feat(mcp): expose sanctioned Codex MCP reconnect request (Closes #678)

Add gitea_request_mcp_reconnect as a report-only callable surface so Codex
and other agent hosts can request host/IDE reconnect with typed blockers
and exact UI steps. Never kills processes or edits config.

Co-Authored-By: Grok 4.5 <[email protected]>

docs/incidents/670-direct-master-commit-2fa97c26.md

@@ -0,0 +1,83 @@
+# Incident #670: bare direct-to-master commit `2fa97c26` (retroactive audit)
+
+Status: verified; disposition recommendation: **accept as-is, no revert** (final
+disposition owned by controller per issue #670).
+
+## Summary
+
+Commit `2fa97c26fbda555a1a83930ca5fdcea9d8e47b50`
+(`fix(mcp): load dotenv relative to project root`) landed on `prgs/master`
+as a single-parent commit with no PR wrapper and no review record, bypassing
+the sanctioned issue → branch → PR → review → merge workflow. It was
+discovered during the PR #654 post-merge audit. PR #654 itself merged
+cleanly via the Gitea API and did **not** introduce this commit.
+
+## Verification evidence (acceptance criteria 1–3)
+
+- **AC1 — present on `prgs/master`: yes.**
+  `git merge-base --is-ancestor 2fa97c26fbda555a1a83930ca5fdcea9d8e47b50 prgs/master` → true.
+- **AC2 — no PR or review record: confirmed.**
+  The commit is a single-parent, non-merge commit sitting directly on
+  first-parent master between the #629 merge (`5ab5fe85`) and the #654
+  merge (`ec903b0d`). A PR landing on master produces a merge commit (or a
+  PR-linked head); neither exists here. The controller audit at issue-create
+  time also found no PR wrapper and no review record for this SHA.
+- **AC3 — changed files and diff summary: confirmed.**
+  `gitea_auth.py | 5 +++--` (+3/−2). Single parent
+  `5ab5fe8583c07134d55dadf09381aecb67df246e`. The change moves
+  `PROJECT_ROOT` derivation above `load_dotenv()` and loads
+  `.env` relative to the project root instead of the process CWD.
+
+## AC4 — why no immediate revert
+
+- The dotenv fix is intentional and required for correct runtime behavior:
+  without it, `load_dotenv()` resolves `.env` against the process working
+  directory, which breaks MCP server launches whose CWD is not the project
+  root.
+- The change is small (+3/−2), self-contained in `gitea_auth.py`, and has
+  been running on master without incident since 2026-07-10.
+- Reverting would re-introduce a real bug to remove a provenance defect —
+  the wrong trade. Provenance is repaired retroactively by this document,
+  issue #670, and the hardening landed under #671.
+- If the controller later judges the change unsafe, a separate
+  revert/repair issue is the sanctioned path (issue #670, recommended
+  disposition option 4).
+
+## AC5 — workflow-hardening linkage
+
+Prevention already landed: **issue #671** (closed)
+*“Block direct pushes to stable branches from MCP workflow sessions”*,
+implemented by commit `5933d87647656643a67a50331c4c7b06ea751dad`
+(`feat(guard): block direct stable-branch pushes from MCP workflow sessions`).
+
+Shipped guardrails include:
+
+- `gitea_record_stable_branch_push_attempt` — classifies proposed commands
+  for direct stable-branch push intent (`git push <remote> master`,
+  refspecs, `HEAD:master`, `--force`, dry-run intent, `:master` delete),
+  plus root/control-checkout local commits not carried by an issue branch,
+  and writes a durable `stable_branch_contamination` marker.
+- `gitea_audit_stable_branch_contamination` — reconciler-only audit/clear
+  path; a contaminated worker session cannot self-clear.
+- Review/merge/close/completion mutations fail closed while a
+  contamination marker is active.
+
+## AC6 — PR #654 was not the source
+
+- `2fa97c26` is the **first parent** of the #654 merge commit
+  `ec903b0d619e7a27d24aed272a890f4e5d381411`; it predates the #654 merge.
+- First-parent history `5ab5fe8..ec903b0`:
+  `2fa97c2 fix(mcp): load dotenv relative to project root` followed by
+  `ec903b0 Merge pull request 'feat: lifecycle role/hazard labels ... (#603)' (#654)`.
+- The #654 merger audit confirmed `ec903b0d` was a valid Gitea-API merge,
+  the `git push prgs master` attempt during that run was a no-op, and the
+  net change `2fa97c2..ec903b0` contained only the reviewed #603
+  lifecycle-label files.
+- Conclusion: #654 merged reviewed content only; the unauthorized-path
+  defect is solely the earlier bare commit `2fa97c26`.
+
+## Explicit non-actions (unchanged by this audit)
+
+- No revert of `2fa97c26`.
+- No force-push or history rewrite.
+- No master mutation from the audit session.

docs/mcp-namespace-eof-recovery.md

@@ -47,18 +47,33 @@ Do the steps in order. Stop as soon as a live **client-namespace** call succeeds
    - Only the Gitea namespace fails → single-namespace transport close. Continue.
    - Every server fails → restart the whole MCP client, not just one namespace.
 
-2. **Reconnect the namespace through the client, not the shell.** Use the IDE /
+2. **Request the sanctioned reconnect surface (#678), then reconnect through
-   client MCP-reconnect action for that server entry (in Claude Code:
+   the client — not the shell.** From a still-reachable Gitea MCP namespace
-   `/mcp` → reconnect the affected `gitea-*` server). Reconnecting forces the
+   (or after host auto-reconnect), call:
-   client to spawn a fresh subprocess and re-open the pipe. This clears the
+
-   closed-client state that a bare `kill`/respawn from a terminal does **not**.
+   ```text
+   gitea_request_mcp_reconnect(
+       namespace="gitea-author",   # or gitea-reviewer / gitea-merger / …
+       reason="transport_eof",
+       client="codex",            # or claude_code / generic
+   )
+   ```
+
+   The tool is **report-only**: it never restarts a process. It returns
+   namespace, profile, pid/session, startup SHA, current master SHA, boundary
+   status, and a **typed blocker** with exact operator UI steps for Codex
+   (Reload Developer Tools / per-server reconnect) or Claude Code (`/mcp`).
+   Then perform the host reconnect those steps describe so the client spawns a
+   fresh subprocess and re-opens the pipe. That clears the closed-client state
+   that a bare `kill`/respawn from a terminal does **not**.
 
 3. **Do not "fix" it by importing the server or poking the process.** Reaching
    for `python -c 'import gitea_mcp_server ...'`, raw JSON-RPC from a shell,
    killing PIDs to force a respawn, or touching MCP config mtimes does **not**
    restore the *client's* view of the namespace and violates the daemon-import
    guard (#558, `docs/mcp-daemon-import-guard.md`). The only sanctioned repair
-   is a **client reconnect / relaunch**.
+   is a **client reconnect / relaunch** (or the typed operator path returned by
+   `gitea_request_mcp_reconnect`).
 
 4. **Verify through the same path the workflow will use.** After reconnect, call
    the specific tool the blocked workflow needs — not just any tool — through

docs/mcp-restart-path-inventory.md

@@ -45,10 +45,11 @@ and *fails closed*.
 | `legacy_auto_restart_helper` | removed | A helper (`_trigger_mcp_auto_restart`) that actively restarted the server from the read-only resolver path. | Removed in #685; kept absent by `assert_auto_restart_helper_absent()`. | #685, #657 |
 | `config_touch_reload` | removed | Touching (utime) the MCP client config to make the host reload the server. | Removed from the resolver in #685: stale detection is report-only, never mutating config, spawning threads, or calling `os._exit`. | #685, #657 |
 | `master_advance_auto_restart` | guarded_fail_closed | On-disk master advancing past the running code. | `master_parity_gate` captures startup parity and blocks mutations while stale, emitting restart guidance; the process never self-restarts. | #420, #591, #657 |
-| `stale_runtime_resolver_reconnect` | guarded_fail_closed | The capability resolver detecting a stale serving process. | Report-only (#685): returns `restart_required`/`stop_required` and an exact reconnect action; no restart, thread, config touch, or `os._exit`. | #685, #657 |
+| `stale_runtime_resolver_reconnect` | guarded_fail_closed | The capability resolver detecting a stale serving process. | Report-only (#685): returns `restart_required`/`stop_required` and an exact reconnect action; no restart, thread, config touch, or `os._exit`. | #685, #657, #678 |
+| `codex_client_reconnect_request` | guarded_fail_closed | `gitea_request_mcp_reconnect` report-only tool for Codex/LLM sessions. | Report-only (#678): returns namespace/profile/pid/startup SHA/master SHA/boundary status and a typed operator blocker with exact client UI steps; never restarts or kills. | #678, #630, #685, #657 |
 | `manual_daemon_kill` | forbidden | Shell kills of the daemon: `pkill -f mcp_server.py`, `killall`, broad `pkill -f python` sweeps, or `kill <pid>` of a daemon pid. | Forbidden (#630): `runtime_recovery_guard` classifies these as contamination and `gitea_record_daemon_process_kill_attempt` writes a durable marker that fails later mutations closed. Operator maintenance authorization is read only from the environment. | #630, #657 |
 | `conflict_marker_infra_stop` | guarded_fail_closed | The daemon entrypoint scans for unresolved merge-conflict markers at startup and stops (`sys.exit(1)`). | Fail-closed startup stop, not a restart: the process exits and waits for the operator to resolve conflicts and relaunch; never loops. | #657 |
-| `ide_client_reconnect` | host_residual | A manual `/mcp reconnect` (or equivalent host action) that recreates the MCP client connection. | Outside this process's control; the sanctioned recovery the gates point operators toward. No in-process code initiates it. | #584, #656, #657 |
+| `ide_client_reconnect` | host_residual | A manual `/mcp reconnect` (or equivalent host action) that recreates the MCP client connection. Agents obtain exact UI steps via `gitea_request_mcp_reconnect` (#678). | Outside this process's control; the sanctioned recovery the gates point operators toward. No in-process code initiates it. | #584, #656, #657, #678 |
 | `profile_switch_runtime` | sanctioned_narrow_recovery | Switching the active execution profile at runtime (dynamic-profile mode). | In-process and restart-free: `runtime_switching_supported` is true, so a switch rebinds capability without recreating the process. | #656, #657 |
 
 ## Guards enforced in CI

docs/mcp-tool-inventory.md

@@ -137,6 +137,7 @@ that gates each call, not which tools exist.
 - `gitea_release_merger_pr_lease`
 - `gitea_release_reviewer_pr_lease`
 - `gitea_release_workflow_lease`
+- `gitea_request_mcp_reconnect`
 - `gitea_request_mcp_restart`
 - `gitea_resolve_task_capability`
 - `gitea_resume_review_draft`

docs/webui-notifications.md

@@ -0,0 +1,81 @@
+# Web Console: Notifications & Human-Attention Routing (#648)
+
+- **Status:** Phase 3 Live
+- **Tracking Issue:** [#648](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/648)
+- **Parent Epic:** [#631](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/631)
+- **Attention Boundary Reference:** [#628](https://gitea.prgs.cc/Scaled-Tech-Consulting/Gitea-Tools/issues/628)
+
+---
+
+## 1. Overview
+
+The **Notifications & Human-Attention Console** (`/notifications`, `/api/v1/notifications`) provides intelligent event classification and human-attention routing for autonomous workflow operations.
+
+To prevent alert fatigue while ensuring critical escalation boundaries are never missed, events are classified into three distinct **Attention Classes**:
+
+1. **`human-required`** (Urgent Escalation Boundary):
+   - Items requiring immediate human intervention or business decisions.
+   - Triggers: Auth failures, hard stops, irrecoverable state, decision locks, failed report validations, critical probe errors.
+   - Display: Highlighted in red (`badge-blocked`) with a `HUMAN REQUIRED` badge.
+
+2. **`operator`** (Operational Inbox):
+   - Items requiring controller or operator review/triage during routine execution.
+   - Triggers: Blocked PRs (merge conflicts), stale leases, duplicate PRs on issues, unassigned ready work.
+   - Display: Displayed in orange/yellow (`badge-claimed`).
+
+3. **`routine`** (Background Workflow Transitions):
+   - Normal, healthy workflow transitions and state progressions.
+   - Triggers: Active PRs/issues in standard state, clean branch creation, routine heartbeats.
+   - Display: Filtered out of default inbox views to eliminate notification spam; viewable on demand via the "Routine" or "All" tab.
+
+---
+
+## 2. API Endpoints
+
+### `GET /api/v1/notifications`
+*Compatibility Alias:* `GET /api/notifications`
+
+#### Query Parameters:
+- `project_id` (optional): Filter notifications by project ID.
+- `attention_class` (optional): `inbox` (default: human-required + operator), `human-required`, `operator`, `routine`, `all`.
+
+#### Example JSON Response:
+```json
+{
+  "project_id": "gitea-tools",
+  "repo_label": "Scaled-Tech-Consulting/Gitea-Tools",
+  "human_required_count": 0,
+  "operator_count": 2,
+  "routine_count": 5,
+  "total_count": 7,
+  "fetch_error": null,
+  "inbox_items": [
+    {
+      "id": "notif-pr-block-742",
+      "attention_class": "operator",
+      "category": "blocker",
+      "title": "Blocked PR #742",
+      "summary": "PR #742 requires merge conflict resolution.",
+      "work_kind": "pr",
+      "work_number": 742,
+      "project_id": "gitea-tools",
+      "repo_label": "Scaled-Tech-Consulting/Gitea-Tools",
+      "created_at": "2026-07-25T16:39:47Z",
+      "deep_link": "/traffic",
+      "requires_human": false,
+      "extra": {}
+    }
+  ],
+  "all_items": [...]
+}
+```
+
+---
+
+## 3. UI Navigation
+
+- Access via the **Traffic** navigation menu: **Traffic → Notifications**.
+- The main view displays:
+  - **Metrics Summary Bar**: Highlighting counts for Human Required, Operator Inbox, and Routine items.
+  - **Attention Filter Tabs**: Toggle between Inbox (Human + Operator), Human Required, Operator, Routine, and All.
+  - **Structured Event Table**: Displays category, title, summary, work item links, and timestamps.

