[spark-compete wave 4] memory / persona / voice (stacked)

evals/compare_personas.py

@@ -179,7 +179,11 @@ def main() -> int:
         print(f"=> candidate {args.candidate} WINS")
         return 0
     print(f"=> baseline {args.baseline} holds")
-    return 1
+    # Exit 0 on a completed comparison regardless of verdict (POSIX: the exit
+    # code signals run success/failure, not a domain result). The verdict is on
+    # stdout: "candidate ... WINS" vs "baseline ... holds". No CI/Makefile/
+    # wrapper in this repo branches on the old exit-1 baseline-holds signal.
+    return 0
 
 
 if __name__ == "__main__":

evals/continuous_eval.py

@@ -42,10 +42,12 @@
 import json
 import sys
 import time
-import traceback
+import logging
 from pathlib import Path
 from statistics import mean as mean_
 
+_logger = logging.getLogger(__name__)
+
 _REPO_ROOT = Path(__file__).resolve().parents[1]
 sys.path.insert(0, str(_REPO_ROOT / "src"))
 
@@ -385,7 +387,7 @@ def main() -> int:
                     last_full = now
             except Exception as exc:
                 print(f"[continuous_eval] eval error on {provider_name}: {exc}", flush=True)
-                traceback.print_exc()
+                _logger.exception("eval error on %s", provider_name)
                 _write_heartbeat(heartbeat_path, "error")
                 if args.once:
                     return 1
@@ -442,7 +444,7 @@ def main() -> int:
             return 0
         except Exception as exc:
             print(f"[continuous_eval] outer error: {exc}", flush=True)
-            traceback.print_exc()
+            _logger.exception("outer eval loop error")
             time.sleep(60)
 
 

evals/evolve.py

@@ -77,7 +77,11 @@ def find_latest_persona() -> tuple[int, PersonaSpec]:
         if match:
             versions.append((int(match.group(1)), path))
     if not versions:
-        raise FileNotFoundError("No persona.vN.md artifacts found.")
+        raise FileNotFoundError(
+            f"No persona.vN.md artifacts found in {ARTIFACTS_DIR}. "
+            f"Seed persona.v1.md in that directory before running evolve, "
+            f"or override the location via ARTIFACTS_DIR before launching."
+        )
     versions.sort(key=lambda x: x[0])
     n, path = versions[-1]
     text = path.read_text(encoding="utf-8")

evals/evolve_persona.py

@@ -115,7 +115,11 @@ def find_latest_persona() -> tuple[int, PersonaSpec]:
         if match:
             versions.append((int(match.group(1)), path))
     if not versions:
-        raise FileNotFoundError("No persona.vN.md artifacts found.")
+        raise FileNotFoundError(
+            f"No persona.vN.md artifacts found in {ARTIFACTS_DIR}. "
+            f"Seed persona.v1.md in that directory before running evolve_persona, "
+            f"or override the location via ARTIFACTS_DIR before launching."
+        )
     versions.sort(key=lambda x: x[0])
     n, path = versions[-1]
     text = path.read_text(encoding="utf-8")

evals/full_pulse.py

@@ -87,7 +87,12 @@ def main() -> int:
             print(f"[{label}] ERROR generating: {exc}")
             t1_rows.append({"label": label, "prompt": prompt, "error": str(exc)})
             continue
-        t1 = score_persona(result.final)
+        try:
+            t1 = score_persona(result.final)
+        except Exception as exc:
+            print(f"[{label}] T1 scorer error: {exc}")
+            t1_rows.append({"label": label, "prompt": prompt, "error": str(exc)})
+            continue
         try:
             t2 = score_distinctiveness(result.final, provider=provider)
         except Exception as exc:

evals/lowest_tier_watch.py

@@ -111,9 +111,10 @@ def _write_heartbeat(path: Path, phase: str) -> None:
 
 def find_lowest_tier(history: list[dict], *, min_runs: int) -> tuple[str | None, float, int]:
     """Return (axis, mean_score, n_samples) for the weakest tier with
-    at least min_runs samples. Returns (None, 1.0, 0) if nothing
-    qualifies."""
-    weakest = (None, 1.0, 0)
+    at least min_runs samples. Returns (None, inf, 0) if nothing
+    qualifies. A perfect-1.0 tier still qualifies as the weakest when
+    all eligible axes tie at the ceiling."""
+    weakest: tuple[str | None, float, int] = (None, float("inf"), 0)
     for axis in TIER_KEYS:
         values = [r.get(axis) for r in history if isinstance(r.get(axis), (int, float))]
         if len(values) < min_runs:

src/spark_character/chip_loader.py

@@ -152,20 +152,38 @@ def _validate_score(value: Any, field_name: str) -> None:
         raise ValueError(f"Personality chip field {field_name} must be a number in [0, 1].")
 
