stage-batch10: #3327 model-id normalize + #3334 RFC slice doc + #3341 profile skill counts

#3327 fix(reasoning): normalize custom-provider model ids for fallback heuristics Co-authored-by: Carry00 <Carry00@users.noreply.github.com> #3334 docs(rfc): mark run-adapter Slice 4f shipped, define Slice 4g gate Co-authored-by: Michaelyklam <Michaelyklam@users.noreply.github.com> #3341 fix(profiles): show enabled vs compatible skill counts Co-authored-by: b3nw <b3nw@users.noreply.github.com>
2026-06-01 18:23:49 +00:00
parent c5aed7654c
commit de22c607bc
7 changed files with 423 additions and 49 deletions
--- a/api/config.py
+++ b/api/config.py
@@ -2094,6 +2094,61 @@ def _strip_provider_hint_for_reasoning(model_id: str) -> str:
    return model


+def _reasoning_name_candidates(model_id: str) -> list[str]:
+    """Return normalized model-name candidates for heuristic capability checks."""
+    bare = str(model_id or "").strip().lower().rsplit("/", 1)[-1]
+    if not bare:
+        return []
+
+    candidates: list[str] = []
+
+    def _add(value: str) -> None:
+        candidate = str(value or "").strip().lower()
+        if candidate and candidate not in candidates:
+            candidates.append(candidate)
+
+    _add(bare)
+
+    dot_parts = [part for part in bare.split(".") if part]
+    if len(dot_parts) > 1:
+        # Try progressively stripping dot-separated vendor namespaces so inputs like
+        # "moonshotai.kimi-k2.5" and "vendor.deepseek.v3.2" both surface the real
+        # model family rather than treating every dot as part of the provider slug.
+        for index in range(1, len(dot_parts)):
+            suffix = ".".join(dot_parts[index:])
+            if any(ch.isalpha() for ch in suffix):
+                _add(suffix)
+
+    for candidate in list(candidates):
+        normalized = re.sub(r"[^a-z0-9]+", "-", candidate).strip("-")
+        _add(normalized)
+
+    return candidates
+
+
+def _candidate_supports_reasoning(candidate: str) -> bool:
+    normalized = re.sub(r"[^a-z0-9]+", "-", str(candidate or "").strip().lower()).strip("-")
+    if not normalized:
+        return False
+
+    tokens = [token for token in normalized.split("-") if token]
+    token_set = set(tokens)
+
+    if "thinking" in token_set or "reasoning" in token_set:
+        return True
+    if normalized in {"o1", "o3", "o4"} or normalized.startswith(("o1-", "o3-", "o4-")):
+        return True
+    if normalized.startswith(("kimi-k2", "kimi-thinking", "claude-3", "claude-4")):
+        return True
+    if normalized.startswith("qwen3") or "qwen3" in token_set:
+        return True
+    if normalized.startswith(("deepseek-v3", "deepseek-v4", "deepseek-r1", "deepseek-r2")):
+        return True
+    if len(tokens) >= 2 and tokens[0] == "deepseek" and tokens[1] in {"v3", "v4", "r1", "r2"}:
+        return True
+    return False
+
+
 def _heuristic_reasoning_efforts(model_id: str, provider_id: str) -> list[str]:
    """Fallback when hermes_cli is unavailable."""
    model = _strip_provider_hint_for_reasoning(model_id).lower()
@@ -2123,31 +2178,11 @@ def _heuristic_reasoning_efforts(model_id: str, provider_id: str) -> list[str]:
    )
    if any(model.startswith(prefix) for prefix in prefixes):
        return list(VALID_REASONING_EFFORTS)
-    # Custom API aggregators (e.g. New API, One API) use non-standard model naming:
-    # bare names like "deepseek-v4-flash" or dot-separated "moonshotai.kimi-k2.5"
-    # rather than the OpenRouter-style "vendor/model" that the prefix list targets.
-    # Strip a dot-vendor prefix (e.g. "moonshotai.kimi-k2.5" → "kimi-k2.5") and
-    # check both the original bare name and the stripped suffix.
-    bare_after_dot = bare.split(".", 1)[-1] if "." in bare else bare
-    thinking_bare_prefixes = (
-        "deepseek-v4",
-        "deepseek-r1",
-        "deepseek-r2",
-        "kimi-k2",
-        "kimi-thinking",
-        "qwen3",
-        "claude-3",
-        "claude-4",
-        "o1-",
-        "o3-",
-        "o4-",
-    )
-    if any(
-        bare.startswith(p) or bare_after_dot.startswith(p)
-        for p in thinking_bare_prefixes
-    ):
-        return list(VALID_REASONING_EFFORTS)
-    if "thinking" in bare or "reasoning" in bare:
+    # Named custom providers often rewrite model ids with dots, underscores, or
+    # extra vendor namespaces. Normalize those shapes before applying family-level
+    # reasoning heuristics so "deepseek.v3.2", "deepseek_v4_flash", and
+    # "vendor.deepseek.v3.2" are treated consistently.
+    if any(_candidate_supports_reasoning(candidate) for candidate in _reasoning_name_candidates(bare)):
        return list(VALID_REASONING_EFFORTS)
    return []

