# Agent Guidelines

Guidelines for AI agents working on this codebase.

## Project Overview

This is a Cloudflare Worker that runs [OpenClaw](https://github.com/openclaw/openclaw) (formerly Moltbot/Clawdbot) in a Cloudflare Sandbox container. It provides:

- Proxying to the OpenClaw gateway (web UI + WebSocket)
- Admin UI at `/_admin/` for device management
- API endpoints at `/api/*` for device pairing
- Debug endpoints at `/debug/*` for troubleshooting

**Note:** The CLI tool and npm package are now named `openclaw`. Config files use `.openclaw/openclaw.json`. Legacy `.clawdbot` paths are supported for backward compatibility during transition.

### Paso4 Extensions (`moltlazy/`)

The `moltlazy/` module is fully owned by the **Paso4 developers** and provides two systems:

1. **Config Module** (`config/` + `moltlazy/index.ts`) — filesystem patcher used by `start-openclaw.sh` at container startup to write `openclaw.json` before the gateway launches.
2. **Features Module** (`sdk/`) — RPC-based config applied to a **running** gateway via the Admin HTTP API. Supported feature packs: Knowledge Graph, CF Unified Billing.

## Project Structure

```
src/
├── index.ts          # Main Hono app, route mounting
├── types.ts          # TypeScript type definitions (AUTOGENERATED by `bun run types`, do not apply changes directly here, instead re-run the command after implementing the rest of the features)
├── config.ts         # Constants (ports, timeouts, paths)
├── auth/             # Cloudflare Access authentication
│   ├── jwt.ts        # JWT verification
│   ├── jwks.ts       # JWKS fetching and caching
│   └── middleware.ts # Hono middleware for auth
├── gateway/          # OpenClaw gateway management
│   ├── process.ts    # Process lifecycle (find, start)
│   ├── env.ts        # Environment variable building
│   ├── r2.ts         # R2 bucket mounting
│   ├── sync.ts       # R2 backup sync logic
│   └── utils.ts      # Shared utilities (waitForProcess)
├── routes/           # API route handlers
│   ├── api.ts        # /api/* endpoints (devices, gateway)
│   ├── admin.ts      # /_admin/* static file serving
│   └── debug.ts      # /debug/* endpoints
└── client/           # React admin UI (Vite)
    ├── App.tsx
    ├── api.ts        # API client
    └── pages/
moltlazy/             # Paso4 proprietary configuration utilities
├── cli.ts            # CLI entry point for patching config
├── index.ts          # Main logic for config patching
├── types.ts          # Type definitions for OpenClaw config
├── config/
│   └── include.ts     # per-section $include injection (session/tools/logging)
├── config/
├── config/           # Configuration templates
├── agents/           # Agent configuration templates
│   ├── defaults.ts   # Default agent settings
│   └── premade/      # Pre-configured agent definitions
├── models/           # Provider-specific model configurations
├── sdk/              # SDK Module — typed RPC client over the Admin HTTP API
│   ├── client.ts     # Typed RPC client
│   └── (agents, channels, config, cron, devices, models, gateway, update, …)
└── tests/            # Tests for configuration utilities
```

## Key Patterns

### Environment Variables

- `DEV_MODE` - Local-dev switch: skips CF Access auth AND bypasses device pairing
- `OPENCLAW_DEV_MODE` - Explicit override for the dev-agents flag. Normally derived by the Worker from the instance's own `-dev` hostname suffix (see `src/dev-mode.ts`), which enables the Fouria Builder/Q&A agents, client-doc materialization, and the opencode-credits automation **without** bypassing auth. Forwarded to the container as `OPENCLAW_DEV_MODE`; falls back to `DEV_MODE` for local development
- `DEBUG_ROUTES` - Enables `/debug/*` routes (disabled by default)
- See `src/types.ts` for full `MoltbotEnv` interface
- See [`docs/schemas/secrets-manifest.schema.json`](./docs/schemas/secrets-manifest.schema.json) for the secrets manifest

### CLI Commands

On OpenClaw 2026.8.2 the CLI **no longer accepts** `--url ws://localhost:18789`
(and most local commands reject `--token` too). Commands run inside the container
(the gateway host) either operate on local state or auto-connect to the local
Gateway — auth resolves from the `OPENCLAW_GATEWAY_TOKEN` env SecretRef, so no
connection flags are passed. Commands that require an owner on a multi-agent
roster take `--agent <id>`. Never hardcode the id: resolve it at runtime via
`openclaw agents list --json` (prefer `main` when present, else the first
configured agent — e.g. the dev-only `fouria-builder`/`fouria-qa` premades when
running with `DEV_MODE=true`; see
`resolveAgentFlag()` in `src/routes/api/shared.ts` and `resolveCliAgentId()`
in `moltlazy/sdk/cli.ts`):

```typescript
// Sandbox SDK 1.0: exec takes argv and resolves on launch, not completion.
startShell(sandbox, 'openclaw devices list --json');
// multi-agent commands need an explicit, dynamically-resolved owner:
startShell(sandbox, `openclaw models list --json${await resolveAgentFlag(sandbox)}`);
```

Gateway-connecting commands (`devices`, `cron`, `secrets reload`) may still accept
`--url`/`--token` when given together, but the 2026.8.2 guidance is to drop `--url`
and let the CLI use the configured local target.

CLI commands take 10-15 seconds due to WebSocket connection overhead. Use `waitForProcess()` helper in `src/routes/api.ts`.

### Success Detection

The CLI outputs "Approved" (capital A). Use case-insensitive checks:

```typescript
stdout.toLowerCase().includes('approved');
```

## Commands

```bash
bun run lint          # Run linting (oxlint)
bun run test          # Run tests (vitest)
bun run test:watch    # Run tests in watch mode
bun run test:client   # Run SPA component/page tests (jsdom, vitest.client.config.ts)
bun run test:ui       # Run the Playwright UI suite (Vite dev server + mocked API)
bun run test:ui:headed # Same, with a visible browser
bun run test:ui:render:local # Local-only plwr admin-UI component render timings
bun run test:gateway:boot     # Local gateway boot smoke test (Docker)
bun run test:gateway:boot:ai  # Same, with real AI Gateway onboard path (.dev.vars.e2e)
bun run build         # Build client (vite)
bun run build:all     # Build client + moltlazy
bun run build:docker  # Build all + install deps (pre-Docker build)
bun run deploy        # Build all and deploy to production
bun run deploy:dev    # Build all and deploy to dev
bun run dev           # Admin SPA Vite dev server (port 8787)
bun run dev:worker    # Worker dev server (Cloudflare Vite plugin, port 8788)
bun run start         # Build all + Worker dev server
bun run typecheck     # TypeScript check
```

## Contribution Guidelines

### Validation Requirements

Before submitting any contribution:

1. **Run lint and tests** - All builds must pass:

   ```bash
   bun run lint
   bun run test
   ```

2. **Environment variable changes** - If your contribution adds, updates, or deletes an environment variable, you MUST update it in all three locations:
   - `README.md` - Add to the secrets reference table
   - `src/gateway/env.ts` - Add to `buildEnvVars()` function if passed to container
   - `.dev.vars.example` - Add to the mapping comments and example

### Testing

Tests use Vitest. Test files are colocated with source files (`*.test.ts`).

Current test coverage:

- `auth/jwt.test.ts` - JWT decoding and validation
- `auth/jwks.test.ts` - JWKS fetching and caching
- `auth/middleware.test.ts` - Auth middleware behavior
- `gateway/env.test.ts` - Environment variable building
- `gateway/process.test.ts` - Process finding logic
- `gateway/r2.test.ts` - R2 mounting logic
- `gateway/sync.test.ts` - R2 backup sync logic
- `moltlazy/tests/index.test.ts` - per-section $include injection and config generation

When adding new functionality, add corresponding tests.

The admin SPA also has a Playwright UI suite (`test/ui/`, config
`playwright.config.ts`) run with `bun run test:ui`. It serves the real dashboard
via the Vite dev server (`bun run dev`) and mocks every `/api/*` call per test,
so it needs no Worker, container, or Cloudflare credentials. It covers the app
shell, role-based navigation/redirects, every admin page's render, and the
notification bell. On NixOS the browsers come from `PLAYWRIGHT_BROWSERS_PATH`
(see `flake.nix`); `playwright.config.ts` selects the newest `chromium-*` there
so the Nix bundle can lag the npm `playwright` version without breaking the run.

A separate **local-only** cctr+`plwr` suite (`test/e2e/ui_components_render.txt`,
`bun run test:ui:render:local`) validates the admin (root) UI components against
the local worker and records their render times (`openMs`/`componentMs`/`waitMs`
plus Navigation Timing/FCP) via `fixture/ui-render-measure`; it self-skips
outside `E2E_LOCAL=true`. See [docs/TESTING.md](/latest/llms/platform/reference/testing/index.md).

## Code Style

- Use TypeScript strict mode
- Prefer explicit types over inference for function signatures
- Keep route handlers thin - extract logic to separate modules
- Use Hono's context methods (`c.json()`, `c.html()`) for responses

## Documentation

- `README.md` - User-facing documentation (setup, configuration, usage)
- `AGENTS.md` - This file, for AI agents

Development documentation goes in AGENTS.md, not README.md.

---

## Architecture

```
Browser
   │
   ▼
┌─────────────────────────────────────┐
│     Cloudflare Worker (index.ts)    │
│  - Starts OpenClaw in sandbox       │
│  - Proxies HTTP/WebSocket requests  │
│  - Passes secrets as env vars       │
└──────────────┬──────────────────────┘
               │
               ▼
┌─────────────────────────────────────┐
│     Cloudflare Sandbox Container    │
│  ┌───────────────────────────────┐  │
│  │     OpenClaw Gateway          │  │
│  │  - Control UI on port 18789   │  │
│  │  - WebSocket RPC protocol     │  │
│  │  - Agent runtime              │  │
│  └───────────────────────────────┘  │
└─────────────────────────────────────┘
```

### Key Files

| File                | Purpose                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                 |
| ------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| `src/index.ts`      | Worker that manages sandbox lifecycle and proxies requests                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                              |
| `src/metrics.ts`    | Cost and usage metrics instrumentation for Analytics Engine                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |
| `Dockerfile`        | Container image based on `cloudflare/sandbox` with Node 22 + OpenClaw                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                   |
| `start-openclaw.sh` | Startup entry point: `start` is the production cold-start (acquire the container-wide startup lock **before** setup, hold fd 9 through the Gateway `exec` — a single owner covers setup **and** launch), `post-start` (dev automation + boot-log reset), `setup`/`launch` (standalone smoke/manual modes only), or `full` (legacy: setup + background gateway launch). The Worker drives production cold-start as one `start-openclaw.sh start` call then a **detached** `post-start` (never awaited, so it cannot block readiness); it no longer invokes `setup`/`launch` separately. `setup`/`post-start` must stay child-free (no `tail -f`/`cloudflared`) — the Worker `waitForExit`s on them. `start`/`setup`/`full` serialize the slow setup body on the exclusive `/tmp/openclaw-setup.lock`, and `start` holds it through the Gateway `exec` (no release before launch); the standalone `launch` takes the dedicated `/tmp/openclaw-gateway-launch.lock`, held for the gateway's whole lifetime (fd inherited by the `exec`'d gateway). Either way exactly one `openclaw gateway run` owner can exist — losers poll `/health` and exit 0 without launching a duplicate; the serial `test/e2e/ac_gateway_startup_concurrency.txt` asserts one owner and the absence of the migration-input-change and SQLite offline-maintenance errors. When the Worker passes `OPENCLAW_RESTORED=true` (a backup was restored), the setup body is skipped entirely and the restored state is launched directly. See [docs/WORKER-DEVELOPMENT.md](/latest/llms/platform/apps/fouria/worker-development/index.md) |
| `wrangler.jsonc`    | Cloudflare Worker + Container configuration                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                                             |

## Cost Observability

The Worker instruments cost metrics using Cloudflare Analytics Engine for per-client cost tracking:

| Event Type         | Dimensions                            | Metrics                        |
| ------------------ | ------------------------------------- | ------------------------------ |
| `ai-query`         | model, provider, customer, actionType | tokensIn, tokensOut, costUsd   |
| `container-active` | customer, instanceType, action        | heartbeat count                |
| `r2-operation`     | customer, opType, opClass             | bytesTransferred, durationMs   |
| `cron-trigger`     | customer, triggerReason               | wokeContainer flag             |
| `ws-session`       | customer, sessionId                   | durationSeconds, messagesCount |

### Analytics Engine Configuration

The `ANALYTICS` binding writes to dataset `fourai_cost_metrics`. Query via GraphQL:

```graphql
query CostMetrics($accountTag: String!) {
  viewer {
    accounts(filter: { accountTag: $accountTag }) {
      analyticsEngineDatasets(filter: { name: "fourai_cost_metrics" }) {
        data(
          query: "SELECT blob1 as event_type, blob2 as customer, blob3 as model, SUM(double1) as tokens_in, SUM(double2) as tokens_out FROM fourai_cost_metrics WHERE timestamp > NOW() - INTERVAL 1 DAY GROUP BY blob1, blob2, blob3"
        ) {
          result
        }
      }
    }
  }
}
```

### Cost API Endpoint

GET `/api/admin/costs` returns pricing constants and query examples for cost dashboards.

## Local Development

```bash
npm install
cp .dev.vars.example .dev.vars
# Edit .dev.vars with your CLOUDFLARE_AI_GATEWAY_API_KEY + CLOUDFLARE_API_TOKEN + CF_AI_GATEWAY_ACCOUNT_ID + CF_AI_GATEWAY_GATEWAY_ID
npm run start
```

### Environment Variables

For local development, create `.dev.vars`:

```bash
CLOUDFLARE_AI_GATEWAY_API_KEY=your-provider-api-key
CLOUDFLARE_API_TOKEN=your-ai-scoped-api-token  # REST transport; use the same AI-scoped credential locally
CF_AI_GATEWAY_ACCOUNT_ID=your-account-id
CF_AI_GATEWAY_GATEWAY_ID=your-gateway-id
DEV_MODE=true           # Skips CF Access auth + device pairing
DEBUG_ROUTES=true       # Enables /debug/* routes
```

### WebSocket Limitations

Local development with the Vite dev server has issues proxying WebSocket connections through the sandbox. HTTP requests work but WebSocket connections may fail. Deploy to Cloudflare for full functionality.

## Docker Image Caching

The Dockerfile includes a cache bust comment. When changing `start-openclaw.sh`, bump the version:

```dockerfile
# Build cache bust: 2026-02-06-v28-openclaw-upgrade
```

## Gateway Configuration

OpenClaw configuration is built at container startup:

1. R2 backup is restored if available (with migration from legacy `.clawdbot` paths)
2. If no config exists, `openclaw onboard --non-interactive` creates one based on env vars
3. Feature flags come from the per-customer `MOLTLAZY_FEATURE_FLAGS` binding (managed via the FourIA API)
4. `start-openclaw.sh start` runs `moltlazy patch` (after the restore/onboard above) which generates per-section include files (`moltlazy-session/tools/logging.json`) and injects them as per-section `$include` directives (no root `$include`); `gateway.*`/`agents.*` are written inline and the gateway token is an env SecretRef
5. The Worker calls `/usr/local/bin/start-openclaw.sh start` **once**: that mode acquires the container-wide startup lock **before** running the setup body, runs the patch while holding it, then `exec`s the supervised Gateway under the same lock (the `exec` replaces the wrapper, so the tracked process _is_ the gateway) and runs `start-openclaw.sh post-start` afterward once the port is ready. The Worker no longer invokes `setup` then `launch` separately; those remain standalone smoke/manual modes only.

### Feature Flag Configuration

Feature flags are managed via the **FourIA API** and applied as a **per-customer binding** (`MOLTLAZY_FEATURE_FLAGS`, a JSON string). The Worker reads the binding and passes it to the container via `buildEnvVars()`. The legacy `CONFIG_BUCKET` / `config-cache.ts` (R2 + ETag) mechanism was removed.

**Feature Flags Example:**

```json
{
  "knowledgeGraph": true,
  "unifiedBilling": false,
  "softwareProvisioning": false
}
```

The container consumes `MOLTLAZY_FEATURE_FLAGS` at startup (`moltlazy patch` generates the per-section include files and injects them as per-section `$include` into `openclaw.json`). There is no worker-side cache — the binding is read directly.

### Software provisioning (`softwareProvisioning`)

Software that agent skills install at runtime (plugins, `apt`/`npm` packages,
downloaded binaries) lives in the container filesystem only, so it is lost on a
container-version upgrade (fresh image) or idle-sleep (container destroyed).

