Generate a suite with one writer per use case, and persona values the platform understands

src/fi/alk/harness/background_noise.py

@@ -0,0 +1,73 @@
+"""Choose the caller-side ambient noise a scenario should be heard through.
+
+A scenario that sets ``background_noise`` wants the agent to handle a caller phoning from somewhere
+real: a car, a street, an office. The clip is chosen here and handed to the voice engine, which
+mixes it under the simulated caller's audio.
+
+Two sources, in order. A run may point ``ALK_BACKGROUND_NOISE_CATALOG`` at a JSON file of clips
+(each with an ``environment`` tag and a ``url`` or ``path``); the catalog stays a local file so its
+asset locations are never committed here. When no catalog matches, a LiveKit builtin clip is used,
+which needs no external asset and always works.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+from pathlib import Path
+
+# LiveKit ships these; they are the reliable default when no custom catalog is configured.
+_BUILTIN_BY_ENVIRONMENT: dict[str, str] = {
+    "street": "CITY_AMBIENCE",
+    "transit": "CITY_AMBIENCE",
+    "vehicle": "CITY_AMBIENCE",
+    "outdoors": "FOREST_AMBIENCE",
+    "retail": "CROWDED_ROOM",
+    "office": "OFFICE_AMBIENCE",
+    "home": "OFFICE_AMBIENCE",
+}
+_DEFAULT_BUILTIN = "OFFICE_AMBIENCE"
+
+
+def enabled() -> bool:
+    """Whether any scenario may be heard through background noise on this run.
+
+    Off unless ``ALK_BACKGROUND_NOISE`` opts in, so a run needs no environment at all to be
+    silent. Continuous ambient audio under the caller competes with endpoint detection, and calls
+    carrying it end earlier and on fewer turns, so silence is the setting a run should fall into
+    rather than the one it has to ask for. Opting in still only permits noise: a scenario that
+    asked for none stays silent either way.
+    """
+    return os.environ.get("ALK_BACKGROUND_NOISE", "0").strip().lower() in (
+        "1",
+        "on",
+        "true",
+        "yes",
+    )
+
+
+def source_for(environment: str = "", seed: str = "") -> str:
+    """A background-noise source for a scenario.
+
+    Returns a ``url``/``path`` from the configured catalog when one matches the environment, else the
+    name of a LiveKit builtin clip. The choice is deterministic in ``seed`` so the same scenario
+    hears the same place across runs.
+    """
+    env = (environment or "").strip().lower()
+    catalog = os.environ.get("ALK_BACKGROUND_NOISE_CATALOG", "").strip()
+    if catalog and Path(catalog).is_file():
+        try:
+            entries = json.loads(Path(catalog).read_text(encoding="utf-8"))
+        except (OSError, ValueError):
+            entries = []
+        if isinstance(entries, list) and entries:
+            pool = [
+                entry
+                for entry in entries
+                if str(entry.get("environment", "")).strip().lower() == env
+            ] or entries
+            chosen = pool[sum(ord(character) for character in (seed or env or "x")) % len(pool)]
+            located = str(chosen.get("url") or chosen.get("path") or "").strip()
+            if located:
+                return located
+    return _BUILTIN_BY_ENVIRONMENT.get(env, _DEFAULT_BUILTIN)

src/fi/alk/harness/build.py

@@ -26,6 +26,7 @@
     permission_gate,
     provider_env,
     provisioning,
+    thinking_config,
 )
 from .contract import AgentContract
 from .session import Stage
@@ -202,6 +203,7 @@ def open_stage(
     options.disallowed_tools = list(UNWANTED)
     options.hooks = gate_hooks(allowed)
     options.can_use_tool = permission_gate(ask, allowed)
+    options.thinking = thinking_config()
     return Stage(options, name=SKILL), destination
 
 

src/fi/alk/harness/config.py

@@ -52,6 +52,24 @@ def chosen_model(model: str | None = None) -> str:
     return model or os.environ.get("ALK_HARNESS_MODEL", DEFAULT_MODEL)
 
 
+def thinking_config() -> dict[str, Any]:
+    """How much the model may think, from ALK_HARNESS_THINKING.
+
+    The Claude Code CLI defaults to adaptive thinking. In this harness the correctness of what a
+    stage produces is re-checked by code gates (a scenario is proved against the real world, a
+    contract is validated), so the model's private reasoning is spent on decisions the gates make
+    again anyway. Left unset, that reasoning was the majority of generated tokens and the majority
+    of wall time. Default to disabled for speed; ``adaptive`` restores the old behaviour, and an
+    integer sets an explicit budget for models that still honour one.
+    """
+    setting = os.environ.get("ALK_HARNESS_THINKING", "disabled").strip().lower()
+    if setting in {"adaptive", "on", "auto"}:
+        return {"type": "adaptive", "display": "omitted"}
+    if setting.isdigit() and int(setting) > 0:
+        return {"type": "enabled", "budget_tokens": int(setting), "display": "omitted"}
+    return {"type": "disabled"}
+
+
 def provisioning(enabled: bool | None = None) -> bool:
     """Compatibility switch for callers selecting the legacy provisioning surface.
 
