  // markersFile / chunksDir are inferred from projectDir if omitted
+});
+```
+
+## Reference
+
+- Issue: https://github.com/a5c-ai/babysitter/issues/477
+- Helper module: `library/processes/shared/forbidden-markers-scanner.js`
+- Origin (cookbook prototype): `cookbook/scripts/check-no-forbidden.mjs` (81 lines)

package/commands/cleanup.md

@@ -6,15 +6,43 @@ allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSea
 
 Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md).
 
-Create and run a cleanup process using the process at `skills\babysit\process\cradle\cleanup-runs.js/processes/cleanup-runs.js`.
+Resolve the active process library with:
+
+```bash
+babysitter process-library:active --json
+```
+
+Read `binding.dir` from that JSON and create/run the cleanup process from `cradle/cleanup-runs.js#process` relative to that active library root. Do not use plugin-cache-relative cradle paths.
 
 Implementation notes (for the process):
 - Parse arguments for `--dry-run` flag (if present, set dryRun: true in inputs) and `--keep-days N` (default: 7)
-- The process scans .a5c/runs/ for completed/failed runs, aggregates insights, writes summaries, then removes old data
-- Always show the user what will be removed before removing (in interactive mode via breakpoints)
-- In non-interactive mode (yolo), proceed with cleanup using defaults
-- The insights file goes to docs/run-history-insights.md
-- Only remove terminal runs (completed/failed) older than the keep-days threshold
-- Never remove active/in-progress runs
-- Remove orphaned process files not referenced by remaining runs
-- After cleanup, show remaining run count and disk usage
+
+CRITICAL: The cleanup MUST follow this exact phase order. Do NOT delete any run before Phase 2 completes.
+
+Phase 1 — Scan:
+- Scan .a5c/runs/ for all runs
+- Classify each as terminal (completed/failed) or active (in-progress/created)
+- Identify terminal runs older than the keep-days threshold as removal candidates
+- Never mark active/in-progress runs for removal
+- Count and report: total runs, terminal, active, removal candidates, disk usage
+
+Phase 2 — Aggregate insights (BEFORE any deletion):
+- For EVERY removal candidate, read its run.json and journal/ events
+- Extract: processId, prompt, status, event count, created date, task summaries
+- Group by process type and extract patterns (retry counts, convergence behavior, failure modes)
+- Append a new dated section to docs/run-history-insights.md with:
+  - Summary statistics (runs removed, disk freed, runs retained)
+  - Run categories with counts and descriptions
+  - Key patterns observed (multi-batch convergence, retry behavior, etc.)
+  - What worked well / what didn't from the run data
+- This file MUST be written and verified before proceeding to Phase 3
+
+Phase 3 — Confirm removal:
+- In interactive mode, show the user what will be removed via a breakpoint
+- In non-interactive mode (yolo), proceed with defaults
+- In dry-run mode, stop here and show what would be removed
+
+Phase 4 — Remove:
+- Delete the terminal runs older than keep-days threshold
+- Identify and remove orphaned process files not referenced by remaining runs
+- Show remaining run count and disk usage after cleanup

package/commands/doctor.md

@@ -156,7 +156,6 @@ If it exists:
 **Goal:** Inspect babysitter session files for health and detect runaway loops.
 
 - Search for session state files using Glob:
-  - `plugins/babysitter/skills/babysit/state/*.md`
   - `.a5c/state/*.md`
   - `.a5c/state/*.json`
 - For each session state file found:
@@ -260,7 +259,7 @@ Mark as PASS if total size < 500MB and no files > 10MB. Mark as WARN if total si
 
 ### 10a. Hook Registration
 
-- Locate the plugin root. Check for `CLAUDE_PLUGIN_ROOT` env var, or search for `plugins/babysitter/hooks/hooks.json` by walking up from the current directory.
+- Locate the plugin root. Check for `CLAUDE_PLUGIN_ROOT` env var first, or search for a babysitter `hooks.json` by walking up from the current directory.
 - If found, read `hooks.json` and verify:
   - A `Stop` hook entry exists with a command referencing `babysitter-stop-hook.sh`.
   - A `SessionStart` hook entry exists with a command referencing `babysitter-session-start-hook.sh`.
@@ -315,7 +314,7 @@ If the stop hook shows NO evidence of execution (no log entries, no journal even
 
 Perform these diagnostic steps in order and report the first failure found:
 
-1. **Plugin not installed**: Check if `plugins/babysitter/` exists relative to the project root and if `CLAUDE_PLUGIN_ROOT` is set. If the plugin directory doesn't exist, report: "Plugin not installed — the babysitter plugin directory is missing."
+1. **Plugin not installed**: Check if `CLAUDE_PLUGIN_ROOT` is set or if a babysitter plugin directory exists relative to the project root. If neither exists, report: "Plugin not installed — the babysitter plugin directory is missing."
 