The per-customer `softwareProvisioning` flag fixes this:

1. The tenant's **manifest** (`customers/<slug>/software.json` in the
   `BACKUP_BUCKET`, uploaded via `PUT /api/admin/software/manifest`) declares each
   package with its actual install command: `{ id, check, install, verify?, timeoutSec? }`.
2. On cold start, `ensureGatewayImpl` materializes the manifest into
   `/home/openclaw/.openclaw/software.json` and — only when the flag is on —
   spawns **`moltlazy software provision`** as an **independent supervised process
   through the Sandbox SDK** (`void spawnSoftwareProvision(...)`), not a step in
   `start-openclaw.sh`. It runs in parallel with the gateway boot, never blocks
   readiness, and is non-fatal (per-package timeouts; failures recorded, not raised).
3. It writes live ticks to `/tmp/software-progress.log` (read by
   `GET /api/startup-progress` / `/api/startup/ws` as `softwareLog`) and a
   structured report to `/tmp/software-report.json` (surfaced by
   `GET /api/admin/software/report` as the homepage **Software** widget). Each
   run **truncates** the progress log first (`resetProgressLog`), because `/tmp`
   survives an in-place restart (`/debug/restart-openclaw`), which would
   otherwise interleave ticks from earlier boots.
4. `POST /api/admin/software/provision` (ADMIN/ROOT) reconciles the manifest
   **on demand**, so a manifest uploaded or corrected mid-life does not need a
   restart. It re-materializes the current R2 manifest, runs the same CLI to
   completion bounded by `SOFTWARE_PROVISION_RUN_TIMEOUT_MS` (5 min), and returns
   the fresh report; `409` when the flag is off (`disabled`) or nothing is
   declared (`no_manifest`), `502` when the run fails.
5. Both `GET` endpoints carry `enabled` (the flag), so the dashboard can tell
   "switched off", "asleep", "not run yet" and "nothing declared" apart instead
   of showing one empty state. The widget exposes a **Run provisioning** button
   to ADMIN/ROOT when the flag is on.

The manifest is arbitrary shell executed inside the tenant container, so the
upload endpoint is ADMIN/ROOT-only (never exposed to readers), `timeoutSec` is
per package, and the on-demand run is bounded by the route.

Key modules: `src/gateway/software-provisioning.ts` (flag + spawn + on-demand
run + report parsing), `src/services/software-manifest/` (R2 store +
materialization), `src/routes/api/software.ts` (manifest read/upload, report,
on-demand provision), `packages/moltlazy/src/software/` (the CLI).

### AI Provider

Cloudflare AI Gateway is the only provisioned AI provider path. The startup script configures
auth when all of these env vars are set:

1. **Cloudflare AI Gateway** (native): `CLOUDFLARE_AI_GATEWAY_API_KEY` + `CF_AI_GATEWAY_ACCOUNT_ID` + `CF_AI_GATEWAY_GATEWAY_ID`

### Container Environment Variables

These are the env vars passed TO the container (internal names):

| Variable                        | Config Path                  | Notes                                                                                                                   |
| ------------------------------- | ---------------------------- | ----------------------------------------------------------------------------------------------------------------------- |
| `CLOUDFLARE_AI_GATEWAY_API_KEY` | (env var)                    | Native AI Gateway key                                                                                                   |
| `CLOUDFLARE_API_TOKEN`          | (env var)                    | AI Gateway REST API Bearer token; tenant binding is derived from the AI key                                             |
| `CF_AI_GATEWAY_ACCOUNT_ID`      | (env var)                    | Account ID for AI Gateway                                                                                               |
| `CF_AI_GATEWAY_GATEWAY_ID`      | (env var)                    | Gateway ID for AI Gateway                                                                                               |
| `OPENCLAW_GATEWAY_TOKEN`        | `--token` flag               | Mapped from `MOLTBOT_GATEWAY_TOKEN`                                                                                     |
| `OPENCLAW_DEV_MODE`             | (dev-agents flag)            | Derived from the `-dev` hostname or an `OPENCLAW_DEV_MODE` override; `DEV_MODE` local                                   |
| `OPENCLAW_RESTORED`             | (zero-data-loss signal)      | `true` when a cold start restored an R2 backup (`hadBackup && restored`); makes `start-openclaw.sh` skip the setup body |
| `TELEGRAM_BOT_TOKEN`            | `channels.telegram.botToken` | **@deprecated** — E2E only. Use integrations dashboard in production.                                                   |
| `DISCORD_BOT_TOKEN`             | `channels.discord.token`     | **@deprecated** — E2E only. Use integrations dashboard in production.                                                   |
| `SLACK_BOT_TOKEN`               | `channels.slack.botToken`    | **@deprecated** — E2E only. Use integrations dashboard in production.                                                   |
| `SLACK_APP_TOKEN`               | `channels.slack.appToken`    | **@deprecated** — E2E only. Use integrations dashboard in production.                                                   |

## OpenClaw Config Schema

OpenClaw has strict config validation. Common gotchas:

- `agents.defaults.model` must be `{ "primary": "model/name" }` not a string
- `gateway.mode` must be `"local"` for headless operation
- No `webchat` channel - the Control UI is served automatically
- `gateway.bind` is not a config option - use `--bind` CLI flag

See [OpenClaw docs](https://docs.openclaw.ai/) for full schema.

## Common Tasks

### Adding a New API Endpoint

1. Add route handler in `src/routes/api.ts`
2. Add types if needed in `src/types.ts`
3. Update client API in `src/client/api.ts` if frontend needs it
4. Add tests

### Adding a New Environment Variable

1. Add to `MoltbotEnv` interface in `src/types.ts`
2. If passed to container, add to `buildEnvVars()` in `src/gateway/env.ts`
3. Update `.dev.vars.example`
4. Document in README.md secrets table

### Debugging

```bash
# View live logs
npx wrangler tail

# Check secrets
npx wrangler secret list
```

Enable debug routes with `DEBUG_ROUTES=true` and check `/debug/processes`.

## R2 Backup Encryption (F4E1)

When `BACKUP_ENCRYPTION_KEY` (base64 32-byte secret) is set, the Worker post-encrypts every
snapshot after `sandbox.createBackup()`: `backups/<id>/data.sqsh` is replaced by
`data.sqsh.enc` (chunked AES-256-GCM, per-object HKDF key) and `meta.json` is flagged
`encrypted: true` (kept readable — the SDK needs it). Restores decrypt transiently and delete
the plaintext in a `finally` block. Config versions under `configs/` are encrypted with
`customMetadata.encrypted='true'`. Legacy plaintext backups remain restorable.
Set `BACKUP_ENCRYPTION_REQUIRED=true` to fail closed. Crypto lives in `src/crypto/`.

workerd's `R2Bucket.put()` rejects `ReadableStream` bodies of unknown length, so uploads wrap
the ciphertext/plaintext stream in a `FixedLengthStream` whose size is computed from the source
object (`encryptedLength`) or the F4E1 header (`decryptedLength`). This keeps backup/restore
memory-bounded (no whole-archive buffering) while satisfying R2. See `src/crypto/r2-backup.ts`.
See `docs/plans/r2-encryption-at-rest.md`.

## R2 Storage Notes

Persistence uses the Sandbox SDK `createBackup()` / `restoreBackup()` API with
the `BACKUP_BUCKET` binding. The snapshot root is `/home/openclaw`; it is not an
s3fs mount and there is no `/data/moltbot` runtime path.

Never delete tenant backup objects outside the retention or decommission
procedure. Hard restarts must checkpoint before killing the gateway.

- **Process status**: The sandbox API's `proc.status` may not update immediately after a process completes. Instead of checking `proc.status === 'completed'`, verify success by checking for expected output (e.g., timestamp file exists after sync).

- **R2 prefix migration**: Backups are now stored under `openclaw/` prefix in R2 (was `clawdbot/`). The startup script handles restoring from both old and new prefixes with automatic migration.

- **Snapshot lease + admin snapshot mode (F5):** the shared R2 `backup-lock` TTL is
  `STEP_DEADLINE_MS` (12 minutes), with a serialized heartbeat every quarter-TTL
  and explicit renewals around slow snapshot/restore boundaries. Renewal events are
  structured; renewal failure is best-effort and does not abort work. The longer
  TTL intentionally favors avoiding false expiry during a live durable step, at the
  cost of a longer stale-holder takeover window. `POST /api/admin/storage/sync`
  enqueues a durable job by default when `ACTIVITY` is available; use `?inline=1`
  only when synchronous completion is required (`?async=1` keeps its prior behavior).
  Without `ACTIVITY`, the route falls back to inline work.

## Cloudflare Tunnel Setup (Manual Flow)

This section documents how to expose an internal service (e.g. the Obsidian sync port 18790) via a Cloudflare Tunnel with a public hostname. Follow this end-to-end when setting up a new tunnel route.

### 1. Register and Add a Domain

You need a domain zone in your Cloudflare account before any DNS record can be created.

1. Register a domain (e.g. via Cloudflare Registrar or a third-party registrar like Spaceship/Namecheap).
2. In the [Cloudflare Dashboard](https://dash.cloudflare.com/), select the **Paso4** account → **Websites** → **Add a site**.
3. Enter the domain name and choose the **Free** plan.
4. Cloudflare will assign two nameservers, e.g.:
   - `albert.ns.cloudflare.com`
   - `priscilla.ns.cloudflare.com`
5. At your registrar, replace the existing nameservers with the Cloudflare ones.
6. Wait for DNS propagation (typically 5–30 minutes). The zone status will change from `pending` to `active`.

**Current domain:** `paso4.io` (zone ID: `6ce23f544f57a1909578032abcda18dd`, account: Paso4)

### 2. Create the Cloudflare Tunnel

If the tunnel does not yet exist, create it with `cloudflared`:

```bash
cloudflared tunnel create obsidian-sync
```

This prints a tunnel UUID and writes credentials to `~/.cloudflared/<UUID>.json`.

**Current tunnel:** `obsidian-sync` — UUID `90792969-7077-40a8-ae10-28e716290d9d`

To get the tunnel token (needed for `CLOUDFLARE_TUNNEL_TOKEN`):

```bash
cloudflared tunnel token 90792969-7077-40a8-ae10-28e716290d9d
```

Set it as a Worker secret:

```bash
npx wrangler secret put CLOUDFLARE_TUNNEL_TOKEN
```

### 3. Create the DNS CNAME Record

A CNAME record must point from your public subdomain to `<UUID>.cfargotunnel.com`.

**Via Cloudflare Dashboard:**

1. Go to **paso4.io** → **DNS** → **Records** → **Add record**.
2. Fill in:

   | Field        | Value                                                   |
   | ------------ | ------------------------------------------------------- |
   | Type         | `CNAME`                                                 |
   | Name         | `obsidian-sync`                                         |
   | Target       | `90792969-7077-40a8-ae10-28e716290d9d.cfargotunnel.com` |
   | Proxy status | **Proxied** (orange cloud — required)                   |
   | TTL          | Auto                                                    |

3. Save. Public URL becomes `https://obsidian-sync.paso4.io`.

**Via MCP API (requires DNS:Edit on the API token):**

```js
cloudflare.request({
  method: 'POST',
  path: `/zones/6ce23f544f57a1909578032abcda18dd/dns_records`,
  body: {
    type: 'CNAME',
    name: 'obsidian-sync',
    content: '90792969-7077-40a8-ae10-28e716290d9d.cfargotunnel.com',
    proxied: true,
    comment: 'Cloudflare Tunnel route for obsidian-sync',
  },
});
```

To grant DNS:Edit to the MCP token: `https://dash.cloudflare.com/profile/api-tokens` → edit the token → add **Zone → DNS → Edit**.

### 4. Configure the Tunnel Ingress

Create or update `~/.cloudflared/config.yml` on the host running `cloudflared`:

```yaml
tunnel: 90792969-7077-40a8-ae10-28e716290d9d
credentials-file: /root/.cloudflared/90792969-7077-40a8-ae10-28e716290d9d.json

ingress:
  - hostname: obsidian-sync.paso4.io
    service: http://localhost:18790
  - service: http_status:404
```

Then run:

```bash
cloudflared tunnel run obsidian-sync
```

### 5. Verify

```bash
# Check tunnel is connected
cloudflared tunnel info obsidian-sync

# Test the public endpoint
curl -I https://obsidian-sync.paso4.io
```

### Gotchas

- The `cfargotunnel.com` CNAME only routes traffic for DNS records in the **same Cloudflare account** as the tunnel. Cross-account routing is blocked.
- The zone must be **active** (nameservers delegated) for public DNS to resolve. A `pending` zone means nameserver propagation is not complete yet.
- If the tunnel stops, the DNS record is not deleted — visitors will see a Cloudflare `1016` error.
- The `CLOUDFLARE_TUNNEL_TOKEN` secret in the Worker is used by `start-openclaw.sh` to run `cloudflared` inside the container, exposing port 18790 to the tunnel.

---

FourIA offers Private Cloud only (Business On Premise is not offered at this time). Pricing is credit-based: Credits are the exclusive billable unit, consumed by completed messages and automations. Credits do NOT include AI costs — AI expenses are managed separately via BYOL or AI Gateway billing. Reader users are unlimited and free. Builder seats (users who can create and edit agents) are plan-limited. The v1.0-beta is delivered through the Design Partner Program with a single engaged partner. Partners pay a fixed credit fee negotiated with sales. See docs/VISION.md for the full pricing definition.

## Active Technologies

- OpenTofu 1.6+ (HCL) + `cloudflare` provider (Terraform/OpenTofu) (001-opentofu-deploy)
- Cloudflare R2 (for FourIA data), OpenTofu Local/Remote State (for infrastructure tracking) (001-opentofu-deploy)

## Recent Changes


- **Gateway memory is observable from the Worker, and a dev instance self-heals before the container limit.** Investigating a `tenant-tarjetero-dev` "reboot" required SSH-ing into the container to read `/proc`, `dmesg` and the gateway log, because nothing exposed the gateway's RSS. **(1) Observability (plan Phase 1):** `src/gateway/memory.ts` reads `VmRSS` + `Threads` from `/proc/<pid>/status` for the gateway **server** (`pgrep -x 'openclaw-gatewa'`, the same exact `comm` the soft reload targets, zombies skipped) in ONE shell exec, throttled to `GATEWAY_RESOURCE_SAMPLE_INTERVAL_MS` (60s) so the frequent health poll adds at most one exec per interval. `attachGatewayResources()` (`routes/public.ts`) attaches it to `/api/health`, `/api/health-snapshot` and `/api/health-ws` as `gateway: { pid, rssBytes, threads }` — only while the snapshot is `running`, so a sleeping tenant is never woken — and writes a `gateway-resources` Analytics Engine row (`double1` = RSS bytes, `double2` = threads). **(2) Watchdog (plan Phase 2.4):** on a **dev** instance (`isDevAgentsEnabled`) at `GATEWAY_RSS_RESTART_BYTES` (3 GiB of the 4 GiB container) the gateway is replaced via `killGateway` + `resetStartupCache` + `ensureGateway({ skipRestore: true, bypassMaintenance: true })` — the only lever that reclaims the per-`worker_thread` V8 heaps — with a 30-minute cooldown claimed **before** acting so concurrent health polls cannot double-restart, dispatched detached so a multi-minute replacement never delays a health response. Deliberate design: the threshold guards the **container limit**, not OpenClaw's 1.5 GiB `rss_threshold` warning (which fires permanently on a healthy ~2.1 GiB dev gateway and would make the watchdog a restart loop), and there is **no soft-reload step** (a measured in-process `SIGUSR2` restart does not reclaim RSS: 1.14 → 1.38 → 1.18 GiB). Production is excluded. Tests: `src/gateway/memory.test.ts`, `src/metrics.test.ts` (`trackGatewayResources`), `src/routes/public.test.ts` (sample attached + AE row, one read per interval, no sample when not running, no watchdog on a non-dev instance); corpus `test/business-rules/operational-modes.txt`; docs `docs/plans/sandbox-memory-usage.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md` (§5.4d).

- OpenClaw migrated to **2026.10.8** (`catalog.openclaw` + the `@openclaw/gateway-client` / `@openclaw/gateway-protocol` pins in the root `package.json`; the plugin repo tracks and publishes the same release). 2026.10.8 is a **bug-fix/security release**: plugin lifecycle repairs (older plugins can process incoming messages again after an upgrade, scheduled plugin work no longer stops after startup/reload, and Gateway startup repairs stale managed plugin links left by an external OpenClaw reinstall — the state a warm R2 restore can produce), session-store fixes (relative session-store paths no longer fork a second database; `openclaw doctor --fix` repairs an already-split store), Doctor robustness on very large retained session histories (stack overflow / retained memory), and a trusted-proxy browser-auth fix that no longer writes `gateway.auth.password` or restarts the Gateway on first browser use. It also hardens POSIX exec allowlists (`test -v` / `[ -v` are no longer auto-trusted; moltlazy writes no exec allowlist, and the dev-only `security.installPolicy allow` is unaffected). No config-schema changes — the moltlazy `openclaw config validate` cctr contract passes unchanged, so no `packages/moltlazy` source change was needed — and the Node engine floor is unchanged at `>=24.22.1 <25 || >=26.12.1`, so no Dockerfile change. No component CalVer bump — the release train handles that.