 
+_CHIP_SCHEMA_HINT = (
+    " See docs/ARCHITECTURE.md (schema: spark-personality-chip.v1) for the "
+    "canonical chip YAML overview."
+)
+
+
 def validate_chip_yaml_spec(spec: Any) -> dict[str, Any]:
     """Validate the minimal chip-lab YAML shape consumed by spark-character."""
     root = _require_mapping(spec, "<root>")
     schema = root.get("schema", "spark-personality-chip.v1")
     if not isinstance(schema, str) or not schema.strip():
-        raise ValueError("Personality chip field schema must be a non-empty string.")
+        raise ValueError(
+            "Personality chip field schema must be a non-empty string (e.g. "
+            "'spark-personality-chip.v1')." + _CHIP_SCHEMA_HINT
+        )
 
     identity = _require_mapping(root.get("identity"), "identity")
     for key in ("id", "name"):
         if not isinstance(identity.get(key), str) or not identity.get(key, "").strip():
-            raise ValueError(f"Personality chip field identity.{key} must be a non-empty string.")
+            got = type(identity.get(key)).__name__
+            raise ValueError(
+                f"Personality chip field identity.{key} must be a non-empty string "
+                f"(got {got}). Example: identity:\\n  {key}: \"founder-operator\"."
+                + _CHIP_SCHEMA_HINT
+            )
     for key in ("archetype", "voice_signature", "tagline"):
         if key in identity and identity[key] is not None and not isinstance(identity[key], str):
-            raise ValueError(f"Personality chip field identity.{key} must be a string.")
+            got = type(identity[key]).__name__
+            raise ValueError(
+                f"Personality chip field identity.{key} must be a string "
+                f"(got {got})." + _CHIP_SCHEMA_HINT
+            )
 
     traits = _require_mapping(root.get("traits", {}), "traits")
     for key in TRAIT_FIELDS:
@@ -177,33 +195,64 @@ def validate_chip_yaml_spec(spec: Any) -> dict[str, Any]:
         if key in emotional_profile:
             _validate_score(emotional_profile[key], f"emotional_profile.{key}")
     if "empathy_style" in emotional_profile and not isinstance(emotional_profile["empathy_style"], str):
-        raise ValueError("Personality chip field emotional_profile.empathy_style must be a string.")
+        got = type(emotional_profile["empathy_style"]).__name__
+        raise ValueError(
+            f"Personality chip field emotional_profile.empathy_style must be a string "
+            f"(got {got}). Example: empathy_style: \"warm, but direct when asked\"." + _CHIP_SCHEMA_HINT
+        )
     emotional_range = _require_mapping(emotional_profile.get("emotional_range", {}), "emotional_profile.emotional_range")
     for key, value in emotional_range.items():
         _validate_score(value, f"emotional_profile.emotional_range.{key}")
     triggers = _require_mapping(emotional_profile.get("triggers", {}), "emotional_profile.triggers")
     for key, value in triggers.items():
         if not isinstance(value, list):
-            raise ValueError(f"Personality chip field emotional_profile.triggers.{key} must be a list.")
+            got = type(value).__name__
+            raise ValueError(
+                f"Personality chip field emotional_profile.triggers.{key} must be a list "
+                f"(got {got}). Example: triggers:\\n  {key}: [\"betrayal\", \"unfairness\"]."
+                + _CHIP_SCHEMA_HINT
+            )
 
     preferences = _require_mapping(root.get("preferences", {}), "preferences")
     for key in ("likes", "dislikes"):
         if key in preferences and not isinstance(preferences[key], list):
-            raise ValueError(f"Personality chip field preferences.{key} must be a list.")
+            got = type(preferences[key]).__name__
+            raise ValueError(
+                f"Personality chip field preferences.{key} must be a list "
+                f"(got {got}). Example: preferences:\\n  {key}: [\"short replies\", \"plain language\"]."
+                + _CHIP_SCHEMA_HINT
+            )
     for key in ("communication", "decision_making"):
         if key in preferences and not isinstance(preferences[key], dict):
-            raise ValueError(f"Personality chip field preferences.{key} must be a mapping.")
+            got = type(preferences[key]).__name__
+            raise ValueError(
+                f"Personality chip field preferences.{key} must be a mapping "
+                f"(got {got}). Example: preferences:\\n  {key}:\\n    style: \"direct\"."
+                + _CHIP_SCHEMA_HINT
+            )
 