 2. **Plugin not enabled**: Check for Claude settings files:
    - `~/.claude/settings.json` — look for `babysitter` in `enabledPlugins`.

package/commands/help.md

@@ -233,7 +233,8 @@ SECONDARY COMMANDS
   How it works: Runs npx @a5c-ai/babysitter-observer-dashboard@latest which watches
   the .a5c/runs/ directory (or a parent directory containing multiple projects) and
   serves a live dashboard. The process is blocking -- it runs until you stop it, and
-  it prints the local URL to share with the user.
+  it prints the local URL to share with the user. Do not use `babysitter observe`
+  as a fallback; the core Babysitter CLI does not expose that subcommand.
 
   Example: /babysitter:observe
   (opens browser showing all runs with live-updating task

package/commands/observe.md

@@ -7,6 +7,11 @@ allowed-tools: Read, Grep, Write, Task, Bash
 Run the babysitter observer dashboard:
 
 1. Determine the watch directory — this is usually the project's container directory (the parent of the project dir), or the current working directory if not specified.
-2. Launch the dashboard: `npx -y @a5c-ai/babysitter-observer-dashboard@latest --watch-dir <dir>`
+2. Launch the standalone dashboard package: `npx -y @a5c-ai/babysitter-observer-dashboard@latest --watch-dir <dir>`.
 3. This is a blocking process — it will keep running until stopped.
 4. Report the URL printed by the dashboard to the user, then open it in the browser.
+
+Do not fall back to `babysitter observe`; the core Babysitter CLI does not expose
+that subcommand. Some harness runtimes may provide a separate
+`agent-platform observe` surface, but this skill uses the verified standalone
+dashboard package.

package/commands/plan.md

@@ -4,4 +4,14 @@ argument-hint: Specific instructions for the run.
 allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSearch, Search, AskUserQuestion, TodoWrite, TodoRead, Skill, BashOutput, KillShell, MultiEdit, LS
 ---
 
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). focus on creating the best process possible, but without creating and running the actual run.
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Focus on creating the best process possible, but without creating and running the actual run.
+
+Before drafting the process, run Phase 0 -- REUSE-AUDIT: extract keyword nouns and verbs from the request, scan for matching existing migrations, API routes, environment variables, SDK dependencies, and imports, honor `.a5c/reuse-audit.json` when present, and put a `Reuse-audit findings (REVIEW BEFORE PROCEEDING)` block before Phase 1 of the plan.
+
+## Process Shape Selection
+
+Choose the process shape before authoring `process.js`:
+
+- Use a flat phase list when the spec is well-defined, the work is wiring or composition, the bug class is already known if this is a fix, and execution should proceed sequentially through clear phases.
+- Use a HYPOTHESES tree when the bug class is unknown, forensics are required, multiple causal models compete, and each hypothesis needs its own observations, falsifying observations, and follow-up phases.
+- Rule of thumb: if the first phase is "investigate", use HYPOTHESES-tree mode. If the first phase is "implement X", use flat-phase-list mode.