- **Software provisioning is now usable mid-life and honest about its state.** Three gaps closed around the `softwareProvisioning` feature. **(1) On-demand reconcile:** `POST /api/admin/software/provision` (ADMIN/ROOT) re-materializes the current R2 manifest, runs `moltlazy software provision` to completion (bounded by `SOFTWARE_PROVISION_RUN_TIMEOUT_MS`, 5 min) and returns the fresh report, so a manifest uploaded or corrected mid-life no longer waits for the next cold start; `409 { code: 'disabled' | 'no_manifest' }` and `502 { code: 'failed', error }` are distinct outcomes. **(2) No cross-boot tick bleed:** `runSoftwareProvision` truncates `/tmp/software-progress.log` before every run (`resetProgressLog`) — `/tmp` survives an in-place `/debug/restart-openclaw`, so the boot terminal used to interleave ticks from earlier boots. **(3) Truthful dashboard states:** `GET /api/admin/software/{manifest,report}` now return `enabled` (the flag), and the homepage widget separates "provisioning is disabled", "instance is asleep", "no report yet" and "no software declared", with a **Run provisioning** button (ADMIN/ROOT) that renders the on-demand report and surfaces its error. `parseSoftwareReport` moved into `src/gateway/software-provisioning.ts` so the report, run and route paths share one shape check. Tests: `src/gateway/software-provisioning.test.ts`, `src/routes/api/software.test.ts`, `src/client/pages/__tests__/HomePage.test.tsx`, `packages/moltlazy/tests/software/run.test.ts` (progress-log reset), E2E `test/e2e/zzz_software_provisioning.txt` (on-demand run must install the new package and **skip** the already-installed one); corpus `test/business-rules/dashboard-api-contract.txt`, `operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- **Lerma deploys are idempotent around Cloudflare DO migration conflicts (#516).** A tenant Worker's Durable Object class set was derived from the script's _bindings_, but Cloudflare validates `new_sqlite_classes` against the script's _migration history_, so the two can diverge: an upgrade redeploy that binds fewer classes (or a rollback that deletes the script while leaving its classes migrated) made every subsequent deploy re-declare `TelegramIngress` and fail with `400 code 10074` — deterministically, so provisioning could not recover on its own. `Lerma.Cloudflare.DispatchNamespace.upload_script/6` now retries the upload without any class Cloudflare reports as already migrated (`10074` names it), switching an implicit full first-deploy set to an explicit `new_sqlite_classes` list; each retry removes at least one class, so it terminates, and the retry is logged. `Lerma.Instances.instance_bindings/2` now re-declares the same DO set as provisioning (`Sandbox`, `ACTIVITY`, `TELEGRAM_INGRESS`, plus `BACKUP_WATCH` while the flag is on) so an upgrade cannot detach a migrated class from its binding. Provisioning cleanup and a failed atomic deploy tear the classes down first via `delete_durable_object_namespaces/2` (never on an upgrade rollback, whose Durable Objects hold the live tenant's state: `preserve_do_namespaces`). Tests: `apps/lerma/test/lerma/cloudflare/dispatch_namespace_test.exs` (migrated-but-unbound re-deploy, resurrected first deploy, unrelated error not retried), `containers_test.exs`, `tenant_provisioner_pipeline_test.exs`; docs `apps/lerma/README.md`.

- OpenClaw migrated to **2026.10.6** (`catalog.openclaw` + the `@openclaw/gateway-client` / `@openclaw/gateway-protocol` pins in the root `package.json`; the plugin repo tracks and publishes the same release). 2026.10.6 is a **security patch release** (eMudhra certificate-compromise remediation, Windows setup/Doctor/Gateway startup fixes, Discord command and iPhone Talk fixes) with no config-schema changes: the moltlazy `openclaw config validate` cctr contract passes unchanged, so no `packages/moltlazy` source change was needed. It does raise OpenClaw's Node engine floor from `>=24.16.0 <25 || >=26.1.0` to **`>=24.22.1 <25 || >=26.12.1`**, so `apps/fouria/Dockerfile` now installs Node **v24.22.1** (from v24.20.0) and its version gate compares major/minor/patch (24.22.0 is rejected). No component CalVer bump — the release train handles that.
- **Idle-sleep no longer re-archives an unchanged tenant, and deprovision snapshots reset the cadence.** `TenantSandbox.onActivityExpired()` used to enqueue a full `deprovision` snapshot on every idle expiry; on a 422 MB tenant the snapshot's own RPCs renewed the DO activity window, so `stopIfIdle()` refused, the next expiry snapshotted again, and the container never stopped (observed on `tenant-tarjetero-dev`: three 422 MB archives at 02:35/02:39/02:44 UTC plus three more at 05:39/05:43/05:49, ~3 GB/day, container `running` since Oct 21). The hook now reads the catalog first (`hasRecentSnapshot()` in `src/persistence.ts`, a pure read): a committed snapshot **inside the idle window** (`resolveSleepAfterMs()` — staging 15m, dev 2h, production 6h) already covers the unchanged filesystem, so it stops inline through `stopIfIdle()` without a new archive; no bucket or a read failure falls back to the durable backup-then-stop job (fail closed). Deprovision `finalize` now passes `stampScheduled: true` too, so a committed archive resets `lastScheduledBackup` and the next cron tick does not archive the same state again. Tests: `src/sandbox-do.test.ts` (recent snapshot stops inline, probe failure snapshots, no bucket skips the probe), `src/persistence.test.ts` (`hasRecentSnapshot`), `src/activity/scheduler.test.ts` (deprovision stamps), `src/build-sandbox-options.test.ts` (`resolveSleepAfterMs`); corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.
- **The dev Fouria Builder heartbeat is reconciled to its intended cadence.** OpenClaw 2026.10.5 retired `agents.entries.*.heartbeat`: heartbeats are ordinary automation jobs, and config reloads/restarts never rewrite an existing job. The migrated `heartbeat-fouria-builder` job therefore kept the **default 30-minute** cadence for 9 days on `tenant-tarjetero-dev` despite moltlazy's `heartbeat.every: 12h`, running full agent turns on the **main** session (~317k tokens each) until the OpenCode subscription hit 0% (`402 Insufficient account funds` → `agent-runner-failure`). `start-openclaw.sh` gains a dev-only `ensure_dev_heartbeat_cadence` (post-start) that reads `openclaw automations get <id> --json`, compares `schedule.everyMs` / `sessionTarget` / `payload.lightContext`, and only then edits the job to `--every 12h --session isolated --light-context` (split so a rejected `--light-context` cannot lose the cadence fix). Idempotent: a matching job is never touched, so the schedule anchor is not reset on every boot. Docs `packages/moltlazy/AGENTS.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Backup catalog reads are **pure**; catalog healing has an explicit entry point (audit finding F3). `GET /api/admin/storage/backups` (`getBackupHistory`) used to run `reconcileOrphanedBackups` on every dashboard poll — listing `backups/`, reading `meta.json`, adopting orphans, deleting over-cap archives, and writing `backup-handle.json` from a `GET`. It now performs three parallel reads (`getPersistenceStore`, `getStableBackup`, `isRecoveryHeld`) and nothing else. The write path is `reconcileCatalogState` (adoption + retention prune inside the etag-CAS mutation, reporting the ids it changed) → `repairBackupCatalog` (sweep metadata-less uploads + in-place encryption) → refreshed history, exposed as `reconcileBackupCatalog`; the durable `repair` step now runs it **before** the `gate` step of every scheduled snapshot, so the catalog still self-heals after a killed invocation (issue #344). Restore/schedule keep their inline reconcile, `debug/reprovision` reconciles before deciding nothing is restorable (a crashed upload must not look unrecoverable), and operators get `POST /api/admin/storage/reconcile` (ADMIN/ROOT). The history panel polls `GET /api/admin/storage/jobs/:id` (`getBackupJobStatus`) while a snapshot is in flight instead of re-listing the catalog every 5s, falling back to the history diff when no job id is returned. Tests: `src/persistence.test.ts`, `src/persistence-encryption.test.ts`, `src/routes/api/storage.test.ts`, `src/activity/scheduler.test.ts`, `src/routes/debug.test.ts`, `src/client/components/__tests__/BackupHistoryPanel.test.tsx`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`, `docs/plans/durable-agents-rules-audit.md`.
- A **config repair is deferred until a restore settles and the gateway is up on the restored state**, and the restored config becomes the stable/latest config version. A restore replaces the writable layer but launches nothing, so a `doctor --fix` applied around it was silently overwritten by the next re-apply of the recovery point (suspected on `tenant-tarjetero-dev` after a config fix). **(1)** Every successful restore (`restoreIfNeeded`, `restoreBackupById`) writes a durable `restore-config-pending.json` marker. **(2)** The boot job, the durable restore job, and the sync/async-fallback admin restore route finish through the shared `launchRestoredState` (`src/post-restore.ts`): `ensureGateway({ waitForReady: true })` on the restored state, then `captureRestoredConfigVersion` writes `configs/<ts>.json` + `stable/config.json` (the newest/stable version) and clears the marker. A gateway-launch failure does not re-run the destructive restore: restore jobs succeed with `detail: "gateway-launch-pending"`, the sync route returns `gatewayReady: false`, and the marker drives the next tick. **(3)** `doctor-repair` checks an active restore job, `getPersistenceBusyReason` (maintenance window / restore in flight), and the pending marker before running; when blocked it throws `MaintenanceJobDeferredError`, and `runMaintenanceJobAlarm` reschedules it every 60s **without consuming an attempt** (previously three deferrals would have failed the job before the restore finished). Tests: `src/activity/maintenance-jobs.test.ts`, `src/activity/scheduler.test.ts`, `src/persistence.test.ts`, `src/routes/api/storage.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`.
- Diagnostics dashboard is **skip-not-park and non-blocking**. `DiagnosticDashboard` only dispatches its container reads (`/api/admin/config/raw`, `/api/admin/config`) while the gateway is actually running (`useGatewayRunning()` — `snapshot.readiness.ok`, ignoring the ROOT breaker bypass) and re-runs them when it comes back; the debug-routes flag (Worker-local) still loads. Every load is deferred after paint (`scheduleAfterPaint` → `requestIdleCallback`, `setTimeout(0)` fallback), so no widget waits on a container response: `ConfigSection` gained a `gatewayReady` prop and now keeps its tabs/editor visible with a non-blocking "Loading configuration…" / "Waiting for the gateway to start…" notice instead of swapping the whole panel for a spinner; devices render a waiting state; config loads never clobber text typed before the response arrived. Gateway-only actions (device list/approve, config save, channel probe) are guarded, and the Debug Tools that only make sense against a live gateway (`gateway-api`, `stop-gateway`) are skipped with an inline explanation — container diagnostics (logs/processes/version/sqlite) and recovery actions (restart/rescue/destroy) stay available so a down gateway can still be inspected and revived. **Repair stays available outside the editor:** `doctor --fix` is the recovery action for a down gateway, so `runRepair` no longer requires a running gateway, a dedicated **Repair Environment** bar (`data-testid="repair-environment-button"`, ADMIN/ROOT) sits above the widgets and is disabled only while a repair is in flight, and `ConfigSection`'s repair button stays enabled while `gatewayReady` is false. Tests: `src/client/pages/__tests__/DiagnosticDashboard.test.tsx`, `src/client/components/__tests__/ConfigSection.test.tsx`, `src/client/GatewayHealthContext.test.tsx`; corpus `test/business-rules/dashboard-api-contract.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/TESTING.md`.
- Billable emission: **staging always emits**, `BILLABLE_EVENTS` opts a dev instance in. Both emit sites (`POST /api/otel/v1/traces` and the `CRON_BILLABLE_DRAIN` cron drain) previously gated on the strict billable cost axis (`resolveInstanceDimensions(env).billable` = production platform + production environment), so staging and dev instances could never exercise the billing pipeline. They now call `isBillableEventsEnabled(env)` (`src/dev-mode.ts`): billable **production always emits and cannot be opted out**, the **staging platform always emits** in every environment (mirroring production so billing is validated before release; the rows land in the staging Analytics Engine dataset), and a **dev environment** emits only with `BILLABLE_EVENTS=true`. Emission only: `getCostContext` keeps the real `platform`/`environment` dimensions (`staging`/`dev` remain non-billable _cost_ axes) and `shouldAutoSleep` is untouched, so staging keeps its 15m idle timeout and an opted-in dev instance its 2h one. The emitting path omits the guard's `billable: false` marker (the local-E2E skip, whose worker is a dev instance via `DEV_MODE`, is unaffected). Tests: `src/dev-mode.test.ts`, `src/routes/otel-traces.test.ts` (staging emits, dev opt-in emits), `src/index.test.ts` (cron drain: dev drops, staging and dev-with-override emit), `src/build-sandbox-options.test.ts` (sleep policy unchanged); corpus `test/business-rules/dev-expense.txt`; docs `docs/ENV-VARIABLES.md`, `docs/TESTING.md`, `apps/fouria/README.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `.dev.vars.example`.
- Request paths never run a destructive archive restore, and a container stops only on verified inactivity (investigated on `tenant-tarjetero-dev`, which appeared to "reboot" while in active use). **(1)** The container instance had **not** restarted (`started_at` 00:18, still `running`); what died was the **gateway process** (`[Gateway] Gateway process … has terminal state: exited`), and the follow-up boot ran a **350 MB** archive restore under request execution contexts. `sandbox.restoreBackup` was aborted at ~30.1 s (`backup.restore error … The operation was aborted (30177ms)`, plus `sandbox.exec … The operation was aborted (30001ms)` for health probes/`umount` and the workerd "promise resolved in a different request context" warning) because Cloudflare cancels a request's background continuations shortly after the response — so the restore never finished and every reload re-attempted it (6 boot jobs + 8 restore errors in one hour). Fix: `restoreIfNeeded(..., { restoreOwner: 'request' })` (the `ensureGateway` default) performs only the cheap non-destructive checks (the `restore-needed` latch + `containerHasState()`) and otherwise throws `RestoreDeferredToBootJobError`; `ensureGateway` requests the durable boot job and throws `GatewayBootPendingError` (proxy answers 503/loading). Durable callers (`ActivityScheduler` boot/wake jobs, the queue consumer) pass `restoreOwner: 'durable'`. The proxy's three inline `restoreIfNeeded` calls (`proxy.request`, `proxy.websocket-recovery`, `proxy.fetch-recovery`) were removed. **(2)** A container is now stopped only after a **stop-time** inactivity check: `TenantSandbox.stopIfIdle()` re-reads the DO's own idle window (renewed by every request) and the process list, and the deprovision job calls it after its multi-minute snapshot instead of `sandbox.stop()`; if the tenant became active meanwhile the job is `skipped` (`active:<reason>`). Refusals: `activity`, `transient-activity`, `probe-failed`, `billable-production`, `not-running`. Tests: `src/sandbox-do.test.ts` (`stopIfIdle` cases), `src/activity/scheduler.test.ts` (stop-time skip), `src/gateway/process.test.ts` (deferral + `restoreOwner`), `src/persistence.test.ts` (request deferral / state-present short-circuit), `src/index.test.ts` (proxy no longer restores); corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`, `webs/fouria-docs/.../worker-development.md`, `operator-guide.md`.

- Fouria Builder build-method skill. The dev-only `fouria-builder` agent now seeds **`fouria-product-build`** as a per-agent workspace skill (`packages/moltlazy/workspace/skills/fouria-product-build.ts` → `<workspace>/fouria-builder/skills/fouria-product-build/SKILL.md`, plus an entry in its `skills` allowlist) that sequences the build method: frame the production roster → tool-first restriction model (each capability names the exact tool; no tool means no capability) → smallest MCP tool surface with `preview`/`apply` pairs → 3–6 channel quick actions (Telegram inline buttons; a callback is untrusted input, never authorization) → per-client tone matrix before writing a client-facing `SOUL.md` → one vertical slice → `cctr` acceptance corpus → `POC.md` handoff. Workspace skills are OpenClaw's highest-precedence skill source and are per-agent, and `workspace/fouria-builder` is excluded from promotions, so no production roster changes. `seedWorkspaceFiles` now creates parent directories so workspace-file keys can be nested paths. This supersedes the pending Skill Workshop proposal `fouria-product-build-20261021-bf543e21c1`: a workspace copy of the same name shadows a workshop skill, so the repo copy is authoritative. Tests: `packages/moltlazy/tests/agents.test.ts`; docs `packages/moltlazy/AGENTS.md`, `packages/moltlazy/src/agents/README.md`, `docs/DEVELOPMENT.md`.

