Gitea-Tools/docs/webui-local-dev.md

# Internal web UI — local development (#426)

Read-only MVP skeleton for the MCP Control Plane operator console. Gitea,
MCP capability gates, and `skills/llm-project-workflow/` remain the source of
truth; this UI only provides route stubs and layout.

## Prerequisites

- Python 3.11+ with project dependencies installed (`pip install -r requirements.txt`)
- No secrets in repo, config, or client bundle

## Start the server

From the repository root (or an issue worktree):

```bash
./scripts/run-webui
```

Or directly:

```bash
python3 -m webui
```

Optional environment variables:

| Variable | Default | Purpose |
|----------|---------|---------|
| `WEBUI_HOST` | `127.0.0.1` | Bind address (keep local for MVP) |
| `WEBUI_PORT` | `8765` | Listen port |
| `WEBUI_REPO_ROOT` | repository root | Prompt library workflow hash root |
| `WEBUI_PROJECT_REGISTRY` | packaged JSON | Project registry path |
| `GITEA_MCP_CONFIG` | unset | Server-side MCP profile config (never sent to browser) |
| `GITEA_MCP_PROFILE` | unset | Active MCP profile name (server-side only) |

See [webui-deployment.md](webui-deployment.md) for internal-only serving,
Cloudflare Access/WARP/VPN guidance, and unsafe bind overrides (#435).

See
[architecture/webui-control-plane-console-architecture-adr.md](architecture/webui-control-plane-console-architecture-adr.md)
for the console architecture: layer and authority boundaries, the redaction
boundary, `/api/v1/...` versioning, the target page map, and the phase gates
that govern when a write path may open (#632, epic #631).

See [webui-project-registry-api.md](webui-project-registry-api.md) for the
versioned project registry contract: registry schema versions 1 and 2, project
status, onboarding checklist state, and the fail-closed error payloads (#635).

## Routes (MVP)

| Path | Description |
|------|-------------|
| `/` | Home / operator overview |
| `/health` | JSON liveness (`status`, `service`, `mode`, `timestamp`, `uptime_seconds`) |
| `/api/v1/system/health` | Structured read-only system health (#634) |
| `/system-health` | System-health dashboard — readiness, version/uptime, dependencies, MCP namespaces, stale-runtime parity (#639) |
| `/queue` | Live PR and issue queue dashboard (#429) |
| `/api/queue` | JSON queue export with pagination metadata |
| `/projects` | Project registry list with status and onboarding progress (#427, #635) |
| `/projects/{id}` | Project detail + onboarding checklist |
| `/api/v1/projects` | Versioned JSON registry export (#635) |
| `/api/v1/projects/{id}` | Versioned JSON project detail (#635) |
| `/api/projects` | JSON registry export — unversioned Phase 1 alias of `/api/v1/projects` |
| `/prompts` | Prompt library with per-prompt copy buttons (#428) |
| `/api/prompts` | JSON prompt export with workflow hashes |
| `/runtime` | MCP runtime health and stale detection (#430) |
| `/api/runtime` | JSON runtime health export |
| `/policy` | Workflow policy and guardrail configuration visibility (#646) |
| `/api/v1/policy` | Versioned JSON guardrail inventory (redacted, read-only) |
| `/audit` | Report audit paste + validator preview (#431) |
| `/api/audit` | JSON validator preview (POST `report_text`, optional `task_kind`) |
| `/worktrees` | Worktree hygiene dashboard (#432) |
| `/api/worktrees` | JSON worktree scan with classifications and anomalies |
| `/actions` | Gated write-action registry — all disabled in MVP (#434) |
| `/api/actions` | JSON action registry with capability metadata |
| `/api/actions/{id}/preview` | Mutation ledger preview (GET, read-only) |
| `/leases` | Lease and collision visibility (#433) |
| `/api/leases` | JSON lease/collision export |
| `/sessions` | Phase 1 shell stub — session inventory (backed by #636) |
| `/inventory` | Phase 1 shell stub — unified inventory (backed by #636) |
| `/timeline` | Phase 1 shell stub — workflow event timeline |
| `/insights` | Phase 1 shell stub — operational insights placeholder |

Most routes are GET-only. POST/PUT/PATCH/DELETE return `405` with
`read-only-mvp`, except `/audit` and `/api/audit` which accept POST for
local validator preview only (no Gitea mutations, no server-side storage).