gitea_mcp_server.py

@@ -2092,6 +2092,7 @@ import root_checkout_guard  # noqa: E402
 import workflow_scope_guard  # noqa: E402  # #683 production scope / force-on guards
 import stable_branch_push_guard  # noqa: E402
 import runtime_recovery_guard  # noqa: E402  # #630 manual daemon-kill contamination
+import mcp_client_reconnect  # noqa: E402  # #678 sanctioned Codex reconnect request
 import remote_repo_guard  # noqa: E402
 import anti_stomp_preflight  # noqa: E402
 import issue_claim_heartbeat  # noqa: E402
@@ -14024,11 +14025,16 @@ def _classify_operation_gate_reasons(reasons: list[str]) -> dict:
 
 
 def _stale_runtime_reconnect_action() -> str:
-    """Sanctioned recovery for a stale daemon — reconnect only (#685/#897)."""
+    """Sanctioned recovery for a stale daemon — reconnect only (#685/#897/#678)."""
     return (
-        "Reconnect the IDE/client MCP session so the server reloads at the "
+        "blocker_kind=runtime_reconnect_required: call "
-        "current master head. Do not call gitea_activate_profile or switch "
+        "gitea_request_mcp_reconnect(namespace=<active gitea-* namespace>, "
-        "MCP role sessions — profile switching does not clear a stale daemon."
+        "reason='stale-runtime', client='codex') for a typed operator "
+        "reconnect blocker with exact UI steps, then reconnect the IDE/client "
+        "MCP session so the server reloads at the current master head. Do not "
+        "call gitea_activate_profile, pkill, touch configs, or switch MCP role "
+        "sessions — profile switching does not clear a stale daemon. After "
+        "reconnect restart from gitea_whoami → gitea_resolve_task_capability."
     )
 
 
@@ -20998,10 +21004,15 @@ def gitea_resolve_task_capability(
     # serving process/profile inventory is stale — even if permission is OK.
     if runtime_stale_blocker:
         next_safe_action = (
-            "blocker_kind=runtime_reconnect_required: reconnect/restart the "
+            "blocker_kind=runtime_reconnect_required: call "
-            "IDE-managed Gitea MCP server for this profile so it reloads current "
+            "gitea_request_mcp_reconnect(namespace=<active gitea-* namespace>, "
-            "master. Do not edit mcp_config.json by hand; the resolver does not "
+            "reason='stale-runtime', client='codex') for a typed operator "
-            "touch config, spawn recovery threads, or terminate the process."
+            "blocker with exact UI steps, then reconnect/reload the IDE-managed "
+            "Gitea MCP server for this profile so it reloads current master. "
+            "Do not edit mcp_config.json by hand, pkill, or touch configs; the "
+            "resolver does not touch config, spawn recovery threads, or "
+            "terminate the process. After reconnect restart from gitea_whoami → "
+            "gitea_resolve_task_capability."
         )
 
     # Task/role alignment guards (#167): the requested task, not the
@@ -22559,6 +22570,129 @@ def gitea_workflow_dashboard(
     return payload
 
 
+@mcp.tool()
+def gitea_request_mcp_reconnect(
+    namespace: str | None = None,
+    reason: str | None = None,
+    client: str = "codex",
+    remote: str = "dadeschools",
+    host: str | None = None,
+    session_id: str | None = None,
+) -> dict:
+    """Request a sanctioned host/IDE MCP reconnect for a named namespace (#678).
+
+    Codex and other agent hosts can detect stale or closed Gitea MCP runtimes
+    (``stop_required`` / ``restart_required`` from capability resolution, transport
+    EOF, missing namespace attachment). The **host owns the transport** — this
+    process cannot reopen the client's stdio pipe. This tool is the callable
+    surface agents use to:
+
+    1. Report reconnect status fields (namespace, profile, pid/session,
+       startup SHA, current master SHA, boundary status).
+    2. Return a **typed blocker** with exact operator UI steps for Codex (or
+       another client) when reconnect is required.
+
+    This tool **never** restarts, kills, reloads, or reconfigures an MCP
+    process. Forbidden recovery paths (pkill, touch/mtime hacks, config/.env
+    edits, session-state edits, raw API) are never recommended.
+
+    After the operator reconnects, workflows must restart from preflight:
+    ``gitea_whoami`` → ``gitea_resolve_task_capability`` → task.
+
+    Args:
+        namespace: MCP namespace to reconnect (e.g. ``gitea-author``). Defaults
+            to the active profile's inferred namespace.
+        reason: Why reconnect is requested: ``stale-runtime``, ``transport_eof``,
+            ``missing_namespace``, ``not_required``, or free-form (normalized).
+        client: Operator UI surface — ``codex`` (default), ``claude_code``, or
+            ``generic``.
+        remote: Known instance — ``dadeschools`` or ``prgs`` (parity context).
+        host: Optional host override for parity context.
+        session_id: Optional session id to echo in the report.
+
+    Returns:
+        dict with reconnect report fields, ``reconnect_performed=False``,
+        ``typed_blocker`` when reconnect is required, and
+        ``forbidden_recovery_paths``.
+    """
+    # Read-only: gitea.read is sufficient. Never a mutation.
+    read_block = _profile_operation_gate("gitea.read")
+    if read_block:
+        return {
+            "success": False,
+            "read_only": True,
+            "reconnect_performed": False,
+            "mutation_performed": False,
+            "reasons": read_block,
+            "permission_report": _permission_block_report("gitea.read"),
+            "forbidden_recovery_paths": list(
+                mcp_client_reconnect.FORBIDDEN_RECOVERY_PATHS
+            ),
+        }
+
+    profile = get_profile()
+    profile_name = (profile.get("profile_name") or "").strip() or None
+    inferred_ns = role_namespace_gate.infer_mcp_namespace(profile_name)
+    ns = (namespace or "").strip() or inferred_ns or "gitea-tools"
+
+    parity = _current_master_parity()
+    startup_sha = (
+        parity.get("daemon_start_head")
+        or parity.get("startup_head")
+        or _process_boot_head_sha
+    )
+    current_sha = parity.get("local_head") or parity.get("current_head")
+    if not current_sha:
+        try:
+            current_sha = master_parity_gate.read_git_head(PROJECT_ROOT)
+        except Exception:  # noqa: BLE001
+            current_sha = None
+
+    boundary = mcp_client_reconnect.classify_boundary_status(
+        startup_sha=startup_sha if isinstance(startup_sha, str) else None,
+        current_master_sha=current_sha if isinstance(current_sha, str) else None,
+        live_stale=bool(parity.get("live_stale")) if parity.get("live_known") else None,
+        in_parity=parity.get("in_parity") if parity.get("determinable") else None,
+    )
+
+    # Infer reason from parity when caller left it unspecified.
+    effective_reason = reason
+    if not (effective_reason or "").strip():
+        if parity.get("restart_required") or parity.get("live_stale"):
+            effective_reason = mcp_client_reconnect.REASON_STALE_RUNTIME
+        elif boundary == mcp_client_reconnect.BOUNDARY_CLEAN:
+            effective_reason = mcp_client_reconnect.REASON_NOT_REQUIRED
+        else:
+            effective_reason = mcp_client_reconnect.REASON_UNSPECIFIED
+
+    payload = mcp_client_reconnect.build_reconnect_request(
+        namespace=ns,
+        profile=profile_name,
+        pid=os.getpid(),
+        session_id=session_id
+        or f"{(profile_name or 'session')}-{os.getpid()}",
+        startup_sha=startup_sha if isinstance(startup_sha, str) else None,
+        current_master_sha=current_sha if isinstance(current_sha, str) else None,
+        boundary_status=boundary,
+        reason=effective_reason,
+        client=client,
+        live_stale=bool(parity.get("live_stale")) if parity.get("live_known") else None,
+        in_parity=parity.get("in_parity") if parity.get("determinable") else None,
+        restart_required=bool(parity.get("restart_required")),
+        stop_required=bool(parity.get("restart_required")),
+        extra={
+            "remote": remote if remote in REMOTES else remote,
+            "host": host,
+            "session_context_audit": session_ctx.mutation_context_audit_fields(),
+            "parity_summary": master_parity_gate.format_parity(parity),
+            "live_stale": parity.get("live_stale"),
+            "live_known": parity.get("live_known"),
+            "in_parity": parity.get("in_parity"),
+        },
+    )
+    return payload
+
+
 @mcp.tool()
 def gitea_request_mcp_restart(
     remote: str = "dadeschools",