- Cloudflare Observability **destination** ingestion (issue #337). Account-level Observability → Destinations now push Workers Logs and Workers Traces to Lerma over OTLP/HTTP JSON, giving event-driven worker diagnostics and uptime statistics that scale with traffic instead of tenant count. **(1) Config:** `apps/fouria/wrangler.jsonc` (`env.staging` → `lerma-worker-logs`/`lerma-worker-traces`, `env.production` → `lerma-worker-logs-prod`/`lerma-worker-traces-prod`) and `Lerma.Cloudflare.Observability.worker_metadata/0` reference the environment's destination names, with per-signal head sampling (10% at rollout) and `persist: false` when a destination is configured (Lerma is the destination of record). The Lerma CD passes the matching names to staging and production, and the token to both; it also reconciles the account-level destination pairs (`mix lerma.observability.destinations --warn-only`) before each deploy so a fresh environment self-heals once the Cloudflare credential has Workers Observability Write. New `WORKER_OBSERVABILITY_LOGS/TRACES_DESTINATIONS`, `*_SAMPLING_RATE` and `*_PERSIST` env vars. **(2) Ingest:** new `POST /api/ingest/otel/v1/logs` and `/v1/traces` (bearer `OTEL_INGEST_TOKEN`, constant-time compare, fail closed; OTLP/HTTP JSON only — protobuf → `415` + hint; `202 {accepted, skipped}` so Cloudflare never retry-loops). `Lerma.Observability.Otlp` is a pure allowlist parser: it keeps script/dispatch namespace/outcome/handler/status/exception class names/cpu+wall/trace+span+invocation ids/timestamps, drops log text, stacks, URLs/query strings, user agents, request/response bodies and DB query text, and keeps only invocation root spans. Rows land in `observability_events` with a unique `dedupe_key` + `ON CONFLICT DO NOTHING`, so replayed batches (at-least-once delivery) add no rows. Cloudflare gzips each OTLP/HTTP JSON batch while keeping `Content-Type: application/json`; `LermaWeb.Plugs.OtlpBodyReader` decompresses it before `Plug.Parsers` (otherwise the destination retries on a `400` forever). **(3) Uptime:** `GET /api/analytics/v1/uptime/:script_name?window=24h|7d|30d|<n>h|<n>d` computes uptime/failure ratios from pushed root spans only; no events observed (including under sampling) → `null` ratios + `status: "unknown"`, never downtime. The instance page gained a **Worker telemetry** widget (window selector + uptime/failure/invocation/CPU/wall), and `Lerma.Observability.Pruner` sweeps 90-day retention daily. `HealthMonitor` and billing flows are untouched. Tests: `test/lerma/observability/otlp_test.exs`, `telemetry_test.exs`, `test/lerma/cloudflare/observability_test.exs`, `test/lerma_web/controllers/api/otlp_ingest_controller_test.exs`, `observability_controller_test.exs`, `test/lerma_web/components/worker_telemetry_widget_test.exs`, `test/lerma_web/live/instance_live_test.exs`; docs `docs/CLOUD-ARCHITECTURE.md`, `docs/ENV-VARIABLES.md`, `docs/TESTING.md`, `apps/fouria/public/articles/PRIVACY.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `SECRETS.example.yaml`, `docs/schemas/secrets-manifest.schema.json`, `apps/lerma/.env.example`.

- OpenClaw migrated to **2026.10.5** (`catalog.openclaw` + the `@openclaw/gateway-client` / `@openclaw/gateway-protocol` pins in the root `package.json`; the plugin repo tracks the same release). One breaking change hit moltlazy: the 2026.10.5 config schema **retired `agents.defaults.compaction.enabled`** (embedded auto-compaction is no longer toggleable) and **retired `compaction.truncateAfterCompaction`** ("byte-triggered compaction now opts in via `maxActiveTranscriptBytes` alone") — `openclaw config validate` rejects both keys with `Invalid input`, which broke the moltlazy cctr `config-save` contract. `AGENT_DEFAULTS.compaction` now emits only `maxActiveTranscriptBytes: '20mb'` + `keepRecentTokens: 20000`, preserving the sandbox OOM transcript bound. The runtime zod schema still accepts the retired keys, so only the `config validate` contract (and the published schema) caught it. Tests: `packages/moltlazy/tests/agents.test.ts` (asserts neither retired key is emitted); docs `packages/moltlazy/AGENTS.md`.

- Local development runs on the **Cloudflare Vite plugin** instead of `wrangler dev`. `bun run dev:worker` / `bun start` start the Worker with `vite dev` (`vite.config.ts`), the same plugin `bun run build` already used; `bun run dev` serves the admin SPA from `vite.ui.config.ts` (:8787) and proxies to the Worker (:8788). Local E2E fixtures (`test/e2e/fixture/server/start-local`, `start-local-pair`) launch `vite dev`, pass the old `--var` overrides through the plugin's `config` hook (`E2E_WORKER_NAME`, `PERSIST_STATE_PATH`, `VITE_CACHE_DIR`, `VITE_INSPECTOR_PORT`), install `test/e2e/.dev.vars` into the Worker config directory for the run (the plugin has no `--env-file`; `stop-local*` restores the developer's file), and fire cron ticks through `/cdn-cgi/local/scheduled` instead of `/__scheduled`. `--env` is gone from the deploy/dry-run scripts: `CLOUDFLARE_ENV=…` selects the environment at build time and `wrangler deploy` discovers the generated `dist/fouria/wrangler.json` via `.wrangler/deploy/config.json`. Docs: `README.md`, `docs/DEVELOPMENT.md`, `docs/TESTING.md`, `apps/fouria/README.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`, `test/e2e*/README.md`.

- `/healthz` is now a real, fast liveness route. It previously fell through to the gateway-proxy catch-all (`app.all('*')`), which ran `restoreIfNeeded` + `ensureGateway` + a `waitForGatewayReady` port wait — on a cold/sleeping instance `GET /healthz` took ~108 s and was killed by the platform before responding (Workers Observability on `tenant-client-dev`: `GET /healthz` → 503 @ ~108 s avg, 348 requests in 2 h). The new `GET /healthz` does exactly one non-waking `isRuntimeActive()` read and returns `200 { ok, ts, status }` (`active` | `idle`) sub-second, never warming the container or probing the gateway; readiness stays on `/api/health`. It sits on the tenant hostname, so it stays behind the host Access app (not public). Tests: `src/routes/public.test.ts` (`/healthz` 200 + no `ensureGateway`/exec/listProcesses; `idle` when the runtime is off); docs `apps/fouria/README.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/WORKER-HEALTH.md`.

- Browser health endpoints moved to Access-safe sibling paths (`/api/health-ws`, `/api/health-snapshot`, `/api/health-components`). On tenant hostnames every browser WebSocket and `fetch()` under `/api/health` or `/api/health/*` failed (`WebSocket connection to 'wss://…/api/health/ws?passive=1' failed`, plus CORS/redirect failures on `GET /api/health/components` and `GET /api/health`): the path-scoped `FourIA Health - <slug>` Access app shadows those paths and its AUD is distinct from the host app's (Cloudflare ignores a requested `aud` on create **and** update — verified against the live API — so the AUDs cannot be unified). A browser logged into the host app therefore gets a cross-origin 302 login redirect to `*.cloudflareaccess.com` on `/api/health*`, which WebSocket handshakes cannot follow and browser `fetch()` rejects on CORS. The sibling paths `/api/health-ws`, `/api/health-snapshot`, and `/api/health-components` sit outside the health app's path prefix, so the host-app session the browser already holds passes (verified against live `staging-client-dev.fouria.io`: `/api/health-*` resolves to the host app `kid`, while `/api/health` and `/api/health/*` resolve to the health app `kid`). The Worker mounts the handlers at both path sets (`healthHandler`, `healthComponentsHandler`, `healthWsHandler` in `src/routes/public.ts`); `api.ts` (`subscribeHealth`, `getHealth`, `getCloudflareComponents`) and `loading.html` use the `-` sibling paths, while `/api/health`, `/api/health/ws`, and `/api/health/components` remain for headless service-token clients (lerma). Tests: `src/routes/public.test.ts`, `src/assets/loading.test.ts`, UI stubs `test/ui/support/api.ts` + `test/ui/cloudflare-components.spec.ts`; docs `apps/fouria/README.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/WORKER-HEALTH.md`, `docs/TESTING.md`; lerma `access_policy.ex` documents the constraint.

- Telegram ingress queue observability: notification-bell entry + home status widget. The per-tenant `TelegramIngress` DO now records every ingress event (bounded counters + last 20) — `queued` (a cold update was enqueued and woke the container), `forwarded` (hot path), `delivered`, and `failed` — via `telegram-ingress/status.ts`, wired from the webhook route (queued/forwarded) and the queue consumer (delivered/failed). `GET /api/admin/telegram-ingress/status` returns the snapshot and `GET /api/admin/telegram-ingress/ws` streams it (a frame per change, `TELEGRAM_INGRESS_WS_POLL_INTERVAL_MS`, rotated). The admin shell's `TelegramIngressProvider` opens the socket only while the feature is in use (an account is registered or activity exists) and raises a notification-bell entry ("Telegram message queued — waking your instance", or a delivery-failure alert); the new `TelegramIngressWidget` on the home page shows state + queued/delivered/failed + last event, hidden until in use. Tests: `src/telegram-ingress/status.test.ts`, `src/routes/api/telegram-ingress.test.ts`, `src/client/__tests__/telegram-ingress-context.test.tsx`, `src/client/components/__tests__/TelegramIngressWidget.test.tsx`, `src/client/pages/__tests__/HomePage.test.tsx`; corpus `test/business-rules/dashboard-api-contract.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Homepage no longer crashes with `Cannot read properties of undefined (reading 'length')`. `GET /api/admin/software/report` ran `cat <report> 2>/dev/null || echo '{}'`, so a tenant whose `moltlazy software provision` had never run got a truthy `report: {}`; the homepage Software widget then read `software.packages.length` and threw on render (only for ADMIN — ROOT lands on diagnostics). The route now returns `report: null` unless the payload is a real report (an object with a `packages` array), and `HomePage` reads `software?.packages ?? []` defensively. Tests: `src/routes/api/software.test.ts` (missing/partial report → `null`), `src/client/pages/__tests__/HomePage.test.tsx` (truthy `{}` renders the empty state instead of crashing).

- Release CD **promotes the already-built artifacts instead of rebuilding**. `deploy-fouria.yml` used to `docker build` the container image from source and re-build the wrangler bundle at release time; it now promotes what CI already built: it resolves the release's develop-mainline build commit (`git rev-list -n1 --first-parent --invert-grep --grep='^chore: release' <tag> -- apps/fouria/ .github/workflows/test.yml`), retags the Cloudflare registry image `fouria:unstable-<sha>` → `fouria:<tag>` + `latest` (minting short-lived pull+push registry credentials via `POST /accounts/{id}/containers/registries/{host}/credentials`, mirroring Lerma's `ImageRegistry`), and copies the CI bundle `fouria-bundles:unstable-<sha>/fouria.mjs` → `<tag>/` + `latest/` in R2. `deploy-dashboard.yml` now tags the develop build with an immutable `lerma:sha-<sha>` in addition to `:latest-pre`, and its `promote` job pins to that per-commit image (resolving the release's lerma build commit the same way) instead of the floating `:latest-pre` (falling back only for the first release after the tag is introduced). Docs: `docs/DEVELOPMENT.md`, `docs/TESTING.md`, `apps/lerma/DEPLOY.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`. Both workflows validated with `actionlint`.

- Durable `doctor-repair`/SQLite jobs start immediately, and their status streams over a WebSocket instead of 2s polling. A `doctor-repair` job enqueued at 12:57:30 only began at 13:02:11 — exactly the previous boot job's 14-minute lease — because the `ActivityScheduler` owns one alarm and a **stale arm** (the alarm handler's trailing re-arm, or a warm-path `refresh`) could read "no pending job", then execute its `setAlarm(later)`/`deleteAlarm()` _after_ the just-enqueued job's `setAlarm(now)`, postponing it. `scheduleCombinedAlarm` now writes through `armWithGenerationGuard` (`activity/alarm.ts`): it bumps a stored generation, and a superseded arm abandons its stale write so the newest arm (fresh read) owns the alarm. The dashboard no longer polls `GET /api/admin/config/fix/jobs/:id` every 2s (one request + Worker log line per tick for the whole multi-minute repair): a new `GET /api/admin/config/fix/jobs/:id/ws` (and the `/debug/sqlite-integrity/jobs/:id/ws` variant) streams the job (`routes/api/maintenance-ws.ts`) — a frame only on status change, close `1000` when terminal, and a `rearm()` **self-heal** while active. The client's `subscribeMaintenanceJob` is WS-first with a slow HTTP fallback only while the socket is down, and `DiagnosticDashboard` renders a live `MaintenanceJobModal` (queued/running/attempt/elapsed, run-in-background) instead of a bare spinner, with a banner to reopen it after backgrounding. Tests: `src/activity/alarm.test.ts`, `src/activity/scheduler.test.ts`, `src/routes/api/maintenance-ws.test.ts`, `src/client/__tests__/maintenance-job-subscription.test.tsx`, `src/client/pages/__tests__/DiagnosticDashboard.test.tsx`; corpus `test/business-rules/dashboard-api-contract.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Sandbox memory OOM mitigation (dev instance). A dev instance grew RSS past OpenClaw's 1.5 GiB `rss_threshold` (worker-thread V8 heaps + native arenas, **not** the main isolate) driven by cron-replayed heavy `agentTurn` jobs. **(1)** `packages/moltlazy` `AGENT_DEFAULTS.compaction` sets `enabled: true`, `maxActiveTranscriptBytes: '20mb'`, `keepRecentTokens: 20000` so a provider-side-managed session can no longer grow its persisted transcript unbounded (the `fouria-builder:main` ~407k-token session inflates the SQLite history and worker heaps). **(2)** The dev-only Fouria Builder entry sets `heartbeat.every: '12h'` plus `isolatedSession: true` / `lightContext: true` — OpenClaw's default 30m heartbeat is a full agent turn ~48×/day that also triggers memory-core dreaming reconciliation. **(3)** `start-openclaw.sh` slows memory-core's dreaming sweep to weekly (`frequency: "0 3 * * 1"`, dev-gated) — the ~6-min daily sweep is replayed on every wake (`cron: running missed jobs after restart`) and produced the 1.6 → 2.08 GiB RSS spike. **(4)** `ensure_opencode_credits_automation` dedupes via `openclaw cron list --all --json` in addition to the legacy `automations list` (which missed a prior boot's job, creating two duplicate `opencode-credits-check` cron jobs). Tests: `packages/moltlazy/tests/agents.test.ts`; docs `packages/moltlazy/AGENTS.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Boot log streams gateway-launch progress (fixes "Launching the gateway — no progress for 6 min"). `start-openclaw.sh` now starts a `start_boot_log_stream` mirror in `launch_gateway_process` that tails the gateway's own `--verbose` stdout (`/tmp/gateway-stdout.log`) into the curated boot log (`/tmp/startup.log`) — so the loading page can follow SQLite migration → plugin load → port bind instead of going blank the moment `start`/`launch` `exec`s the gateway. `post-start`/`full` stop the mirror (`stop_boot_log_stream`) and truncate the buffer at readiness. The Worker's `/api/startup/ws` stream now reads the boot log while `processType === 'gateway'` (not just `'script'`), and `loading.html` shows the log panel during gateway startup — restyled as a vintage CRT terminal (title bar + traffic lights + phosphor text + scanlines) — instead of hiding it behind the logo. Tests: `src/assets/loading.test.ts`, `src/routes/public.test.ts`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Loading page no longer gives up on its own. `loading.html` had two client-side budget specs — a 5-error retry cap (`maxConsecutiveErrors`, `Retrying... (n/5)`) and an 8-minute wall-clock deadline (`MAX_WAIT_MS` / `deadline` → `runDoctorAndShow()`). Since the cold boot moved to the durable `ActivityScheduler` boot job (which owns retries and terminal failure), the page's self-imposed give-up only interrupted a legitimately long startup: it now waits indefinitely (WebSocket push, with an unlimited 5s HTTP-poll fallback) and reports terminal failure solely from `boot.status === 'failed'` and permanent `degraded` container errors. Removed the now-dead `runDoctorAndShow`/`showStructuredError`/`showError` helpers. Tests: `src/assets/loading.test.ts`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.
- Health endpoints no longer require the ROOT role at the edge. A tenant ADMIN opening `/_admin` was 302'd to the Cloudflare Access login for `GET /api/health` and `GET /api/health/ws` (the dashboard's health stream), because the per-tenant `FourIA Health - <slug>` Access app on `<slug>.fouria.io/api/health` allowed only the `Paso4 Root Users` group plus lerma's health-monitor service token. The health app (and the platform-wide `*.fouria.io/api/health` / `*.dev.fouria.io/api/health` IaC apps) now carries **no role requirement**: an `allow` policy on `everyone` (any authenticated user) plus the existing `non_identity` service token, so the dashboard and lerma's headless probe both work. `AccessPolicy.health_policies/1` drops the root-group policy (the host app keeps it); `access.tf` `public_health` matches. Tests: `test/lerma/cloudflare/access_policy_test.exs`, `apps/fouria/iac/tests/access_app.tftest.hcl`; docs `apps/lerma/README.md`, `apps/fouria/iac/README.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/ENV-VARIABLES.md`. Existing tenants were reconciled in place via the Cloudflare API (aud preserved).
- Restore decrypt no longer exhausts the Worker CPU budget, and interrupted restores are labelled honestly. A large recovery archive (`tenant-tarjetero-dev`) was killed with `Worker exceeded CPU time limit` mid-`Decrypting backup …` and retried in a loop. `encryptStream`/`decryptStream` (`src/crypto/backup-encryption.ts`) grew each 4 MiB chunk with `concat` on every ~64 KB R2 read, recopying the pending chunk ~64× over; a `ByteQueue` now copies each byte once (96 MiB decrypt: 1.5 s → 0.14 s at 64 KB reads, 5.6 s → 0.14 s at 16 KB). The R2 binding delivers the archive in **4 KiB reads** (64 MiB → 16,385 reads), which is why the per-read copy dominated; each restore now logs `[crypto] Decrypt read profile` (reads, bytes, max/avg read, chunks) so the pattern can be checked on a live tenant. The loading bar also showed the last durable phase (`Waiting for the backup lock 3%`) as if still live: a failed attempt is now labelled `Restore attempt failed during "<phase>" — retrying` (`recordRestoreFailure`), and an idle record gets `— no progress for N min` (`describeRestoreProgressIdle`); the fraction never moves backward. Tests: `src/crypto/backup-encryption.test.ts`, `src/activity/restore-progress.test.ts`, `src/activity/scheduler.test.ts`; docs `docs/WORKER-DEVELOPMENT.md`.
- Recovery progress bar synchronized with the durable object's events. During a restore the loading page showed a generic spinner; it now shows a **progress bar driven by the `ActivityScheduler` DO's own phase events**, not a client-side timer. `restoreIfNeeded`/`restoreBackupById` call their `onProgress(phase)` hook on every `persistence.phase`; the DO wires it to `recordRestoreProgress`, which maps the phase to a label + a monotonic fraction (`describeRestoreProgress` in `activity/restore-progress.ts`, cumulative phase weights) and exposes it via `getRestoreProgress()`. `/api/startup/ws` (and non-passive `/api/health`, plus `/api/admin/boot-state`) push `restoreProgress: { phase, label, fraction }`; `loading.html` renders the bar and hides the stepper while a restore is in flight, then switches back to the container's `startup-progress.json` stepper once the gateway launch begins (an indeterminate bar covers a request-scoped restore with no DO feed). Also declares `getBootJob`/`getRestoreProgress` on `ActivitySchedulerStub` (fixing the `public.ts` type error). Tests: `src/activity/restore-progress.test.ts`, `src/activity/scheduler.test.ts`, `src/routes/public.test.ts`, `src/routes/api/misc.test.ts`, `src/assets/loading.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`.