--- a/api/profiles.py
+++ b/api/profiles.py
@@ -891,6 +891,7 @@ def switch_profile(name: str, *, process_wide: bool = True) -> dict:
            raise ValueError(f"Profile '{name}' does not exist.")

    with _profile_lock:
+        _SKILLS_STATS_CACHE.clear()
        if process_wide:
            global _active_profile
            _active_profile = name
@@ -989,6 +990,77 @@ def switch_profile(name: str, *, process_wide: bool = True) -> dict:
    }


+_SKILLS_STATS_CACHE: dict[Path, tuple[int, int, float]] = {}
+_SKILLS_STATS_CACHE_TTL = 8.0  # seconds
+
+
+def _get_profile_skills_stats(profile_dir: Path) -> tuple[int, int]:
+    """Calculate (enabled_count, compatible_count) for a profile directory."""
+    import time
+    profile_dir = Path(profile_dir).resolve()
+    now = time.time()
+    if profile_dir in _SKILLS_STATS_CACHE:
+        enabled, compat, expiry = _SKILLS_STATS_CACHE[profile_dir]
+        if now < expiry:
+            return enabled, compat
+
+    skills_dir = profile_dir / "skills"
+    if not skills_dir.is_dir():
+        res = (0, 0)
+        _SKILLS_STATS_CACHE[profile_dir] = (res[0], res[1], now + _SKILLS_STATS_CACHE_TTL)
+        return res
+
+    disabled = set()
+    config_path = profile_dir / "config.yaml"
+    if config_path.exists():
+        try:
+            import yaml as _yaml
+            cfg = _yaml.safe_load(config_path.read_text(encoding="utf-8"))
+            if isinstance(cfg, dict):
+                skills_cfg = cfg.get("skills")
+                if isinstance(skills_cfg, dict):
+                    # Align with get_disabled_skill_names(platform="webui") behavior:
+                    platform_disabled = (skills_cfg.get("platform_disabled") or {}).get("webui")
+                    if platform_disabled is not None:
+                        disabled_val = platform_disabled
+                    else:
+                        disabled_val = skills_cfg.get("disabled")
+                    
+                    if disabled_val is not None:
+                        if isinstance(disabled_val, str):
+                            disabled_val = [disabled_val]
+                        disabled = {str(v).strip() for v in disabled_val if str(v).strip()}
+        except Exception:
+            pass
+
+    from agent.skill_utils import iter_skill_index_files, parse_frontmatter, skill_matches_platform
+    
+    seen_names = set()
+    enabled_count = 0
+    compatible_count = 0
+    
+    for skill_md in iter_skill_index_files(skills_dir, "SKILL.md"):
+        try:
+            content = skill_md.read_text(encoding="utf-8")[:4000]
+            frontmatter, _ = parse_frontmatter(content)
+            if not skill_matches_platform(frontmatter):
+                continue
+            name = frontmatter.get("name", skill_md.parent.name)[:64]
+            if name in seen_names:
+                continue
+            seen_names.add(name)
+            
+            compatible_count += 1
+            if name not in disabled:
+                enabled_count += 1
+        except Exception:
+            pass
+            
+    res = (enabled_count, compatible_count)
+    _SKILLS_STATS_CACHE[profile_dir] = (res[0], res[1], now + _SKILLS_STATS_CACHE_TTL)
+    return res
+
+
 def list_profiles_api() -> list:
    """List all profiles with metadata, serialized for JSON response."""
    try:
@@ -1001,6 +1073,7 @@ def list_profiles_api() -> list:
    active = get_active_profile_name()
    result = []
    for p in infos:
+        enabled_count, total_count = _get_profile_skills_stats(p.path)
        result.append({
            'name': p.name,
            'path': str(p.path),
@@ -1010,13 +1083,16 @@ def list_profiles_api() -> list:
            'model': p.model,
            'provider': p.provider,
            'has_env': p.has_env,
-            'skill_count': p.skill_count,
+            'skill_count': enabled_count,
+            'enabled_skills': enabled_count,
+            'total_skills': total_count,
        })
    return result