mcp_client_reconnect.py

@@ -0,0 +1,328 @@
+"""Sanctioned MCP client reconnect request surface for Codex/LLM sessions (#678).
+
+Codex and other agent hosts can detect stale or closed Gitea MCP runtimes, but
+the host owns the transport. This module never restarts, kills, or reloads a
+daemon. It builds:
+
+1. A **callable reconnect request** result agents can invoke via
+   ``gitea_request_mcp_reconnect`` (report-only, side-effect free).
+2. A **typed blocker** with exact operator UI steps when recovery must be
+   performed by the host/operator.
+
+Forbidden recovery paths (must never be recommended):
+
+* ``pkill`` / ``kill`` / ``killall`` of MCP daemons
+* ``touch`` / mtime config reload hacks
+* ``.env`` or MCP config edits as recovery
+* session-state file edits
+* raw Gitea API / direct server-import fallbacks
+
+After the operator reconnects, workflows restart from identity / runtime /
+capability preflight (``gitea_whoami`` → ``gitea_resolve_task_capability`` →
+task).
+"""
+
+from __future__ import annotations
+
+from typing import Any, Mapping
+
+# --- Reason vocabulary -------------------------------------------------------
+
+REASON_STALE_RUNTIME = "stale-runtime"
+REASON_TRANSPORT_EOF = "transport_eof"
+REASON_MISSING_NAMESPACE = "missing_namespace"
+REASON_NOT_REQUIRED = "not_required"
+REASON_UNSPECIFIED = "unspecified"
+
+VALID_REASONS = frozenset(
+    {
+        REASON_STALE_RUNTIME,
+        REASON_TRANSPORT_EOF,
+        REASON_MISSING_NAMESPACE,
+        REASON_NOT_REQUIRED,
+        REASON_UNSPECIFIED,
+    }
+)
+
+# Boundary statuses reported to callers (match review_workflow_boundary style).
+BOUNDARY_CLEAN = "clean"
+BOUNDARY_MISMATCH = "mismatch"
+BOUNDARY_STALE = "stale"
+BOUNDARY_UNKNOWN = "unknown"
+
+# Typed blocker kinds
+BLOCKER_OPERATOR_RECONNECT = "operator_mcp_reconnect_required"
+BLOCKER_NONE = "none"
+
+FORBIDDEN_RECOVERY_PATHS: tuple[str, ...] = (
+    "pkill / kill / killall of mcp_server.py, gitea_mcp_server, or broad python sweeps",
+    "touch / mtime-based MCP config reload hacks",
+    ".env edits as recovery",
+    "MCP config file edits as recovery",
+    "session-state file edits as recovery",
+    "raw Gitea API or direct MCP server-import fallbacks",
+)
+
+# Client-specific operator UI steps. Keep Codex first (issue title surface).
+OPERATOR_UI_STEPS: dict[str, tuple[str, ...]] = {
+    "codex": (
+        "In Codex, open the MCP / Developer tools panel for this workspace.",
+        "Locate the named Gitea MCP server entry (namespace) that needs reconnect "
+        "(e.g. gitea-author, gitea-reviewer, gitea-merger, gitea-tools, "
+        "gitea-controller, gitea-reconciler).",
+        "Click 'Reload Developer Tools' or the server reconnect/reload control "
+        "for that entry so the client spawns a fresh MCP subprocess.",
+        "If per-server reconnect is unavailable, fully restart the Codex client "
+        "(quit and relaunch) so all MCP namespaces reattach.",
+        "After reconnect, rerun the blocked workflow from preflight: "
+        "gitea_whoami → gitea_resolve_task_capability → the original task. "
+        "Do not resume mid-mutation.",
+    ),
+    "claude_code": (
+        "Run `/mcp` (or open the MCP servers UI) in Claude Code.",
+        "Reconnect the affected gitea-* server entry so the client reopens stdio.",
+        "If reconnect fails, relaunch the Claude Code session entirely.",
+        "After reconnect, restart the workflow from gitea_whoami → "
+        "gitea_resolve_task_capability → task.",
+    ),
+    "generic": (
+        "Use the host/IDE MCP reconnect or reload control for the named namespace.",
+        "If no per-namespace control exists, restart the MCP client/editor.",
+        "After reconnect, restart the workflow from identity/capability preflight.",
+    ),
+}
+
+DEFAULT_CLIENT = "codex"
+
+
+def normalize_reason(reason: str | None) -> str:
+    """Map free-form reason strings onto the closed vocabulary."""
+    raw = (reason or "").strip().lower()
+    if not raw:
+        return REASON_UNSPECIFIED
+    if raw in VALID_REASONS:
+        return raw
+    text = raw.replace(" ", "_").replace("-", "_")
+    aliases = {
+        "stale_runtime": REASON_STALE_RUNTIME,
+        "staleruntime": REASON_STALE_RUNTIME,
+        "runtime_stale": REASON_STALE_RUNTIME,
+        "stale": REASON_STALE_RUNTIME,
+        "transport_eof": REASON_TRANSPORT_EOF,
+        "transport_closed": REASON_TRANSPORT_EOF,
+        "eof": REASON_TRANSPORT_EOF,
+        "client_is_closing": REASON_TRANSPORT_EOF,
+        "missing_namespace": REASON_MISSING_NAMESPACE,
+        "namespace_missing": REASON_MISSING_NAMESPACE,
+        "not_required": REASON_NOT_REQUIRED,
+        "healthy": REASON_NOT_REQUIRED,
+        "ok": REASON_NOT_REQUIRED,
+        "unspecified": REASON_UNSPECIFIED,
+    }
+    if text in aliases:
+        return aliases[text]
+    hyphenated = text.replace("_", "-")
+    if hyphenated in VALID_REASONS:
+        return hyphenated
+    return REASON_UNSPECIFIED
+
+
+def normalize_client(client: str | None) -> str:
+    """Return a known client key for operator UI steps."""
+    text = (client or "").strip().lower().replace(" ", "_").replace("-", "_")
+    if text in ("codex", "openai_codex", "openai"):
+        return "codex"
+    if text in ("claude", "claude_code", "claude_desktop", "anthropic"):
+        return "claude_code"
+    if text in OPERATOR_UI_STEPS:
+        return text
+    return DEFAULT_CLIENT
+
+
+def classify_boundary_status(
+    *,
+    startup_sha: str | None,
+    current_master_sha: str | None,
+    live_stale: bool | None = None,
+    in_parity: bool | None = None,
+) -> str:
+    """Derive boundary_status from parity evidence."""
+    if live_stale is True or in_parity is False:
+        return BOUNDARY_STALE
+    start = (startup_sha or "").strip().lower()
+    current = (current_master_sha or "").strip().lower()
+    if start and current and start != current:
+        return BOUNDARY_MISMATCH
+    if start and current and start == current:
+        return BOUNDARY_CLEAN
+    if in_parity is True:
+        return BOUNDARY_CLEAN
+    return BOUNDARY_UNKNOWN
+
+
+def operator_ui_steps(client: str | None, *, namespace: str | None = None) -> list[str]:
+    """Exact operator UI steps for the named client."""
+    key = normalize_client(client)
+    steps = list(OPERATOR_UI_STEPS.get(key) or OPERATOR_UI_STEPS[DEFAULT_CLIENT])
+    ns = (namespace or "").strip()
+    if ns:
+        steps = [
+            s.replace("named Gitea MCP server entry (namespace)", f"namespace '{ns}'")
+            .replace("affected gitea-* server entry", f"server entry '{ns}'")
+            .replace("named namespace", f"namespace '{ns}'")
+            for s in steps
+        ]
+    return steps
+
+
+def build_reconnect_request(
+    *,
+    namespace: str,
+    profile: str | None = None,
+    pid: int | str | None = None,
+    session_id: str | None = None,
+    startup_sha: str | None = None,
+    current_master_sha: str | None = None,
+    boundary_status: str | None = None,
+    reason: str | None = None,
+    client: str | None = DEFAULT_CLIENT,
+    live_stale: bool | None = None,
+    in_parity: bool | None = None,
+    restart_required: bool | None = None,
+    stop_required: bool | None = None,
+    extra: Mapping[str, Any] | None = None,
+) -> dict[str, Any]:
+    """Build the structured reconnect-request / typed-blocker payload (#678).
+
+    Never mutates process, config, or session state. Always side-effect free.
+    """
+    ns = (namespace or "").strip() or "unknown"
+    normalized_reason = normalize_reason(reason)
+    boundary = (boundary_status or "").strip() or classify_boundary_status(
+        startup_sha=startup_sha,
+        current_master_sha=current_master_sha,
+        live_stale=live_stale,
+        in_parity=in_parity,
+    )
+
+    reconnect_needed = True
+    if normalized_reason == REASON_NOT_REQUIRED and boundary == BOUNDARY_CLEAN:
+        reconnect_needed = False
+    if restart_required is False and stop_required is False and boundary == BOUNDARY_CLEAN:
+        # Explicit healthy probe
+        if normalized_reason in (REASON_NOT_REQUIRED, REASON_UNSPECIFIED):
+            reconnect_needed = False
+            normalized_reason = REASON_NOT_REQUIRED
+
+    if restart_required is True or stop_required is True:
+        reconnect_needed = True
+        if normalized_reason in (REASON_NOT_REQUIRED, REASON_UNSPECIFIED):
+            normalized_reason = REASON_STALE_RUNTIME
+
+    client_key = normalize_client(client)
+    steps = operator_ui_steps(client_key, namespace=ns)
+
+    result: dict[str, Any] = {
+        "success": True,
+        "read_only": True,
+        "reconnect_performed": False,
+        "mutation_performed": False,
+        "reconnect_needed": reconnect_needed,
+        "namespace": ns,
+        "profile": (profile or "").strip() or None,
+        "pid": pid,
+        "session_id": (session_id or "").strip() or None,
+        "startup_sha": (startup_sha or "").strip() or None,
+        "current_master_sha": (current_master_sha or "").strip() or None,
+        "boundary_status": boundary,
+        "reason": normalized_reason,
+        "client": client_key,
+        "forbidden_recovery_paths": list(FORBIDDEN_RECOVERY_PATHS),
+        "post_reconnect_preflight": [
+            "gitea_whoami",
+            "gitea_resolve_task_capability",
+            "original_task",
+        ],
+        "exact_safe_next_action": None,
+        "blocker_kind": BLOCKER_NONE,
+        "operator_ui_steps": steps,
+        "typed_blocker": None,
+    }
+
+    if reconnect_needed:
+        result["blocker_kind"] = BLOCKER_OPERATOR_RECONNECT
+        result["stop_required"] = True
+        result["restart_required"] = True
+        result["exact_safe_next_action"] = (
+            f"blocker_kind={BLOCKER_OPERATOR_RECONNECT}: operator must reconnect "
+            f"MCP namespace '{ns}' via the host UI (client={client_key}). "
+            "Do not pkill, touch configs, edit session state, or use raw API. "
+            "After reconnect, restart from gitea_whoami → "
+            "gitea_resolve_task_capability → task."
+        )
+        result["typed_blocker"] = {
+            "blocker_kind": BLOCKER_OPERATOR_RECONNECT,
+            "namespaces": [ns],
+            "why_reconnect_required": normalized_reason,
+            "operator_ui_steps": steps,
+            "client": client_key,
+            "forbidden_recovery_paths": list(FORBIDDEN_RECOVERY_PATHS),
+            "instruction_after_reconnect": (
+                "Rerun the blocked workflow from preflight "
+                "(gitea_whoami → gitea_resolve_task_capability → task). "
+                "Do not continue mid-mutation from pre-reconnect state."
+            ),
+        }
+    else:
+        result["stop_required"] = False
+        result["restart_required"] = False
+        result["exact_safe_next_action"] = (
+            f"Reconnect not required for namespace '{ns}' "
+            f"(boundary_status={boundary}). Proceed with the original task."
+        )
+
+    if extra:
+        for key, value in extra.items():
+            if key not in result:
+                result[key] = value
+
+    return result
+
+
+def reasons_never_suggest_forbidden(text: str) -> bool:
+    """Return True when *text* does not recommend a forbidden recovery path.
+
+    Mentions that *ban* a path (e.g. ``Do not pkill`` / ``never edit session
+    state``) are allowed. Positive recommendations such as ``use pkill`` or
+    ``run killall`` fail.
+    """
+    import re
+
+    lowered = (text or "").lower()
+    # Strip common ban prefixes so "do not pkill" does not trip positive checks.
+    scrubbed = re.sub(
+        r"\b(?:do not|don't|never|must not|forbid(?:den)?|ban(?:ned)?)\b"
+        r"[^.!;\n]{0,80}",
+        " ",
+        lowered,
+    )
+    # Positive imperative / advisory forms that would tell an agent to do harm.
+    positive_suggestions = (
+        "use pkill",
+        "run pkill",
+        "try pkill",
+        "pkill -f",
+        "use killall",
+        "run killall",
+        "killall mcp",
+        "use kill ",
+        "run kill ",
+        "touch the mcp",
+        "touch mcp config",
+        "utime(",
+        "edit the mcp config to recover",
+        "edit .env to recover",
+        "import gitea_mcp_server",
+        "python -c 'import gitea_mcp",
+    )
+    return not any(frag in scrubbed for frag in positive_suggestions)

mcp_restart_paths.py

@@ -225,8 +225,33 @@ _RESTART_PATHS: tuple[RestartPath, ...] = (
             "exact_safe_next_action pointing at IDE/client reconnect; performs "
             "no restart, thread spawn, config touch, or os._exit."
         ),
-        locations=("gitea_mcp_server.py (gitea_resolve_task_capability)",),
+        locations=(
-        references=("#685", "#657"),
+            "gitea_mcp_server.py (gitea_resolve_task_capability)",
+            "gitea_mcp_server.py (gitea_request_mcp_reconnect)",
+            "mcp_client_reconnect.py",
+        ),
+        references=("#685", "#657", "#678"),
+    ),
+    RestartPath(
+        path_id="codex_client_reconnect_request",
+        title="Sanctioned Codex/LLM reconnect request tool",
+        mechanism=(
+            "gitea_request_mcp_reconnect: agents invoke a report-only tool that "
+            "returns namespace/profile/pid/startup SHA/master SHA/boundary "
+            "status plus a typed operator blocker with exact client UI steps."
+        ),
+        classification=CLASS_GUARDED_FAIL_CLOSED,
+        guard=(
+            "Report-only (#678): never restarts, kills, reloads, or edits "
+            "config; recovery is always host/operator reconnect. Forbidden "
+            "paths (pkill, touch, .env/config/session-state hacks) are listed "
+            "and never recommended."
+        ),
+        locations=(
+            "mcp_client_reconnect.py",
+            "gitea_mcp_server.py (gitea_request_mcp_reconnect)",
+        ),
+        references=("#678", "#630", "#685", "#657"),
     ),
     RestartPath(
         path_id="manual_daemon_kill",
@@ -271,7 +296,8 @@ _RESTART_PATHS: tuple[RestartPath, ...] = (
         title="Host/IDE MCP reconnect",
         mechanism=(
             "A manual `/mcp reconnect` (or equivalent host action) that the "
-            "IDE performs to recreate the MCP client connection."
+            "IDE performs to recreate the MCP client connection. Agents obtain "
+            "exact UI steps via gitea_request_mcp_reconnect (#678)."
         ),
         classification=CLASS_HOST_RESIDUAL,
         guard=(
@@ -279,8 +305,12 @@ _RESTART_PATHS: tuple[RestartPath, ...] = (
             "gates point operators toward; documented as residual host "
             "behavior. No in-process code initiates it."
         ),
-        locations=("host/IDE",),
+        locations=(
-        references=("#584", "#656", "#657"),
+            "host/IDE",
+            "mcp_client_reconnect.py",
+            "gitea_mcp_server.py (gitea_request_mcp_reconnect)",
+        ),
+        references=("#584", "#656", "#657", "#678"),
         residual_host=True,
     ),
     RestartPath(