## System health API (#634)

`GET /api/v1/system/health` is the structured, read-only health surface for
automated readiness checks. It is the first console API under the `/api/v1`
prefix; the unversioned MVP exports remain as compatibility aliases.

`/health` is unchanged for existing consumers — every MVP key is still present
— and now also carries `started_at`, `uptime_seconds`, and a
`system_health_api` pointer. It stays deliberately cheap and runs no dependency
probe, because answering readiness costs real work.

**Status codes.** `200` when ready, `503` when a required dependency failed or
was never probed. Automation can branch on the code without parsing the body.

**Query flags.** The Gitea check is a network call, so it is opt-in:
`GET /api/v1/system/health?deep=1` runs it and caches the result for
`WEBUI_HEALTH_PROBE_TTL_SECONDS` (default 15s) so dashboard polling does not
amplify into remote load. Without the flag that probe reports `skipped`.

**Dependencies.** `control_plane_db` and `repository` are required and drive
readiness. `gitea` is optional: when it fails the overall `status` degrades but
`readiness.ready` stays true, because local inventory is still serveable. Each
entry carries `status`, `detail`, `required`, and `latency_ms`.

Two honesty rules are worth knowing before reading the payload:

* `stale_runtime.mutation_safe` is true only when the runtime, checkout, and
  remote-tracking commits are all known and equal. An unfetched remote is
  reported as indeterminate, never as safe.
* `mcp_namespaces` entries are always `unproven`. A web process runs outside
  the IDE-managed MCP client and cannot invoke a namespace tool, so per #543
  only a `client_namespace` probe can prove that path.

Sample response (abridged, healthy):

```json
{
  "status": "ok",
  "service": "mcp-control-plane-webui",
  "mode": "read-only",
  "api": "/api/v1/system/health",
  "timestamp": "2026-07-22T11:04:18.512034+00:00",
  "readiness": { "ready": true, "complete": true, "reasons": [] },
  "version": {
    "git_sha": "620ed6e9a9550b8da2ceb82d9ab8744e8920490f",
    "git_describe": "v1.1.0-898-g620ed6e",
    "control_plane_schema_version": 4,
    "python_version": "3.14.5",
    "known": true
  },
  "process": { "started_at": "2026-07-22T10:58:02.114+00:00", "uptime_seconds": 376.4 },
  "deep_probes_requested": false,
  "dependencies": [
    {
      "name": "control_plane_db",
      "kind": "sqlite",
      "status": "ok",
      "detail": "schema v4 readable",
      "required": true,
      "healthy": true,
      "latency_ms": 1.482,
      "metadata": { "schema_version": 4, "active_leases": 3 }
    },
    { "name": "repository", "kind": "git", "status": "ok", "required": true, "healthy": true },
    { "name": "gitea", "kind": "http", "status": "skipped", "required": false, "healthy": false }
  ],
  "mcp_namespaces": [
    { "namespace": "gitea-author", "required_tool": "gitea_whoami", "status": "unproven" }
  ],
  "stale_runtime": { "stale": false, "determinable": true, "mutation_safe": true, "reasons": [] },
  "probe_errors": []
}
```

No restart, reload, or process-kill control is exposed here: those are Phase 2
at the earliest, and #630 forbids process-kill recovery outright. Every probe
opens its subject read-only — the control-plane database is opened through a
`mode=ro` URI so a health check can never create or migrate a schema.

## Report audit (#431)

Paste an LLM final report at `/audit` or POST JSON to `/api/audit`. The UI
reuses `final_report_validator` and review schema checks to surface missing
proof fields, wrong validation vocabulary, mutation contradictions, and a
suggested next prompt or issue-comment draft. Task kind can be auto-detected
or selected explicitly.

## Project registry (#427)

Versioned registry file: `webui/data/projects.registry.json` (schema version `1`).

Override path with `WEBUI_PROJECT_REGISTRY` when operators keep a machine-local
copy outside git. The registry stores repo identity, remotes, profile names,
workflow/schema path references, and onboarding checklist steps — never tokens
or credentials.

Seed entry: **Gitea-Tools** on `https://gitea.prgs.cc` with `prgs-author`,
`prgs-reviewer`, and `prgs-reconciler` profiles.