 def _default_profile_dict() -> dict:
    """Fallback profile dict when hermes_cli is not importable."""
+    enabled_count, compatible_count = _get_profile_skills_stats(_DEFAULT_HERMES_HOME)
    return {
        'name': 'default',
        'path': str(_DEFAULT_HERMES_HOME),
@@ -1026,7 +1102,9 @@ def _default_profile_dict() -> dict:
        'model': None,
        'provider': None,
        'has_env': (_DEFAULT_HERMES_HOME / '.env').exists(),
-        'skill_count': 0,
+        'skill_count': enabled_count,
+        'enabled_skills': enabled_count,
+        'total_skills': compatible_count,
    }


@@ -1437,6 +1515,7 @@ def create_profile_api(name: str, clone_from: str = None,

    # Invalidate cached root-profile-name lookup; create_profile may have added
    # a new profile that flips is_default semantics on the agent side (#1612).
+    _SKILLS_STATS_CACHE.clear()
    _invalidate_root_profile_cache()

    # Find and return the newly created profile info.
@@ -1456,6 +1535,8 @@ def create_profile_api(name: str, clone_from: str = None,
        'provider': None,
        'has_env': (profile_path / '.env').exists(),
        'skill_count': 0,
+        'enabled_skills': 0,
+        'total_skills': 0,
    }


@@ -1488,5 +1569,6 @@ def delete_profile_api(name: str) -> dict:
            raise ValueError(f"Profile '{name}' does not exist.")

    # Drop cached root-profile-name lookup — list_profiles_api() shape changed.
+    _SKILLS_STATS_CACHE.clear()
    _invalidate_root_profile_cache()
    return {'ok': True, 'name': name}
--- a/docs/rfcs/hermes-run-adapter-contract.md
+++ b/docs/rfcs/hermes-run-adapter-contract.md
@@ -956,28 +956,34 @@ Non-goals for Slice 4e:

 #### Slice 4f: Supervised local runner client backend gate

-Status as of 2026-05-28: client transport proposed in #3073 behind
-`HERMES_WEBUI_RUNNER_BASE_URL`; it should be described as under review until a
-release PR actually ships it.
-`runner-local` still remains default-off and returns the bounded not-configured
-path unless that endpoint is explicitly configured. When configured, WebUI uses a
-JSON HTTP client boundary for start / observe / status / controls and bridges
-observed runner events through the existing SSE stream route rather than adding
-main-process runner-owned maps.
+Status as of 2026-05-31: shipped in v0.51.188 via #3073 / #3274. The client
+transport is now implemented behind `HERMES_WEBUI_RUNNER_BASE_URL` and remains
+default-off. With no endpoint configured, `runner-local` still returns the
+bounded not-configured path and the live in-process `_run_agent_streaming` path
+is unchanged. When configured, WebUI uses a JSON HTTP client boundary for start /
+observe / status / controls and bridges observed runner events through the
+existing SSE stream route rather than adding main-process runner-owned maps.
+
+The release added two security hardening checks while absorbing #3073:
+`HttpRunnerClient` rejects non-`http(s)` base URL schemes and uses an opener that
+does not follow redirects, so a misconfigured or compromised runner cannot leak a
+Bearer token to a redirected host. The release gate reported full pytest passing
+and independent default-off/inert-path review.
+
 This bridge is intentionally a WebUI consumer transport seam: the configured
 runner must emit events that are already compatible with the browser SSE event
 names/payloads, or a later runner-owned normalization layer must translate
 Hermes runtime families such as `token.delta`, `tool.started`, and `done` before
 they reach this route.

-After the route-selection harness ships, the next reviewable step is not to make
-`runner-local` the default. It is to define the first concrete supervised/local
-runner client backend that can replace the bounded 501 path under the existing
-feature flag and prove execution ownership has moved out of the main WebUI
-request process.
+After the configured runner-client boundary ships, the next reviewable step is
+not to make `runner-local` the default. It is to define the first supervised
+runner process harness that can actually own `AIAgent` execution behind that
+client boundary and prove restart/reattach with a real local runner, not just a
+configured external endpoint or fake-runner fixture.

-This slice is a contract gate before backend code lands. The goal is to pin the
-minimum runner client behavior so the implementation cannot become a renamed
+This slice was the client-boundary implementation gate. The goal was to pin the
+minimum runner client behavior so the implementation could not become a renamed
 `STREAMS` / `CANCEL_FLAGS` / cached `AIAgent` surrogate inside `api/routes.py`.

 Scope:
@@ -1027,6 +1033,59 @@ Non-goals for Slice 4f:
 - no permanent WebUI-owned active-run discovery cache that duplicates runner or
  future Hermes Runtime API responsibility.

+#### Slice 4g: Supervised local runner process harness gate
+
+After #3073 / #3274, WebUI has an explicit configured-runner HTTP client and SSE
+consumer bridge, but it still does not ship the supervised runner process itself.
+The next gate should define the smallest local runner harness that can own
+`AIAgent` execution outside the main WebUI request process while being consumed
+through the already-shipped `runner-local` client boundary.
+
+Scope:
+
+- define the local runner process lifecycle: spawn/start, health check, run
+  ownership, graceful shutdown, crash classification, and cleanup;
+- keep WebUI as a client of `HERMES_WEBUI_RUNNER_BASE_URL`, not the owner of
+  process-local runner execution state;
+- persist run/session lookup, ordered events, terminal state, and active controls
+  in runner-owned or journal-backed state that a restarted WebUI can discover;
+- carry explicit profile, workspace, attachments, provider/model, toolset,
+  source, and metadata payloads into the runner without WebUI process-global
+  environment mutation;
+- prove cancel as the first live control for active runner-owned runs, with
+  approval, clarify, goal, and queue either mapped to explicit runner
+  capabilities or returned as bounded unsupported/conflict `ControlResult`
+  values.
+
+Acceptance tests for Slice 4g:
+
+1. **Process ownership moved.** A local runner process, not `hermes-webui`, owns
+   `AIAgent` construction/reuse and active run execution for `runner-local` runs.
+2. **Restart/reattach with a real runner.** Start a non-trivial `runner-local`
+   run, restart only `hermes-webui`, reload the session, rediscover the active or
+   terminal runner-owned run, replay/catch up from cursor without duplicate
+   transcript/tool/reasoning state, and preserve cancel if still active.
+3. **No runtime-surrogate globals in WebUI.** The main WebUI server still does not
+   gain new module-level maps for runner-owned streams, cancel flags,
+   approval/clarify callbacks, cached agents, child process run registries, goal
+   state, or queue schedulers.
+4. **Default-off and reversible.** Unset `HERMES_WEBUI_RUNNER_BASE_URL` or switch
+   the adapter mode back to legacy and the existing in-process path remains
+   available without session or journal migration.
+5. **Runner health and failure are observable.** A missing, unhealthy, or crashed
+   runner returns bounded diagnostics and terminal/interrupted state rather than
+   silently falling back to WebUI-owned execution for a runner-selected run.
+
+Non-goals for Slice 4g:
+
+- no default-on runner mode;
+- no removal of `legacy-direct` or `legacy-journal`;
+- no server-side queue scheduler just for adapter symmetry;
+- no broad WebUI product-surface migration;
+- no claim that this is the canonical Hermes Agent Runtime API; if Hermes Agent
+  later ships `/v1/runs`, this local runner remains a replaceable backend behind
+  the same adapter/client boundary.
+
 ## First Meaningful Success Criteria