package/commands/yolo.md

@@ -1,7 +1,11 @@
----
-description: Orchestrate a babysitter run. use this command to start babysitting a complex workflow in a non-interactive mode, without any user interaction or breakpoints in the run.
-argument-hint: Specific instructions for the run.
-allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSearch, Search, AskUserQuestion, TodoWrite, TodoRead, Skill, BashOutput, KillShell, MultiEdit, LS
----
-
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). but without any user interaction or breakpoints in the run.
+---
+description: Orchestrate a babysitter run. use this command to start babysitting a complex workflow in a non-interactive mode, without any user interaction or breakpoints in the run.
+argument-hint: Specific instructions for the run.
+allowed-tools: Read, Grep, Write, Task, Bash, Edit, Grep, Glob, WebFetch, WebSearch, Search, AskUserQuestion, TodoWrite, TodoRead, Skill, BashOutput, KillShell, MultiEdit, LS
+---
+
+Run the Babysitter orchestration instructions directly through the CLI, without any user interaction or breakpoints. In Claude Code, use Bash to run `babysitter instructions:babysit-skill --harness claude-code --no-interactive`; in Codex, run `babysitter instructions:babysit-skill --harness codex --no-interactive`; in other harnesses, use the same command with that harness id. Then follow the returned instructions in this same turn until completion proof is produced. Do not stop after reading the instructions, do not invoke the Skill tool first, and use the non-interactive/no-breakpoints path when the instructions offer a mode choice.
+
+User arguments for this command:
+
+$ARGUMENTS

package/hooks/babysitter-proxied-post-tool-use.ps1

@@ -0,0 +1,12 @@
+# PowerShell hook wrapper — sets env vars and delegates to bash
+$env:HOOK_TYPE = 'post-tool-use'
+$env:ADAPTER_NAME = 'copilot'
+$env:PLUGIN_ROOT = Split-Path -Parent (Split-Path -Parent $PSScriptRoot)
+
+$input_data = [Console]::In.ReadToEnd()
+$result = $input_data | & bash "$PSScriptRoot/../$($MyInvocation.MyCommand.Name -replace '\.ps1$','.sh')" 2>$null
+if ($LASTEXITCODE -eq 0 -and $result) {
+  Write-Output $result
+} else {
+  Write-Output '{}'
+}

package/hooks/babysitter-proxied-post-tool-use.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+set -euo pipefail
+babysitter hook:run --harness unified --hook-type post-tool-use --json

package/hooks/babysitter-proxied-pre-compact.ps1

@@ -0,0 +1,12 @@
+# PowerShell hook wrapper — sets env vars and delegates to bash
+$env:HOOK_TYPE = 'pre-compact'
+$env:ADAPTER_NAME = 'copilot'
+$env:PLUGIN_ROOT = Split-Path -Parent (Split-Path -Parent $PSScriptRoot)
+
+$input_data = [Console]::In.ReadToEnd()
+$result = $input_data | & bash "$PSScriptRoot/../$($MyInvocation.MyCommand.Name -replace '\.ps1$','.sh')" 2>$null
+if ($LASTEXITCODE -eq 0 -and $result) {
+  Write-Output $result
+} else {
+  Write-Output '{}'
+}

package/hooks/babysitter-proxied-pre-compact.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+set -euo pipefail
+babysitter hook:run --harness unified --hook-type pre-compact --json

package/hooks/babysitter-proxied-pre-tool-use.ps1

@@ -0,0 +1,12 @@
+# PowerShell hook wrapper — sets env vars and delegates to bash
+$env:HOOK_TYPE = 'pre-tool-use'
+$env:ADAPTER_NAME = 'copilot'
+$env:PLUGIN_ROOT = Split-Path -Parent (Split-Path -Parent $PSScriptRoot)
+
+$input_data = [Console]::In.ReadToEnd()
+$result = $input_data | & bash "$PSScriptRoot/../$($MyInvocation.MyCommand.Name -replace '\.ps1$','.sh')" 2>$null
+if ($LASTEXITCODE -eq 0 -and $result) {
+  Write-Output $result
+} else {
+  Write-Output '{}'
+}

package/hooks/babysitter-proxied-pre-tool-use.sh

@@ -0,0 +1,3 @@
+#!/bin/bash
+set -euo pipefail
+babysitter hook:run --harness unified --hook-type pre-tool-use --json