tests/test_issue_678_mcp_client_reconnect.py

@@ -0,0 +1,323 @@
+"""Tests for sanctioned Codex MCP reconnect request surface (#678)."""
+
+from __future__ import annotations
+
+import os
+import unittest
+from unittest import mock
+
+import mcp_client_reconnect as mcr
+
+
+class NormalizeReasonTests(unittest.TestCase):
+    def test_stale_runtime_aliases(self):
+        self.assertEqual(mcr.normalize_reason("stale-runtime"), mcr.REASON_STALE_RUNTIME)
+        self.assertEqual(mcr.normalize_reason("stale_runtime"), mcr.REASON_STALE_RUNTIME)
+        self.assertEqual(mcr.normalize_reason("STALE"), mcr.REASON_STALE_RUNTIME)
+
+    def test_transport_eof_aliases(self):
+        self.assertEqual(mcr.normalize_reason("transport_eof"), mcr.REASON_TRANSPORT_EOF)
+        self.assertEqual(mcr.normalize_reason("EOF"), mcr.REASON_TRANSPORT_EOF)
+        self.assertEqual(
+            mcr.normalize_reason("client_is_closing"), mcr.REASON_TRANSPORT_EOF
+        )
+
+    def test_missing_namespace(self):
+        self.assertEqual(
+            mcr.normalize_reason("missing_namespace"), mcr.REASON_MISSING_NAMESPACE
+        )
+
+    def test_empty_is_unspecified(self):
+        self.assertEqual(mcr.normalize_reason(None), mcr.REASON_UNSPECIFIED)
+        self.assertEqual(mcr.normalize_reason(""), mcr.REASON_UNSPECIFIED)
+
+
+class BoundaryClassificationTests(unittest.TestCase):
+    def test_clean_when_shas_match(self):
+        self.assertEqual(
+            mcr.classify_boundary_status(
+                startup_sha="abc", current_master_sha="abc"
+            ),
+            mcr.BOUNDARY_CLEAN,
+        )
+
+    def test_mismatch_when_shas_differ(self):
+        self.assertEqual(
+            mcr.classify_boundary_status(
+                startup_sha="aaa", current_master_sha="bbb"
+            ),
+            mcr.BOUNDARY_MISMATCH,
+        )
+
+    def test_stale_when_live_stale(self):
+        self.assertEqual(
+            mcr.classify_boundary_status(
+                startup_sha="aaa",
+                current_master_sha="aaa",
+                live_stale=True,
+            ),
+            mcr.BOUNDARY_STALE,
+        )
+
+
+class BuildReconnectRequestTests(unittest.TestCase):
+    def test_stale_runtime_returns_typed_blocker_with_codex_steps(self):
+        result = mcr.build_reconnect_request(
+            namespace="gitea-author",
+            profile="prgs-author",
+            pid=1234,
+            session_id="sess-1",
+            startup_sha="aaa111",
+            current_master_sha="bbb222",
+            reason="stale-runtime",
+            client="codex",
+            restart_required=True,
+            stop_required=True,
+        )
+        self.assertTrue(result["success"])
+        self.assertTrue(result["read_only"])
+        self.assertFalse(result["reconnect_performed"])
+        self.assertFalse(result["mutation_performed"])
+        self.assertTrue(result["reconnect_needed"])
+        self.assertEqual(result["namespace"], "gitea-author")
+        self.assertEqual(result["profile"], "prgs-author")
+        self.assertEqual(result["pid"], 1234)
+        self.assertEqual(result["session_id"], "sess-1")
+        self.assertEqual(result["startup_sha"], "aaa111")
+        self.assertEqual(result["current_master_sha"], "bbb222")
+        self.assertEqual(result["boundary_status"], mcr.BOUNDARY_MISMATCH)
+        self.assertEqual(result["blocker_kind"], mcr.BLOCKER_OPERATOR_RECONNECT)
+        self.assertIsNotNone(result["typed_blocker"])
+        blocker = result["typed_blocker"]
+        self.assertEqual(blocker["namespaces"], ["gitea-author"])
+        self.assertEqual(blocker["why_reconnect_required"], mcr.REASON_STALE_RUNTIME)
+        self.assertTrue(any("Codex" in s or "Reload" in s for s in blocker["operator_ui_steps"]))
+        self.assertIn("pkill", " ".join(result["forbidden_recovery_paths"]).lower())
+        self.assertTrue(
+            mcr.reasons_never_suggest_forbidden(result["exact_safe_next_action"] or "")
+        )
+        # Must not recommend forbidden recovery.
+        for step in blocker["operator_ui_steps"]:
+            self.assertTrue(mcr.reasons_never_suggest_forbidden(step), step)
+
+    def test_transport_eof_typed_blocker(self):
+        result = mcr.build_reconnect_request(
+            namespace="gitea-reviewer",
+            reason="transport_eof",
+            client="claude_code",
+        )
+        self.assertTrue(result["reconnect_needed"])
+        self.assertEqual(result["reason"], mcr.REASON_TRANSPORT_EOF)
+        self.assertEqual(result["client"], "claude_code")
+        steps = " ".join(result["operator_ui_steps"]).lower()
+        self.assertIn("/mcp", steps)
+
+    def test_missing_namespace_typed_blocker(self):
+        result = mcr.build_reconnect_request(
+            namespace="gitea-merger",
+            reason="missing_namespace",
+            client="codex",
+        )
+        self.assertTrue(result["reconnect_needed"])
+        self.assertEqual(result["reason"], mcr.REASON_MISSING_NAMESPACE)
+        self.assertEqual(
+            result["typed_blocker"]["blocker_kind"], mcr.BLOCKER_OPERATOR_RECONNECT
+        )
+
+    def test_healthy_not_required(self):
+        result = mcr.build_reconnect_request(
+            namespace="gitea-tools",
+            startup_sha="deadbeef",
+            current_master_sha="deadbeef",
+            reason="not_required",
+            client="codex",
+            in_parity=True,
+            restart_required=False,
+            stop_required=False,
+        )
+        self.assertFalse(result["reconnect_needed"])
+        self.assertEqual(result["blocker_kind"], mcr.BLOCKER_NONE)
+        self.assertIsNone(result["typed_blocker"])
+        self.assertFalse(result["stop_required"])
+        self.assertFalse(result["restart_required"])
+        self.assertIn("not required", (result["exact_safe_next_action"] or "").lower())
+
+    def test_successful_reconnect_report_fields_present(self):
+        """AC2: reconnect result reports required fields (even when needed)."""
+        result = mcr.build_reconnect_request(
+            namespace="gitea-controller",
+            profile="prgs-controller",
+            pid=99,
+            session_id="sid",
+            startup_sha="s" * 40,
+            current_master_sha="c" * 40,
+            reason="stale-runtime",
+        )
+        for key in (
+            "namespace",
+            "profile",
+            "pid",
+            "session_id",
+            "startup_sha",
+            "current_master_sha",
+            "boundary_status",
+        ):
+            self.assertIn(key, result)
+            self.assertIsNotNone(result[key], key)
+
+
+class ToolSurfaceTests(unittest.TestCase):
+    """Exercise gitea_request_mcp_reconnect with a stubbed server context."""
+
+    def test_tool_is_registered_and_side_effect_free(self):
+        import gitea_mcp_server as srv
+
+        self.assertTrue(hasattr(srv, "gitea_request_mcp_reconnect"))
+        with mock.patch.object(srv, "_profile_operation_gate", return_value=None):
+            with mock.patch.object(
+                srv,
+                "get_profile",
+                return_value={
+                    "profile_name": "prgs-author",
+                    "role_kind": "author",
+                    "role": "author",
+                },
+            ):
+                with mock.patch.object(
+                    srv,
+                    "_current_master_parity",
+                    return_value={
+                        "startup_head": "a" * 40,
+                        "current_head": "a" * 40,
+                        "daemon_start_head": "a" * 40,
+                        "local_head": "a" * 40,
+                        "in_parity": True,
+                        "stale": False,
+                        "restart_required": False,
+                        "determinable": True,
+                        "live_stale": False,
+                        "live_known": True,
+                        "reasons": [],
+                    },
+                ):
+                    with mock.patch.object(
+                        srv.master_parity_gate,
+                        "format_parity",
+                        return_value="in parity",
+                    ):
+                        with mock.patch.object(
+                            srv.role_namespace_gate,
+                            "infer_mcp_namespace",
+                            return_value="gitea-author",
+                        ):
+                            with mock.patch.object(
+                                srv.session_ctx,
+                                "mutation_context_audit_fields",
+                                return_value={"session_profile": "prgs-author"},
+                            ):
+                                result = srv.gitea_request_mcp_reconnect(
+                                    namespace="gitea-author",
+                                    reason="not_required",
+                                    client="codex",
+                                    remote="prgs",
+                                )
+        self.assertTrue(result.get("success"))
+        self.assertFalse(result.get("reconnect_performed"))
+        self.assertFalse(result.get("mutation_performed"))
+        self.assertEqual(result.get("namespace"), "gitea-author")
+        self.assertEqual(result.get("profile"), "prgs-author")
+        self.assertEqual(result.get("pid"), os.getpid())
+        self.assertIn("startup_sha", result)
+        self.assertIn("current_master_sha", result)
+        self.assertIn("boundary_status", result)
+        self.assertTrue(
+            mcr.reasons_never_suggest_forbidden(
+                result.get("exact_safe_next_action") or ""
+            )
+        )
+
+    def test_tool_stale_returns_typed_blocker(self):
+        import gitea_mcp_server as srv
+
+        with mock.patch.object(srv, "_profile_operation_gate", return_value=None):
+            with mock.patch.object(
+                srv,
+                "get_profile",
+                return_value={
+                    "profile_name": "prgs-reconciler",
+                    "role_kind": "reconciler",
+                    "role": "reconciler",
+                },
+            ):
+                with mock.patch.object(
+                    srv,
+                    "_current_master_parity",
+                    return_value={
+                        "startup_head": "a" * 40,
+                        "current_head": "b" * 40,
+                        "daemon_start_head": "a" * 40,
+                        "local_head": "b" * 40,
+                        "in_parity": False,
+                        "stale": True,
+                        "restart_required": True,
+                        "determinable": True,
+                        "live_stale": True,
+                        "live_known": True,
+                        "reasons": ["stale"],
+                    },
+                ):
+                    with mock.patch.object(
+                        srv.master_parity_gate,
+                        "format_parity",
+                        return_value="stale",
+                    ):
+                        with mock.patch.object(
+                            srv.role_namespace_gate,
+                            "infer_mcp_namespace",
+                            return_value="gitea-reconciler",
+                        ):
+                            with mock.patch.object(
+                                srv.session_ctx,
+                                "mutation_context_audit_fields",
+                                return_value={},
+                            ):
+                                result = srv.gitea_request_mcp_reconnect(
+                                    reason="stale-runtime",
+                                    client="codex",
+                                )
+        self.assertTrue(result["reconnect_needed"])
+        self.assertEqual(
+            result["blocker_kind"], mcr.BLOCKER_OPERATOR_RECONNECT
+        )
+        self.assertIsNotNone(result["typed_blocker"])
+        self.assertIn("gitea-reconciler", result["typed_blocker"]["namespaces"])
+        self.assertTrue(result["stop_required"])
+        self.assertTrue(result["restart_required"])
+        self.assertTrue(
+            mcr.reasons_never_suggest_forbidden(
+                result.get("exact_safe_next_action") or ""
+            )
+        )
+
+
+class InventoryRegistrationTests(unittest.TestCase):
+    def test_reconnect_path_in_restart_inventory(self):
+        import mcp_restart_paths as mrp
+
+        ids = {p.path_id for p in mrp.iter_restart_paths()}
+        self.assertIn("codex_client_reconnect_request", ids)
+        self.assertIn("ide_client_reconnect", ids)
+
+    def test_tool_name_in_documented_inventory(self):
+        import mcp_tool_inventory as inv
+
+        doc_path = os.path.join(
+            os.path.dirname(os.path.dirname(__file__)), inv.INVENTORY_DOC_PATH
+        )
+        with open(doc_path, encoding="utf-8") as handle:
+            documented = inv.parse_documented_inventory(handle.read())
+        self.assertIn("gitea_request_mcp_reconnect", documented)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_mcp_restart_governance.py

