docs: describe the current state, not the change history

Sweep of all tracked docs for a new user: facts now match the de-forked
stack, the 7-step setup wizard, and the release tooling.

- README: Pi row/node/update-flow row de-forked (npm at LPB_PI_VERSION);
  fork table drops the retired pi row; CI tags 4 repos; Dockerfile +
  stack.env structure comments; wizard steps include MCP servers;
  related-repo pi entry marked retired
- CONTRIBUTING: pi row -> upstream npm (not a stack repo)
- lpb-devstack: docs review via local http.server (mike serve serves the
  stale gh-pages branch); promote repo list drops pi
- config-repo: skills tree + list gain lemonade-model-params;
  vision-analysis example no longer names a specific model
- lpb-config / lpb-cli: setup wizard step list includes MCP servers
- lpb-devstack promote + install-browser: formatting fix (orphaned
  use-case list gets its own heading)
- install-openspec: 5-repo stack
- manual-tagging: point-in-time design-record banner with pointers to
  the current lpb-devstack/lpb-config references
- fork-improvements: MTP/FLM model ids marked as examples
- repo-workflow skill: review command aligned with docs skill

Author: lpb-docs

.pi/skills/localpibox-repo-workflow/SKILL.md

@@ -132,7 +132,9 @@ docs sync + review happens at release time via `docs-ready`.
 
 ```bash
 # 0. Flag docs as reviewed for the release (merge dev→docs, build site,
-#    review with `cd ~/.lpb-stack/docs-preview && mike serve`, confirm →
+#    review the local build with
+#    `cd ~/.lpb-stack/docs-preview && python3 -m http.server 8000 -d site` →
+#    http://localhost:8000, confirm →
 #    commits DOCS_READY=<stable-version> on the docs branch + pushes)
 lpb-devstack release docs-ready
 

CONTRIBUTING.md

@@ -18,7 +18,7 @@ has its own scope:
 |---|---|
 | [devstack](https://github.com/lpb-stack/devstack) | Container image (`Dockerfile`), `lpb` launcher (`scripts/`), entrypoints (`support/`), CI |
 | [config](https://github.com/lpb-stack/config) | Pi settings, MCP servers, custom skills, subagents presets |
-| [pi](https://github.com/lpb-stack/pi) | Qwen `reasoning_effort` + context-overflow patches (fork) |
+| [pi](https://github.com/earendil-works/pi) | Upstream of the coding agent core — installed from npm at `LPB_PI_VERSION`, not a stack repo since the 2026-08-31 de-fork; patches go upstream |
 | [lemonade-pi-plugin](https://github.com/lpb-stack/lemonade-pi-plugin) | Qwen thinking + vision support (fork) |
 | [pi-subagents](https://github.com/lpb-stack/pi-subagents) | Centralized subagent model registry (fork) |
 | [lpb-memory](https://github.com/lpb-stack/lpb-memory) | Persistent memory / session search extension |

README.md

@@ -34,8 +34,9 @@ On the **first run** — in every mode (`lpb`, `--shell`, `--ssh`, `--web`) —
 `lpb` checks the model-provider state (stored credentials + a live probe of
 the Lemonade server). If it's missing or broken and you're on a terminal,
 the **setup wizard** runs before the container starts: config repo, server
-URL, API key, default model (from the server's live model list) and
-lpb-memory config — every step validated, with errors shown and re-asked
+URL, API key, default model (from the server's live model list), MCP
+servers, and lpb-memory config — every step validated, with errors shown and
+re-asked
 until they're right. The configuration is written straight into the state
 volume, so it's persisted **before first use**; the launch summary ends with
 a model status line.
@@ -68,7 +69,7 @@ lpb --help       # full usage
 
 | Component | Role |
 |---|---|
-| **Pi** (forked) | Coding agent CLI — local fork with Qwen reasoning + context-overflow patches |
