review: per-surface rider, loosen for portfolio commentary

Reviewer was rejecting legitimate IT portfolio analyses, citing
descriptive risk language as actionable advice:

  reason: "Allocation guidance throughout: 'concentrazione gestibile',
  'non eliminabile', 'bassa esposizione', 'va monitorato'. Treats
  portfolio construction as actionable."

These phrases describe portfolio state (manageable concentration,
non-eliminable risk, low exposure, warrants monitoring) without
directing the user to take action. They are exactly the kind of
prose a portfolio commentary surface is supposed to produce. The
reviewer's generic "no financial advice" rule is too broad here.

Add a `surface` parameter to review_read() with a per-surface rider
mechanism (_SURFACE_RIDERS). The "portfolio" rider:

- Lists DESCRIPTIVE phrasings that are EXPLICITLY permitted:
  attribute naming ("high concentration", "currency exposure"),
  thesis invalidation conditions, impersonal observations about a
  position's sensitivity.
- Tightens the reject list to EXPLICIT calls to action: imperative
  verbs aimed at the reader, "you should", "consider X-ing",
  specific allocation prescriptions, price-target predictions.

portfolio_analysis.analyse() now passes surface="portfolio". All
other reviewer call sites (indicator summary, log, chat, digest)
default to surface=None and keep the generic rules.

tests/conftest.py's autouse review_read stub picks up **_kw so
adding new keyword arguments to review_read doesn't keep breaking
the locale-integration tests.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

app/routers/chat.py

@@ -21,7 +21,7 @@ from app.db import get_session, utcnow
 from app.jobs._market_context import REFERENCE_LINE
 from app.models import AICall, Headline, Quote, StrategicLog
 from app.routers.api import _md_to_html
-from app.services.i18n import respond_in_clause
+from app.services.i18n import language_directive_lead, respond_in_clause
 from app.services.llm_prompts import build_chat_system_prompt
 from app.services.openrouter import call_llm, month_start
 from app.services.output_review import review_read
@@ -165,13 +165,17 @@ async def chat(
         headlines=headlines,
         reference_line=REFERENCE_LINE,
     )
-    # Respect the user's interface language preference: append a single
+    # Respect the user's interface language preference. The tail
-    # localized "respond in" nudge so the assistant answers in IT when
+    # "Respond in X" clause is easy for the model to drop when the
-    # the user has lang=it. The prompt + history (which includes the
+    # rest of the prompt is English (long log content, English
-    # user's own question, often in their language) are usually enough,
+    # market data, English headlines), so we ALSO prepend a stronger
-    # but the nudge guarantees the first reply lands correctly.
+    # language directive at the top — see services/i18n.
     user_lang = principal.user.lang if principal and principal.user else "en"
-    system_prompt = system_prompt + respond_in_clause(user_lang)
+    system_prompt = (
+        language_directive_lead(user_lang)
+        + system_prompt
+        + respond_in_clause(user_lang)
+    )
 
     msgs = [{"role": "system", "content": system_prompt}]
     for m in history:

app/routers/universe.py

@@ -362,10 +362,19 @@ async def analyze_portfolio(
     except Exception:
         raise HTTPException(status_code=400, detail="malformed JSON body")
 
-    user_lang = (
+    # Resolve lang. The frontend sends the live toggle state in
+    # payload["lang"]; that's what the user is *looking at* right now
+    # and is the most up-to-date value. user.lang from the DB is the
+    # persisted preference and is used as a fallback when the frontend
+    # didn't send anything (older clients, scripts, direct curl).
+    db_lang = (
         principal.user.lang if (principal.user and principal.user.lang) else "en"
     )
-    payload["lang"] = user_lang
+    incoming = (payload.get("lang") or "").strip().lower()
+    payload["lang"] = incoming or db_lang
+    log.info("analyze.lang_resolved",
+             payload_lang=incoming or None, db_lang=db_lang,
+             final=payload["lang"])
 
     try:
         req = portfolio_analysis.parse_request(payload)