- Restore legibility: name the backup + date being loaded (and mark the recovery point). A cold restore showed only a generic "Restoring data from backup…" spinner, so an operator could not tell _which_ snapshot was loading. **(1) Durable recovery state.** `restoreIfNeeded`/`restoreBackupById` now write `restore-target.json` (the archive being applied) just before `sandbox.restoreBackup` and clear it on success; `readRecoveryState(bucket)` returns `{ restoring, committed }` (`committed` = `stable-backup.json`). It is read from **R2**, not the per-isolate `restoreStatus`, because the durable boot/restore runs in the `ActivityScheduler` DO whose in-memory state the request isolate cannot see. **(2) Surfaced on the loading page.** `/api/health` (non-passive, not-yet-ready) and `/api/startup/ws` attach `recovery` (bounded by `RECOVERY_STATE_TIMEOUT_MS`, fail-soft to null; lerma's passive poll never pays the read); `loading.html` renders "Restoring backup <id> — <date>" (or "Recovery point …") and its `isRestoring` also honours `recovery.restoring`. **(3) Dashboard.** `BackupsPage` threads `stableBackupId` to `BackupHistoryPanel`, which marks the recovery point with a **recovery point** badge + a "A cold boot restores <id> — <date>" summary. Tests: `src/persistence.test.ts` (`readRecoveryState`, restore-target publish/clear), `src/assets/loading.test.ts`, `src/client/components/__tests__/BackupHistoryPanel.test.tsx`, `src/routes/public.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`.

- Manual-restore hardening: recovery-point flag, writer quiesce, and no restore loop. Restoring an older archive from the Backups page could not be applied without either making it permanent or being silently undone. Three fixes in `src/persistence.ts` + the storage route/durable job. **(1) `setCurrent` flag.** `POST /api/admin/storage/restore/:id?setCurrent=0` (also `?async=1&setCurrent=0`) passes `{ setRecoveryPoint: false }` to `restoreBackupById` so the archive is applied **without** force-promoting it to `stable-backup.json` (the default still promotes; the durable `BackupJobRequest`/`BackupJobRecord` carry the flag so a retried job keeps the intent). The Backups page exposes it as a **Set as recovery point** checkbox (checked by default). **(2) Stop writers on the lease.** `restoreBackupById` now opens a cross-isolate maintenance window (`maintenance/replacement.json`, operation `restore`) and stops the gateway the moment it takes the shared restore lease (the route no longer pre-kills before the lock), closing the window in a `finally`; a live gateway/automation/channel therefore cannot keep writing while `sandbox.restoreBackup` replaces the writable layer. **(3) No remount loop.** `restoreIfNeeded` refuses to run while any maintenance window is active (and `ensureGateway` already refuses to launch), so a health/proxy refetch cannot restart the gateway and re-run the restore against the committed recovery point — the restore → stop → refetch → remount-the-latest loop. Tests: `src/persistence.test.ts` (set-recovery-point, maintenance window, deferred restore), `src/persistence-encryption.test.ts`, `src/routes/api/storage.test.ts`, `src/activity/scheduler.test.ts`, `src/activity/backup-jobs.test.ts`, `src/client/components/__tests__/BackupHistoryPanel.test.tsx`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`.

- Durable cold-boot job + slim health probe (observed on `tenant-client-dev`: `start-openclaw.sh` never ran; `phase=restore, age=686365ms`; 141 `waitUntil() tasks did not complete`; `/api/health?passive=1` at error level ×98). **(1) Durable boot.** The cold boot (R2 restore → gateway launch) was driven from a request's `ctx.waitUntil()`, which Cloudflare cancels shortly after the response — so a restore that outlived it was killed mid-flight (`Decrypting backup …` ×14) and `start-openclaw.sh` was never reached while every poll parked on the stuck start promise. It is now a durable job on the `ActivityScheduler` DO (`src/activity/boot-job.ts`): `/api/warmup`, non-passive `/api/health`, and `/api/startup/ws` only _request_ it (deduped) and return; the DO alarm performs `restoreIfNeeded` → `ensureGateway` and retries on a bounded lease. Falls back to the request-scoped boot when `ACTIVITY` is absent. **(2)** `sandbox.restoreBackup` is bounded by `RESTORE_BACKUP_TIMEOUT_MS` (a hung control channel fails fast instead of holding the `restore` phase for 11 min). **(3)** The restore reports sub-phase progress so `isGatewayStartStuck` judges `restore` on `GATEWAY_RESTORE_NO_PROGRESS_MS` of **no progress** (`setGatewayStartProgress`) — a slow-but-advancing restore is not killed. **(4)** `withLease` releases the shared R2 lease when its request aborts (`PersistenceDiagnosticContext.signal`), so a cancelled operation can't leave `restore-needed` + a held `backup-lock` re-triggering the loop. Also: `/api/health` now issues **one** gateway `/health` exec shared by readiness + liveness (was a duplicate curl pair per poll) and the **passive** path skips the container + R2 component probe (lerma's frequent probe never reads `components`). Tests: `src/activity/boot-job.test.ts`, `src/gateway/process.test.ts`, `src/gateway/health-probes.test.ts`, `src/routes/public.test.ts`, `src/persistence.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Slimmer `/api/health` (frequent lerma probe). Two per-poll costs removed. **(1)** Readiness and liveness each ran the same gateway `/health` `curl`, so every poll spawned a **pair** of identical `curl` processes (visible as paired entries in `/debug/processes`); `getHealthSnapshot` now fetches the process scan and the `/health` probe **once**, in parallel, and reuses both across the probes (new `HealthProbeOptions.health`). **(2)** The component probe (an `isRuntimeActive` read + an R2 HEAD of `backup-handle.json`) was built on **every** `/api/health`, including lerma's frequent `?passive=1` poll — which never reads `components`; the passive path now skips `componentsForHealth` entirely (the non-passive path and the dedicated `GET /api/health/components` are unchanged). Tests: `src/gateway/health-probes.test.ts` (single shared exec), `src/routes/public.test.ts` (passive omits components); corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`.

- Restore config guard + recovery-point promotion (observed on `tenant-tarjetero-dev`). After the `v2026.10.13` container rollout, an explicit restore of an older known-good backup did not stick: the redeploy changed the container runtime mid-restore, and a concurrent scheduled snapshot captured the partially-restored state and was auto-promoted to the stable recovery point, so the next cold boot restored the wrong archive and launched the gateway with no config. Three fixes: **(1)** `start-openclaw.sh` only honours `OPENCLAW_RESTORED=true` when `$CONFIG_FILE` exists — otherwise it runs the setup body — so a "restored" boot with no materialized config (interrupted restore, or `restoreIfNeeded`'s "container already holds tenant state" short-circuit) can no longer launch OpenClaw configless. **(2)** `restoreBackupById` now **force-promotes** the restored archive to `stable-backup.json` (`promoteStableBackup(bucket, entry, { force: true })`) so an explicit restore becomes the recovery point even when it predates the current one; the default pointer still only advances forward in time. **(3)** the pre-backup config cleanup used `.${CONFIG_DIR}` with `CONFIG_DIR='.openclaw'`, building `/home/openclaw/..openclaw/…` (a never-existent path, so the cleanup silently no-oped) — it now uses `${CONFIG_DIR}`. Tests: `src/persistence.test.ts` (force-promote; cleanup path); corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `apps/fouria/docs/OPERATOR-GUIDE.md`.

- Cold-start latency + boot legibility (a full non-restored boot on `staging-client-dev` took ~14 min). Two causes, both fixed, plus a loading-page alert. **(1) Non-owner killed the booting gateway.** A non-owner isolate that observed the `start-openclaw.sh start` wrapper waited a _fresh_ `STARTUP_TIMEOUT_MS` (300s) from its own observation, but the wrapper's real boot is `setup (~290s) + gateway bootstrap (~45s) ≈ 335s` — so the waiter timed out ~35s early and `kill()`ed the process just as it had `exec`ed into the gateway, releasing the startup lock and making the waiting second wrapper re-run the whole setup. Confirmed via Workers Observability: `Startup script present but gateway never became reachable … killing and restarting` = 1 while the warm-probe start/catch = 0/0 (ruling out the 30s warm probe). Fixed with `STARTUP_WRAPPER_STUCK_BUDGET_MS` (`STARTUP_TIMEOUT_MS + GATEWAY_START_STUCK_GRACE_MS`): the non-owner's stuck check and readiness wait are measured from the wrapper's `startedAt` (mirroring the owner's watchdog) and a `/health` re-probe runs immediately before any kill. Tests: `src/gateway/process.test.ts`, `src/gateway/process-extra.test.ts`. **(2) Plugin detection read the hanging registry.** `plugin_is_installed` ran `openclaw plugins list --json`, which hangs for its full `CLI_TIMEOUT` and reads as "not installed" on a hung/partial list, so the pre-installed guard fell through and triggered a needless network install (ai-gateway ~57s + otel ~74s). Replaced with `plugin_extension_present` (the `extensions/<id>` dir the gateway actually loads). **(3) Bootup-UI alert.** The loading page now shows a non-fatal slow-step notice (step name + elapsed seconds) once a step overruns `SLOW_STEP_MS` (120s), so a slow-but-progressing boot is distinguishable from a stuck one without opening logs. Tests: `src/assets/loading.test.ts`. Corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/plans/cold-start-latency.md`.

- The tenant **Usage & Billing** widget no longer 403s on Analytics Engine. `GET /api/admin/costs/v2/summary` resolved its AE SQL token as `ANALYTICS_ENGINE_API_TOKEN ?? CLOUDFLARE_API_TOKEN`, but Lerma never provisioned the dedicated token, so the query fell back to `CLOUDFLARE_API_TOKEN` — the AI Gateway credential alias, which lacks Account Analytics:Read — and logged `[cost-summary] Analytics Engine query failed with HTTP 403` on every load (observed on `tenant-client-dev`), zeroing the widget. The token is now a Lerma **platform secret** (`tenant_provisioner.ex` `@platform_secrets`), read from Lerma's environment and deployed per-tenant as a `secret_text` binding; the deploy path wires it through `mix lerma.deploy`, Ansible `app-config`, `docker-compose.prod.yml`/`.pre.yml`, and `deploy-dashboard.yml`, and the secrets manifest. Optional: when unset the worker still degrades to an all-zero summary. Tests: `tenant_provisioner_pipeline_test.exs`, `tenant_provisioner_integration_test.exs`; docs `docs/ENV-VARIABLES.md`, `apps/lerma/README.md`, `apps/fouria/docs/cost-observability.md`, `SECRETS.example.yaml`, `docs/schemas/secrets-manifest.schema.json`.

- Zero-data-loss: a restored container skips the setup body. The R2 restore runs in the Worker (`restoreIfNeeded` → `sandbox.restoreBackup()`) _before_ `start-openclaw.sh`; when that restore actually restored a backup, `ensureGateway` now passes `OPENCLAW_RESTORED=true` in the launch env (gated on `getRestoreStatus().hadBackup && restored`, so a brand-new tenant with no backup still onboards), and the script skips the whole setup body — onboard → plugin installs → `moltlazy patch` → config validate — emitting those phases as skipped and launching the gateway directly from the restored filesystem. The restored state is authoritative: the cold-start budget is spent launching it, not re-mutating it (which could diverge from or clobber the snapshot). Each boot rewrites a durable receipt at `/tmp/openclaw-restored-setup-skipped` (written on the skip path, removed on the setup path). Tests: `src/gateway/process.test.ts`; E2E `test/e2e-backup/zzz_backup_restore_skip_setup.txt` (opt-in `E2E_BACKUP_RESTORE=1`; sync → destroy → restore, asserts the receipt + unchanged restored setup state); corpus `test/business-rules/operational-modes.txt`; docs `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `test/e2e-backup/README.md`, `test/e2e/BACKUP-E2E-DESIGN.md`.

- The dashboard **Usage & Billing** widget now reflects emitted billable actions. `GET /api/admin/costs/v2/summary` was a stub returning `totalCostUsd: 0`; it now aggregates the tenant's `billable-action` Analytics Engine rows via the AE SQL API (`src/cost-summary.ts` → `getCostSummary()`), returning `totalCostUsd`, `totalCredits`, `actionCount`, `byTriggerType`, and `byCostCategory`. Config mirrors Lerma: `ANALYTICS_DATASET` (default `fourai_cost_metrics`), `ANALYTICS_ENGINE_BASE_URL`, and `ANALYTICS_ENGINE_API_TOKEN` (falls back to `CLOUDFLARE_API_TOKEN`; needs Account Analytics:Read). A missing credential/transport error/non-200 degrades to an all-zero summary, so the widget never breaks. The widget surfaces the billable action/credit rollup when data is present. Tests: `src/cost-summary.test.ts`, `src/routes/api/costs.test.ts`, `src/client/pages/__tests__/HomePage.test.tsx`, `src/client/pages/__tests__/AccountPage.test.tsx`, `test/ui/billing-widget.spec.ts`; corpus `test/business-rules/dashboard-api-contract.txt`; docs `apps/fouria/docs/cost-observability.md`, `docs/ENV-VARIABLES.md`, `docs/TESTING.md`.

- The version surfaces now report the Worker's own build identity. `GET /api/admin/version` and `GET /debug/version` return `fouria_version` (component CalVer from `apps/fouria/package.json`) and `fouria_commit` (short git SHA) alongside the container's OpenClaw/Node versions; the DiagnosticDashboard **Version Info** tool surfaces them. Both values are injected at build time by `vite.config.ts` (`__FOURIA_VERSION__` / `__FOURIA_COMMIT__`, consumed via `src/build-info.ts`; `GITHUB_SHA` wins in CI, `git rev-parse --short HEAD` locally). Tests: `src/routes/debug.test.ts`, `src/routes/api/misc.test.ts`; docs `docs/WORKER-DEVELOPMENT.md`, `docs/OPERATOR-GUIDE.md`.

- Expected CLI non-zero exits are no longer logged as failures. `waitForProcess` (like `runShell`) now takes `reportNonZeroExit`; the knowledge `wiki status` probe sets it `false` (operation `knowledge.wiki-status`) because `openclaw wiki status` exits non-zero for an uninitialized vault — which `/api/admin/knowledge/tree` handles as a normal `initialized: false` result — yet every load emitted `Error [SandboxProcessError]: sandbox.waitForProcess failed (non-zero-exit, exit=1)`. Timeouts still log. Tests: `src/gateway/utils.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`.

- The prelaunch gateway port probe no longer logs a false failure. `isGatewayPortOpen` ran `curl -sf …/health` through `runShell`, whose `logCommandFailure` emitted a structured `sandbox.process.failed` **error** for the expected closed-port exit (`curl` 7 = connection refused). Every cold start therefore surfaced a scary `Error [SandboxProcessError]: shell failed (non-zero-exit, exit=7)` for the normal "gateway down, start it" outcome. It now passes `reportNonZeroExit: false` with `operation: 'gateway.port-probe'` (matching the existing `gateway.health-probe`), so a closed port is quiet while a probe **timeout** (a hung gateway) still logs. Test: `src/gateway/process.test.ts` (`does not log a structured failure for the expected closed-port exit`); corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`.