@@ -0,0 +1,478 @@
+"""Concurrent-session MCP restart safety & dogfooding test suite (#666).
+
+Automated test suite proving all 10 dogfooding bullets required by Issue #666:
+1. One LLM cannot restart MCP unilaterally (role-based restart authorization matrix).
+2. New work stops during drain (assignments_stopped gate enforcement).
+3. Active safe work can finish (ack collection / graceful completion before restart).
+4. Unsafe mutations block restart (in-flight author/reviewer mutation gates).
+5. Session state is durably checkpointed (checkpoints_complete validation).
+6. Leases/locks not silently orphaned (lease lifecycle & post-restart lease audit).
+7. Sessions resume or receive canonical next action (reconcile proof canonical next action).
+8. Failed drain creates durable incident work (durable incident descriptor & bridge integration).
+9. Restart of one component does not unnecessarily interrupt unrelated work (scoped restart impact).
+10. Restart/upgrade workflows do not require manual chat reconstruction (state handoff ledger & completion proof).
+
+Links parent #655, vision #652, roadmap #653, #658, #659, #660, #661, #662, #663.
+"""
+
+from __future__ import annotations
+
+import os
+import unittest
+from datetime import datetime, timedelta, timezone
+
+import drain_proof as dp
+import mcp_restart_paths as rp
+import post_restart_reconcile as prr
+import restart_coordinator as rc
+from restart_coordinator import RestartClass
+
+NOW = datetime(2026, 7, 25, 12, 0, 0, tzinfo=timezone.utc)
+SECRET = b"test-secret-dogfooding-issue-666-0123456789"
+
+
+def _live_pid() -> int:
+    return os.getpid()
+
+
+def _clean_drain_state() -> dict:
+    return {
+        "assignments_stopped": True,
+        "checkpoints_complete": True,
+        "handoffs_verified": True,
+        "leases_handled": True,
+        "acks": {},
+        "ack_timeout_policy_applied": False,
+    }
+
+
+def _clean_inventory() -> dict:
+    return {
+        "service_health": {"healthy": True},
+        "clients": [],
+        "sessions": [
+            {
+                "session_id": "prgs-controller-1",
+                "role": "controller",
+                "profile": "prgs-controller",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            }
+        ],
+        "checkpoints": [],
+        "leases": [],
+        "capabilities": {},
+        "worktree_bindings": [],
+        "pending_mutations": [],
+        "inventory_complete": True,
+    }
+
+
+class TestBullet1UnilateralRestartForbidden(unittest.TestCase):
+    """Bullet 1: One LLM cannot restart MCP unilaterally."""
+
+    def test_worker_role_unilateral_full_restart_denied(self):
+        policy = rc.RESTART_CLASS_POLICIES[RestartClass.FULL_MCP_RESTART]
+        for worker_role in ("author", "reviewer", "merger", "reconciler"):
+            self.assertNotIn(
+                worker_role,
+                policy.request_roles,
+                f"Worker role '{worker_role}' must not unilaterally authorize FULL_MCP_RESTART",
+            )
+
+    def test_privileged_role_full_restart_authorized(self):
+        policy = rc.RESTART_CLASS_POLICIES[RestartClass.FULL_MCP_RESTART]
+        for priv_role in ("controller", "operator", "admin"):
+            self.assertIn(
+                priv_role,
+                policy.request_roles,
+                f"Privileged role '{priv_role}' must be authorized for FULL_MCP_RESTART",
+            )
+
+    def test_evaluate_impact_records_unauthorized_worker_request(self):
+        report = rc.evaluate_restart_impact(
+            {"sessions": [], "leases": [], "inventory_complete": True},
+            now=NOW,
+            restart_class=RestartClass.FULL_MCP_RESTART,
+            requester_role="author",
+            requesting_session_id="prgs-author-123",
+        )
+        self.assertFalse(report.role_authorized)
+        self.assertEqual(report.verdict, rc.VERDICT_UNSAFE)
+        self.assertTrue(any("may not request" in r.lower() or "authorization denied" in r.lower() for r in report.reasons))
+
+
+class TestBullet2NewWorkStopsDuringDrain(unittest.TestCase):
+    """Bullet 2: New work stops during drain."""
+
+    def test_assignments_stopped_false_blocks_drain_proof(self):
+        state = _clean_drain_state()
+        state["assignments_stopped"] = False
+
+        impact = rc.evaluate_restart_impact(
+            {"sessions": [], "leases": [], "inventory_complete": True},
+            now=NOW,
+        ).as_dict()
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=impact,
+            drain_state=state,
+            now=NOW,
+        )
+
+        self.assertFalse(proof.clean)
+        check = next(c for c in proof.checks if c.name == dp.CHECK_ASSIGNMENTS_STOPPED)
+        self.assertFalse(check.passed)
+
+        gate = dp.gate_apply_restart(proof=proof.as_dict(), secret=SECRET, now=NOW)
+        self.assertEqual(gate.verdict, dp.GATE_DENY)
+        self.assertFalse(gate.allow)
+        self.assertTrue(any("drain proof invalid" in r.lower() or "assignments_stopped" in r.lower() for r in gate.reasons))
+
+
+class TestBullet3ActiveSafeWorkCanFinish(unittest.TestCase):
+    """Bullet 3: Active safe work can finish."""
+
+    def test_active_safe_sessions_ack_allows_clean_drain(self):
+        sessions = [
+            {
+                "session_id": "prgs-controller-1",
+                "role": "controller",
+                "profile": "prgs-controller",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+            {
+                "session_id": "prgs-reviewer-42",
+                "role": "reviewer",
+                "profile": "prgs-reviewer",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+        ]
+        leases = [
+            {
+                "lease_id": "lease-ro",
+                "session_id": "prgs-reviewer-42",
+                "role": "reviewer",
+                "phase": "reviewing",
+                "is_mutating": False,
+                "expires_at": (NOW + timedelta(minutes=5)).isoformat(),
+                "pid": _live_pid(),
+            }
+        ]
+
+        impact = rc.evaluate_restart_impact(
+            {"sessions": sessions, "leases": leases, "inventory_complete": True},
+            now=NOW,
+            requesting_session_id="prgs-controller-1",
+        ).as_dict()
+
+        state = _clean_drain_state()
+        state["acks"] = {"prgs-reviewer-42": "ack"}
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=impact,
+            drain_state=state,
+            now=NOW,
+        )
+
+        self.assertTrue(proof.clean)
+        gate = dp.gate_apply_restart(proof=proof.as_dict(), secret=SECRET, now=NOW)
+        self.assertTrue(gate.allow)
+        self.assertEqual(gate.verdict, dp.GATE_ALLOW)
+
+
+class TestBullet4UnsafeMutationsBlockRestart(unittest.TestCase):
+    """Bullet 4: Unsafe mutations block restart."""
+
+    def test_inflight_unsafe_mutation_yields_unsafe_verdict(self):
+        sessions = [
+            {
+                "session_id": "prgs-controller-1",
+                "role": "controller",
+                "profile": "prgs-controller",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+            {
+                "session_id": "prgs-author-99",
+                "role": "author",
+                "profile": "prgs-author",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+        ]
+        leases = [
+            {
+                "lease_id": "lease-mutating",
+                "session_id": "prgs-author-99",
+                "role": "author",
+                "phase": "implementing",
+                "worktree_path": "/Users/jasonwalker/Development/Gitea-Tools/branches/feat-test",
+                "freshness": {"freshness": "active"},
+                "expires_at": (NOW + timedelta(minutes=5)).isoformat(),
+                "pid": _live_pid(),
+            }
+        ]
+
+        report = rc.evaluate_restart_impact(
+            {"sessions": sessions, "leases": leases, "inventory_complete": True},
+            now=NOW,
+            requesting_session_id="prgs-controller-1",
+        )
+
+        self.assertEqual(report.verdict, rc.VERDICT_UNSAFE)
+        self.assertFalse(report.allow_restart)
+        self.assertGreater(len(report.mutations), 0)
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=report.as_dict(),
+            drain_state=_clean_drain_state(),
+            now=NOW,
+        )
+
+        self.assertFalse(proof.clean)
+        check = next(c for c in proof.checks if c.name == dp.CHECK_NO_INFLIGHT_MUTATIONS)
+        self.assertFalse(check.passed)
+
+        gate = dp.gate_apply_restart(proof=proof.as_dict(), secret=SECRET, now=NOW)
+        self.assertEqual(gate.verdict, dp.GATE_DENY)
+        self.assertFalse(gate.allow)
+
+
+class TestBullet5DurableSessionCheckpoints(unittest.TestCase):
+    """Bullet 5: Session state is durably checkpointed."""
+
+    def test_incomplete_checkpoints_blocks_drain_proof(self):
+        state = _clean_drain_state()
+        state["checkpoints_complete"] = False
+
+        impact = rc.evaluate_restart_impact(
+            {"sessions": [], "leases": [], "inventory_complete": True},
+            now=NOW,
+        ).as_dict()
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=impact,
+            drain_state=state,
+            now=NOW,
+        )
+
+        self.assertFalse(proof.clean)
+        check = next(c for c in proof.checks if c.name == dp.CHECK_CHECKPOINTS_COMPLETE)
+        self.assertFalse(check.passed)
+
+    def test_post_restart_reconcile_audits_checkpoint_dimension(self):
+        inv = _clean_inventory()
+        inv["checkpoints_available"] = True
+        inv["checkpoints"] = [
+            {
+                "session_id": "prgs-author-99",
+                "checkpoint_id": "chk-1",
+                "stale": True,
+            }
+        ]
+
+        proof = prr.reconcile_after_restart(inv, now=NOW, mode=prr.MODE_ENFORCE)
+        chk_item = next(i for i in proof.items if i.dimension == prr.DIM_CHECKPOINTS)
+        self.assertIn(chk_item.status, (prr.ITEM_UNRESOLVED, prr.ITEM_DEGRADED, prr.ITEM_SKIPPED))
+
+
+class TestBullet6LeasesNotSilentlyOrphaned(unittest.TestCase):
+    """Bullet 6: Leases/locks not silently orphaned."""
+
+    def test_unhandled_leases_block_drain_proof(self):
+        state = _clean_drain_state()
+        state["leases_handled"] = False
+
+        impact = rc.evaluate_restart_impact(
+            {"sessions": [], "leases": [], "inventory_complete": True},
+            now=NOW,
+        ).as_dict()
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=impact,
+            drain_state=state,
+            now=NOW,
+        )
+
+        self.assertFalse(proof.clean)
+        check = next(c for c in proof.checks if c.name == dp.CHECK_LEASES_HANDLED)
+        self.assertFalse(check.passed)
+
+    def test_post_restart_reconcile_audits_all_leases(self):
+        inv = _clean_inventory()
+        inv["leases"] = [
+            {
+                "lease_id": "lease-orphaned-1",
+                "session_id": "prgs-author-dead",
+                "role": "author",
+                "status": "active",
+                "freshness": "expired",
+                "expires_at": (NOW - timedelta(minutes=10)).isoformat(),
+            }
+        ]
+
+        proof = prr.reconcile_after_restart(inv, now=NOW, mode=prr.MODE_LOG_ONLY)
+        lease_item = next(i for i in proof.items if i.dimension == prr.DIM_LEASES)
+        self.assertIsNotNone(lease_item)
+        self.assertTrue(lease_item.summary)
+
+
+class TestBullet7SessionsResumeOrReceiveNextAction(unittest.TestCase):
+    """Bullet 7: Sessions resume or receive canonical next action."""
+
+    def test_reconcile_provides_canonical_next_action_for_unresolved(self):
+        inv = _clean_inventory()
+        inv["pending_mutations"] = [
+            {
+                "mutation_id": "mut-404",
+                "session_id": "prgs-author-77",
+                "phase": "implementing",
+                "issue_number": 666,
+            }
+        ]
+
+        proof = prr.reconcile_after_restart(inv, now=NOW, mode=prr.MODE_ENFORCE)
+        self.assertEqual(proof.overall_status, prr.STATUS_DEGRADED)
+        self.assertTrue(proof.mutation_hold)
+        self.assertTrue(proof.note)
+        self.assertGreater(len(proof.proposed_follow_ups), 0)
+
+
+class TestBullet8FailedDrainCreatesIncidentWork(unittest.TestCase):
+    """Bullet 8: Failed drain creates durable incident work."""
+
+    def test_denied_drain_gate_mints_durable_incident_descriptor(self):
+        impact = rc.evaluate_restart_impact(
+            {"sessions": [], "leases": [], "inventory_complete": True},
+            now=NOW,
+        ).as_dict()
+
+        state = _clean_drain_state()
+        state["assignments_stopped"] = False
+
+        proof = dp.build_drain_proof(
+            secret=SECRET,
+            impact_report=impact,
+            drain_state=state,
+            now=NOW,
+        )
+
+        gate = dp.gate_apply_restart(proof=proof.as_dict(), secret=SECRET, now=NOW)
+        self.assertEqual(gate.verdict, dp.GATE_DENY)
+
+        incident = gate.incident
+        self.assertIsNotNone(incident)
+        self.assertEqual(incident["kind"], "restart_drain_gate_denied")
+        self.assertTrue(any("assignments_stopped" in r for r in incident["reasons"]))
+
+
+class TestBullet9ScopedRestartNonInterference(unittest.TestCase):
+    """Bullet 9: Restart of one component does not unnecessarily interrupt unrelated work."""
+
+    def test_scoped_role_restart_impacts_only_target_role(self):
+        sessions = [
+            {
+                "session_id": "prgs-controller-1",
+                "role": "controller",
+                "profile": "prgs-controller",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+            {
+                "session_id": "prgs-author-10",
+                "role": "author",
+                "profile": "prgs-author",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+            {
+                "session_id": "prgs-reviewer-20",
+                "role": "reviewer",
+                "profile": "prgs-reviewer",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+        ]
+
+        policy = rc.RESTART_CLASS_POLICIES[RestartClass.ROLE_RUNTIME_RESTART]
+        report = rc.evaluate_restart_impact(
+            {"sessions": sessions, "leases": [], "inventory_complete": True},
+            now=NOW,
+            restart_class=RestartClass.ROLE_RUNTIME_RESTART,
+            target_role="reviewer",
+            requesting_session_id="prgs-controller-1",
+            requester_role="controller",
+            requester_permissions=list(policy.request_roles),
+            controller_approved=True,
+        )
+
+        self.assertTrue(report.role_authorized)
+
+    def test_scoped_connector_restart_limits_blast_radius(self):
+        sessions = [
+            {
+                "session_id": "prgs-author-10",
+                "role": "author",
+                "connector": "gitea-author",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+            {
+                "session_id": "prgs-reviewer-20",
+                "role": "reviewer",
+                "connector": "gitea-reviewer",
+                "pid": _live_pid(),
+                "status": "active",
+                "last_heartbeat_at": NOW.isoformat(),
+            },
+        ]
+
+        policy = rc.RESTART_CLASS_POLICIES[RestartClass.CONNECTOR_RESTART]
+        report = rc.evaluate_restart_impact(
+            {"sessions": sessions, "leases": [], "inventory_complete": True},
+            now=NOW,
+            restart_class=RestartClass.CONNECTOR_RESTART,
+            target_connector="gitea-author",
+            requesting_session_id="prgs-controller-1",
+            requester_role="controller",
+            requester_permissions=list(policy.request_roles),
+            controller_approved=True,
+        )
+
+        self.assertIsNotNone(report)
+
+
+class TestBullet10NoManualChatReconstruction(unittest.TestCase):
+    """Bullet 10: Restart/upgrade workflows do not require manual chat reconstruction."""
+
+    def test_end_to_end_restart_reconcile_handoff_proof(self):
+        inv = _clean_inventory()
+        proof = prr.reconcile_after_restart(inv, now=NOW, mode=prr.MODE_LOG_ONLY)
+
+        proof_dict = proof.as_dict()
+        self.assertEqual(proof_dict["overall_status"], prr.STATUS_COMPLETE)
+        self.assertFalse(proof_dict["mutation_hold"])
+        self.assertTrue(proof_dict["note"])
+        self.assertIn("links", proof_dict)
+        self.assertEqual(proof_dict["links"]["umbrella"], 655)
+
+
+if __name__ == "__main__":
+    unittest.main()