app/services/i18n.py

@@ -46,3 +46,31 @@ def respond_in_clause(lang: str | None) -> str:
     if not lang or lang == "en" or lang not in LANGUAGES:
         return ""
     return f"\n\nRespond in {LANGUAGES[lang]}."
+
+
+def language_directive_lead(lang: str | None) -> str:
+    """Strong, top-of-prompt language directive for callers that
+    generate user-facing prose in real time (portfolio analysis,
+    chat) and need the output to actually land in the user's
+    preferred language. A single tail clause like
+    ``respond_in_clause`` is easy for the model to ignore when the
+    rest of the prompt + user message are entirely in English; this
+    leads with an explicit "all output in X" block, kept verbatim
+    rules for symbols/numbers, and is intended to be prepended to
+    the system prompt so the model anchors on the target language
+    before reading the rest. Combined with respond_in_clause at the
+    tail it gives a belt-and-suspenders defence.
+
+    Empty string for English or unknown codes so callers can paste
+    it in unconditionally.
+    """
+    if not lang or lang == "en" or lang not in LANGUAGES:
+        return ""
+    language = LANGUAGES[lang]
+    return (
+        f"# LANGUAGE — write everything in {language}\n"
+        f"All output — section headers, prose, lists, and any inline "
+        f"labels — must be written in {language}. Do NOT mix English in. "
+        f"Ticker symbols (AAPL, MSFT, VOD.L), ISO currency codes "
+        f"(USD, EUR, GBP), and numeric values stay unchanged.\n\n"
+    )

app/services/output_review.py

@@ -47,6 +47,18 @@ reply, or an email digest — and decide if it is publishable as-is.
 Mark CLEAN only if the text reads like finished editorial commentary
 a reader could see on a public dashboard without confusion.
 
+Editorial framework you should KNOW about (don't flag these):
+This dashboard's voice deliberately contrasts a "rational" read
+(fundamentals, policy regime, valuation) with an "irrational" read
+(positioning, narrative momentum, flows) and names the gap between
+them. Section labels like "Rational:" / "Irrational:" (or "Bull /
+Bear", or any explicit "X vs Y" contrast) are STRUCTURAL DEVICES,
+not the author thinking on the page. Treat them as finished prose.
+The Italian / Spanish / French / German equivalents
+("Razionalmente / Irrazionalmente", "Racionalmente / Irracionalmente",
+"Rationnellement / Irrationnellement", "Rational / Irrational") are
+the same device translated and equally fine.
+
 Mark UNCLEAN if the text contains ANY of:
 - Chain-of-thought / scratchpad markers — the author thinking on the
   page rather than presenting finished commentary. Phrases like
@@ -56,7 +68,9 @@ Mark UNCLEAN if the text contains ANY of:
   front of the reader (self-questioning) are not.
 - Self-questioning parentheticals: "Q1 2026? Actually Q4 2025?",
   "is it X or Y?", any place where the author appears to be working
-  out the answer in front of the reader.
+  out the answer in front of the reader. The "rational vs irrational"
+  contrast above is NOT self-questioning — the author is presenting
+  both reads as parallel takes, not asking which one is correct.
 - Meta-commentary about the task, output format, word limits, or
   instructions — e.g. "as required by the constraints", "the prompt
   asks", "let me address each".
@@ -87,6 +101,38 @@ No preamble, no markdown fences, no other fields.
 """
 
 
+# Surface-specific rider appended to the system prompt when the caller
+# passes a known `surface` to review_read(). Lets us relax or tighten
+# rules per editorial context without rewriting the whole prompt.
+_SURFACE_RIDERS = {
+    "portfolio": """\
+
+# Surface: portfolio commentary
+This text describes a real investor's holdings. DESCRIPTIVE risk
+language is the whole point of this surface and must NOT be flagged
+as financial advice. The following ARE fine:
+- Naming portfolio attributes: "high concentration", "single-name
+  exposure", "currency risk is unhedged", "FX exposure", "elevated
+  risk", "stretched valuations", "concentration is manageable", "low
+  diversification".
+- Stating what would invalidate the posture: "this view fails if
+  rates retrace", "the thesis depends on X holding".
+- Impersonal observation about a position's behaviour or sensitivity:
+  "the position warrants monitoring", "carries vulnerability to a
+  policy shock", "is sensitive to rate moves".
+
+ONLY flag EXPLICIT calls to action where a verb or directive is
+aimed at the reader:
+- Imperative verbs in the second person: "buy X", "sell Y",
+  "trim Z", "hedge", "rotate into".
+- "You should", "investors should", "consider X-ing", "we recommend".
+- Specific allocation prescriptions: "go 20% bonds", "overweight
+  tech", "underweight defensives".
+- Price-target predictions: "will reach $X by year-end".
+""",
+}
+
+
 @dataclass(frozen=True)
 class Verdict:
     clean: bool
@@ -94,18 +140,32 @@ class Verdict:
     cost_usd: float | None  # cost of the review call itself, for the ledger
 
 
-async def review_read(client: httpx.AsyncClient, candidate: str) -> Verdict:
+async def review_read(
+    client: httpx.AsyncClient,
+    candidate: str,
+    surface: str | None = None,
+) -> Verdict:
     """Ask the LLM whether `candidate` is a publishable read.
 