 The first meaningful milestones are deliberately split.
--- a/static/panels.js
+++ b/static/panels.js
@@ -5035,7 +5035,7 @@ async function loadProfilesPanel() {
      const meta = [];
      if (p.model) meta.push(p.model.split('/').pop());
      if (p.provider) meta.push(p.provider);
-      if (p.skill_count) meta.push(t('profile_skill_count', p.skill_count));
+      if (p.total_skills && p.total_skills > 0) meta.push(t('profile_skill_count', p.total_skills).replace(String(p.total_skills), `${p.enabled_skills} / ${p.total_skills}`));
      const gwDot = p.gateway_running
        ? `<span class="profile-opt-badge running" title="${esc(t('profile_gateway_running'))}"></span>`
        : `<span class="profile-opt-badge stopped" title="${esc(t('profile_gateway_stopped'))}"></span>`;
@@ -5109,7 +5109,7 @@ function _renderProfileDetail(p, activeName){
  if (p.provider) rows.push(`<div class="detail-row"><div class="detail-row-label">Provider</div><div class="detail-row-value">${esc(p.provider)}</div></div>`);
  if (p.base_url) rows.push(`<div class="detail-row"><div class="detail-row-label">Base URL</div><div class="detail-row-value"><code>${esc(p.base_url)}</code></div></div>`);
  rows.push(`<div class="detail-row"><div class="detail-row-label">API key</div><div class="detail-row-value">${p.has_env ? esc(t('profile_api_keys_configured')) : '<span style="color:var(--muted)">Not configured</span>'}</div></div>`);
-  if (typeof p.skill_count === 'number') rows.push(`<div class="detail-row"><div class="detail-row-label">Skills</div><div class="detail-row-value">${esc(t('profile_skill_count', p.skill_count))}</div></div>`);
+  if (p.total_skills && p.total_skills > 0) rows.push(`<div class="detail-row"><div class="detail-row-label">Skills</div><div class="detail-row-value">${esc(t('profile_skill_count', p.total_skills).replace(String(p.total_skills), `${p.enabled_skills} / ${p.total_skills}`))}</div></div>`);
  if (p.default_workspace) rows.push(`<div class="detail-row"><div class="detail-row-label">Default space</div><div class="detail-row-value"><code>${esc(p.default_workspace)}</code></div></div>`);
  body.innerHTML = `
    <div class="main-view-content">
@@ -5199,7 +5199,7 @@ function renderProfileDropdown(data) {
    opt.className = 'profile-opt' + (p.name === active ? ' active' : '');
    const meta = [];
    if (p.model) meta.push(p.model.split('/').pop());
-    if (p.skill_count) meta.push(t('profile_skill_count', p.skill_count));
+    if (p.total_skills && p.total_skills > 0) meta.push(t('profile_skill_count', p.total_skills).replace(String(p.total_skills), `${p.enabled_skills} / ${p.total_skills}`));
    const gwDot = `<span class="profile-opt-badge ${p.gateway_running ? 'running' : 'stopped'}"></span>`;
    const checkmark = p.name === active ? ' <svg width="12" height="12" viewBox="0 0 24 24" fill="none" stroke="var(--link)" stroke-width="3" style="vertical-align:-1px"><polyline points="20 6 9 17 4 12"/></svg>' : '';
    const defaultBadge = p.is_default ? ` <span style="opacity:.5;font-weight:400">${esc(t('profile_default_label'))}</span>` : '';
--- a/tests/test_custom_provider_bare_model_reasoning.py
+++ b/tests/test_custom_provider_bare_model_reasoning.py
@@ -10,6 +10,8 @@ these combinations, hiding the reasoning effort selector in the UI even though
 the underlying models fully support thinking/reasoning.
 """