tests/test_webui_notifications.py

@@ -0,0 +1,465 @@
+"""Unit tests for Phase 3 Notifications and Human-Attention Console (#648)."""
+
+from __future__ import annotations
+
+import pytest
+from starlette.testclient import TestClient
+
+from webui.app import create_app
+from webui.notifications import (
+    ATTENTION_HUMAN_REQUIRED,
+    ATTENTION_OPERATOR,
+    ATTENTION_ROUTINE,
+    CATEGORY_AUTH,
+    CATEGORY_BLOCKER,
+    CATEGORY_LEASE,
+    CATEGORY_SYSTEM,
+    CATEGORY_VALIDATION,
+    CATEGORY_WORKFLOW,
+    NotificationItem,
+    NotificationSnapshot,
+    classify_attention_event,
+    load_notifications_snapshot,
+    snapshot_to_dict,
+)
+from webui.notification_views import render_notifications_page
+from webui.project_registry import load_registry
+from webui.queue_loader import QueueItem, QueueSnapshot
+from webui.lease_loader import CollisionWarning, LeaseSnapshot
+from webui.system_health import DependencyProbe, SystemHealthSnapshot, VersionInfo, StaleRuntime
+
+
+def test_classify_attention_event_rules():
+    # 1. Critical escalation boundaries -> human-required
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_AUTH, "Auth error", "Unauthorized access attempt", is_auth_failure=True
+    )
+    assert att_cls == ATTENTION_HUMAN_REQUIRED
+    assert req_human is True
+
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_SYSTEM, "Hard stop", "Hard stop triggered", is_hard_stop=True
+    )
+    assert att_cls == ATTENTION_HUMAN_REQUIRED
+    assert req_human is True
+
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_VALIDATION, "Validation Error", "Report validation failed", is_validation_failure=True
+    )
+    assert att_cls == ATTENTION_HUMAN_REQUIRED
+    assert req_human is True
+
+    # 2. Operational issues -> operator
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_BLOCKER, "PR Blocked", "Merge conflict detected", is_blocker=True
+    )
+    assert att_cls == ATTENTION_OPERATOR
+    assert req_human is False
+
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_LEASE, "Lease Expired", "Session lease expired", is_stale=True
+    )
+    assert att_cls == ATTENTION_OPERATOR
+    assert req_human is False
+
+    # 3. Routine workflow transitions -> routine
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_WORKFLOW, "PR Active", "PR in review"
+    )
+    assert att_cls == ATTENTION_ROUTINE
+    assert req_human is False
+
+
+def test_notification_snapshot_aggregation():
+    reg = load_registry()
+    proj_id = reg.projects[0].id if reg.projects else "gitea-tools"
+
+    mock_queue = QueueSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        prs=(
+            QueueItem(
+                number=101,
+                title="Blocked PR",
+                badges=("blocked",),
+                extra={},
+            ),
+            QueueItem(
+                number=102,
+                title="Normal PR",
+                badges=("in-review",),
+                extra={},
+            ),
+        ),
+        issues=(),
+        pr_pagination=None,
+        issue_pagination=None,
+    )
+
+    mock_leases = LeaseSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        issue_lock=None,
+        claim_inventory={},
+        reviewer_leases=(
+            {
+                "pr_number": 101,
+                "status": "expired",
+                "is_expired": True,
+            },
+        ),
+        duplicate_prs=(
+            CollisionWarning(
+                kind="duplicate_pr",
+                message="Multiple open PRs for issue #101",
+                issue_number=101,
+                pr_numbers=(101, 103),
+            ),
+        ),
+        duplicate_branches=(),
+        collision_history=(),
+        fetch_error=None,
+    )
+
+    mock_version = VersionInfo(
+        git_sha="abc1234",
+        git_describe="v1.0.0",
+        control_plane_schema_version=1,
+        python_version="3.11",
+        known=True,
+    )
+
+    mock_stale = StaleRuntime(
+        daemon_head="abc1234",
+        checkout_head="abc1234",
+        remote_head="abc1234",
+        stale=False,
+        determinable=True,
+        mutation_safe=True,
+        reasons=(),
+    )
+
+    mock_health = SystemHealthSnapshot(
+        status="degraded",
+        ready=False,
+        readiness_complete=True,
+        readiness_reasons=("Auth failure",),
+        service="webui",
+        mode="test",
+        version=mock_version,
+        started_at="2026-07-25T00:00:00Z",
+        uptime_seconds=100.0,
+        timestamp="2026-07-25T00:00:00Z",
+        deep_probes_requested=True,
+        dependencies=(
+            DependencyProbe(
+                name="auth_service",
+                kind="auth",
+                status="unauthorized",
+                detail="Token expired",
+                required=True,
+            ),
+        ),
+        mcp_namespaces=(),
+        stale_runtime=mock_stale,
+        probe_errors=(),
+    )
+
+    snapshot = load_notifications_snapshot(
+        proj_id,
+        load_queue=lambda _id: mock_queue,
+        load_leases=lambda **_kwargs: mock_leases,
+        load_health=lambda **_kwargs: mock_health,
+    )
+
+    assert snapshot.project_id == proj_id
+    assert snapshot.total_count == 5
+    assert snapshot.human_required_count >= 1  # auth probe failure
+    assert snapshot.operator_count >= 3  # blocked PR + expired lease + duplicate PR collision
+    assert snapshot.routine_count >= 1  # normal PR
+
+    # Inbox items should include operator and human-required items only
+    inbox_classes = {item.attention_class for item in snapshot.inbox_items}
+    assert ATTENTION_ROUTINE not in inbox_classes
+    assert ATTENTION_OPERATOR in inbox_classes
+    assert ATTENTION_HUMAN_REQUIRED in inbox_classes
+
+
+def test_snapshot_to_dict_and_redaction():
+    item = NotificationItem(
+        id="notif-1",
+        attention_class=ATTENTION_HUMAN_REQUIRED,
+        category=CATEGORY_AUTH,
+        title="Auth Error",
+        summary="Failed auth header: Bearer secret_token_12345",
+        work_kind="system",
+        work_number=None,
+        project_id="test-proj",
+        repo_label="org/repo",
+        created_at="2026-07-25T16:00:00Z",
+        requires_human=True,
+    )
+    snap = NotificationSnapshot(
+        project_id="test-proj",
+        repo_label="org/repo",
+        items=(item,),
+        human_required_count=1,
+        operator_count=0,
+        routine_count=0,
+        total_count=1,
+    )
+
+    data = snapshot_to_dict(snap)
+    assert data["project_id"] == "test-proj"
+    assert data["human_required_count"] == 1
+    assert len(data["inbox_items"]) == 1
+
+    # Redaction test
+    summary = data["inbox_items"][0]["summary"]
+    assert "secret_token_12345" not in summary
+    assert "<redacted>" in summary or "Bearer" in summary
+
+
+def test_notifications_html_views():
+    item = NotificationItem(
+        id="notif-1",
+        attention_class=ATTENTION_HUMAN_REQUIRED,
+        category=CATEGORY_AUTH,
+        title="Critical Auth Failure",
+        summary="Auth failure details",
+        work_kind="issue",
+        work_number=42,
+        project_id="test-proj",
+        repo_label="org/repo",
+        created_at="2026-07-25T16:00:00Z",
+        requires_human=True,
+    )
+    snap = NotificationSnapshot(
+        project_id="test-proj",
+        repo_label="org/repo",
+        items=(item,),
+        human_required_count=1,
+        operator_count=0,
+        routine_count=0,
+        total_count=1,
+    )
+
+    html = render_notifications_page(snap, filter_class="inbox")
+    assert "Notifications &amp; Attention Inbox" in html or "Notifications & Attention Inbox" in html
+    assert "Critical Auth Failure" in html
+    assert "HUMAN REQUIRED" in html
+    assert "Human Required" in html
+
+
+def test_notifications_app_routes():
+    app = create_app()
+    client = TestClient(app)
+
+    # 1. HTML Route
+    res = client.get("/notifications")
+    assert res.status_code == 200
+    assert "Notifications" in res.text
+    assert "Attention Inbox" in res.text
+
+    # 2. API Route /api/v1/notifications
+    res_api = client.get("/api/v1/notifications")
+    assert res_api.status_code == 200
+    json_data = res_api.json()
+    assert "human_required_count" in json_data
+    assert "operator_count" in json_data
+    assert "routine_count" in json_data
+    assert "inbox_items" in json_data
+
+    # 3. Compatibility Alias /api/notifications
+    res_alias = client.get("/api/notifications")
+    assert res_alias.status_code == 200
+    assert res_alias.json()["project_id"] == json_data["project_id"]
+
+
+def test_classify_ignores_human_authored_title_and_summary_keywords():
+    """B1: keywords in human-authored titles must not escalate routine work (#905)."""
+    # Routine transition whose title/summary mention critical-boundary words
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_WORKFLOW,
+        "record irrecoverable decision lock provenance",
+        "PR #999 'record irrecoverable decision lock provenance' is in routine state in-review.",
+    )
+    assert att_cls == ATTENTION_ROUTINE
+    assert req_human is False
+
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_WORKFLOW,
+        "fix unauthorized token path",
+        "Issue #1 'fix unauthorized token path' state: claimed. hard stop docs only.",
+    )
+    assert att_cls == ATTENTION_ROUTINE
+    assert req_human is False
+
+    # Structured flags still escalate (machine-driven)
+    att_cls, req_human = classify_attention_event(
+        CATEGORY_SYSTEM,
+        "anything",
+        "anything with hard stop in text",
+        is_hard_stop=True,
+    )
+    assert att_cls == ATTENTION_HUMAN_REQUIRED
+    assert req_human is True
+
+
+def test_notification_ids_are_unique_across_probe_errors_and_collisions():
+    """B2: published notification ids must be unique within a snapshot (#905)."""
+    reg = load_registry()
+    proj_id = reg.projects[0].id if reg.projects else "gitea-tools"
+
+    mock_queue = QueueSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        prs=(),
+        issues=(),
+        pr_pagination=None,
+        issue_pagination=None,
+    )
+    mock_leases = LeaseSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        issue_lock=None,
+        claim_inventory={},
+        reviewer_leases=(),
+        duplicate_prs=(
+            CollisionWarning(
+                kind="duplicate_pr",
+                message="Multiple open PRs for issue #10",
+                issue_number=10,
+                pr_numbers=(10, 11),
+            ),
+            CollisionWarning(
+                kind="duplicate_branch",
+                message="Another collision without issue",
+                issue_number=None,
+                pr_numbers=(12, 13),
+            ),
+            CollisionWarning(
+                kind="duplicate_pr",
+                message="Second issue collision",
+                issue_number=10,
+                pr_numbers=(14, 15),
+            ),
+        ),
+        duplicate_branches=(),
+        collision_history=(),
+        fetch_error=None,
+    )
+    mock_version = VersionInfo(
+        git_sha="abc1234",
+        git_describe="v1.0.0",
+        control_plane_schema_version=1,
+        python_version="3.11",
+        known=True,
+    )
+    mock_stale = StaleRuntime(
+        daemon_head="abc1234",
+        checkout_head="abc1234",
+        remote_head="abc1234",
+        stale=False,
+        determinable=True,
+        mutation_safe=True,
+        reasons=(),
+    )
+    mock_health = SystemHealthSnapshot(
+        status="degraded",
+        ready=False,
+        readiness_complete=True,
+        readiness_reasons=(),
+        service="webui",
+        mode="test",
+        version=mock_version,
+        started_at="2026-07-25T00:00:00Z",
+        uptime_seconds=100.0,
+        timestamp="2026-07-25T00:00:00Z",
+        deep_probes_requested=True,
+        dependencies=(),
+        mcp_namespaces=(),
+        stale_runtime=mock_stale,
+        probe_errors=("error alpha", "error beta"),
+    )
+
+    snapshot = load_notifications_snapshot(
+        proj_id,
+        load_queue=lambda _id: mock_queue,
+        load_leases=lambda **_kwargs: mock_leases,
+        load_health=lambda **_kwargs: mock_health,
+    )
+    ids = [item.id for item in snapshot.items]
+    assert len(ids) == len(set(ids)), f"duplicate notification ids: {ids}"
+    assert any(i.startswith(f"notif-sys-err-{proj_id}-") for i in ids)
+    assert any(i.startswith("notif-collision-") for i in ids)
+
+
+def test_probe_errors_do_not_set_fetch_error():
+    """B3: probe_errors must not be reported as fetch_error (#905)."""
+    reg = load_registry()
+    proj_id = reg.projects[0].id if reg.projects else "gitea-tools"
+
+    mock_queue = QueueSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        prs=(),
+        issues=(),
+        pr_pagination=None,
+        issue_pagination=None,
+        fetch_error=None,
+    )
+    mock_leases = LeaseSnapshot(
+        project_id=proj_id,
+        repo_label="org/repo",
+        issue_lock=None,
+        claim_inventory={},
+        reviewer_leases=(),
+        duplicate_prs=(),
+        duplicate_branches=(),
+        collision_history=(),
+        fetch_error=None,
+    )
+    mock_version = VersionInfo(
+        git_sha="abc1234",
+        git_describe="v1.0.0",
+        control_plane_schema_version=1,
+        python_version="3.11",
+        known=True,
+    )
+    mock_stale = StaleRuntime(
+        daemon_head="abc1234",
+        checkout_head="abc1234",
+        remote_head="abc1234",
+        stale=False,
+        determinable=True,
+        mutation_safe=True,
+        reasons=(),
+    )
+    mock_health = SystemHealthSnapshot(
+        status="degraded",
+        ready=False,
+        readiness_complete=True,
+        readiness_reasons=(),
+        service="webui",
+        mode="test",
+        version=mock_version,
+        started_at="2026-07-25T00:00:00Z",
+        uptime_seconds=100.0,
+        timestamp="2026-07-25T00:00:00Z",
+        deep_probes_requested=True,
+        dependencies=(),
+        mcp_namespaces=(),
+        stale_runtime=mock_stale,
+        probe_errors=("probe blew up",),
+    )
+
+    snapshot = load_notifications_snapshot(
+        proj_id,
+        load_queue=lambda _id: mock_queue,
+        load_leases=lambda **_kwargs: mock_leases,
+        load_health=lambda **_kwargs: mock_health,
+    )
+    assert snapshot.fetch_error is None
+    # probe errors still appear as items
+    assert any("probe blew up" in item.summary for item in snapshot.items)

webui/app.py

@@ -80,6 +80,11 @@ from webui.system_health import (
     snapshot_to_dict as system_health_to_dict,
 )
 from webui.system_health_views import render_system_health_page
+from webui.notifications import (
+    load_notifications_snapshot,
+    snapshot_to_dict as notifications_snapshot_to_dict,
+)
+from webui.notification_views import render_notifications_page
 from webui import request_service
 from webui.request_views import render_requests_page
 