     Returns Verdict(clean, reason, cost). Any error — provider failure,
     JSON parse failure, missing field, wrong type — yields a CONSERVATIVE
     verdict (clean=False) so the caller drops the candidate. The
-    previously cached good summary stays visible on the dashboard."""
+    previously cached good summary stays visible on the dashboard.
+
+    `surface` selects a surface-specific rider that's appended to the
+    base system prompt — see _SURFACE_RIDERS. Currently only the
+    "portfolio" surface uses one (descriptive risk language is the
+    whole point there and shouldn't be flagged as advice). Unknown
+    or None surfaces fall back to the generic rules."""
     if not candidate or not candidate.strip():
         return Verdict(clean=False, reason="empty candidate", cost_usd=0.0)
 
+    system_prompt = _SYSTEM_PROMPT
+    if surface and surface in _SURFACE_RIDERS:
+        system_prompt = system_prompt + _SURFACE_RIDERS[surface]
+
     messages = [
-        {"role": "system", "content": _SYSTEM_PROMPT},
+        {"role": "system", "content": system_prompt},
         # Sent as a fenced user turn so the model can't confuse the
         # candidate with instructions, even if the candidate happens to
         # contain prompt-like prose.

app/services/portfolio_analysis.py

@@ -31,7 +31,7 @@ from app.config import get_settings
 from app.db import utcnow
 from app.logging import get_logger
 from app.models import AICall
-from app.services.i18n import LANGUAGES, respond_in_clause
+from app.services.i18n import LANGUAGES, language_directive_lead, respond_in_clause
 from app.services.llm_prompts import build_system_prompt
 from app.services.output_review import review_read
 from app.services.openrouter import (
@@ -257,18 +257,14 @@ implies X under scenario Y"), not advice ("buy X" / "sell Y" are forbidden).
 - ~350 words. No bullet lists. No buy/sell recommendations.
 - Do not repeat the input data verbatim — interpret it.
 
-# Rational vs irrational lens (mandatory)
+# DO NOT include in this surface (overrides the base prompt)
-Carry the base prompt's rational-vs-irrational framing through to every
+- No "Rational vs irrational" framing, no "Rational:" / "Irrational:"
-paragraph of the portfolio read. For each section above, contrast:
+  section labels, no parallel contrast lists. The base prompt asks
-- The RATIONAL read: what the underlying factors (fundamentals,
+  for this framework elsewhere; this surface is plain declarative
-  macro/policy regime, valuation, currency dynamics) justify for this
+  commentary on the holdings, not a comparative essay.
-  exposure;
+- No "System temperature:" closing line. That artefact belongs to the
-- The IRRATIONAL read: what positioning, narrative momentum, sentiment
+  daily strategic log; here the analysis ends with the last paragraph.
-  or flows are doing to that same exposure right now.
+- No "Update mode" headers, no anchor-date callouts, no watch list.
-Then name the GAP — does the holder's posture line up with the rational
-read, or is it riding the irrational one? A paragraph that names only
-the pie's numbers or only the macro backdrop, without placing the
-holding on this rational-vs-irrational axis, is incomplete.
 """
 
 
@@ -282,7 +278,18 @@ def build_prompt(req: AnalysisRequest) -> tuple[str, str]:
     head = enriched[:MAX_POSITIONS_INLINED]
     tail_count = max(0, len(enriched) - MAX_POSITIONS_INLINED)
 