     safety = _require_mapping(root.get("safety", {}), "safety")
     if "harm_avoidance" in safety and not isinstance(safety["harm_avoidance"], list):
-        raise ValueError("Personality chip field safety.harm_avoidance must be a list.")
+        got = type(safety["harm_avoidance"]).__name__
+        raise ValueError(
+            f"Personality chip field safety.harm_avoidance must be a list "
+            f"(got {got}). Example: harm_avoidance: [\"no medical advice\", \"no legal advice\"]." + _CHIP_SCHEMA_HINT
+        )
 
     for key in TOP_LEVEL_LIST_FIELDS:
         if key in root and not isinstance(root[key], list):
-            raise ValueError(f"Personality chip field {key} must be a list.")
+            got = type(root[key]).__name__
+            raise ValueError(
+                f"Personality chip field {key} must be a list (got {got})."
+                + _CHIP_SCHEMA_HINT
+            )
     for key in TOP_LEVEL_DICT_FIELDS:
         if key in root and not isinstance(root[key], dict):
-            raise ValueError(f"Personality chip field {key} must be a mapping.")
+            got = type(root[key]).__name__
+            raise ValueError(
+                f"Personality chip field {key} must be a mapping (got {got})."
+                + _CHIP_SCHEMA_HINT
+            )
     return root
 
 
@@ -337,8 +386,11 @@ def load_chip_by_id(
                 continue
             if chip.id == safe_chip_id:
                 return chip
+    # Report only the basenames of the labs we actually searched: enough to
+    # debug a misplaced chip without leaking the full filesystem layout.
+    searched = ", ".join(sorted({p.name for p in paths}))
     raise FileNotFoundError(
-        f"Personality chip '{safe_chip_id}' not found in: {[str(p) for p in paths]}"
+        f"Personality chip '{safe_chip_id}' not found in labs: {searched}"
     )
 
 

src/spark_character/codex_provider.py

@@ -25,7 +25,12 @@
 def _default_codex_binary() -> str:
     explicit = os.environ.get("CODEX_PATH") or os.environ.get("SPARK_CODEX_PATH")
     if explicit:
-        return explicit
+        # Validate the path exists and is a regular file to prevent
+        # arbitrary binary execution via malicious env vars
+        expanded = os.path.expanduser(explicit)
+        if not os.path.isfile(expanded):
+            raise FileNotFoundError(f"Codex binary not found: {expanded}")
+        return expanded
     if sys.platform.startswith("win"):
         return "codex.cmd"
     return "codex"
@@ -75,15 +80,35 @@ def call_codex(
             "--output-last-message", str(out_path),
             "-",
         ]
-        result = subprocess.run(
-            cmd,
-            input=combined.encode("utf-8"),
-            capture_output=True,
-            timeout=spec.timeout_seconds,
-        )
+        try:
+            result = subprocess.run(
+                cmd,
+                input=combined.encode("utf-8"),
+                capture_output=True,
+                timeout=spec.timeout_seconds,
+            )
+        except FileNotFoundError as exc:
+            # Guards the eval/judge driver against a raw stack trace when the
+            # codex CLI is not installed or CODEX_PATH points at a removed
+            # binary. Preserves the operator's next move (install codex or
+            # set CODEX_PATH) instead of leaking the OSError text.
+            raise RuntimeError(
+                f"codex binary not found at {spec.binary!r}. Install the codex CLI "
+                f"or set CODEX_PATH / SPARK_CODEX_PATH to its absolute path."
+            ) from exc
+        except subprocess.TimeoutExpired as exc:
+            # Closes the silent-hang window when codex exec exceeds the
+            # configured timeout; surfaces the actual budget so the operator
+            # can raise CodexSpec.timeout_seconds rather than guess.
+            raise RuntimeError(
+                f"codex exec timed out after {spec.timeout_seconds:.0f}s. "
+                f"Increase CodexSpec.timeout_seconds or check that the codex "
+                f"CLI is responsive."
+            ) from exc
         if result.returncode != 0:
-            stderr = result.stderr.decode("utf-8", errors="replace") if result.stderr else ""
-            raise RuntimeError(f"codex exec failed (rc={result.returncode}): {stderr.strip()[:300]}")
+            # Redact raw stderr (may carry internal paths / prompt fragments);
+            # keep the return code so operators can still triage.
+            raise RuntimeError(f"codex exec failed (rc={result.returncode})")
         if not out_path.exists():
             raise RuntimeError("codex exec did not write the expected output file.")
         text = out_path.read_text(encoding="utf-8", errors="replace").strip()

src/spark_character/critic.py

@@ -41,7 +41,7 @@ class CritiqueResult:
 def load_critic(version: str = DEFAULT_CRITIC_VERSION) -> CriticSpec:
     path = ARTIFACTS_DIR / f"critic.{version}.md"
     if not path.exists():
-        raise FileNotFoundError(f"Critic artifact not found: {path}")
+        raise FileNotFoundError("Critic artifact not found")
     return CriticSpec(version=version, text=path.read_text(encoding="utf-8"))
 