+| **Pi** | Coding agent CLI — mainstream pi from the npm registry (pinned in `lpb.stack.env`); Qwen thinking is handled by the lemonade plugin |
 | **VSCodium** | Web-based editor (`:web` image), connects over the OpenVSCode protocol |
 | **lemonade-pi-plugin** (forked) | Model provider for the local **Lemonade** server — Qwen thinking + vision support |
 | **lpb-memory** | Persistent memory + session search for the agent |
@@ -91,7 +92,7 @@ flowchart LR
 
     subgraph Image["ghcr.io/lpb-stack/devstack<br/>(Ubuntu 26.04 + Node.js 24)"]
         direction TB
-        RT["runtime: Pi (forked, patched) + VSCodium + Chrome"]
+        RT["runtime: Pi (npm) + VSCodium + Chrome"]
         CF["extensions (lemonade, memory, subagents) + config preset"]
     end
 
@@ -194,26 +195,29 @@ flowchart TB
 | `lpb` launcher | `lpb --update` self-update (branch follows the selected tag) |
 | Extensions | Runtime — `pi update --extensions` (unpinned packages) |
 | Config preset | `lpb-config update` (git pull of the config repo) |
-| Pi core + patches | Baked into the image — rebuilt by CI |
+| Pi core | Baked into the image from the npm registry (`LPB_PI_VERSION`) — a new version means a CI rebuild |
 
 ## Forked Repos & Upstream Policy
 
-Fork URLs and branches are tracked in `lpb.stack.env` at the repo root.
-Each fork keeps its LocalPibox work as clean commits on top of upstream
-merges, so the delta vs upstream can always be extracted as one patch.
+Pi itself is no longer a fork: the image installs mainstream pi from the
+npm registry at `LPB_PI_VERSION` (the `lpb-stack/pi` fork is retired — see
+[Fork improvements](doc/fork-improvements.md) for the retirement note).
+
+The remaining fork URLs and branches are tracked in `lpb.stack.env` at the
+repo root. Each fork keeps its LocalPibox work as clean commits on top of
+upstream merges, so the delta vs upstream can always be extracted as one
+patch.
 
 | Repo | Upstream | LocalPibox work | Update policy |
 |---|---|---|---|
-| **pi** | `earendil-works/pi` (v0.84.3) | Qwen `reasoning_effort` + context-overflow patches | rebase onto new upstream releases |
 | **lemonade-pi-plugin** | `lemonade-sdk/lemonade-pi-plugin` (no stable release) | Qwen thinking + vision support | follow upstream `main`, check periodically |
 | **pi-subagents** | `tintinweb/pi-subagents` (v0.16.1) | centralized local-first subagent model registry | follow upstream; merge + repair as needed |
 | **lpb-memory** | *(independent project)* | Pi memory extension (subprocess reviews) | no upstream to track |
 | **config** / **devstack** | — | own | own |
 
 Patches are **candidate upstream contributions** — they go upstream only if
 generally useful and not too opinionated for this stack. See
-[Fork improvements](doc/fork-improvements.md) for the full patch-by-patch
-breakdown.
+[Fork improvements](doc/fork-improvements.md) for what each fork adds.
 
 ### Forking & repointing this stack
 
@@ -234,7 +238,7 @@ GitHub Actions (`.github/workflows/build-and-publish.yml`) runs on:
 Versioning is **manual**: `lpb-devstack bump` commits a new `VERSION`, and CI
 builds + tags only when VERSION changed in the pushed commit. Pipeline jobs:
 **VERSION check** → **test** (always) → **build & publish images** →
-**tag repos** (CI tags the other 5 stack repos on their pipeline branches) →
+**tag repos** (CI tags the other 4 stack repos on their pipeline branches) →
 **docs publish** (main pipeline only — publishes the stable docs version,
 gated on the docs being flagged ready via `lpb-devstack release docs-ready`
 before promotion) → **status**. Devstack itself is tracked by its `VERSION`
@@ -285,8 +289,8 @@ lpb --update      # self-update launcher + pull the latest image for your pipeli
 