+import pytest
+
 import api.config as cfg


@@ -33,6 +35,26 @@ def test_deepseek_r1_bare_name_custom_provider():
    assert set(efforts) >= {"low", "medium", "high"}


+@pytest.mark.parametrize(
+    "model_id",
+    [
+        "deepseek.v3.2",
+        "deepseek_v3_2",
+        "vendor.deepseek.v3.2",
+        "deepseek.v4-flash",
+        "deepseek_v4_flash",
+    ],
+)
+def test_deepseek_separator_variants_custom_provider(model_id):
+    efforts = cfg.resolve_model_reasoning_efforts(
+        model_id,
+        provider_id="custom:newapi",
+    )
+    assert set(efforts) >= {"low", "medium", "high"}, (
+        f"{model_id} via custom provider should expose reasoning efforts"
+    )
+
+
 # ── dot-separated model names (vendor.model) ─────────────────────────────────

 def test_kimi_dot_separated_custom_provider():
@@ -91,6 +113,20 @@ def test_plain_llm_dot_separated_custom_provider_no_reasoning():
    ) == []


+@pytest.mark.parametrize(
+    "model_id",
+    [
+        "thinkinghub.llama-3.1-70b",
+        "reasoninghub.llama-3.1-70b",
+    ],
+)
+def test_vendor_prefix_keyword_does_not_trigger_reasoning(model_id):
+    assert cfg.resolve_model_reasoning_efforts(
+        model_id,
+        provider_id="custom:newapi",
+    ) == []
+
+
 # ── slash-prefixed names must still work (no regression) ─────────────────────