-    system = build_system_prompt(req.tone, req.analysis) + "\n\n" + _SYSTEM_OVERRIDES + respond_in_clause(req.lang)
+    # Language directive both prepended (so the model anchors on the
+    # target language before reading the long English instruction
+    # block) and appended (defence in depth — a tail nudge alone
+    # was being ignored by deepseek-v4-flash when most of the
+    # context is English).
+    system = (
+        language_directive_lead(req.lang)
+        + build_system_prompt(req.tone, req.analysis)
+        + "\n\n"
+        + _SYSTEM_OVERRIDES
+        + respond_in_clause(req.lang)
+    )
 
     user_parts = [
         f"# Portfolio commentary request — {utcnow().strftime('%Y-%m-%d')}",
@@ -333,7 +340,17 @@ async def analyse(
                     {"role": "system", "content": system},
                     {"role": "user", "content": user},
                 ],
-                max_tokens=2000,
+                # 4000 not 2000. Italian / Spanish / French / German
+                # output runs ~25-35% longer in tokens than English; on
+                # top of that DeepSeek-V4-flash bills its internal
+                # reasoning against the same budget. At 2000 we
+                # repeatedly hit finish_reason=length mid-sentence,
+                # which the reviewer agent then correctly flags as
+                # truncated and rejects — the user ends up looking at
+                # whatever stale row was last cached. 4000 leaves
+                # ample headroom; we only pay for tokens actually
+                # emitted, not the cap itself.
+                max_tokens=4000,
             )
             status = "ok"
             error_msg = None
@@ -348,8 +365,14 @@ async def analyse(
         # buy/sell or allocation language is a regulatory hazard. Drop
         # the response on a reject and surface a retry-able error to the
         # caller; no analysis is ever persisted server-side anyway.
+        # surface="portfolio" applies a rider that loosens the generic
+        # "no advice" rule to permit descriptive risk language
+        # ("concentration is high", "currency exposure is unhedged",
+        # "the position warrants monitoring") which is the actual
+        # purpose of this surface, while keeping explicit
+        # buy/sell/allocation directives forbidden.
         if llm is not None:
-            verdict = await review_read(client, llm.content)
+            verdict = await review_read(client, llm.content, surface="portfolio")
             review_cost = verdict.cost_usd or 0.0
             if not verdict.clean:
                 status = "leaked"

app/static/css/layout.css

@@ -246,6 +246,20 @@ body.drawer-open .drawer-backdrop { opacity: 1; }
     grid-template-columns: 1fr;
     grid-template-areas: "header" "indicators" "portfolio" "log" "news";
   }
+  /* Single-column layout — the log panel no longer shares a row with
+     indicators + portfolio, so the height-constraint dance above
+     would just collapse the panel to nothing. Drop the constraint and
+     let the log expand to its natural content height; page scroll
+     takes over. */
+  #log-panel {
+    contain: none;
+    display: block;
+  }
+  #log-panel .panel-body {
+    flex: none;
+    min-height: auto;
+    overflow-y: visible;
+  }
 }
 
 #dash-header-container { grid-area: header; }
@@ -253,10 +267,25 @@ body.drawer-open .drawer-backdrop { opacity: 1; }
 #portfolio-panel  { grid-area: portfolio; }
 #log-panel        {
   grid-area: log;
-  /* Stretch (default align-self) so the log panel's border reaches the
+  /* Bottom-align with the portfolio panel WITHOUT padding the inside
-     bottom of the portfolio next to it — the two right-hand panels
+     of either box. The key is `contain: size`: a grid item with this
-     align cleanly. The log body itself sits at the top of the panel;
+     contracts to declare "my contents do not contribute to my own
-     any height beyond its content is empty padding inside the box. */
+     intrinsic size." The outer grid therefore sizes rows 2-3 from
+     indicators + portfolio alone; this panel stretches to that
+     combined height; if the log content is taller it scrolls inside.
+     The flex column + min-height:0 chain lets .panel-body fill the
+     remaining height below the header and scroll instead of
+     overflowing the panel. */
+  contain: size;
+  display: flex;
+  flex-direction: column;
+  min-height: 0;
+}
+#log-panel .panel-header { flex-shrink: 0; }
+#log-panel .panel-body {
+  flex: 1;
+  min-height: 0;
+  overflow-y: auto;
 }
 #news-panel       { grid-area: news; }
 

app/static/css/portfolio.css

@@ -86,6 +86,13 @@
 .pf-actions button:disabled { opacity: 0.5; cursor: not-allowed; }
 .pf-secondary { color: var(--muted); }
 .pf-secondary:hover { color: var(--negative); border-color: var(--negative); }
+/* "Forget this pie" is destructive — only show it while the user is
+   in edit mode (the same mode that surfaces the per-row delete × and
+   the add-position form). Outside of edit mode the row stays in the
+   DOM so existing handlers and any future surface that wants to
+   toggle it can do so without re-rendering. */
+#pf-forget { display: none; }
+#portfolio-panel.pf-editing #pf-forget { display: inline-block; }
 