## Prompt library (#428)

Prompts are generated at load time from canonical workflow files under
`skills/llm-project-workflow/workflows/`. SHA-256 hashes are computed from
`WEBUI_REPO_ROOT` (defaults to the repository root). Prompt bodies are short
copy/paste starters; canonical workflow files remain the only full policy
source.

## Live queue dashboard (#429)

`/queue` loads open PRs and issues for the default registry project (seed:
**Gitea-Tools** on `https://gitea.prgs.cc`) using existing `gitea_auth` read
credentials. The UI surfaces pagination proof (returned count, pages fetched,
`has_more`, `inventory_complete`) and classification badges (`claimed`,
`blocked`, `in-review`, `duplicate`) when evidence exists.

If credentials are missing or the fetch fails, the page shows an explicit error
instead of an empty queue (fail closed).

## Gated actions (#434)

`/actions` registers future write actions (claim, comment, review, merge,
delete branch, create PR/issue). Each entry declares the MCP tool, required
permission, and profile role from `task_capability_map.py` — aligned with
`gitea_resolve_task_capability`. Buttons are disabled; previews always render
a mutation ledger. Direct `attempt_action` calls fail closed without invoking
MCP tools.

## Worktree hygiene (#432)

`/worktrees` scans local `branches/` directories and registered git worktrees.
Each entry is classified (`active-pr`, `active-issue`, `dirty`, `stale-clean`,
`detached-review`, `unsafe-unknown`, `orphan`). Missing preserved worktrees
referenced by the issue lock file are flagged as anomalies (#404). The page
includes a copy/paste canonical cleanup prompt only — no deletion actions.

Override scan root with `WEBUI_REPO_ROOT` (defaults to repository root).

## Lease visibility (#433)

`/leases` surfaces read-only lease and collision state: local issue lock file,
in-progress claim inventory (#268), reviewer PR lease comments when present
(`<!-- mcp-review-lease:v1 -->`, #407), duplicate open PRs per issue (#400),
and duplicate local branches per issue. Links to collision-history backend
issues (#267, #268, #400, #407) are included. No lease acquire/release from UI.

## Runtime health (#430)

`/runtime` surfaces read-only MCP/runtime diagnostics for the default registry
project: active profile and role kind, authenticated identity (when credentials
are available), config model/mode, local vs remote `master` SHA sync, shell
health, workflow/schema SHA-256 hashes, and stale-runtime warnings when the
checkout is behind merged safety-gate changes. Restart guidance links to #420;
no tokens or MCP restart actions are exposed.

## Policy & guardrail visibility (#646)

`/policy` (HTML) and `/api/v1/policy` (JSON) surface a **read-only** projection
of the major workflow guardrails — role separation/RBAC, lease lifecycle,
author worktree binding, merge confirmation, secret redaction, contamination
containment, allocator policy, audit logging, and mutation gating. Each entry
carries source pointers to the file/module/doc that owns it, a compact active
value derived from the existing safe policy accessors, and — where a documented
default is declared — a diff of active vs documented. The whole payload is run
through the console redaction pass before it is emitted, so a planted or
accidental secret degrades to the placeholder rather than reaching a client.
The view never edits policy and exposes no gate-weakening toggle.

## Application shell — Phase 1 (#638)

The console shell (`webui/layout.py`) renders a grouped navigation driven by a
single nav-config module, `webui/nav.py`. Nav groups follow the epic #631
Phase 1 information architecture: **Health, Traffic, Runtime/Sessions,
Projects, Inventory, Timeline, Policy** (live via #646), and **Insights**
(placeholder). Live views and Phase 1 placeholders (`stub`) are declared in one
place so the layout and the route table cannot drift.

The header carries two read-only status badges — an **environment** badge
(`local` for loopback binds, `remote` otherwise, derived from `WEBUI_HOST`) and
a **mode: read-only** badge — plus a **Docs** link to this document. No
privileged action controls are present in the Phase 1 shell.

Not-yet-implemented surfaces (`/sessions`, `/inventory`, `/timeline`,
`/insights`) resolve to graceful read-only stub pages instead of 404s; their
backing views land in later child issues of #631 (the inventory surfaces are
backed by #636). `/policy` is a live read-only surface (#646), not a stub.
Mutating methods on stub routes still fail closed with `read-only-mvp`.