package/hooks.json

@@ -17,6 +17,30 @@
         "timeoutSec": 15
       }
     ],
+    "preToolUse": [
+      {
+        "type": "command",
+        "bash": "./hooks/babysitter-proxied-pre-tool-use.sh",
+        "powershell": "./hooks/babysitter-proxied-pre-tool-use.ps1",
+        "timeoutSec": 30
+      }
+    ],
+    "postToolUse": [
+      {
+        "type": "command",
+        "bash": "./hooks/babysitter-proxied-post-tool-use.sh",
+        "powershell": "./hooks/babysitter-proxied-post-tool-use.ps1",
+        "timeoutSec": 30
+      }
+    ],
+    "PreCompact": [
+      {
+        "type": "command",
+        "bash": "./hooks/babysitter-proxied-pre-compact.sh",
+        "powershell": "./hooks/babysitter-proxied-pre-compact.ps1",
+        "timeoutSec": 30
+      }
+    ],
     "sessionEnd": [
       {
         "type": "command",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@a5c-ai/babysitter-github",
-  "version": "5.0.1-staging.ddee3ed9",
+  "version": "5.0.1-staging.e195a4980892",
   "description": "Orchestrate complex, multi-step workflows with event-sourced state management, hook-based extensibility, and human-in-the-loop approval",
   "scripts": {
     "deploy": "npm publish --access public",
@@ -36,15 +36,14 @@
     "access": "public"
   },
   "dependencies": {
-    "@a5c-ai/babysitter-sdk": "5.0.1-staging.ddee3ed9"
+    "@a5c-ai/babysitter-sdk": "5.0.1-staging.e195a4980892"
   },
   "repository": {
     "type": "git",
-    "url": "git+https://github.com/a5c-ai/babysitter.git",
-    "directory": "plugins/babysitter-github"
+    "url": "git+https://github.com/a5c-ai/babysitter-github-copilot.git"
   },
-  "homepage": "https://github.com/a5c-ai/babysitter/tree/main/plugins/babysitter-github#readme",
+  "homepage": "https://github.com/a5c-ai/babysitter-github-copilot#readme",
   "bugs": {
-    "url": "https://github.com/a5c-ai/babysitter/issues"
+    "url": "https://github.com/a5c-ai/babysitter-github-copilot/issues"
   }
 }

package/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "babysitter",
-  "version": "5.0.1-staging.ddee3ed9",
+  "version": "5.0.1-staging.e195a4980892",
   "description": "Orchestrate complex, multi-step workflows with event-sourced state management, hook-based extensibility, and human-in-the-loop approval",
   "author": {
     "name": "a5c.ai"

package/scripts/create-release-tag.mjs

@@ -0,0 +1,18 @@
+#!/usr/bin/env node
+import { spawnSync } from 'node:child_process';
+import { existsSync, readFileSync } from 'node:fs';
+
+function run(command, args) {
+  const result = spawnSync(command, args, { encoding: 'utf8', stdio: 'inherit' });
+  if (result.status !== 0) process.exit(result.status || 1);
+}
+
+const branch = process.env.GITHUB_REF_NAME || 'develop';
+const sha = (process.env.GITHUB_SHA || '').slice(0, 12);
+const version = existsSync('package.json') ? JSON.parse(readFileSync('package.json', 'utf8')).version : JSON.parse(readFileSync('versions.json', 'utf8')).sdkVersion;
+const normalized = String(version).replace(/[^0-9A-Za-z._-]/g, '-');
+const tag = 'release/' + branch + '/v' + normalized + '-' + sha;
+run('git', ['config', 'user.name', 'github-actions[bot]']);
+run('git', ['config', 'user.email', 'github-actions[bot]@users.noreply.github.com']);
+run('git', ['tag', tag]);
+run('git', ['push', 'origin', tag]);

package/scripts/publish-from-tag.mjs

@@ -0,0 +1,41 @@
+#!/usr/bin/env node
+import { spawnSync } from 'node:child_process';
+import { readFileSync } from 'node:fs';
+
+function run(command, args, options = {}) {
+  const result = spawnSync(command, args, { stdio: options.stdio || 'inherit', encoding: options.encoding });
+  if (result.status !== 0 && !options.allowFailure) process.exit(result.status || 1);
+  return result;
+}
+
+function npmView(packageSpec) {
+  return run('npm', ['view', packageSpec, 'version'], { allowFailure: true, stdio: 'pipe', encoding: 'utf8' }).status === 0;
+}
+
+const pkg = JSON.parse(readFileSync('package.json', 'utf8'));
+const ref = process.env.GITHUB_REF_NAME || '';
+const branch = ref.split('/')[1] || 'develop';
+const tag = branch === 'main' ? 'latest' : branch;
+
+if (!process.env.NODE_AUTH_TOKEN) {
+  console.log('NODE_AUTH_TOKEN is not configured; skipping npm publish.');
+  process.exit(0);
+}
+
+if (npmView(pkg.name + '@' + pkg.version)) {
+  console.log(pkg.name + '@' + pkg.version + ' already exists; ensuring dist-tag ' + tag + '.');
+  run('npm', ['dist-tag', 'add', pkg.name + '@' + pkg.version, tag], { allowFailure: true });
+  process.exit(0);
+}
+
+for (const field of ['dependencies', 'peerDependencies', 'optionalDependencies']) {
+  for (const [name, version] of Object.entries(pkg[field] || {})) {
+    if (!name.startsWith('@a5c-ai/') || version.startsWith('^') || version.startsWith('~') || version === '*' || version.startsWith('workspace:')) continue;
+    if (!npmView(name + '@' + version)) {
+      console.log('Required internal dependency ' + name + '@' + version + ' is not published yet; skipping npm publish.');
+      process.exit(0);
+    }
+  }
+}
+
+run('npm', ['publish', '--access', 'public', '--tag', tag]);