 ```
 devstack/
-├── Dockerfile            # image build (pi clone, VSCodium, Chrome)
-├── lpb.stack.env         # fork URLs, image names, container identity
+├── Dockerfile            # image build (pi from npm, VSCodium, Chrome)
+├── lpb.stack.env         # stack identity: config repo, pi version, images, container name
 ├── lpb.conf.env          # runtime defaults (baked into the image)
 ├── .env.example          # template for per-project .env (LPB_ vars)
 ├── scripts/              # CLIs + shared package (single source, baked to
@@ -308,7 +312,7 @@ devstack/
 
 - [Documentation site](https://lpb-stack.github.io/devstack/) — one version
   per stable release (served from the `docs` branch)
-- [lpb-stack/pi](https://github.com/lpb-stack/pi) — Pi monorepo fork
+- [lpb-stack/pi](https://github.com/lpb-stack/pi) — retired pi fork (reference only; last state `pre-defork-0.0.71`)
 - [lpb-stack/config](https://github.com/lpb-stack/config) — agent config preset
 - [lpb-stack/lemonade-pi-plugin](https://github.com/lpb-stack/lemonade-pi-plugin) — Lemonade provider plugin
 - [lpb-stack/pi-subagents](https://github.com/lpb-stack/pi-subagents) — subagent model registry

doc/config-repo.md

@@ -22,6 +22,7 @@ boot by `start.sh`.
 ├── skills/
 │   ├── agent-browser-mcp-integration/
 │   ├── browser-validation/
+│   ├── lemonade-model-params/
 │   └── mcp-vision-analysis/
 ├── agents/
 │   ├── vision-analysis.md    → visual analysis subagent
@@ -128,8 +129,8 @@ The `agents/` directory defines **subagent model configurations**. Each
 - The system prompt and tools available
 - How the subagent should behave
 
-For example, `vision-analysis.md` defines a subagent that uses the
-local Qwen3.6 vision model to analyze browser screenshots.
+For example, `vision-analysis.md` defines a subagent that opens a page,
+takes a screenshot, and analyzes the image with the session's vision model.
 
 ## Skills Directory
 
@@ -139,6 +140,7 @@ and the pitfalls. The config repo ships:
 
 - `agent-browser-mcp-integration` — browser automation with agent-browser MCP
 - `browser-validation` — automated browser validation pipeline with JSON reports
+- `lemonade-model-params` — tuning lemonade-served models (wire IDs, per-model params, vendor settings)
 - `mcp-vision-analysis` — visual page analysis with the local vision model
 
 (Workspace-level skills, such as `localpibox-repo-workflow`, live in the

doc/fork-improvements.md

@@ -142,8 +142,8 @@ The delta is always visible as the diff between upstream and `lpb-dev`.
 
 | Backend | Reasoning Support | Why |
 |---|---|---|
-| **MTP** (`Qwen3.6-35B-A3B-MTP-GGUF`) | ✅ Yes | Uses newer chat template that accepts `developer` role |
-| **FLM** (`qwen3.5-9b-FLM`, `qwen3.6-moe-35b-a3b-FLM`) | ❌ No | Chat template only accepts `system/user/assistant/tool` roles |
+| **MTP** (e.g. `Qwen3.6-35B-A3B-MTP-GGUF`) | ✅ Yes | Uses newer chat template that accepts `developer` role |
+| **FLM** (e.g. `qwen3.5-9b-FLM`) | ❌ No | Chat template only accepts `system/user/assistant/tool` roles |
 