## System-health dashboard (#639)

`/system-health` renders the same snapshot the `/api/v1/system/health` API
returns, so the page and the API can never disagree. Cards: overall readiness,
stale-runtime parity, version and uptime, dependency probes, MCP namespaces,
probe errors (only when present), and recovery pointers. `?deep=1` opts into
the network probe exactly as the API does; the plain page load stays cheap.

Field authority and honesty rules:

* `ready` and `readiness_complete` are shown separately. A snapshot whose
  required probes never ran is not the same as one that ran them and passed,
  and the page never collapses the two into an unproven green.
* A probe that did not run appears under **Not probed**, never as healthy.
* `stale_runtime.mutation_safe` is displayed verbatim from the API. When the
  runtime is stale, or when parity is indeterminate, the page warns and does
  not claim mutation safety.
* MCP namespaces are reported `unproven`: the web process runs outside the
  IDE-managed MCP client and cannot prove that path (#543).

Redaction is split by field kind. Free text — probe details, readiness and
parity reasons, probe errors — passes through `system_health.redact`.
Structured fields — commit SHAs, probe names, statuses, timestamps — are
HTML-escaped only, because `redact`'s opaque-token rule matches any run of 32
or more characters and would otherwise blank every 40-character git SHA, which
is precisely the evidence the parity view exists to show.

The dashboard is read-only: no restart, reload, or process-kill control. Those
arrive in Phase 2 (#642). Recovery guidance points at the sanctioned client
reconnect / operator restart path — never a manual daemon kill (#630).

## Deployment boundary (#435)

MVP serves on loopback by default. Binding `0.0.0.0` or `::` is **refused**
unless `WEBUI_ALLOW_PUBLIC_BIND=1`. Non-loopback hosts log a warning unless
`WEBUI_ALLOW_REMOTE_BIND=1`. `GET /health` exposes `deployment` metadata.

## Tests

```bash
pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_gated_actions.py -q
pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_audit.py tests/test_webui_worktree_hygiene.py -q
pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_lease_visibility.py tests/test_webui_runtime_health.py -q

pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_deployment_boundary.py -q

## Tests (#436)

Run the full hermetic web UI suite (all `test_webui_*.py` modules):

```bash
./scripts/test-webui
```

CI / Jenkins multibranch can call the path-filtered gate (runs only when the
diff touches `webui/`, `tests/test_webui_*`, or web UI docs/scripts):

```bash
./scripts/ci-webui-check
WEBUI_CI_FORCE=1 ./scripts/ci-webui-check   # always run
```

`scripts/test-webui` sets `WEBUI_TEST_OFFLINE=1` by default. In that mode the
queue, lease, and runtime routes use empty offline snapshots instead of Gitea
credentials, so CI can run without MCP daemon credential access. Set
`WEBUI_TEST_OFFLINE=0` only when deliberately validating live fetch behavior.

Or invoke unittest directly:

```bash
python3 -m unittest discover -s tests -p 'test_webui_*.py' -q
```

## Lease visibility (#433)

`/leases` surfaces read-only lease and collision state: local issue lock file,
in-progress claim inventory (#268), reviewer PR lease comments when present
(`<!-- mcp-review-lease:v1 -->`, #407), duplicate open PRs per issue (#400),
and duplicate local branches per issue. Links to collision-history backend
issues (#267, #268, #400, #407) are included. No lease acquire/release from UI.

## Runtime health (#430)

`/runtime` surfaces read-only MCP/runtime diagnostics for the default registry
project: active profile and role kind, authenticated identity (when credentials
are available), config model/mode, local vs remote `master` SHA sync, shell
health, workflow/schema SHA-256 hashes, and stale-runtime warnings when the
checkout is behind merged safety-gate changes. Restart guidance links to #420;
no tokens or MCP restart actions are exposed.

## Workflow-event timeline (#637)

`GET /api/v1/timeline` is a read-only, versioned aggregation of workflow
events from every available source into one normalised, filterable stream. It
is the model layer for the Phase 1 timeline console view (a later child issue
of #631); this issue ships the schema, adapters, and read API only.

### Schema (versioned)

`webui/timeline.py` declares `TIMELINE_SCHEMA_VERSION` (currently `1`) and the
frozen `WorkflowEvent` record. Every response carries `schema_version` so a
consumer can branch on shape. One event:

```json
{
  "source": "control_plane",
  "event_type": "lease.renew",
  "event_key": "cp:1421",
  "timestamp": "2026-07-23T02:00:00Z",
  "actor": null,
  "role": null,
  "issue_number": 637,
  "pr_number": null,
  "session_id": null,
  "tool_name": null,
  "decision": null,
  "message": "lease renewed",
  "correlation_id": "issue#637",
  "evidence_refs": [],
  "sensitive": true
}
```

`event_key` is stable and unique per source (`cp:<event_id>`,
`cth:<kind>:<number>:<comment_id>`), so pagination and dedup are deterministic.

### Sources and field authority

| Source | Adapter | Authority |
|---|---|---|
| Control-plane `events` ⋈ `work_items` | `adapt_cp_events` | `event_type`, `message`, `timestamp`, issue/PR scope come from the CP database, read through a `mode=ro` URI (never creates the DB or runs migrations) |
| Gitea Canonical Thread Handoff comments | `adapt_cth_comments` | `actor`, `role` (next owner), `decision`, `evidence_refs`, `timestamp` come from the parsed CTH comment body (`canonical_thread_handoff`) |

Handoff comments are thread-scoped: they are only read when the request filters
by a single `issue` or `pr`. Otherwise the handoff source reports `not run`
with a reason — it is never rendered as empty-and-healthy. Each source degrades
independently: an unavailable control-plane DB or a failed comment fetch is a
`sources[]` entry with `ok:false` and a `reason`, never a dropped timeline.

### Query parameters

`issue`, `pr`, `session` (conjunctive filters); `limit` (default 50, max 500)
and `offset` for pagination; `remote`, `org`, `repo` to override the default
registry-project scope. Events sort ascending by
`(timestamp, source_rank, event_key)`; missing timestamps sort last.

### Filter authority, and refusing what cannot be answered

A filter dimension is only meaningful for a source whose records carry it.
Each source declares its own support in `_SOURCE_FILTER_SUPPORT` and reports it
per response as `supported_filters` / `unsupported_filters`:

| Source | issue | pr | session |
|---|---|---|---|
| `control_plane` | yes | yes | **no** — the `events` table is `(event_id, work_item_id, event_type, message, created_at)` and records no session |
| `gitea_handoff` | yes | yes | yes — a CTH comment declares its own `Session:` field |

`session_id` is read only from that declared CTH field. It is never inferred
from a work item, an actor, or message text, and a value that is
redaction-altering or bare-secret-shaped is dropped rather than emitted.

When **no source that ran** can carry a requested dimension, the request is
refused rather than answered: the response is `422` with `ok:false` and a
structured `error` naming `unsupported_filters` and the per-source reason. A
`200` with zero events would tell an operator that no such activity exists,
which is a stronger — and false — claim than "this cannot be answered here".
A source that *can* answer the dimension and simply matched nothing still
returns `200` with `ok:true` and an empty page.

### Redaction

Every free-text field (event messages, decision/proof text, roles, actors) is
passed through the console redaction policy (`webui.console_redaction`, backed
by `gitea_audit.redact`) before it leaves the module, failing closed to the
placeholder. No unredacted tool arguments or secrets are ever emitted, and a
generation error never drops raw data to a caller or a log.

Redaction also runs *before* any structured value is derived from free text.
`evidence_refs` are extracted from already-redacted proof/decision text, and a
commit reference is recognised only where the text declares one (`commit`,
`head`, `base`, `sha`, …). An undeclared 40-character hex run has the exact
shape of a Gitea access token, so it is never lifted out of prose into a
structured field. Every reference is then independently revalidated against an
allowed shape and a second redaction pass immediately before serialization;
anything unproven is dropped and the event is flagged `sensitive`.

### Tests

```bash
pytest tests/test_webui_timeline.py -q
```

## Tests

```bash
pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_audit.py tests/test_webui_worktree_hygiene.py -q
pytest tests/test_webui_skeleton.py tests/test_webui_project_registry.py tests/test_webui_prompt_library.py tests/test_webui_queue_dashboard.py tests/test_webui_lease_visibility.py tests/test_webui_runtime_health.py -q
```