 .pf-analysis {
   margin-top: 14px;
@@ -107,6 +114,24 @@
   list-style: none;        /* hide native marker in Firefox */
 }
 .pf-analysis__head::-webkit-details-marker { display: none; }
+.pf-analysis__head-right {
+  display: inline-flex;
+  align-items: center;
+  gap: 12px;
+}
+.pf-regen {
+  background: transparent;
+  border: 1px solid var(--border);
+  color: var(--muted);
+  padding: 3px 9px;
+  font: inherit;
+  font-size: 10.5px;
+  letter-spacing: inherit;
+  cursor: pointer;
+  text-transform: inherit;
+}
+.pf-regen:hover { color: var(--accent); border-color: var(--accent); }
+.pf-regen:disabled { opacity: 0.5; cursor: not-allowed; }
 .pf-analysis__head-left::before {
   content: "▸ ";
   display: inline-block;

app/static/js/portfolio.js

@@ -372,13 +372,24 @@
         '</tr></thead>' +
         '<tbody>' + rows + '</tbody>' +
       '</table>' +
+      // The "Generate" button only renders when there's no cached
+      // analysis yet. Once one exists, regeneration moves inside the
+      // collapsible analysis box (see showAnalysis below). The "Forget
+      // this pie" button is destructive enough that it lives in
+      // edit-mode only — CSS in portfolio.css hides it when the
+      // portfolio panel isn't carrying the .pf-editing class.
       '<div class="pf-actions">' +
-        '<button id="pf-analyze" type="button">Generate AI analysis</button>' +
+        (pie.analysis && pie.analysis.content
+          ? ''
+          : '<button id="pf-analyze" type="button">Generate AI analysis</button>') +
         '<button id="pf-forget" type="button" class="pf-secondary">Forget this pie</button>' +
       '</div>' +
       '<div id="pf-analysis" class="pf-analysis" hidden></div>';
 
-    document.getElementById('pf-analyze').addEventListener('click', () => runAnalysis(pie, enriched));
+    const analyzeBtn = document.getElementById('pf-analyze');
+    if (analyzeBtn) {
+      analyzeBtn.addEventListener('click', () => runAnalysis(pie, enriched));
+    }
     document.getElementById('pf-forget').addEventListener('click', () => {
       if (confirm('Remove the saved pie from this browser? The server holds nothing — this is local.')) {
         clearPie();
@@ -390,13 +401,16 @@
     // wipe it. Rendered expanded so the user keeps seeing the body they
     // just generated — collapsing it under their cursor every minute
     // reads as "the analysis disappeared". They can still click the
-    // header to collapse manually within a single refresh window.
+    // header to collapse manually within a single refresh window. The
+    // regenerate callback closes over the current pie/enriched so a
+    // click rebuilds the analysis with the same context that drove
+    // the initial render.
     if (pie.analysis && pie.analysis.content) {
-      showAnalysis(pie.analysis, { open: true });
+      showAnalysis(pie.analysis, { open: true }, () => runAnalysis(pie, enriched));
     }
   }
 
-  function showAnalysis(analysis, opts) {
+  function showAnalysis(analysis, opts, onRegenerate) {
     const out = document.getElementById('pf-analysis');
     if (!out) return;
     const openAttr = (opts && opts.open) ? ' open' : '';
@@ -407,20 +421,43 @@
           '<span class="pf-analysis__head-left">' +
             'AI analysis' +
           '</span>' +
+          '<span class="pf-analysis__head-right">' +
             '<span class="pf-as-of">' +
               esc((analysis.generated_at || '').slice(0, 19).replace('T', ' ')) +
               ' UTC</span>' +
+            (onRegenerate
+              ? '<button id="pf-regen" type="button" class="pf-regen"' +
+                ' title="Run the analysis again on the current portfolio">' +
+                'Regenerate</button>'
+              : '') +
+          '</span>' +
         '</summary>' +
         '<pre class="pf-analysis__body">' + esc(analysis.content) + '</pre>' +
       '</details>';
+    if (onRegenerate) {
+      const regen = document.getElementById('pf-regen');
+      if (regen) {
+        regen.addEventListener('click', (e) => {
+          // The button lives inside <summary>; clicking it would
+          // normally toggle the <details> open/closed. Suppress the
+          // default toggle and the bubble so only our regen runs.
+          e.preventDefault();
+          e.stopPropagation();
+          onRegenerate();
+        });
+      }
+    }
   }
 