 ---
 

doc/install-browser.md

@@ -23,6 +23,9 @@ Running `agent-browser install --with-deps` directly downloads a system Chrome a
 - **Exec bit self-healing** — Python's `zipfile.extractall()` drops Unix exec bits; `install-browser` detects and restores them automatically (Chrome crashes at startup if `chrome_crashpad_handler` is not executable)
 - **Container config merging** — writes `~/.agent-browser/config.json` with container-safe launch args, **merged** into any existing config (preserves user customizations)
 - **Version tracking** — installs from the same version JSON API that Playwright uses, keeping Chrome in sync with agent-browser's expectations
+
+### Use cases
+
 - Visual testing of web apps
 - Browser-based validation (login flows, form submission, UI testing)
 - Capturing screenshots and accessibility audits

doc/install-openspec.md

@@ -277,7 +277,7 @@ OpenSpec enables spec-driven development for the LocalPibox stack:
 - Audit trail of all changes
 
 This is particularly valuable for:
-- Managing the 6-repo stack (each repo has specs)
+- Managing the 5-repo stack (each repo has specs)
 - Tracking breaking changes across upstream updates
 - Ensuring consistency between fork patches and upstream
 - Documenting rationale for design decisions

doc/lpb-cli.md

@@ -102,7 +102,7 @@ HTTP probe of the Lemonade server), not a one-shot first-boot flag:
 
 - **healthy** → silent; the stored URL/key are passed to the container.
 - **missing or broken + TTY** → the **setup wizard** runs (config repo →
-  server URL → API key → default model → lpb-memory config). Every step is
+  server URL → API key → default model → MCP servers → lpb-memory config). Every step is
   validated live (real HTTP probes) and failures loop back with the raw
   error visible; `a` continues anyway (flagged), `q` aborts (nothing is
   written, no container starts). The wizard writes `auth.json`,

doc/lpb-config.md

@@ -51,7 +51,7 @@ see [lpb-devstack reference](lpb-devstack.md).
 
 | Command | Description |
 |---|---|
-| `lpb-config setup` | The **same wizard** as `lpb setup` on the host: config repo, lemonade provider (auth.json), default model (settings.json), lpb-memory config. Every step is validated live (HTTP probes); failures loop with the error visible. Re-run any time to correct a bad configuration — current values are pre-filled. `--reconfigure` forces a re-run |
+| `lpb-config setup` | The **same wizard** as `lpb setup` on the host: config repo, lemonade provider (auth.json), default model (settings.json), MCP servers (mcp.json), lpb-memory config. Every step is validated live (HTTP probes); failures loop with the error visible. Re-run any time to correct a bad configuration — current values are pre-filled. `--reconfigure` forces a re-run |
 | `lpb-config setup --non-interactive` | Same, no prompts — `LEMONADE_BASE_URL` / `LEMONADE_API_KEY` env + defaults. First failure aborts with a precise error and writes nothing. This is the `start.sh` fallback for boots where the host wizard did not run |
 | `lpb-config check` | Validate the installation (read-only checklist: config repo, settings, credentials, live server probe) — the in-container twin of `lpb doctor` |
 

doc/lpb-devstack.md

@@ -75,7 +75,8 @@ promoting:
 
 1. `lpb-devstack release docs-ready` — merges `origin/dev` → `docs`, builds
    the site, and (after your local review via
-   `cd ~/.lpb-stack/docs-preview && mike serve`) commits
+   `cd ~/.lpb-stack/docs-preview && python3 -m http.server 8000 -d site` →
+   http://localhost:8000) commits
    `DOCS_READY=<stable-version>` on the `docs` branch and pushes it
 2. `release status` shows the docs verdict: `READY` / `MISSING` /
    `WRONG-VERSION` (flag for another version) / `STALE` (doc content changed
@@ -87,7 +88,7 @@ promoting:
    `https://lpb-stack.github.io/devstack/<version>/`
 
 `promote` does per repo (dev branch → stable branch: `dev` → `main` for
-devstack/config/lpb-memory, `lpb-dev` → `lpb` for pi/pi-subagents/
+devstack/config/lpb-memory, `lpb-dev` → `lpb` for pi-subagents/
 lemonade-pi-plugin):
 - **Fast-forward / clean merge**: resets the local stable branch to
   `origin/<stable>`, merges `origin/<dev>`, pushes