@@ -889,6 +894,22 @@ async def api_v1_analytics_ingest(request: Request) -> JSONResponse:
     )
 
 
+async def notifications_route(request: Request) -> HTMLResponse:
+    project_id = request.query_params.get("project_id")
+    attention_class = request.query_params.get("attention_class") or "inbox"
+    snap = load_notifications_snapshot(project_id)
+    html = render_notifications_page(
+        snap, filter_class=attention_class, filter_project=project_id
+    )
+    return HTMLResponse(html)
+
+
+async def api_notifications(request: Request) -> JSONResponse:
+    project_id = request.query_params.get("project_id")
+    snap = load_notifications_snapshot(project_id)
+    data = notifications_snapshot_to_dict(snap)
+    return JSONResponse(data)
+
 def _default_request_scope() -> dict[str, str]:
     """Resolve remote/org/repo from the project registry for request forms.
 
@@ -1020,6 +1041,9 @@ def create_app(*, bind_host: str | None = None) -> Starlette:
             Route("/api/queue", api_queue, methods=["GET"]),
             Route("/traffic", traffic, methods=["GET"]),
             Route("/api/traffic", api_traffic, methods=["GET"]),
+            Route("/notifications", notifications_route, methods=["GET"]),
+            Route("/api/notifications", api_notifications, methods=["GET"]),
+            Route("/api/v1/notifications", api_notifications, methods=["GET"]),
             Route("/projects", projects, methods=["GET"]),
             Route("/projects/{project_id}", project_detail, methods=["GET"]),
             Route("/api/projects", api_projects, methods=["GET"]),

webui/nav.py

@@ -46,6 +46,7 @@ NAV_GROUPS: tuple[NavGroup, ...] = (
         NavItem("/queue", "Queue"),
         NavItem("/leases", "Leases"),
         NavItem("/actions", "Actions"),
+        NavItem("/notifications", "Notifications"),
         NavItem("/requests", "Requests"),
     )),
     NavGroup("Runtime/Sessions", (

webui/notification_views.py

@@ -0,0 +1,158 @@
+"""HTML rendering for Phase 3 Notifications and Human-Attention Console (#648)."""
+
+from __future__ import annotations
+
+from html import escape
+from typing import Sequence
+
+from webui.layout import render_page
+from webui.notifications import (
+    ATTENTION_HUMAN_REQUIRED,
+    ATTENTION_OPERATOR,
+    ATTENTION_ROUTINE,
+    NotificationItem,
+    NotificationSnapshot,
+)
+
+
+def _render_attention_badge(attention_class: str) -> str:
+    cls = "badge"
+    if attention_class == ATTENTION_HUMAN_REQUIRED:
+        cls += " badge-blocked"
+    elif attention_class == ATTENTION_OPERATOR:
+        cls += " badge-claimed"
+    else:
+        cls += " muted"
+    return f'<span class="{cls}">{escape(attention_class)}</span>'
+
+
+def _render_notification_row(item: NotificationItem) -> str:
+    category_label = escape(item.category.upper())
+    id_str = escape(item.id)
+    title_str = escape(item.title)
+    summary_str = escape(item.summary)
+    att_badge = _render_attention_badge(item.attention_class)
+
+    work_item_html = "—"
+    if item.work_number and item.work_kind:
+        kind_label = escape(item.work_kind.upper())
+        num_str = f"#{item.work_number}"
+        link = item.deep_link or "#"
+        work_item_html = f'<a href="{escape(link)}"><code>{kind_label} {num_str}</code></a>'
+
+    requires_human_label = (
+        '<span class="badge badge-blocked" style="font-size:0.75rem;">HUMAN REQUIRED</span>'
+        if item.requires_human
+        else ""
+    )
+
+    return f"""<tr>
+  <td><code>{category_label}</code><br><span class="muted" style="font-size:0.75rem;">{id_str}</span></td>
+  <td>
+    <div><strong>{title_str}</strong> {att_badge} {requires_human_label}</div>
+    <div class="muted" style="font-size:0.85rem; margin-top:0.25rem;">{summary_str}</div>
+  </td>
+  <td>{work_item_html}</td>
+  <td><span class="muted" style="font-size:0.8rem;">{escape(item.created_at[:19])}</span></td>
+</tr>"""
+
+
+def _render_notifications_table(items: Sequence[NotificationItem], empty_message: str) -> str:
+    if not items:
+        return f'<p class="muted" style="padding:1rem 0;">{escape(empty_message)}</p>'
+
+    rows = "".join(_render_notification_row(item) for item in items)
+    return f"""<table class="registry">
+  <thead>
+    <tr>
+      <th style="width: 18%;">Category & ID</th>
+      <th style="width: 52%;">Title & Attention Summary</th>
+      <th style="width: 15%;">Work Item</th>
+      <th style="width: 15%;">Time</th>
+    </tr>
+  </thead>
+  <tbody>
+    {rows}
+  </tbody>
+</table>"""
+
+
+def render_notifications_page(
+    snapshot: NotificationSnapshot,
+    *,
+    filter_class: str = "inbox",
+    filter_project: str | None = None,
+) -> str:
+    """Render the notifications and attention inbox page."""
+    title = "Notifications & Attention Inbox"
+
+    err_html = ""
+    if snapshot.fetch_error:
+        err_html = f'<div class="stub" style="border-color:#e53e3e; background:#fff5f5; color:#c53030; margin-bottom:1rem;"><p><strong>Fetch Warning:</strong> {escape(snapshot.fetch_error)}</p></div>'
+
+    # Determine items to render based on filter_class
+    if filter_class == ATTENTION_HUMAN_REQUIRED:
+        display_items = snapshot.human_required_items
+        active_tab_title = "Human-Required Escalations"
+    elif filter_class == ATTENTION_OPERATOR:
+        display_items = snapshot.operator_items
+        active_tab_title = "Operator Inbox Items"
+    elif filter_class == ATTENTION_ROUTINE:
+        display_items = snapshot.routine_items
+        active_tab_title = "Routine Workflow Transitions"
+    elif filter_class == "all":
+        display_items = snapshot.items
+        active_tab_title = "All Events (including Routine)"
+    else:  # "inbox" default
+        display_items = snapshot.inbox_items
+        active_tab_title = "Attention Inbox (Human + Operator)"
+
+    hr_cls = "badge-blocked" if snapshot.human_required_count > 0 else "muted"
+    op_cls = "badge-claimed" if snapshot.operator_count > 0 else "muted"
+
+    metrics_html = f"""<div style="display:flex; gap:1rem; margin-bottom:1.5rem;">
+  <div class="health-card" style="flex:1;">
+    <span class="muted" style="font-size:0.85rem;">Human Required</span>
+    <h2 style="margin:0.2rem 0;"><span class="badge {hr_cls}" style="font-size:1.4rem;">{snapshot.human_required_count}</span></h2>
+    <p class="muted" style="font-size:0.8rem; margin:0;">Critical escalation boundary</p>
+  </div>
+  <div class="health-card" style="flex:1;">
+    <span class="muted" style="font-size:0.85rem;">Operator Inbox</span>
+    <h2 style="margin:0.2rem 0;"><span class="badge {op_cls}" style="font-size:1.4rem;">{snapshot.operator_count}</span></h2>
+    <p class="muted" style="font-size:0.8rem; margin:0;">Operational items needing review</p>
+  </div>
+  <div class="health-card" style="flex:1;">
+    <span class="muted" style="font-size:0.85rem;">Routine Transitions</span>
+    <h2 style="margin:0.2rem 0;"><span class="badge muted" style="font-size:1.4rem;">{snapshot.routine_count}</span></h2>
+    <p class="muted" style="font-size:0.8rem; margin:0;">Background transitions (filtered)</p>
+  </div>
+</div>"""
+
+    # Filter navigation links
+    def _tab_link(target_class: str, label: str) -> str:
+        is_active = (filter_class == target_class)
+        style = "font-weight:bold; border-bottom:2px solid currentColor;" if is_active else "color:#4a5568;"
+        return f'<a href="/notifications?attention_class={target_class}" style="margin-right:1.25rem; text-decoration:none; padding-bottom:0.25rem; {style}">{label}</a>'
+
+    tabs_html = f"""<div style="margin-bottom:1.25rem; border-bottom:1px solid #e2e8f0; padding-bottom:0.5rem;">
+  {_tab_link("inbox", f"Attention Inbox ({snapshot.human_required_count + snapshot.operator_count})")}
+  {_tab_link("human-required", f"Human Required ({snapshot.human_required_count})")}
+  {_tab_link("operator", f"Operator ({snapshot.operator_count})")}
+  {_tab_link("routine", f"Routine ({snapshot.routine_count})")}
+  {_tab_link("all", f"All Events ({snapshot.total_count})")}
+</div>"""
+
+    table_html = _render_notifications_table(
+        display_items,
+        f"No items match attention filter '{filter_class}'.",
+    )
+
+    body = f"""<h2>{escape(title)}</h2>
+<p class="muted">Phase 3 console surface for human-attention routing (#648). Routine workflow transitions are filtered by default to eliminate notification fatigue.</p>
+{err_html}
+{metrics_html}
+{tabs_html}
+<h3>{escape(active_tab_title)}</h3>
+{table_html}"""
+
+    return render_page(title=title, body_html=body)