package/skills/babysit/SKILL.md

@@ -1,14 +1,12 @@
 ---
 name: babysit
 description: >-
-  Orchestrate via @babysitter. Use this skill when asked to babysit a run,
-  orchestrate a process or whenever it is called explicitly. (babysit,
-  babysitter, orchestrate, orchestrate a run, workflow, etc.)
+  Execute via @babysitter. Use this skill when asked to babysit a task, do anything that is structured process-driven (even a loop) or whenever it is called explicitly. (babysit, babysitter, with a process, orchestrate, orchestrate a run, workflow, loop until, etc.)
 ---
 
 # babysit
 
-Orchestrate `.a5c/runs/<runId>/` through iterative execution.
+Follow the commands below exactly. but apply judgement and creativity to get the job done. no matter how small or simple the task may seem, you must use the babysitter SDK and run lifecycle to execute it. This is non-negotiable. If you are asked to do something that seems like it could be done with a simple tool call, you must still use the babysitter SDK to create a run, define tasks, execute them, and complete the run. This is how you will learn and demonstrate mastery of the babysitter system. Always follow the full process, even for trivial tasks.
 
 ## Dependencies
 
@@ -20,10 +18,14 @@ Read the SDK version from `versions.json` to ensure version compatibility:
 SDK_VERSION=$(node -e "try{console.log(JSON.parse(require('fs').readFileSync('${PLUGIN_ROOT}/versions.json','utf8')).sdkVersion||'latest')}catch{console.log('latest')}")
 npm i -g @a5c-ai/babysitter-sdk@$SDK_VERSION
 