 

src/spark_character/memory_grounded.py

@@ -23,8 +23,8 @@
 
     from spark_character.memory_grounded import build_t7_probes_from_state, latest_user_states
     probes = build_t7_probes_from_state(
-        sib_home="C:/Users/USER/Desktop/.../tmp-home",
-        human_id="human:telegram:8319079055",
+        sib_home=Path.home() / ".spark" / "sib-home",
+        human_id="human:telegram:<your-user-id>",
     )
     for p in probes:
         result = run_deep_probe(p, provider=..., persona=...)
@@ -68,8 +68,10 @@ class UserStateObservation:
 def _open_state(sib_home: str | Path) -> sqlite3.Connection:
     db = Path(sib_home) / "state.db"
     if not db.exists():
-        raise FileNotFoundError(f"state.db not found in {sib_home}")
-    return sqlite3.connect(str(db))
+        raise FileNotFoundError("State database not found")
+    # Open read-only via URI so the probe builder can never accidentally
+    # mutate SIB's authoritative state.db (user_instructions, personality_observations).
+    return sqlite3.connect(f"file:{db}?mode=ro", uri=True)
 
 
 def latest_user_instructions(

src/spark_character/persona.py

@@ -102,7 +102,7 @@ def set_latest_persona_version(
     resolved = validate_persona_version(version)
     artifact_path = artifacts_dir / f"persona.{resolved}.md"
     if not artifact_path.exists():
-        raise FileNotFoundError(f"Persona artifact not found: {artifact_path}")
+        raise FileNotFoundError(f"Persona artifact not found: persona.{resolved}.md")
 
     previous = pointer_path.read_text(encoding="utf-8").strip() if pointer_path.exists() else ""
     pointer_path.parent.mkdir(parents=True, exist_ok=True)
@@ -180,7 +180,7 @@ def load_persona(
     resolved = version or resolve_latest_persona_version()
     path = ARTIFACTS_DIR / f"persona.{resolved}.md"
     if not path.exists():
-        raise FileNotFoundError(f"Persona artifact not found: {path}")
+        raise FileNotFoundError(f"Persona artifact not found: persona.{resolved}.md")
     base_text = sanitize_prompt_text(path.read_text(encoding="utf-8"))
     parts = [base_text.rstrip()]
     overlay_text = load_overlay(provider_kind)
@@ -200,6 +200,6 @@ def load_persona(
 def load_persona_from_path(path: str | Path) -> PersonaSpec:
     p = Path(path)
     if not p.exists():
-        raise FileNotFoundError(f"Persona artifact not found: {p}")
+        raise FileNotFoundError(f"Persona artifact not found: {p.name}")
     version = p.stem.split(".", 1)[-1] if "." in p.stem else "custom"
     return PersonaSpec(version=version, text=sanitize_prompt_text(p.read_text(encoding="utf-8")))

src/spark_character/prompt_guard.py

@@ -28,7 +28,7 @@
         re.compile(PROMPT_BOUNDARY_PREFIX + r"(system|developer)\s+(prompt|message|instruction)s?\b.*\b(override|replace|ignore)\b", re.I),
     ),
     ("hidden-html", re.compile(r"<!--|<\s*(?:div|span)[^>]*(?:display\s*:\s*none|visibility\s*:\s*hidden)", re.I)),
-    ("secret-exfiltration", re.compile(r"\b(curl|wget|fetch)\b.*\b(\.env|secret|token|api[_-]?key|password)\b", re.I)),
+    ("secret-exfiltration", re.compile(r"\b(curl|wget|fetch)\b.*(?<!\w)(\.env|secret|token|api[_-]?key|password)\b", re.I)),
     ("secret-file-request", re.compile(r"\b(read|open|print|cat|get-content)\b.*(\.env|secrets\.local\.json|id_rsa|\.ssh|api[_-]?key)\b", re.I)),
     ("private-key", re.compile(r"-----BEGIN [A-Z0-9 ]*PRIVATE KEY-----", re.I)),
 )

src/spark_character/registry.py

@@ -95,6 +95,8 @@ def promote_evolved_persona_to_chip_lab(
         try:
             base_spec = validate_chip_yaml_spec(yaml.safe_load(base_yaml_path.read_text(encoding="utf-8")) or {})
         except (OSError, ValueError, yaml.YAMLError) as exc:
+            # Keep the basename (not the full path) so the operator knows which
+            # chip file failed without leaking the surrounding directory layout.
             raise ValueError(
                 f"Base personality chip YAML is invalid: {base_yaml_path.name}"
             ) from exc