webui/notifications.py

@@ -0,0 +1,486 @@
+"""Notifications and human-attention routing module for Phase 3 web console (#648).
+
+Defines attention classes, event classification rules, and inbox aggregation so
+operators receive direct alerts only for human-required escalation boundaries
+(#628) while routine workflow transitions remain available for pull-based review.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass, field
+from datetime import datetime, timezone
+from typing import Any, Callable
+
+from webui import console_redaction
+from webui.project_registry import load_registry
+from webui.queue_loader import QueueSnapshot, load_queue_snapshot
+from webui.lease_loader import LeaseSnapshot, load_lease_snapshot
+from webui.system_health import SystemHealthSnapshot, load_system_health
+
+# Attention class definitions (#628, #648)
+ATTENTION_ROUTINE = "routine"
+ATTENTION_OPERATOR = "operator"
+ATTENTION_HUMAN_REQUIRED = "human-required"
+
+ATTENTION_CLASSES = (
+    ATTENTION_ROUTINE,
+    ATTENTION_OPERATOR,
+    ATTENTION_HUMAN_REQUIRED,
+)
+
+# Notification categories
+CATEGORY_AUTH = "auth"
+CATEGORY_BLOCKER = "blocker"
+CATEGORY_LEASE = "lease"
+CATEGORY_VALIDATION = "validation"
+CATEGORY_WORKFLOW = "workflow"
+CATEGORY_SYSTEM = "system"
+
+CATEGORIES = (
+    CATEGORY_AUTH,
+    CATEGORY_BLOCKER,
+    CATEGORY_LEASE,
+    CATEGORY_VALIDATION,
+    CATEGORY_WORKFLOW,
+    CATEGORY_SYSTEM,
+)
+
+
+@dataclass(frozen=True)
+class NotificationItem:
+    """A single notification or inbox event."""
+
+    id: str
+    attention_class: str  # "routine", "operator", "human-required"
+    category: str  # "auth", "blocker", "lease", "validation", etc.
+    title: str
+    summary: str
+    work_kind: str | None  # "issue", "pr", "session", "system"
+    work_number: int | None
+    project_id: str
+    repo_label: str
+    created_at: str
+    deep_link: str | None = None
+    requires_human: bool = False
+    extra: dict[str, Any] = field(default_factory=dict)
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "id": self.id,
+            "attention_class": self.attention_class,
+            "category": self.category,
+            "title": self.title,
+            "summary": console_redaction.redact_text(self.summary),
+            "work_kind": self.work_kind,
+            "work_number": self.work_number,
+            "project_id": self.project_id,
+            "repo_label": self.repo_label,
+            "created_at": self.created_at,
+            "deep_link": self.deep_link,
+            "requires_human": self.requires_human,
+            "extra": self.extra,
+        }
+
+
+@dataclass(frozen=True)
+class NotificationSnapshot:
+    """Snapshot of notifications and attention inbox state."""
+
+    project_id: str
+    repo_label: str
+    items: tuple[NotificationItem, ...]
+    human_required_count: int
+    operator_count: int
+    routine_count: int
+    total_count: int
+    fetch_error: str | None = None
+
+    @property
+    def inbox_items(self) -> tuple[NotificationItem, ...]:
+        """Items requiring operator or human attention (excluding routine)."""
+        return tuple(
+            item
+            for item in self.items
+            if item.attention_class in {ATTENTION_OPERATOR, ATTENTION_HUMAN_REQUIRED}
+        )
+
+    @property
+    def human_required_items(self) -> tuple[NotificationItem, ...]:
+        return tuple(
+            item for item in self.items if item.attention_class == ATTENTION_HUMAN_REQUIRED
+        )
+
+    @property
+    def operator_items(self) -> tuple[NotificationItem, ...]:
+        return tuple(
+            item for item in self.items if item.attention_class == ATTENTION_OPERATOR
+        )
+
+    @property
+    def routine_items(self) -> tuple[NotificationItem, ...]:
+        return tuple(
+            item for item in self.items if item.attention_class == ATTENTION_ROUTINE
+        )
+
+    def as_dict(self) -> dict[str, Any]:
+        return {
+            "project_id": self.project_id,
+            "repo_label": self.repo_label,
+            "human_required_count": self.human_required_count,
+            "operator_count": self.operator_count,
+            "routine_count": self.routine_count,
+            "total_count": self.total_count,
+            "fetch_error": self.fetch_error,
+            "inbox_items": [item.as_dict() for item in self.inbox_items],
+            "all_items": [item.as_dict() for item in self.items],
+        }
+
+
+def classify_attention_event(
+    category: str,
+    title: str,
+    summary: str,
+    *,
+    is_hard_stop: bool = False,
+    is_auth_failure: bool = False,
+    is_irrecoverable: bool = False,
+    is_decision_lock: bool = False,
+    is_validation_failure: bool = False,
+    is_stale: bool = False,
+    is_blocker: bool = False,
+) -> tuple[str, bool]:
+    """Classify an event into an attention class and human requirement flag.
+
+    Rules (#628, #648):
+    1. Critical boundaries (hard stop, auth failure, irrecoverable state,
+       decision lock, validation failure) -> ATTENTION_HUMAN_REQUIRED (requires_human=True).
+    2. Operational queues (blocker, stale lease, unassigned ready work, queue collision)
+       -> ATTENTION_OPERATOR (requires_human=False).
+    3. Routine state transitions (clean progression, healthy heartbeats) -> ATTENTION_ROUTINE (requires_human=False).
+
+    Classification uses structured flags and category only. Human-authored
+    ``title`` / ``summary`` text is never substring-matched for escalation
+    (PR #905 review B1) — callers that need text signals must set flags from
+    machine-generated status/detail fields before calling this function.
+    """
+    del title, summary  # kept for API stability; never used for classification
+    if (
+        is_hard_stop
+        or is_auth_failure
+        or is_irrecoverable
+        or is_decision_lock
+        or is_validation_failure
+        or category in {CATEGORY_AUTH, CATEGORY_VALIDATION}
+    ):
+        return ATTENTION_HUMAN_REQUIRED, True
+
+    if is_stale or is_blocker or category in {CATEGORY_BLOCKER, CATEGORY_LEASE}:
+        return ATTENTION_OPERATOR, False
+
+    return ATTENTION_ROUTINE, False
+
+
+def load_notifications_snapshot(
+    project_id: str | None = None,
+    *,
+    load_queue: Callable[..., QueueSnapshot] | None = None,
+    load_leases: Callable[..., LeaseSnapshot] | None = None,
+    load_health: Callable[..., SystemHealthSnapshot] | None = None,
+) -> NotificationSnapshot:
+    """Load and classify attention notifications across queue, leases, and system health."""
+    registry = load_registry()
+    project = None
+    if project_id:
+        for entry in registry.projects:
+            if entry.id == project_id:
+                project = entry
+                break
+    else:
+        project = registry.projects[0] if registry.projects else None
+
+    if project is None:
+        return NotificationSnapshot(
+            project_id=project_id or "",
+            repo_label="",
+            items=(),
+            human_required_count=0,
+            operator_count=0,
+            routine_count=0,
+            total_count=0,
+            fetch_error="project not found in registry",
+        )
+
+    queue_loader_fn = load_queue or load_queue_snapshot
+    lease_loader_fn = load_leases or load_lease_snapshot
+    health_loader_fn = load_health or load_system_health
+
+    try:
+        queue_snap = queue_loader_fn(project.id)
+    except TypeError:
+        queue_snap = queue_loader_fn(project_id=project.id)
+
+    try:
+        lease_snap = lease_loader_fn(project_id=project.id)
+    except TypeError:
+        lease_snap = lease_loader_fn(project.id)
+
+    try:
+        health_snap = health_loader_fn(project_id=project.id)
+    except TypeError:
+        try:
+            health_snap = health_loader_fn(project.id)
+        except TypeError:
+            health_snap = health_loader_fn()
+
+    items: list[NotificationItem] = []
+    now_iso = datetime.now(timezone.utc).isoformat()
+
+    # 1. System health alerts (highest priority)
+    for err_idx, probe_err in enumerate(getattr(health_snap, "probe_errors", ())):
+        att_cls, req_human = classify_attention_event(
+            CATEGORY_SYSTEM,
+            "System Health Probe Error",
+            probe_err,
+            is_blocker=True,
+        )
+        items.append(
+            NotificationItem(
+                id=f"notif-sys-err-{project.id}-{err_idx}",
+                attention_class=att_cls,
+                category=CATEGORY_SYSTEM,
+                title="System Health Error",
+                summary=f"System health error: {probe_err}",
+                work_kind="system",
+                work_number=None,
+                project_id=project.id,
+                repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                created_at=now_iso,
+                deep_link="/system",
+                requires_human=req_human,
+            )
+        )
+
+    for probe in getattr(health_snap, "dependencies", ()):
+        if probe.status not in ("ok", "healthy"):
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_SYSTEM,
+                f"Probe Failure: {probe.name}",
+                probe.detail or probe.status,
+                is_hard_stop=("stop" in probe.status or "fatal" in probe.status),
+                is_auth_failure=("auth" in probe.name.lower() or "unauthorized" in probe.status.lower()),
+                is_blocker=True,
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-probe-{probe.name}",
+                    attention_class=att_cls,
+                    category=CATEGORY_AUTH if "auth" in probe.name.lower() else CATEGORY_SYSTEM,
+                    title=f"Health Probe Alert: {probe.name}",
+                    summary=f"Probe '{probe.name}' reported status '{probe.status}': {probe.detail}",
+                    work_kind="system",
+                    work_number=None,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link="/system",
+                    requires_human=req_human,
+                )
+            )
+
+    # 2. Queue items (PRs and Issues)
+    for pr in queue_snap.prs:
+        if "blocked" in pr.badges:
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_BLOCKER,
+                f"PR #{pr.number} Blocked",
+                f"PR #{pr.number} '{pr.title}' is blocked or has merge conflicts.",
+                is_blocker=True,
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-pr-block-{pr.number}",
+                    attention_class=att_cls,
+                    category=CATEGORY_BLOCKER,
+                    title=f"Blocked PR #{pr.number}",
+                    summary=f"PR #{pr.number} ({pr.title}) requires merge conflict resolution.",
+                    work_kind="pr",
+                    work_number=pr.number,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link=f"/traffic",
+                    requires_human=req_human,
+                )
+            )
+        elif "stale" in pr.badges:
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_WORKFLOW,
+                f"PR #{pr.number} Stale",
+                f"PR #{pr.number} '{pr.title}' has had no activity for over 14 days.",
+                is_stale=True,
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-pr-stale-{pr.number}",
+                    attention_class=att_cls,
+                    category=CATEGORY_WORKFLOW,
+                    title=f"Stale PR #{pr.number}",
+                    summary=f"PR #{pr.number} ({pr.title}) is stale.",
+                    work_kind="pr",
+                    work_number=pr.number,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link=f"/queue",
+                    requires_human=req_human,
+                )
+            )
+        else:
+            # Routine PR transition
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_WORKFLOW,
+                f"PR #{pr.number} Active",
+                f"PR #{pr.number} '{pr.title}' is in routine state {', '.join(pr.badges)}.",
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-pr-routine-{pr.number}",
+                    attention_class=att_cls,
+                    category=CATEGORY_WORKFLOW,
+                    title=f"Routine PR #{pr.number}",
+                    summary=f"PR #{pr.number} ({pr.title}) state: {', '.join(pr.badges)}.",
+                    work_kind="pr",
+                    work_number=pr.number,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link=f"/queue",
+                    requires_human=req_human,
+                )
+            )
+
+    for issue in queue_snap.issues:
+        if "duplicate" in issue.badges:
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_BLOCKER,
+                f"Issue #{issue.number} Duplicate PRs",
+                f"Issue #{issue.number} has multiple linked PRs.",
+                is_blocker=True,
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-issue-dup-{issue.number}",
+                    attention_class=att_cls,
+                    category=CATEGORY_BLOCKER,
+                    title=f"Duplicate PRs on Issue #{issue.number}",
+                    summary=f"Issue #{issue.number} ({issue.title}) linked to multiple PRs.",
+                    work_kind="issue",
+                    work_number=issue.number,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link=f"/traffic",
+                    requires_human=req_human,
+                )
+            )
+        elif "claimed" in issue.badges or "in-review" in issue.badges:
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_WORKFLOW,
+                f"Issue #{issue.number} Active",
+                f"Issue #{issue.number} '{issue.title}' in state {', '.join(issue.badges)}.",
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-issue-routine-{issue.number}",
+                    attention_class=att_cls,
+                    category=CATEGORY_WORKFLOW,
+                    title=f"Routine Issue #{issue.number}",
+                    summary=f"Issue #{issue.number} ({issue.title}) state: {', '.join(issue.badges)}.",
+                    work_kind="issue",
+                    work_number=issue.number,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link=f"/queue",
+                    requires_human=req_human,
+                )
+            )
+
+    # 3. Leases / Collisions
+    for lease in lease_snap.reviewer_leases:
+        if lease.get("is_expired") or lease.get("status") == "expired":
+            pr_num = lease.get("pr_number") or lease.get("work_item_number")
+            att_cls, req_human = classify_attention_event(
+                CATEGORY_LEASE,
+                f"Reviewer Lease Expired for PR #{pr_num}",
+                f"Reviewer lease for PR #{pr_num} has expired.",
+                is_stale=True,
+            )
+            items.append(
+                NotificationItem(
+                    id=f"notif-lease-exp-pr-{pr_num}",
+                    attention_class=att_cls,
+                    category=CATEGORY_LEASE,
+                    title=f"Expired Reviewer Lease (PR #{pr_num})",
+                    summary=f"Reviewer lease for PR #{pr_num} expired.",
+                    work_kind="pr",
+                    work_number=pr_num,
+                    project_id=project.id,
+                    repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                    created_at=now_iso,
+                    deep_link="/leases",
+                    requires_human=req_human,
+                )
+            )
+
+    for col_idx, collision in enumerate(lease_snap.duplicate_prs):
+        att_cls, req_human = classify_attention_event(
+            CATEGORY_BLOCKER,
+            f"Duplicate PR Collision ({collision.kind})",
+            collision.message,
+            is_blocker=True,
+        )
+        issue_part = collision.issue_number if collision.issue_number is not None else "none"
+        kind_part = (collision.kind or "unknown").replace(" ", "-")
+        items.append(
+            NotificationItem(
+                id=f"notif-collision-{kind_part}-{issue_part}-{col_idx}",
+                attention_class=att_cls,
+                category=CATEGORY_BLOCKER,
+                title=f"Collision Alert ({collision.kind})",
+                summary=collision.message,
+                work_kind="issue" if collision.issue_number else "pr",
+                work_number=collision.issue_number,
+                project_id=project.id,
+                repo_label=f"{project.gitea_owner}/{project.repo_name}",
+                created_at=now_iso,
+                deep_link="/leases",
+                requires_human=req_human,
+            )
+        )
+
+    human_req_count = sum(1 for i in items if i.attention_class == ATTENTION_HUMAN_REQUIRED)
+    operator_count = sum(1 for i in items if i.attention_class == ATTENTION_OPERATOR)
+    routine_count = sum(1 for i in items if i.attention_class == ATTENTION_ROUTINE)
+
+    # Fetch errors are transport/load failures only — not probe results that
+    # already surface as first-class notification items (PR #905 review B3).
+    fetch_err = queue_snap.fetch_error or lease_snap.fetch_error
+    if isinstance(fetch_err, (tuple, list)):
+        fetch_err = "; ".join(fetch_err) if fetch_err else None
+
+    return NotificationSnapshot(
+        project_id=project.id,
+        repo_label=f"{project.gitea_owner}/{project.repo_name}",
+        items=tuple(items),
+        human_required_count=human_req_count,
+        operator_count=operator_count,
+        routine_count=routine_count,
+        total_count=len(items),
+        fetch_error=fetch_err,
+    )
+
+
+def snapshot_to_dict(snapshot: NotificationSnapshot) -> dict[str, Any]:
+    """JSON-serializable export for /api/v1/notifications."""
+    return snapshot.as_dict()