-CLI="npx -y @a5c-ai/babysitter-sdk@$SDK_VERSION"
+if command -v babysitter >/dev/null 2>&1 && babysitter --version >/dev/null 2>&1; then
+  CLI="babysitter"
+else
+  CLI="npm exec --yes --package @a5c-ai/babysitter-sdk@$SDK_VERSION -- babysitter"
+fi
 ```
 
-If `babysitter` is already installed globally at the correct version, you may use `CLI="babysitter"` instead.
+If a stale or broken global shim fails with `MODULE_NOT_FOUND`, repair it with `npm rm -g @a5c-ai/babysitter @a5c-ai/babysitter-sdk && npm i -g @a5c-ai/babysitter-sdk@$SDK_VERSION`, then re-run `babysitter --version`.
 
 ### jq
 

package/skills/call/SKILL.md

@@ -5,4 +5,8 @@ description: Orchestrate a babysitter run. use this command to start babysitting
 
 # call
 
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md).
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Then continue executing the returned instructions in this same turn. Do not stop after the Skill tool returns; carry the requested run through to completion proof.
+
+User arguments for this command:
+
+$ARGUMENTS

package/skills/check-forbidden-markers/SKILL.md

@@ -0,0 +1,69 @@
+---
+name: check-forbidden-markers
+description: Pre-deploy gate that scans built JS chunks for forbidden substring markers (saga-era / obsolete code paths) listed in a project-local forbidden-markers.txt
+---
+
+# check-forbidden-markers
+
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Compose the gate from the shared helper at `library/processes/shared/forbidden-markers-scanner.js` (issue #477).
+
+## What this gate does
+
+Reads a list of literal substring markers from `scripts/forbidden-markers.txt` (blank lines and `#`-prefixed comments stripped) and greps every `.js` chunk under `.vercel/output/static/_next/static/chunks/` (Next.js / Vercel default; configurable) for any occurrence. Reports structured hits per `(marker, chunk)` pair with occurrence counts. Designed to chain between `vercel build --prod` and `vercel deploy --prod`.
+
+Use this gate when a refactor or restart-from-baseline replaced load-bearing code paths and you need a structural guarantee the obsolete symbols never re-ship. Burned-in evidence: cookbook VI-9 / VI-12 near-miss revivals during the 2026-05 iOS-Safari saga; the prototype lives at `cookbook/scripts/check-no-forbidden.mjs` and shipped two upstream contributions before being generalized as this gate.
+
+## When to use
+
+- **Pre-deploy.** Insert after build, before deploy. Block the deploy when `ok: false`.
+- **Post-restart.** After a baseline rollback + step-by-step re-add, snapshot the saga-era markers in `forbidden-markers.txt` and let CI hold the line.
+- **Post-refactor.** When old helper / handler / module names must not coexist with the new ones in the same bundle.
+
+## Expected config locations
+
+- `scripts/forbidden-markers.txt` — one marker per line, `#` for comments. The list is the contract; the gate is mechanical. Commit this file to source control.
+- `.vercel/output/static/_next/static/chunks/` — default scan target. Override for non-Vercel frameworks via the `--chunks-dir` flag or the `chunksDir` task input.
+
+A missing markers file is a no-op (`ok: true`, `reason: 'missing-markers-file'`) — misconfiguration is never a deploy block. A missing chunks directory is likewise a no-op (`reason: 'missing-chunks-dir'`) so the gate is safe to chain into `check:all` before the build runs.
+
+## Exit semantics
+
+| Reason                  | `ok`   | Deploy decision                |
+|-------------------------|--------|--------------------------------|
+| `missing-markers-file`  | true   | Pass (no gate active)          |
+| `missing-chunks-dir`    | true   | Pass (run before build)        |
+| `empty-markers`         | true   | Pass (list is empty)           |
+| `no-chunks`             | true   | Pass (nothing to scan)         |
+| `clean`                 | true   | Pass — proceed to deploy       |
+| `hits`                  | false  | **BLOCK** — surface hits, ask for triage |
+
+For each hit, the gate emits `{ marker, chunk, count }` so the operator sees the exact marker string, the absolute chunk path, and the number of occurrences in that chunk. Multiple hits across chunks for the same marker are reported separately.
+
+## Programmatic surface
+
+```js
+import { scanForbiddenMarkers, checkForbiddenMarkersTask } from '@a5c-ai/babysitter-library/processes/shared';
+
+// Direct call:
+const result = await scanForbiddenMarkers({
+  markersFile: 'scripts/forbidden-markers.txt',
+  chunksDir:   '.vercel/output/static/_next/static/chunks',
+});
+if (!result.ok) {
+  // result.hits: Array<{ marker, chunk, count }>
+  // result.reason === 'hits'
+  process.exit(1);
+}
+
+// Or dispatched as a babysitter task:
+const gate = await ctx.task(checkForbiddenMarkersTask, {
+  projectDir: '.',
+  // markersFile / chunksDir are inferred from projectDir if omitted
+});
+```
+