- OpenClaw is installed globally with **bun**, not npm. `bun install --global --trust openclaw` runs with `BUN_INSTALL=/usr/local`: bun's default global prefix is `$HOME/.bun/bin`, which is not on `PATH` in this image, so `BUN_INSTALL=/usr/local` puts the binary at `/usr/local/bin/openclaw` (on `PATH`) and the package under `/usr/local/install/global/node_modules/openclaw`. `bun pm cache rm` resolves the current project and hard-fails without a `package.json` — the base image's `WORKDIR /container-server` has none — so it runs from `/tmp` (which holds the copied repo `package.json`). The dead-weight trim now targets bun's hoisted global layout (`<prefix>/install/global/node_modules`), and `start-openclaw.sh` resolves the package by following the installed binary (`command -v openclaw` + `readlink -f`) instead of `npm root -g`, so the runtime-version log stays accurate.

- `post-start` no longer blocks readiness. After the port wait succeeds the gateway is already serving `/health` 200, but `ensureGatewayImpl` still `await`ed `start-openclaw.sh post-start` (dev automation + boot-log reset, bounded by `GATEWAY_POST_START_TIMEOUT_MS` = 60s). On dev instances that phase runs `openclaw automations list/create` (30–60s), so every request logged `Gateway start already in progress … (phase=post-start)` and waited on the shared start promise even though the gateway was up. It is now dispatched **detached** (`void runGatewayPostStart(...)`, not awaited) right after `markGatewayReady`; `startShell` dispatches the `sandbox.exec` synchronously before its first `await`, so the work still runs server-side if the isolate is evicted, and `runGatewayPostStart` swallows its own errors. The `post-start` phase is no longer set on the start tracker (the start is complete). Test: `src/gateway/process.test.ts` (`does not extend the shared start promise with the detached post-start phase`); docs `docs/WORKER-DEVELOPMENT.md`.

- ClawHub plugin pre-install is best-effort so the image still builds. The unpinned `clawhub:@openclaw/cloudflare-ai-gateway-provider` resolved to `2026.9.9`, which requires a newer plugin API than the pinned `2026.9.8` runtime, so `openclaw plugins install` failed and aborted the whole Docker build (`Plugin … requires plugin API >=2026.9.9, but this OpenClaw runtime exposes 2026.9.8`). The Dockerfile now wraps each pre-install in a warn-and-continue guard, and `start-openclaw.sh`'s single enablement `config patch` only enables a plugin whose extension directory is present (`CF_AI_GATEWAY_PLUGIN_PRESENT` / `UNIFIED_BILLING_PLUGIN_PRESENT`), so a failed install can neither fail the build nor invalidate the patch. Root-cause follow-up: bump `catalog.openclaw` to the plugin-compatible release (openclaw-migration skill) or pin the specs; until then a build on the pinned runtime ships without the Cloudflare provider plugins. Corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`, `docs/OPENCLAW-PLUGINS.md`.

- Gateway start watchdog + process-count hygiene. A cold start is one shared `ensureGateway` promise (restore → supervised `start-openclaw.sh start` → port wait); a wrapper that never `exec`s the gateway, or a control-channel drop, left it "in progress" for ~20 min while every request logged `Gateway start already in progress … waiting...` and waited on it (observed on `tenant-client-dev`: 64 processes, port closed, curl health probes exiting 7 every ~5s, `[PROXY] Failed to start gateway` stack-only). The start is now tracked per sandbox key with a phase (`probe` → `restore` → `launch` → `port-wait`; `post-start` runs detached after readiness, not tracked) and a per-phase budget; `isGatewayStartStuck` trips when a phase overruns its budget + `GATEWAY_START_STUCK_GRACE_MS`. Before resetting `port-wait`, `recoverStuckGatewayStart` rechecks `/health`: a late-positive response emits `gateway.start.late-ready` and is adopted without killing, while a failed/unknown recheck emits `gateway.start.health-recheck-failed` and defers reset. Confirmed stuck starts emit `gateway.start.stuck`, kill the process (`killGateway`), clear startup/readiness caches, and reprovision after `GATEWAY_START_STUCK_MAX_ATTEMPTS` resets (`recoverStaleSandbox`). `findExistingGatewayProcess` also uses the pure `classifyGatewayStart` to kill a `start-openclaw.sh start` wrapper older than `STARTUP_TIMEOUT_MS` with the port closed, instead of waiting out another readiness budget. Because `listProcesses()` includes `exited`/`error` entries the Worker never GCs (the SDK exposes no cleanup), the raw count is append-only and must not gate health: the log now reads `listProcesses returned <total> processes (<running> running)` and every decision considers only `running` entries. Emits `gateway.start.begin`/`phase`/`stuck`/`late-ready`/`health-recheck-failed`. Tests: `src/gateway/classify.test.ts`, `src/gateway/process.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`, `docs/OPERATOR-GUIDE.md`.

- Legacy bridge excludes the image-provided shared OpenClaw package. `extensions/<id>/node_modules/openclaw` is a symlink to the image's `/usr/local/lib/node_modules/openclaw`; archiving it is pointless and (when an old dereferencing stage expanded it into a full package copy) made the restore materialization a multi-minute copy that blocked the gateway launch, left `restore-needed` set, and re-ran the restore on every boot (observed on `staging-client-dev`: `cp -a … /root/openclaw/.openclaw/` running 8m45s, 65 Worker probe processes, no `start-openclaw.sh`/gateway). Both `stageLegacyRootsForBackup` and `materializeLegacyRootsAfterRestore` now use a symlink-preserving `tar` pipe with `--exclude='*/node_modules/openclaw'` (a copy-then-`rm -rf` is equally slow — deleting a squashfs-lower tree requires copy-up), so old and new archives restore fast. `restoreIfNeeded` also deletes the `restore-needed` marker immediately after a successful materialization (before the analytics bookkeeping) so an isolate eviction can no longer leave a stale marker. Tests: `src/persistence.test.ts`; corpus `test/business-rules/operational-modes.txt`.

- Legacy restore bridge preserves symlinks (backup/restore fix). The legacy state bridge staged `/root/openclaw/.openclaw` with `cp -aL`, which dereferenced the plugin `extensions/*/node_modules/openclaw` symlinks (→ `/usr/local/lib/node_modules/openclaw`) into full OpenClaw package copies. That bloated every archive (the package three times over), timed out the pre-snapshot stage (600s) — surfacing to the dashboard as a backup-create `Internal Server Error` — and made restore fail because `cp` cannot overwrite the image's symlink with a directory (`cp: cannot overwrite non-directory … with directory`), leaving the tenant looping on a failed restore (`staging-client-dev`). The stage and the restore materialization now use `cp -a` (preserve symlinks); the restore additionally removes any destination symlink the stage replaces with a real directory first, so backups created by the old dereferencing stage still restore. Tests: `src/persistence.test.ts`; corpus `test/business-rules/operational-modes.txt`. Also: the admin API client (`src/client/api.ts`) now reads the response body once and surfaces a plain-text platform error (e.g. `Internal Server Error`) instead of throwing `Unexpected token 'I', "Internal S"... is not valid JSON`; tests `src/client/api.test.ts`.

- Recurring-error hardening: credential redaction, bounded-RPC cleanup, live-gateway logs, and idempotent WS teardown. Investigation of the recurring `ReadableStream received over RPC disconnected prematurely` / `[WS] Client error` telemetry found four concrete causes. (1) **Credential logging:** the catch-all proxy logged the full forwarded header map (`Object.fromEntries(httpHeaders.entries())`), writing CF Access JWT/assertions, cookies, and the tenant metadata blob to Workers Logs; it now logs only an allowlisted subset (`safeHeaderLog`, `src/utils/logging.ts`). (2) **Late container RPCs:** a bounded `containerFetch`/`wsConnect` whose timeout won could still settle later with a live body/socket that was never drained/closed; `withTimeoutAndCleanup` (`src/utils/timeout.ts`) plus `containerFetchWithTimeout`/`wsConnectWithTimeout` (`src/utils/container-rpc.ts`) cancel the late body / close the late socket, applied to the public proxy, diagnostics, and the `gateway-api` probes. (3) **Live gateway logs:** a running gateway's SDK log stream never settles, so reading it opened an RPC stream that was later disconnected; a running `GatewayProcess.getLogs()` now tails `/tmp/gateway-stdout.log`, and `/debug/processes?logs=true` / `/api/admin/logs?id=` refuse a live stream for a still-running process. (4) **WS teardown:** both peer close handlers recorded the same session metric twice, and a client tab disconnect was logged at `error`; a single `settleSession()` now records the metric once and client disconnects log at `warn` with the `sessionId`. Tests: `src/utils/container-rpc.test.ts`, `src/utils/timeout.test.ts`, `src/logging.test.ts`, `src/index.test.ts`, `src/gateway/process.test.ts`, `src/routes/api/misc.test.ts`, `src/routes/debug.test.ts`; local observability canary `test/e2e/ad_observability_canary.txt` + `test/e2e/fixture/observability-query`; corpus `test/business-rules/operational-modes.txt`; docs `docs/TESTING.md`.

- Two local-E2E self-skip fixes. `provision_and_dispatch`'s first scenario (`%require`) echoed a `SKIP:` line and `exit 1` in local mode, so it _failed_ instead of skipping — and because it is `%require`, that failure skipped every later scenario in the file. It now emits the same `PASS: … (skipped - local E2E mode)` line and `exit 0` the other local skips in the file use. `p_otel_billing_queue` asserted `emitted >= 1`, but a dev-environment/staging-platform instance is non-billable by design (`resolveInstanceDimensions`, `src/routes/otel.ts`): the receiver accepts + ignores approved spans (200) and returns `billable: false` without emitting. The local worker (`DEV_MODE=true`, no routable `WORKER_URL`) is always non-billable, so the test now skips on `billable == false` (like its existing missing-secret/`401` skip); the emit path stays asserted by the cloud profile and `src/routes/otel-traces.test.ts`. Corpus: `test/business-rules/operational-modes.txt`, `test/business-rules/dev-expense.txt`; docs `docs/TESTING.md`.

- Local E2E readiness gate accepts a reachable-but-starting Worker. `fixture/server/start-local` waited for `/api/health` with `curl -sf -m 5` (60×2s), but a cold boot legitimately answers **503** while the container/gateway starts (the non-waking-health semantics) and the non-passive probe can block on container provisioning — so the gate misread a live-but-starting Worker as dead. Worse, the non-passive probe on every 2s poll re-triggered `restoreIfNeeded`/`ensureGateway`, launching overlapping R2 restores that abort each other with `OperationInterruptedError: ... runtime changed`. The gate now probes the fast, **non-waking** `GET /api/health?passive=1` and accepts **any** HTTP status (`200`/`503`/…) via the shared `e2e_worker_reachable` helper (`test/lib/e2e-helpers.sh`), failing only on a connection failure (status `000`); the window is overridable with `E2E_STARTUP_TIMEOUT_SECS` (default 120s) and the "local dev server died" check still fails fast. The real gateway wait (`startup.ok` + `readiness.ok`) stays in the `_setup.txt` readiness step. Tests: `src/e2e-helpers.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `docs/TESTING.md`.

- Restore failures are logged as structured `persistence.restore.failed` events with `backupId`, `errorName`, and `errorMessage`; manual route failures also include the Cloudflare Ray ID as `storage.restore.failed`. This keeps the backup identity searchable and correlates async failures to the originating request. Tests: `src/persistence.test.ts`, `src/routes/api/storage.test.ts`.

- Backup retention is enforced on catalog reads, and restores survive a failed pre-restore `umount`. Live `fouria-backup-tarjetero-dev` held 7 archives under `maxVersions: 3`: every snapshot was killed after `sandbox.createBackup()` (260 MB) but before its commit, so a later read adopted all 7 as orphans, and every scheduled snapshot after that failed at the legacy stage before its own prune — the cap was only ever applied on the happy snapshot path. `reconcileOrphanedBackups` now runs `cleanupOldBackups` after merging, so the invariant holds on every read (`getBackupHistory` included) and no write/delete happens when the catalog is already within the cap. Restores failed in a loop on `[persistence] Restore of … failed: … runShell … restoreIfNeeded`: the best-effort `umount` (`|| true` guards the shell, not the exec RPC) threw a cold-start/platform interruption and aborted the restore before `restoreBackup`, keeping `restore-needed` and failing the startup barrier. The umount is now caught + logged in `restoreIfNeeded` and `restoreBackupById`; an empty catalog read no longer deletes `restore-needed` (a transient reconcile `list` failure silently dropped an operator restore) — the marker clears only after a successful restore. Tests: `src/persistence.test.ts` (over-cap prune, adopted-orphan prune, no-delete within cap, marker retention, umount-failure restore); corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`, `docs/OPERATOR-GUIDE.md`.

- Async admin backup/restore + non-blocking Backups page. A manual snapshot/restore can wait up to the 120s R2 lease and then transfer a large archive; holding the admin HTTP request open surfaced as a bare `500` after ~132s (restore) with the error dropped. `POST /api/admin/storage/sync?async=1` and `POST /api/admin/storage/restore/:id?async=1` now acknowledge `202 { accepted: true }` and run in `executionCtx.waitUntil` (default stays synchronous for the E2E checkpoint). `BackupsPage` paints every section independently and reads the fast `GET /api/admin/config/raw` instead of the slow structured `/api/admin/config` (which awaits `ensureGateway` + up to two 30s shell waits) — so no single request blocks the page. `BackupHistoryPanel` fires the async endpoints, pushes a "started" bell notification, then polls `GET /storage/backups` (snapshot) / `/api/health` (restore) and pushes success/failure via `NotificationProvider` (survives navigation). Restore failures are reported as structured `storage.restore.failed` events. Tests: `src/routes/api/storage.test.ts`, `src/client/components/__tests__/BackupHistoryPanel.test.tsx`, `src/client/pages/__tests__/BackupsPage.test.tsx`; corpus `test/business-rules/operational-modes.txt`; docs `docs/WORKER-DEVELOPMENT.md`.

- Startup-stream and dashboard polling back-off. `/api/startup/ws` now drops the 2-exec startup-progress read and backs off to `STARTUP_WS_READY_POLL_INTERVAL_MS` (30s) once the gateway is ready — the 5s cadence plus a per-push `listProcesses`/exec probe burst is a large share of the Sandbox DO control-channel churn (`Establishing RPC connection`) during a boot. `useGatewayHealth` now stops its HTTP fallback polling when a `/api/health/ws` snapshot arrives (it previously kept double-polling after the socket recovered). `transformErrorMessage` maps the gateway's `1008 device token mismatch` close to a "re-open `/_admin/` to re-approve this browser" message instead of a reconnect loop. Tests: `src/client/hooks/useGatewayHealth.test.tsx`; corpus `test/business-rules/operational-modes.txt`.

- Backup local-bucket mode is decoupled from `DEV_MODE`. `isLocalBucketDevMode()` keyed `createBackup({ localBucket })` on `process.env.DEV_MODE`, and with `nodejs_compat_populate_process_env` the cloud E2E's `DEV_MODE=true` var (set only to bypass Access on a bare `workers.dev` host) made every E2E snapshot land in the Sandbox **local binding bucket** — invisible to the Worker's `BACKUP_BUCKET` HEAD, so restore logged `Backup <id> missing from R2` and booted without tenant state. It is now `isLocalBucketBackupMode(env)`: an explicit `BACKUP_LOCAL_BUCKET=true` opt-in that falls back to `DEV_MODE` only when unset; the E2E deploy pins `BACKUP_LOCAL_BUCKET=false`. `resolveRestoreEntry` threads `env` through. Tests: `src/persistence.test.ts` (explicit override wins); corpus `test/business-rules/operational-modes.txt`; docs `docs/ENV-VARIABLES.md`, `docs/OPERATOR-GUIDE.md`, `docs/WORKER-DEVELOPMENT.md`, `.dev.vars.example`.

- A dropped port probe no longer spawns a duplicate gateway. When the DO↔container control channel flapped (`Network connection lost.` on every `exec`), `ensureGatewayImpl`'s prelaunch `isGatewayPortOpenResilient` throw was treated as "gateway absent", so it re-ran `start-openclaw.sh start` on each request even though `listProcesses` matched the gateway `running` — piling lock contention and RPC load onto the unhealthy channel. A transport-classified probe failure is now "unknown": the Worker logs `Port probe transport drop …` and returns without spawning, retrying on the next request. Test: `src/gateway/process.test.ts` (`ensureGateway transport-drop resilience`); corpus `test/business-rules/operational-modes.txt`.

- E2E `ac_gateway_startup_concurrency` accepts the 503 cold-boot health response. The concurrent non-passive `/api/health` fan-out used `curl -f`, which fails on any non-2xx, but `/api/health` answers **503 while `starting`** (JSON snapshot present) since the non-waking-health change — so a healthy cold boot always reported `probes_ok=0`. The probes now accept `200`/`503` with a shaped `{startup, readiness}` body; the ready-wait and single-owner assertions are unchanged. Docs/corpus updated.