@@ -75,10 +93,20 @@ def provider_env(model: str | None = None) -> dict[str, str]:
     Claude Code resolves the GCP project from ``GOOGLE_CLOUD_PROJECT``, the credential file, or
     the active gcloud configuration, in that order, so an unset project id is not an error here.
     """
+    # Every model a session can reach is pinned to the same one. Naming only the main model
+    # leaves the sub-agent and fast-path settings to the CLI's own preference, and a suite written
+    # by twenty writers then runs on whatever that preference happens to be rather than on the
+    # model the run asked for.
+    chosen = chosen_model(model)
     env = {
         "CLAUDE_CODE_USE_VERTEX": "1",
         "CLOUD_ML_REGION": os.environ.get("CLOUD_ML_REGION", "global"),
-        "ANTHROPIC_MODEL": chosen_model(model),
+        "ANTHROPIC_MODEL": chosen,
+        "ANTHROPIC_DEFAULT_SONNET_MODEL": chosen,
+        "ANTHROPIC_DEFAULT_OPUS_MODEL": chosen,
+        "ANTHROPIC_DEFAULT_HAIKU_MODEL": chosen,
+        "ANTHROPIC_SMALL_FAST_MODEL": chosen,
+        "CLAUDE_CODE_SUBAGENT_MODEL": chosen,
     }
     for passthrough in (
         "ANTHROPIC_VERTEX_PROJECT_ID",
@@ -123,6 +151,7 @@ def read_only_session(
     options.disallowed_tools = list(UNWANTED)
     options.hooks = gate_hooks(allowed)
     options.can_use_tool = permission_gate(granted=allowed)
+    options.thinking = thinking_config()
     return options
 
 

src/fi/alk/harness/data/persona_vocabulary.json

@@ -0,0 +1,111 @@
+{
+  "GenderChoices": [
+    "male",
+    "female"
+  ],
+  "AgeGroupChoices": [
+    "18-25",
+    "25-32",
+    "32-40",
+    "40-50",
+    "50-60",
+    "60+"
+  ],
+  "LocationChoices": [
+    "United States",
+    "Canada",
+    "United Kingdom",
+    "Australia",
+    "India"
+  ],
+  "ProfessionChoices": [
+    "Student",
+    "Teacher",
+    "Engineer",
+    "Doctor",
+    "Nurse",
+    "Business Owner",
+    "Manager",
+    "Sales Representative",
+    "Customer Service",
+    "Technician",
+    "Consultant",
+    "Accountant",
+    "Marketing Professional",
+    "Retired",
+    "Homemaker",
+    "Freelancer",
+    "Other"
+  ],
+  "PersonalityChoices": [
+    "Friendly and cooperative",
+    "Professional and formal",
+    "Cautious and skeptical",
+    "Impatient and direct",
+    "Detail-oriented",
+    "Easy-going",
+    "Anxious",
+    "Confident",
+    "Analytical",
+    "Emotional",
+    "Reserved",
+    "Talkative"
+  ],
+  "CommunicationStyleChoices": [
+    "Direct and concise",
+    "Detailed and elaborate",
+    "Casual and friendly",
+    "Formal and polite",
+    "Technical",
+    "Simple and clear",
+    "Questioning",
+    "Assertive",
+    "Passive",
+    "Collaborative"
+  ],
+  "AccentChoices": [
+    "American",
+    "Australian",
+    "Indian",
+    "Canadian",
+    "Neutral"
+  ],
+  "LanguageChoices": [
+    "Arabic",
+    "Bulgarian",
+    "Chinese Simplified",
+    "Czech",
+    "Danish",
+    "Dutch",
+    "English",
+    "Finnish",
+    "French",
+    "German",
+    "Greek",
+    "Hindi",
+    "Hungarian",
+    "Indonesian",
+    "Italian",
+    "Japanese",
+    "Korean",
+    "Malay",
+    "Norwegian",
+    "Polish",
+    "Portuguese",
+    "Romanian",
+    "Russian",
+    "Slovak",
+    "Spanish",
+    "Swedish",
+    "Turkish",
+    "Ukrainian",
+    "Vietnamese"
+  ],
+  "ConversationSpeedChoices": [
+    "0.5",
+    "0.75",
+    "1.0",
+    "1.25",
+    "1.5"
+  ]
+}