   async function runAnalysis(pie, enriched) {
     const out = document.getElementById('pf-analysis');
-    const btn = document.getElementById('pf-analyze');
+    // First-run click is on pf-analyze; the regenerate path is pf-regen
+    // inside the details summary. Either may be the live trigger.
+    const btn = document.getElementById('pf-analyze') ||
+                document.getElementById('pf-regen');
     out.hidden = false;
     out.innerHTML = '<div class="empty">generating…</div>';
-    btn.disabled = true;
+    if (btn) btn.disabled = true;
 
     // Build the prices payload from the universe cache so the server
     // doesn't have to re-fetch.
@@ -432,6 +469,13 @@
       }
     }
 
+    // The language toggle's data-lang attribute is the user's LIVE
+    // pick — newer than user.lang in the DB if the user toggled and
+    // hit Generate/Regenerate before the toggle-PATCH committed.
+    // Backend prefers this value if provided (see universe.py).
+    const langPill = document.getElementById('lang-toggle');
+    const userLang = (langPill && langPill.dataset.lang) || 'en';
+
     try {
       const r = await fetch('/api/analyze', {
         method: 'POST',
@@ -441,6 +485,7 @@
           positions: pie.positions,
           prices: prices,
           base_currency: pie.base_currency || 'GBP',
+          lang: userLang,
         }),
       });
       const data = await r.json();
@@ -458,11 +503,17 @@
       }
       // Persist before rendering so auto-refresh can re-hydrate.
       saveAnalysis(data);
-      showAnalysis(data, { open: true });
+      // Pass the regenerate callback so the in-details "Regenerate"
+      // button shows up on the freshly-rendered analysis too.
+      showAnalysis(data, { open: true }, () => runAnalysis(pie, enriched));
     } catch (e) {
       out.innerHTML = '<div class="pf-warn">' + esc(e.message) + '</div>';
     } finally {
-      btn.disabled = false;
+      // The original button may have been replaced by showAnalysis →
+      // re-fetch its handle (or null if neither id is on the page now).
+      const liveBtn = document.getElementById('pf-analyze') ||
+                      document.getElementById('pf-regen');
+      if (liveBtn) liveBtn.disabled = false;
     }
   }
 

app/templates/settings.html

@@ -98,10 +98,8 @@
     <details class="settings-section" id="import" open>
       <summary class="settings-section__head">Import portfolio (CSV)</summary>
       <p class="settings-section__lede">
-        Drop a portfolio CSV from any broker &mdash; Trading 212 is recognised
+        Drop a portfolio CSV from any broker. We&rsquo;ll parse it and show
-        natively and other formats (IBKR, Fidelity, Schwab&hellip;) are
+        a preview before importing anywhere.
-        auto-detected. We&rsquo;ll parse it and show a preview before importing
-        anywhere.
         <br><span class="muted">T212 export path:
         <span class="neu">Investing &rarr; Your Pie &rarr; &middot;&middot;&middot; &rarr; Export</span>.</span>
       </p>
@@ -332,8 +330,8 @@
   </div>
 </div>
 
-<script src="{{ url_for(‘static’, path=’/js/portfolio-sync.js’) }}" defer></script>
+<script src="{{ url_for('static', path='/js/portfolio-sync.js') }}?v={{ ASSET_VERSION }}" defer></script>
-<script src="{{ url_for(‘static’, path=’/js/settings-sync.js’) }}" defer></script>
+<script src="{{ url_for('static', path='/js/settings-sync.js') }}?v={{ ASSET_VERSION }}" defer></script>
 {% endif %}
 
 <script>

tests/conftest.py

@@ -36,7 +36,7 @@ def stub_reviewer(monkeypatch):
     """
     from app.services.output_review import Verdict
 
-    async def _clean(_client, _candidate):
+    async def _clean(_client, _candidate, **_kw):
         return Verdict(clean=True, reason="stubbed-by-conftest", cost_usd=0.0)
 
     for mod_path in (