- Health is not a routing gate; ROOT operators can always reach a tenant. Tenant containers are loaded lazily by the User Worker, so a failed health probe is not an outage — yet the dispatch route gate treated `clients.status`/`instances.status = error` as non-operational (and lerma's `HealthMonitor` / `HealthPoller` wrote that `error`), which locked operators out of a reachable tenant (observed on `tenant-tarjetero-dev`). `workers/fouria-dispatch` `tenantRouteGate` now never blocks on a health-derived `error` (production client or dev instance), keeps `frozen`/`stopped` gated only for **non-ROOT** callers (a ROOT operator is exempt so they can revive a stopped tenant), and runs **after** authentication so the exemption can be applied. `apps/lerma` `HealthMonitor` and `TenantProvisioner.verify_health` are now **observability-only** on failure: they record the down transition for the instance uptime widget but never change instance/client status (a successful probe still marks the instance `running` and recovers a client left in `error` by a provisioning failure). The now-meaningless `HEALTH_MONITOR_FAILURE_THRESHOLD` env var is removed. Tests: `workers/fouria-dispatch/test/integration/resolution.test.ts`, `apps/lerma/test/lerma/provisioning/health_monitor_test.exs`, `tenant_provisioner_pipeline_test.exs`; corpus `test/business-rules/operational-modes.txt`; docs `workers/fouria-dispatch/README.md`, `docs/CLOUD-ARCHITECTURE.md`, `docs/ENV-VARIABLES.md`.

- Sandbox→R2 bridges resolve the tenant from the per-tenant Worker env, not the fixed sandbox id. The `leads.r2.internal` and `docs.r2.internal` bridges (`src/sandbox-leads-bridge.ts`, `src/sandbox-docs-bridge.ts`) scoped every path to the caller's `<slug>/` using `ctx.containerId`, but production provisions the sandbox under the fixed id `openclaw` (`getSandbox(..., 'openclaw')` in `src/index.ts`, `src/tasks/index.ts`, `src/activity/scheduler.ts`), so listings resolved to `openclaw/` and always returned empty. Added `resolveTenantSlug` / `tenantSlugFromEnv` in `src/sandbox-tenant.ts`: an explicit `customer-<slug>` container id still wins (legacy multi-tenant), otherwise the slug is derived from the per-tenant Worker env — `CUSTOMER_SLUG`, `WORKER_NAME` via `deriveCustomerSlug`, or the always-provisioned `WORKER_URL` hostname (strips `tenant-` / `staging-` prefixes and the `-dev` suffix, so `tenant-tarjetero-dev` / `https://tarjetero-dev.fouria.io` → `tarjetero`). Tests: `src/sandbox-tenant.test.ts`, `src/sandbox-leads-bridge.test.ts`, `src/sandbox-docs-bridge.test.ts`; corpus `test/business-rules/tenant-isolation.txt`; docs `docs/WORKER-DEVELOPMENT.md`.

- Dev-agent personas source the client proposal via the `fouria-leads` skill. The Fouria Builder and Q&A personas (`packages/moltlazy/workspace/fouria-builder.ts` + `fouria-qa.ts`) now read the skill first and fetch the proposal through the sandbox `leads.r2.internal` bridge (global `paso4-leads` R2 bucket) instead of the legacy `/home/openclaw/clawd/client-docs/` path; the Worker's dev-only docs materialization (`materializeClientDocsIfDev`) is no longer the personas' input path. Docs updated alongside (`packages/moltlazy/src/agents/README.md`, `packages/moltlazy/AGENTS.md`, `docs/DEVELOPMENT.md`).
- Crash-consistent backup cataloging (issue #344). The scheduled `backup` control action runs the full snapshot orchestration inside one tenant request and could exceed the Worker CPU budget after `sandbox.createBackup()` had uploaded `backups/<id>/…` but before the handle was written — leaving an orphaned snapshot, a missing `backup-handle.json` (and empty `configs/`), a live `backup-lock`, and a dispatch `[cron] Failed to trigger backup`. `createSnapshotUnlocked` now commits the catalog entry **immediately after the archive is created** (after encryption, before the config-version capture and retention prune), and `reconcileOrphanedBackups` rebuilds entries from `backups/*/meta.json` on read (`getBackupHistory`, `restoreIfNeeded`, `scheduleBackupIfNeeded`) so pre-existing orphans self-heal and stay restorable. The Dispatch Worker gives the `backup` action a dedicated CPU budget (`BACKUP_CONTROL_CPU_MS`) and logs the underlying error on a thrown control fetch. Tests: `src/persistence.test.ts` (commit ordering, config-save failure, reconciliation), `workers/fouria-dispatch/test/unit/control.test.ts`, `cron-fanout.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `docs/OPERATOR-GUIDE.md`, `docs/WORKER-DEVELOPMENT.md`.

- Transport-resilient backup stage (issue #344 follow-up). The pre-snapshot legacy-stage shell now waits via `runShellResilient` (→ `waitForExitResilient`): a transient Sandbox DO↔container transport drop mid-stage reattaches to the process by id and resumes within the stage deadline, instead of ending the `/control` backup request (observed on `staging-client-dev`: the stage ran ~5 min then the request ended with a `Network connection lost.` drop and no `Creating backup...`). The Dispatch `backup` CPU budget (`BACKUP_CONTROL_CPU_MS`) is raised 300s → 600s to match the 10-minute stage bound. Tests: `src/persistence.test.ts` (transport-drop retry); corpus `test/business-rules/operational-modes.txt`; docs `docs/OPERATOR-GUIDE.md`, `docs/WORKER-DEVELOPMENT.md`.

- Legacy-stage backup robustness + diagnosability (issue #344 follow-up). The scheduled `backup` stage shell (`stageLegacyRootsForBackup` → `runLegacyStageShell`) copies `/root/openclaw/.openclaw` + `/root/clawd` with `cp -aL` and runs a `node:sqlite` online backup; on a large tenant it exceeded the old 300 s bound and killed the backup (`tenant-tarjetero`, observed at exactly 300 s). `LEGACY_STAGE_TIMEOUT_MS` is now env-overridable (default `600000` / 10 min), the idempotent stage is retried **once** on a transient timeout/transport failure, and stage failures log `exit`/`timedOut`/stderr plus a structured `sandbox.process.failed` Worker event with `processId` and a stable operation label. The process id joins Worker Logs to the Containers dataset without exposing the raw command/output. Tests: `src/persistence.test.ts` (retry, no-retry, timeout override, error logging); corpus `test/business-rules/operational-modes.txt`; docs `docs/OPERATOR-GUIDE.md`, `docs/WORKER-DEVELOPMENT.md`, `docs/ENV-VARIABLES.md`.

- Sleeping tenants + non-waking health. `/api/health` now answers **503 unless the gateway is ready**, and a **passive** probe (`?passive=1`) short-circuits on the non-waking `isRuntimeActive()` read and reports `status: "sleeping"` (HTTP 200, healthy-idle) for an idle container instead of running the exec RPCs that auto-start it — so Lerma's `HealthMonitor` (which now probes `?passive=1`) never wakes a tenant. Non-billable tenants (staging 15m, dev 2h) sleep per a **derived** idle timeout (the Sandbox SDK 1.0 `setInactivityTimeout()` platform cap; there is no `SANDBOX_SLEEP_AFTER` binding; the value comes from the instance hostname via `src/dev-mode.ts`), and the `ActivityScheduler` DO alarm wakes/heals ahead of the next automation and runs an interval-gated R2 checkpoint after the wake. **Billable production** (`platform === 'production' && environment === 'production'`) does **not** auto-sleep; the container stays awake and is stopped only by an explicit lerma freeze (`POST /control` → `sandbox.destroy()`), so a warmed gateway never incurs a cold-start/downtime window after an idle period. Runtime cutover is gated by `SANDBOX_RUNTIME` (`legacy` default | `v1`); `SandboxV1` is still scaffold-only, so `resolveSandboxRuntime()` reports `v1` as `pending` and the legacy class keeps serving (`src/sandbox-runtime.ts`). Tests: `src/routes/public.test.ts`, `src/gateway/health-probes.test.ts`, `src/activity/alarm.test.ts`, `src/sandbox-runtime.test.ts`, `src/sandbox-do.test.ts`, `src/build-sandbox-options.test.ts`; corpus `test/business-rules/operational-modes.txt`; lerma `health_monitor_test.exs`, `health_poller_test.exs`, `tenant_provisioner_pipeline_test.exs`. See [docs/WORKER-DEVELOPMENT.md](/latest/llms/platform/apps/fouria/worker-development/index.md).