 def test_deepseek_slash_prefix_still_works():
--- a/tests/test_profile_skills_stats.py
+++ b/tests/test_profile_skills_stats.py
@@ -0,0 +1,142 @@
+from pathlib import Path
+import yaml
+from api import profiles
+from tests.conftest import requires_agent_modules
+
+
+def _write_skill(root: Path, name: str, platforms=None):
+    skill_dir = root / "skills" / name
+    skill_dir.mkdir(parents=True, exist_ok=True)
+    platforms_line = f"platforms: {platforms}\n" if platforms else ""
+    (skill_dir / "SKILL.md").write_text(
+        f"---\nname: {name}\ndescription: {name} skill\n{platforms_line}---\n\n# {name}\n",
+        encoding="utf-8",
+    )
+
+def _write_config(home: Path, disabled):
+    home.mkdir(parents=True, exist_ok=True)
+    (home / "config.yaml").write_text(
+        yaml.safe_dump({"skills": {"disabled": list(disabled)}}, sort_keys=False),
+        encoding="utf-8",
+    )
+
+@requires_agent_modules
+def test_get_profile_skills_stats(tmp_path):
+    # Setup skills directory with:
+    # 1 compatible & enabled skill ("alpha")
+    # 1 compatible & disabled skill ("beta")
+    # 1 incompatible skill ("gamma" - macos only on linux test run)
+    profile_home = tmp_path / "auditor"
+    _write_skill(profile_home, "alpha")
+    _write_skill(profile_home, "beta")
+    _write_skill(profile_home, "gamma", platforms=["macos"])
+    _write_config(profile_home, ["beta"])
+
+    # Explicitly clear the stats cache to ensure we compute fresh
+    profiles._SKILLS_STATS_CACHE.clear()
+
+    enabled, compatible = profiles._get_profile_skills_stats(profile_home)
+    assert enabled == 1
+    assert compatible == 2
+
+@requires_agent_modules
+def test_list_profiles_api_contains_formatted_skills(monkeypatch, tmp_path):
+    class FakeProfile:
+        def __init__(self, name, path):
+            self.name = name
+            self.path = Path(path)
+            self.is_default = name == "default"
+            self.gateway_running = False
+            self.model = "gpt-4"
+            self.provider = "openai"
+            self.has_env = False
+            self.skill_count = 3  # Raw on-disk count from hermes_cli
+
+    p_default = tmp_path / "default"
+    p_fintech = tmp_path / "profiles" / "fintech"
+
+    _write_skill(p_default, "a1")
+    _write_skill(p_default, "a2")
+    _write_config(p_default, ["a2"])
+
+    _write_skill(p_fintech, "f1")
+    _write_skill(p_fintech, "f2")
+    _write_skill(p_fintech, "f3")
+    _write_config(p_fintech, ["f2", "f3"])
+
+    fake_infos = [
+        FakeProfile("default", p_default),
+        FakeProfile("fintech", p_fintech)
+    ]
+
+    monkeypatch.setattr(profiles, "get_active_profile_name", lambda: "default")
+    
+    # We patch the upstream list_profiles call to return our FakeProfile list
+    try:
+        import hermes_cli.profiles as cli_p
+        monkeypatch.setattr(cli_p, "list_profiles", lambda: fake_infos)
+    except ImportError:
+        pass
+
+    profiles._SKILLS_STATS_CACHE.clear()
+
+    results = profiles.list_profiles_api()
+    by_name = {p["name"]: p for p in results}
+
+    assert "default" in by_name
+    assert "fintech" in by_name
+
+    # backward-compatible skill_count as an integer:
+    assert by_name["default"]["skill_count"] == 1
+    assert by_name["fintech"]["skill_count"] == 1
+
+    # new enabled_skills and total_skills integer fields:
+    assert by_name["default"]["enabled_skills"] == 1
+    assert by_name["default"]["total_skills"] == 2
+    assert by_name["fintech"]["enabled_skills"] == 1
+    assert by_name["fintech"]["total_skills"] == 3
+
+@requires_agent_modules
+def test_no_skills_dir(tmp_path):
+    """Profile with no skills/ directory should return (0, 0)."""
+    profiles._SKILLS_STATS_CACHE.clear()
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert enabled == 0 and compat == 0
+
+@requires_agent_modules
+def test_corrupt_config(tmp_path):
+    """Corrupt config.yaml should not crash — disabled set stays empty."""
+    profiles._SKILLS_STATS_CACHE.clear()
+    _write_skill(tmp_path, "a")
+    (tmp_path / "config.yaml").write_text("not: [valid: yaml: {{", encoding="utf-8")
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert compat == 1 and enabled == 1  # no disabled parsing, all enabled
+
+@requires_agent_modules
+def test_platform_disabled_webui(tmp_path):
+    """platform_disabled.webui list should be used when present."""
+    profiles._SKILLS_STATS_CACHE.clear()
+    _write_skill(tmp_path, "web-only")
+    cfg = {"skills": {"platform_disabled": {"webui": ["web-only"]}, "disabled": []}}
+    (tmp_path / "config.yaml").write_text(yaml.safe_dump(cfg), encoding="utf-8")
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert compat == 1 and enabled == 0
+
+@requires_agent_modules
+def test_skills_stats_cache(tmp_path):
+    """Verify that caching works and has short TTL behavior."""
+    profiles._SKILLS_STATS_CACHE.clear()
+    
+    _write_skill(tmp_path, "alpha")
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert enabled == 1 and compat == 1
+
+    # Add a skill but since cache is active (TTL 8s), we should still get old values
+    _write_skill(tmp_path, "beta")
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert enabled == 1 and compat == 1
+
+    # Force clear or override the mock TTL, or clear cache manually to see changes
+    profiles._SKILLS_STATS_CACHE.clear()
+    enabled, compat = profiles._get_profile_skills_stats(tmp_path)
+    assert enabled == 2 and compat == 2
--- a/tests/test_runtime_adapter_seam.py
+++ b/tests/test_runtime_adapter_seam.py
@@ -574,12 +574,14 @@ def test_rfc_defines_slice4f_supervised_local_runner_client_gate():
    rfc = (routes.Path(__file__).parent.parent / "docs" / "rfcs" / "hermes-run-adapter-contract.md").read_text(encoding="utf-8")