+## Reference
+
+- Issue: https://github.com/a5c-ai/babysitter/issues/477
+- Helper module: `library/processes/shared/forbidden-markers-scanner.js`
+- Origin (cookbook prototype): `cookbook/scripts/check-no-forbidden.mjs` (81 lines)

package/skills/cleanup/SKILL.md

@@ -7,15 +7,43 @@ description: Clean up .a5c/runs and .a5c/processes directories. Aggregates insig
 
 Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md).
 
-Create and run a cleanup process using the process at `skills\babysit\process\cradle\cleanup-runs.js/processes/cleanup-runs.js`.
+Resolve the active process library with:
+
+```bash
+babysitter process-library:active --json
+```
+
+Read `binding.dir` from that JSON and create/run the cleanup process from `cradle/cleanup-runs.js#process` relative to that active library root. Do not use plugin-cache-relative cradle paths.
 
 Implementation notes (for the process):
 - Parse arguments for `--dry-run` flag (if present, set dryRun: true in inputs) and `--keep-days N` (default: 7)
-- The process scans .a5c/runs/ for completed/failed runs, aggregates insights, writes summaries, then removes old data
-- Always show the user what will be removed before removing (in interactive mode via breakpoints)
-- In non-interactive mode (yolo), proceed with cleanup using defaults
-- The insights file goes to docs/run-history-insights.md
-- Only remove terminal runs (completed/failed) older than the keep-days threshold
-- Never remove active/in-progress runs
-- Remove orphaned process files not referenced by remaining runs
-- After cleanup, show remaining run count and disk usage
+
+CRITICAL: The cleanup MUST follow this exact phase order. Do NOT delete any run before Phase 2 completes.
+
+Phase 1 — Scan:
+- Scan .a5c/runs/ for all runs
+- Classify each as terminal (completed/failed) or active (in-progress/created)
+- Identify terminal runs older than the keep-days threshold as removal candidates
+- Never mark active/in-progress runs for removal
+- Count and report: total runs, terminal, active, removal candidates, disk usage
+
+Phase 2 — Aggregate insights (BEFORE any deletion):
+- For EVERY removal candidate, read its run.json and journal/ events
+- Extract: processId, prompt, status, event count, created date, task summaries
+- Group by process type and extract patterns (retry counts, convergence behavior, failure modes)
+- Append a new dated section to docs/run-history-insights.md with:
+  - Summary statistics (runs removed, disk freed, runs retained)
+  - Run categories with counts and descriptions
+  - Key patterns observed (multi-batch convergence, retry behavior, etc.)
+  - What worked well / what didn't from the run data
+- This file MUST be written and verified before proceeding to Phase 3
+
+Phase 3 — Confirm removal:
+- In interactive mode, show the user what will be removed via a breakpoint
+- In non-interactive mode (yolo), proceed with defaults
+- In dry-run mode, stop here and show what would be removed
+
+Phase 4 — Remove:
+- Delete the terminal runs older than keep-days threshold
+- Identify and remove orphaned process files not referenced by remaining runs
+- Show remaining run count and disk usage after cleanup

package/skills/doctor/SKILL.md

@@ -157,7 +157,6 @@ If it exists:
 **Goal:** Inspect babysitter session files for health and detect runaway loops.
 
 - Search for session state files using Glob:
-  - `plugins/babysitter/skills/babysit/state/*.md`
   - `.a5c/state/*.md`
   - `.a5c/state/*.json`
 - For each session state file found:
@@ -261,7 +260,7 @@ Mark as PASS if total size < 500MB and no files > 10MB. Mark as WARN if total si
 
 ### 10a. Hook Registration
 
-- Locate the plugin root. Check for `CLAUDE_PLUGIN_ROOT` env var, or search for `plugins/babysitter/hooks/hooks.json` by walking up from the current directory.
+- Locate the plugin root. Check for `CLAUDE_PLUGIN_ROOT` env var first, or search for a babysitter `hooks.json` by walking up from the current directory.
 - If found, read `hooks.json` and verify:
   - A `Stop` hook entry exists with a command referencing `babysitter-stop-hook.sh`.
   - A `SessionStart` hook entry exists with a command referencing `babysitter-session-start-hook.sh`.
@@ -316,7 +315,7 @@ If the stop hook shows NO evidence of execution (no log entries, no journal even
 
 Perform these diagnostic steps in order and report the first failure found:
 
-1. **Plugin not installed**: Check if `plugins/babysitter/` exists relative to the project root and if `CLAUDE_PLUGIN_ROOT` is set. If the plugin directory doesn't exist, report: "Plugin not installed — the babysitter plugin directory is missing."
+1. **Plugin not installed**: Check if `CLAUDE_PLUGIN_ROOT` is set or if a babysitter plugin directory exists relative to the project root. If neither exists, report: "Plugin not installed — the babysitter plugin directory is missing."
 
 2. **Plugin not enabled**: Check for Claude settings files:
    - `~/.claude/settings.json` — look for `babysitter` in `enabledPlugins`.

package/skills/help/SKILL.md

@@ -234,7 +234,8 @@ SECONDARY COMMANDS
   How it works: Runs npx @a5c-ai/babysitter-observer-dashboard@latest which watches
   the .a5c/runs/ directory (or a parent directory containing multiple projects) and
   serves a live dashboard. The process is blocking -- it runs until you stop it, and
-  it prints the local URL to share with the user.
+  it prints the local URL to share with the user. Do not use `babysitter observe`
+  as a fallback; the core Babysitter CLI does not expose that subcommand.
 
   Example: /babysitter:observe
   (opens browser showing all runs with live-updating task

package/skills/observe/SKILL.md

@@ -8,6 +8,11 @@ description: Launch the babysitter observer dashboard. Installs and runs the rea
 Run the babysitter observer dashboard:
 
 1. Determine the watch directory — this is usually the project's container directory (the parent of the project dir), or the current working directory if not specified.
-2. Launch the dashboard: `npx -y @a5c-ai/babysitter-observer-dashboard@latest --watch-dir <dir>`
+2. Launch the standalone dashboard package: `npx -y @a5c-ai/babysitter-observer-dashboard@latest --watch-dir <dir>`.
 3. This is a blocking process — it will keep running until stopped.
 4. Report the URL printed by the dashboard to the user, then open it in the browser.
+
+Do not fall back to `babysitter observe`; the core Babysitter CLI does not expose
+that subcommand. Some harness runtimes may provide a separate
+`agent-platform observe` surface, but this skill uses the verified standalone
+dashboard package.

package/skills/plan/SKILL.md

@@ -5,4 +5,14 @@ description: Plan a babysitter run. use this command to plan a complex workflow,
 
 # plan
 
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). focus on creating the best process possible, but without creating and running the actual run.
+Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). Focus on creating the best process possible, but without creating and running the actual run.
+
+Before drafting the process, run Phase 0 -- REUSE-AUDIT: extract keyword nouns and verbs from the request, scan for matching existing migrations, API routes, environment variables, SDK dependencies, and imports, honor `.a5c/reuse-audit.json` when present, and put a `Reuse-audit findings (REVIEW BEFORE PROCEEDING)` block before Phase 1 of the plan.
+
+## Process Shape Selection
+
+Choose the process shape before authoring `process.js`:
+
+- Use a flat phase list when the spec is well-defined, the work is wiring or composition, the bug class is already known if this is a fix, and execution should proceed sequentially through clear phases.
+- Use a HYPOTHESES tree when the bug class is unknown, forensics are required, multiple causal models compete, and each hypothesis needs its own observations, falsifying observations, and follow-up phases.
+- Rule of thumb: if the first phase is "investigate", use HYPOTHESES-tree mode. If the first phase is "implement X", use flat-phase-list mode.

package/skills/yolo/SKILL.md

@@ -5,4 +5,8 @@ description: Orchestrate a babysitter run. use this command to start babysitting
 
 # yolo
 
-Invoke the babysitter:babysit skill (using the Skill tool) and follow its instructions (SKILL.md). but without any user interaction or breakpoints in the run.
+Run the Babysitter orchestration instructions directly through the CLI, without any user interaction or breakpoints. In Claude Code, use Bash to run `babysitter instructions:babysit-skill --harness claude-code --no-interactive`; in Codex, run `babysitter instructions:babysit-skill --harness codex --no-interactive`; in other harnesses, use the same command with that harness id. Then follow the returned instructions in this same turn until completion proof is produced. Do not stop after reading the instructions, do not invoke the Skill tool first, and use the non-interactive/no-breakpoints path when the instructions offer a mode choice.
+
+User arguments for this command:
+
+$ARGUMENTS

package/versions.json

@@ -1,4 +1,4 @@
 {
-  "sdkVersion": "5.0.1-staging.ddee3ed9",
-  "extensionVersion": "5.0.0"
+  "sdkVersion": "5.0.1-staging.e195a4980892",
+  "extensionVersion": "5.0.1-staging.e195a4980892"
 }

package/scripts/sync-command-surfaces.js

@@ -1,20 +0,0 @@
-#!/usr/bin/env node
-'use strict';
-
-const path = require('path');
-const { spawnSync } = require('child_process');
-
-const REPO_ROOT = path.resolve(__dirname, '..', '..');
-const syncScript = path.join(REPO_ROOT, 'scripts', 'sync-plugin-commands.cjs');
-const args = [syncScript, '--target', 'github-copilot'];
-
-if (process.argv.includes('--check')) {
-  args.push('--check');
-}
-
-const result = spawnSync(process.execPath, args, {
-  cwd: REPO_ROOT,
-  stdio: 'inherit',
-});
-
-process.exit(result.status ?? 1);