- Single-owner gateway launch (`start-openclaw.sh start`): the production cold-start mode acquires the container-wide startup lock **before** config setup and holds fd 9 through the Gateway `exec`, so one owner covers setup **and** launch. The lock is inherited by the gateway process and released exactly when it exits, so only **one** owner can run — the OpenClaw 2026.9.8 state-DB single-owner migration otherwise caused concurrent launches (from different Worker isolates) to lose the migration, exit before binding the port, and leave the boot forever unhealthy (`state.sqlite is undergoing offline maintenance`). Losers poll the local `/health` (bounded) and exit 0 without a duplicate launch, or take over if the owner dies and releases the lock; the log-mirror `tail -f` closes fd 9 (`9>&-`) so it cannot hold the lock. The `launch` mode remains a standalone smoke/manual entry point with its dedicated `/tmp/openclaw-gateway-launch.lock`; `full` routes through the same lock in `background` mode. Regression: the serial `test/e2e/ac_gateway_startup_concurrency.txt` (`phaseD_startup`) asserts exactly one owner and the absence of both the migration-input-change error and the SQLite offline-maintenance error. See [docs/WORKER-DEVELOPMENT.md](/latest/llms/platform/apps/fouria/worker-development/index.md).
- Fixed the develop cold-boot stall introduced by #318 (`startup.ok=false / readiness.ok=false` for the whole E2E readiness budget; 68 processes, `start-openclaw.sh setup` exited 0, `curl /health` always exit 7, **no** `openclaw gateway run` ever present). Root cause: #318 had split the boot so the **Worker** launched the gateway _after_ `start-openclaw.sh setup` exited (supervised exec) — this was the former behavior; current production collapses setup and launch into the atomic `start-openclaw.sh start` described above (startup lock acquired before setup, held through the Gateway `exec`). `/api/health` is bounded to a 5s snapshot, so the request returned long before the ~24s setup finished; when the response was sent, workerd cancelled the unregistered `ensureGateway` continuation, so the container-side setup completed but the Worker never resumed to launch the gateway. Fix: `/api/health` (and `/api/warmup`) now register the in-flight boot with `c.executionCtx.waitUntil(ensureGateway(sandbox, env, { waitForReady: true }))` so the setup → gateway-launch sequence survives past the response; `waitForReady: true` returns the shared start promise unwrapped and dedupes onto the one the snapshot already started. Also hardened the RPC-first health probe (`getRpcHealthSnapshot`, `src/gateway/health-probes.ts`) to run only once a gateway process is detected — `sandbox.containerFetch` auto-starts the container and blocks on the closed gateway port during a cold boot, piling port-waits on the Sandbox DO control channel. `buildHealthSnapshot` (`src/routes/public.ts`) passes its already-fetched `findGatewayProcess` result as `{ proc }`. Bisected: production E2E boot green at `ad9ad4a5` (#317), red at `9ae5f66c` (#318). See `apps/fouria/docs/WORKER-DEVELOPMENT.md` §Admin RPC.
- Startup resilience to Sandbox DO code updates: a Worker/DO redeploy (e.g. a Lerma redeploy of a tenant script) resets the DO↔container control session while the container keeps running, so the supervised `waitForPort` on the atomic `start-openclaw.sh start` process rejected with `Network connection lost.` and the cold start was aborted — killing a gateway that would have completed. The atomic `start` path has **no separate setup-process wait** (one supervised process spans config setup through the gateway's whole life), so only the port wait is transport-resilient: `ensureGateway` waits with `waitForPortResilient` (`src/gateway/sandbox-exec.ts`), which reattaches to the process by id via `getProcess(id)` and resumes the wait within the shared `STARTUP_TIMEOUT_MS` deadline; only a non-transport error, an un-reattachable process, or an exhausted deadline fails. The generic `waitForExitResilient` helper remains in `sandbox-exec.ts` for the legacy/boot-smoke supervised waits. The loading page's `GET /api/startup/ws` stream also treats a transport-drop snapshot as transient: it re-sends the last known state (or a synthetic `starting` view) and backs off exponentially (`STARTUP_WS_POLL_INTERVAL_MS` → `STARTUP_WS_MAX_BACKOFF_MS`, 30s) instead of re-probing and re-logging the same error every 5s.
- Backups run only on running instances; health is observability-only. `GET /api/health` no longer schedules snapshots (passive or active) — the lerma `HealthMonitor` is a pure probe and MUST NOT alter the R2 catalog. The **Dispatch Worker cron** (`workers/fouria-dispatch`, 15 min prod / hourly staging) fans a `backup` control action (`POST /control`, `X-FourIA-Control: backup`) out to every non-frozen tenant; the tenant asserts `sandbox.isRuntimeActive()` (a **non-waking** bridge read) and skips a sleeping container rather than waking it (an idle instance has no workload or filesystem changes, and waking would restart billing), then runs the interval/lock-gated `scheduleBackupIfNeeded`. `POST /debug/reprovision` now checkpoints only a live runtime, bounded by `REPROVISION_CHECKPOINT_TIMEOUT_MS`, verifies restorability by HEADing the newest catalog entries' `meta.json`, and **fails closed** (`409`) unless a restorable backup exists; `{"force":true}` overrides for an already-unrecoverable container. Root cause of the empty buckets: the old health-triggered path scheduled a snapshot regardless of `readiness.ok`, so a booting/stopped container could start `sandbox.createBackup()`, hang, and leave a stale `backup-lock` (observed in `fouria-backup-tarjetero`); dev instances the HealthMonitor does not monitor had no trigger at all. Tests: `src/routes/public.test.ts`, `src/routes/debug.test.ts`, `src/index.test.ts`, `workers/fouria-dispatch/test/unit/control.test.ts`; corpus `test/business-rules/operational-modes.txt`; docs `docs/OPERATOR-GUIDE.md`, `apps/fouria/docs/WORKER-DEVELOPMENT.md`, `docs/CLOUD-ARCHITECTURE.md`, `workers/fouria-dispatch/README.md`. E2E: `test/e2e-backup/zz_reprovision_guard.txt` + `test/e2e-backup/zzz_backup_cold_restore.txt` + `test/e2e-backup/zzz_backup_semantics.txt` (separate `test/e2e-backup/` corpus, excluded from the default `test/e2e` run; run locally with `bun run test:e2e:backup`; registered in `phaseE_restore`, reprovision guard first), the dispatch fan-out unit test `workers/fouria-dispatch/test/unit/cron-fanout.test.ts`, deterministic E2E backup env in `test/e2e/fixture/server/deploy`, and `POST /_control/:slug/backup` on the dispatch worker; design in `test/e2e/BACKUP-E2E-DESIGN.md`. Dispatch note: `vitest.config.ts` now has `unit` (node) + `integration` (workerd) projects so the fan-out logic is testable without Miniflare.
- PR E2E scope: the `e2e` job now sets `E2E_BOOT_ONLY=1` for **pull requests**, so `run-parallel.sh` provisions the worker and only asserts the gateway reaches `startup.ok + readiness.ok`, then tears down without running any cctr group. `develop` and a manual `workflow_dispatch` (new `full_tests` input, default checked) run the full corpus. See `apps/fouria/test/e2e-parallel/run-parallel.sh` and `docs/TESTING.md`.
- Paso4-exclusive instances + agent: lerma can now provision internal `scope: "paso4"` instances (nullable `client_id`, no billing plan, `@paso4.io`-only Access), and moltlazy ships an opt-in `paso4-dev` agent (`PASO4_AGENTS`, gated by `MOLTLAZY_PASO4_AGENTS=true` and named in `MOLTLAZY_AGENTS`) that picks up GitHub issues via the bundled `github` skill (`gh` CLI, installed in the image) authenticated by `GITHUB_TOKEN`. The Dispatch Worker resolves paso4 tenants from `instances` (no `clients` row) and rejects non-`@paso4.io` emails. See [apps/lerma/README.md](/latest/llms/platform/apps/lerma/index.md#paso4-exclusive-instances) and `docs/ENV-VARIABLES.md`.
- Setup serialization (`start-openclaw.sh`): `start`, `setup`, and `full` take an exclusive `flock` on `/tmp/openclaw-setup.lock` (~900s wait; `OPENCLAW_SETUP_LOCK_TIMEOUT` override) right after the gateway fast-path, then re-check the fast-path and exit 0 if a concurrent invocation brought the gateway up while they waited. In production, `start` acquires this lock **before** the setup body and holds it through the Gateway `exec`, so one owner serializes setup and launch. The Worker's `gatewayStartPromises` guard is per-isolate, so an E2E cold-start burst lands several runs in different isolates at once; without the lock they ran the slow setup body concurrently, contended on OpenClaw's native config/state locks, and stalled cold start for ~18 min (run 36692775633 — `[SandboxRecovery]` never fired because the admission-failure streak is also per-isolate). `full` releases fd 9 before backgrounding the gateway so the child does not inherit the lock; `post-start` is outside it. Degrades to the previous unserialized behaviour when `flock` is absent.
- Native supervised gateway launch (`src/gateway/process.ts`): the Worker's production cold start is one atomic supervised `sandbox.exec` of `start-openclaw.sh start`, which acquires the container-wide startup lock **before** the setup body (R2 restore → onboard → plugin installs → `moltlazy patch`) and holds it through the gateway `exec` — `/bin/bash -lc 'touch /tmp/gateway-stdout.log; pgrep … || ( tail -f … & ); exec openclaw gateway run … >> /tmp/gateway-stdout.log 2>&1'` — instead of the script backgrounding it with `&`. `exec` replaces the bash wrapper so the supervised process _is_ the gateway (`listProcesses`/`getProcess` track it for the life of the container; `kill()` targets it directly); the guarded `tail` starts the container-stdout log mirror as a sibling (safe because the Worker never `waitForExit`s on the gateway handle), and the redirect preserves the ops-readable log file. After the port is ready the Worker dispatches `start-openclaw.sh post-start` (dev automation + boot-log reset, best-effort) **detached** (`void`, not awaited) on every cold start — the gateway is already serving, so post-start must not extend the shared start promise (dev instances run `openclaw automations list/create` there, 30–60s, and awaiting it left every request logging `Gateway start already in progress … waiting...` while `/health` already returned 200); because it is dispatched unconditionally, a `/api/health`-triggered boot (`waitForReady:false`) still creates the dev automation even though its caller returned early. `start-openclaw.sh` gains `setup`/`post-start` modes; the default `full` mode keeps the legacy background launch for manual runs and `test:gateway:boot`. A shared `STARTUP_TIMEOUT_MS` deadline bounds setup + port wait; `launchWithRecovery` centralizes the stale/orphan/transport retry policy shared by both phases. **The `setup`/`post-start` phases must stay child-free**: the SDK's `waitForExit` only settles once the supervised terminal exits, which never happens while the script leaves a live background child (a `tail -f` log mirror or a background `cloudflared` there stalled every cold start for the full `STARTUP_TIMEOUT_MS`, re-running setup until the ~11-minute mark). The mirror is started by the native launch; cloudflared is detached with `setsid`. **Never call `output()`/`runShell` on the live gateway handle** (it never exits): read diagnostics from the `/tmp/gateway-stdout.log` / `/tmp/startup.log` tails instead (`readLogTail`).
- All three ClawHub plugins (`cloudflare-ai-gateway` provider, `diagnostics-otel`, `cloudflare-unified-billing`) are **pre-installed into the container image at Docker build time** (unpinned → latest ClawHub release; capability consent recorded in the image). This moves the per-boot network installs off the latency-critical startup path: `start-openclaw.sh` detects them via the extension dir + `plugins list` and only falls back to a ClawHub install when that state is missing (e.g. a warm R2 restore replaced it). The Dockerfile does **not** copy the Paso4 plugin source — it downloads the published ClawHub artifact — so the image stays source-free. The plugin's source and tests now live in [Paso4/openclaw-plugins](https://github.com/Paso4/openclaw-plugins) (removed from this repo); see [docs/OPENCLAW-PLUGINS.md](/latest/llms/platform/libraries/openclaw-plugins/index.md). Runtime enablement is written by a single working `config patch` that also enables `admin-http-rpc` and disables `diagnostics-otel` unless its endpoint is set (the patch must not carry `plugins.bundledDiscovery` — removed in 2026.9.x, it fails validation and drops the whole patch — nor `plugins.allow`, which disables every bundled provider not listed). Memory search is disabled via the top-level `memory.search.enabled=false` (the old `agents.defaults.memorySearch` key is gone in 2026.9.x). Measured local cold boot: 35s vs 56s before
- v0.3.x $include Configuration: immutable per-section $include files (session/tools/logging); gateway/agents inline; gateway token via env SecretRef
- Gateway restart split: `POST /api/admin/gateway/restart` is the **soft reload** — sends `SIGUSR2` to the live `openclaw gateway run` process (in-process, no connection drop, no port rebind) via `softReloadGateway` and falls back to `signalRestoreNeeded` only when no live gateway is found; `POST /debug/restart-openclaw` (DEBUG_ROUTES=true, ROOT) is the **hard restart** — `killGateway` + `resetStartupCache` + `signalRestoreNeeded` + `start-openclaw.sh` reprovision (in-place, no `startGatewayDirect`) so existing client tokens keep working. `SIGUSR1` is **not** a restart signal (it opens Node's inspector on `127.0.0.1:9229`); the E2E `n_gateway_soft_reload.txt` asserts the inspector stays closed after a soft reload. The reload targets the **server**, not the `openclaw gateway run` launcher: `buildSoftReloadScript` signals the exact `comm` (`pgrep -x 'openclaw-gatewa'`, Linux 15-char truncation, plus a defensive `openclaw-gateway`), because signalling the launcher kills it and orphans the server without restarting it (readiness never dips while the endpoint still reports `reloaded: true`). The E2E therefore requires both a stable `/api/status.processId` **and** an increased `http server listening` count, since `processId` alone survives a broken reload. Issue #463: `OPENCLAW_NO_RESPAWN` stays — removing it cannot enable process-level restarts inside a container (see the decision record in `docs/plans/sandbox-memory-usage.md`)
- `OPENCLAW_NO_RESPAWN=1` is set in the Dockerfile as explicit intent only: inside a container OpenClaw already resolves restart mode to in-process (`isContainerEnvironment()` → `disabled` / `unmanaged`), so removing the flag cannot enable process-level restarts or reclaim memory (#463); `softReloadGateway` only signals live processes and skips zombies
- Sandbox SDK 1.0 preview (`@cloudflare/sandbox@0.13.0-next.776.1` + `cloudflare/sandbox:0.13.0-next.776.1` image): `startProcess`/`execStream`/`getLogs()` and `SANDBOX_TRANSPORT` are gone. All command execution goes through `sandbox.exec(argv)` (which resolves on **launch**, not completion); `src/gateway/sandbox-exec.ts` centralizes the shapes (`startShell` for a handle, `runShell` for buffered output, `outputOf`, `collectLogs`). `listProcesses()` now returns `ProcessStatus[]` (`.state`, argv `.command`) and a live handle is fetched via `getProcess(id)`; `findExistingGatewayProcess` returns a `GatewayProcess` wrapper. `waitForProcess` delegates to `waitForExit({timeout})` and kills on timeout. `SandboxOptions.labels` was removed upstream. Deploy cutover requires `wrangler deploy --containers-rollout=immediate` (stable/1.0 control protocols are incompatible)
- Stale sandbox recovery (`src/gateway/sandbox-recovery.ts`): when the platform terminates the container behind a live Sandbox DO, every RPC fails with `sandbox.exec error … HTTP error! status: 500` — `isStaleContainerError` classifies these and `recoverStaleSandbox` destroys the dead instance + resets startup/persistence caches so the next RPC auto-provisions a fresh container; wired into `findExistingGatewayProcess`, `checkGatewayHealth` and the `ensureGateway` `startShell` retry; cooldown-guarded (60s), never fires for deploy `OperationInterruptedError` or transient provisioning errors. It also matches the **control-protocol mismatch** (`'utils.getRuntimeMetadata' is not a function`, `activateControlSession`, `RuntimeControlProtocolError`) seen during a deploy cutover while the still-running container predates the new SDK's control handshake — the SDK logs a `sandbox.exec error …` line but re-throws the bare `TypeError`, so the message patterns are matched directly and the stale instance is reprovisioned instead of failing until the platform rollout replaces it
- Dev-only Fouria agents + client docs: moltlazy ships no production premade agents (vanilla `researcher`/`coder` removed) and injects **Fouria Builder** (`fouria-builder`) + **Fouria Q&A agent** (`fouria-qa`) only when the instance is a dev environment (`OPENCLAW_DEV_MODE=true` on the container, derived by the Worker from the instance's `-dev` hostname — see `src/dev-mode.ts`); per-tenant client documentation lives in R2 under `customers/<slug>/docs/` and is materialized by the Worker into `/home/openclaw/clawd/client-docs/` before gateway start (`src/services/fouria-docs/`, `materializeClientDocsIfDev`), with a dev-only admin API at `/api/admin/fouria-docs` (404 outside dev); the Dockerfile installs the `cctr` binary, platform docs are fetched from the docs site (`https://fouria.io`, fallback `https://fouria-docs.pages.dev`, starting at `/llms.txt`), and `start-openclaw.sh` keeps `security.installPolicy allow` in dev so the Builder can self-install plugins
- Orphaned-instance recovery (`src/gateway/sandbox-recovery.ts`): a container destroyed out-of-band (Cloudflare Containers dashboard "Destroy", failed force-stop) leaves the DO alive pointing at a dead placement and surfaces the SDK's _retryable_ admission errors (`no container instance available`, `max container instances exceeded`) instead of the 500 signature. `isContainerUnavailableError` + `trackProvisioningFailure` escalate to a recovery only once the streak reaches `ORPHANED_INSTANCE_MIN_FAILURES` **and** spans `ORPHANED_INSTANCE_GRACE_MS` (120s, > the SDK's ≈110s provisioning budget, so slow-but-healthy cold starts are never cancelled); `maybeRecoverSandbox` unifies both failure classes and clears the streak whenever a sandbox RPC succeeds. Operator lever: `POST /debug/reprovision` (DEBUG_ROUTES, ROOT) destroys the instance + clears caches + resets the streak **without** touching the gateway, so it works when the container is entirely gone
- Activity-driven wake (`src/activity/`): automations run inside the Gateway, so it must be running for schedules to fire. A per-tenant `ActivityScheduler` Durable Object (binding `ACTIVITY`, SQLite-backed, **separate** from the SDK `Sandbox` DO whose alarm drives `sleepAfter`) mirrors the tenant's schedule (`openclaw automations list --json`, refreshed on warm `/api/health` when `readiness.ok`, throttled to 5 min) and arms one alarm `CRON_WAKE_AHEAD_MINUTES` (default 10) before the next enabled job. On fire it wakes + heals via `ensureGateway` (recovering an orphaned instance) then re-arms; no enabled jobs → alarm cleared → container sleeps. Healing/wakeup is independent of user traffic, gated on job activity, and free when idle (`src/activity/alarm.ts` state machine is unit-tested; the DO is a thin adapter). Requires the `ACTIVITY` binding in `wrangler.jsonc` + lerma (`tenant_provisioner.ex` `resource_bindings`, `dispatch_namespace.ex` `@do_sqlite_classes`)
- Container image build profiles: the Dockerfile takes a `BUILD_PROFILE` build arg (default `production` → slim, only `xz-utils`; `dev` → adds the optional `pandoc` + `weasyprint` PDF/document stack, ~760 MB, used by the dev-only Fouria Builder `fouria-leads` skill). The profile is selected per wrangler environment via `image_vars` (equivalent to `docker build --build-arg`): `env.staging` (dev/test) sets `dev`; `env.production` and the default/local config keep the slim default. The release CD builds the slim default, and the E2E `basic` (PR) and `production` (develop) profiles both provision the slim production image on the same `standard-2` instance (differ by env only)
- Sandbox → leads R2 bridge (`src/sandbox-leads-bridge.ts`, `leads.r2.internal`): the `fouria-leads` skill now reaches the global `paso4-leads` bucket through the Sandbox `outboundByHost` bridge (plain HTTP resolved in the Workers runtime with the `LEADS_BUCKET` binding — no R2 creds in the container, no public route) instead of the public `wss://<worker>/api/leads/ws` proxy. The proxy is intercepted by Cloudflare Access on tenant hostnames (a per-tenant Access app shadows the `*.fouria.io/<path>` apps — confirmed via the login redirect's `kid`) and is retained only for the host-side E2E fixture. Related: lerma's `HealthMonitor` now probes with `redirect: false`, so a shadowed `/api/health` 302 (Access login page) is surfaced instead of being followed and misreported as healthy
- Client docs hardening + operations (dev-only): tenant-isolated the `leads.r2.internal` bridge (`src/sandbox-leads-bridge.ts` resolves the caller slug via the shared `src/sandbox-tenant.ts` resolver and treats every agent path as relative to `<slug>/`, so a sandboxed agent can only reach its own proposal; the shared `build-pdf.sh` stays read-only). Added the Admin **Client Docs** page (`/_admin/docs`, ADMIN, dev-only) + the extended `src/routes/api/fouria-docs.ts` managing docs (`customers/<slug>/docs/`) and leads **directly from the Worker** (no CLI) with markdown rendering (`marked` + `DOMPurify`), native-browser PDF opening via `/raw`, and multipart uploads; HTML/SVG raw downloads are forced to `attachment` + sandbox CSP. Added Worker-side git-like versioning (`DocsGitStore`: content-addressed blobs/commits/refs, `GET /commits`, `POST /rollback/:sha` as a non-destructive revert) and backups (`DocsBackupStore`: `POST /backup`, `GET /backups`, `POST /backups/:id/restore`, retention 10). See `docs/WORKER-DEVELOPMENT.md` §"Client Docs (dev-only admin surface)"
- `ECONNRESET` from the sandbox R2 bridges fixed: the tenant's `LEADS_BUCKET` was provisioned as a **plain-text var** (the bucket name) instead of an `r2_bucket` binding, so `env.LEADS_BUCKET` was a string, `bucket.list()` threw inside the `outboundByHost` handler, and the escaping exception surfaced to the container as an opaque `ECONNRESET`. Two fixes: (1) the bridges (`leads`/`docs`/`releases`) now validate the binding with `isR2Bucket` (`src/r2-binding.ts`) and wrap their bodies in try/catch, returning a structured JSON `503`/`500` instead of resetting the socket; (2) lerma (`instances.ex` `instance_bindings/2`, config `LEADS_R2_BUCKET` default `paso4-leads`) now provisions `LEADS_BUCKET` as an `r2_bucket` binding on every tenant worker (redeploy existing tenants to replace the stray var)
- Admin RPC-first read path (`src/gateway/admin-rpc.ts`): the hot dashboard reads/writes now go through OpenClaw's bundled `admin-http-rpc` plugin (`POST /api/v1/admin/rpc`) over `sandbox.containerFetch`, reusing the typed `moltlazy/sdk` HTTP `RpcClient` (which gained an injectable `fetcher`). This removes the ~10–15s per-call CLI spawn/WS-reconnect on `GET /api/admin/devices`, `POST .../approve`, `GET /api/admin/models`, and `GET /api/admin/channels/probe` (measured locally: ~1.3–2.5s CLI → ~30–40ms RPC), each of which keeps its CLI path as a fallback (`createAdminRpcClient` returns null / throws → existing parser runs). `start-openclaw.sh` enables the plugin with a dedicated `plugins.entries["admin-http-rpc"].enabled=true` patch — NOT `plugins.allow`, which disables every bundled provider not listed (and the nearby batched plugin patch is rejected upstream for the removed `plugins.bundledDiscovery` key). `src/index.ts` shadows `/api/v1/admin/rpc` with a 404 before the public catch-all proxy so the operator surface stays container-local. The home-grown `moltlazy/sdk` WebSocket client is intentionally not used for admin RPC
- Gateway WebSocket RPC (`src/gateway/gateway-rpc.ts`): methods outside the `admin-http-rpc` allowlist (notably `audit.activity.list`) are reached with OpenClaw's official `@openclaw/gateway-client/browser` reference client + `@openclaw/gateway-protocol` types, bridged over `sandbox.wsConnect` and authenticated with the shared gateway token as an `operator`. The wire-v4 handshake (`connect.challenge` → `connect` → `hello-ok`) is owned by the official client; the connect frame's `client.version` comes from `__OPENCLAW_VERSION__`, injected by `vite.config.ts` from `catalog.openclaw`. RPC-first with CLI fallback: `GET /api/admin/audit-logs` (`audit.activity.list`) and the agent board (`agents.create`/`agents.delete`/`agents.update`, model via `config.patch`). `GET /api/admin/integrations` is deliberately **WebSocket-free** (the gateway drops these sockets): channels/agents over the admin HTTP RPC, plugin enabled-state from `config.get`, toggles/config via `config.patch`, secrets reload via the CLI. `theme`, agent bindings, `secrets.apply`, and `channels list --all` have no RPC method and stay on the CLI. The two packages are pinned in `catalog.@openclaw/gateway-client` / `catalog.@openclaw/gateway-protocol` and must track `catalog.openclaw`
- Realtime device-pairing notifications (`src/gateway/gateway-rpc.ts` `openGatewayEventStream`, `src/routes/api/devices.ts` `GET /ws`): the dashboard `NotificationBell` now subscribes to `GET /api/admin/devices/ws`, which pushes an initial `{ pending, paired }` snapshot then relays Gateway `device.pair.*` events (reconnecting via the official gateway client over `sandbox.wsConnect`). Admins approve/reject pending devices inline from the bell via `POST /api/admin/devices/:requestId/approve` / `.../reject` (RPC-first; the reject route was added). Also fixed `secrets apply` invocations (`integrations`/`dreaming`): current OpenClaw removed `--yes` (only `--from`, `--dry-run`, `--allow-exec`, `--json`)
- CLI→RPC sweep (finishing the RPC-first migration): `config` set/unset use `config.patch`; `dreaming` config writes use `config.patch` (shared `patchConfigViaRpc` in `src/routes/api/shared.ts`); `secrets.reload` uses the WS `secrets.reload`; and `resolveAgentFlag` (used by `models`/`knowledge`) resolves the roster via `agents.list` instead of `openclaw agents list`. CLI fallbacks are retained everywhere. Remaining CLI-only paths have no Gateway RPC method: `secrets apply/audit`, `openclaw wiki *`, agent `bind`/`unbind`/`bindings`, and `openclaw --version`