    assert "#### Slice 4f: Supervised local runner client backend gate" in rfc
-    assert "Status as of 2026-05-28: client transport proposed in #3073 behind" in rfc
-    assert "it should be described as under review until a\nrelease PR actually ships it" in rfc
-    assert "replace the bounded 501 path under the existing\nfeature flag" in rfc
-    assert "durable runner-owned run id plus session-to-run lookup" in rfc
+    assert "Status as of 2026-05-31: shipped in v0.51.188 via #3073 / #3274" in rfc
+    assert "The client\ntransport is now implemented behind `HERMES_WEBUI_RUNNER_BASE_URL`" in rfc
+    assert "`HttpRunnerClient` rejects non-`http(s)` base URL schemes" in rfc
+    assert "uses an opener that\ndoes not follow redirects" in rfc
    assert "the configured\nrunner must emit events that are already compatible with the browser SSE event\nnames/payloads" in rfc
    assert "a later runner-owned normalization layer must translate\nHermes runtime families such as `token.delta`, `tool.started`, and `done`" in rfc
+    assert "After the configured runner-client boundary ships" in rfc
+    assert "configured external endpoint or fake-runner fixture" in rfc
    assert "cancel as the first required live control" in rfc
    assert "501 path replaced only when configured" in rfc
    assert "Restart/reattach proves ownership moved" in rfc
@@ -588,6 +590,24 @@ def test_rfc_defines_slice4f_supervised_local_runner_client_gate():
    assert "Unsupported runner controls return safe\n   `unsupported`, `not-active`, or `conflict` results" in rfc
    assert "no permanent WebUI-owned active-run discovery cache" in rfc

+
+def test_rfc_defines_slice4g_supervised_local_runner_process_gate():
+    routes = importlib.import_module("api.routes")
+    rfc = (routes.Path(__file__).parent.parent / "docs" / "rfcs" / "hermes-run-adapter-contract.md").read_text(encoding="utf-8")
+
+    assert "#### Slice 4g: Supervised local runner process harness gate" in rfc
+    assert "After #3073 / #3274, WebUI has an explicit configured-runner HTTP client" in rfc
+    assert "still does not ship the supervised runner process itself" in rfc
+    assert "own\n`AIAgent` execution outside the main WebUI request process" in rfc
+    assert "keep WebUI as a client of `HERMES_WEBUI_RUNNER_BASE_URL`" in rfc
+    assert "without WebUI process-global\n  environment mutation" in rfc
+    assert "Process ownership moved" in rfc
+    assert "Restart/reattach with a real runner" in rfc
+    assert "No runtime-surrogate globals in WebUI" in rfc
+    assert "Default-off and reversible" in rfc
+    assert "Runner health and failure are observable" in rfc
+    assert "no claim that this is the canonical Hermes Agent Runtime API" in rfc
+
 def test_runtime_runner_client_factory_stays_bounded_until_endpoint_configured(monkeypatch):
    routes = importlib.import_module("api.routes")