@heretek-ai/epistemic-swarm 0.4.7 → 0.4.8

package/runner/refinement.py

@@ -18,6 +18,7 @@ V(c_i) has never been what the runtime computes. Tier weighting becomes an
 opt-in constitution — nothing flips without an explicit pack.
 """
 
+import math
 from dataclasses import dataclass, field
 from typing import Any, Dict, Iterable, List, Optional, Tuple
 
@@ -154,6 +155,59 @@ def claim_verdict(
     }
 
 
+def _check_verified_and_inferred(c: ClaimWitness, hasher: Any, out: List[Violation]) -> None:
+    if c.tag == TAG_VERIFIED or c.kind == KIND_CLAIM:
+        if not c.source_hash:
+            out.append(Violation(c.claim_id, "VERIFIED_REQUIRES_HASH", "source_hash missing"))
+        if not c.verbatim_quote:
+            out.append(Violation(c.claim_id, "VERIFIED_REQUIRES_QUOTE", "verbatim_quote missing"))
+        if hasher is not None and c.source_hash and c.verbatim_quote:
+            passed, _conf, msg = witness_check(c, hasher)
+            if not passed:
+                out.append(Violation(c.claim_id, "WITNESS_CHECK_FAILED", msg or "quote not in source"))
+
+    if c.tag == TAG_INFERRED or c.kind == KIND_INFERENCE:
+        if not c.parent_claims:
+            out.append(Violation(c.claim_id, "INFERRED_REQUIRES_PARENTS", "parent_claims empty"))
+        if not c.deductive_logic:
+            out.append(Violation(c.claim_id, "INFERRED_REQUIRES_LOGIC", "deductive_logic missing"))
+
+
+def _check_hypotheses_and_neg_knowledge(c: ClaimWitness, out: List[Violation]) -> None:
+    if (c.tag == TAG_HYPOTHESIS or c.kind == KIND_HYPOTHESIS) and not c.falsification:
+        out.append(Violation(c.claim_id, "HYPOTHESIS_REQUIRES_FALSIFICATION", "falsification missing"))
+
+    if c.tag == TAG_NEGATIVE_KNOWLEDGE or c.kind == KIND_NEGATIVE_KNOWLEDGE:
+        if not c.query:
+            out.append(Violation(c.claim_id, "NEG_KNOWLEDGE_REQUIRES_QUERY", "query missing"))
+        if not c.finding:
+            out.append(Violation(c.claim_id, "NEG_KNOWLEDGE_REQUIRES_FINDING", "finding missing"))
+
+
+def _check_constitution_and_parents(
+    c: ClaimWitness,
+    known_ids: set,
+    constitution: Constitution,
+    out: List[Violation],
+) -> None:
+    for parent in c.parent_claims:
+        if parent not in known_ids:
+            out.append(
+                Violation(c.claim_id, "PARENT_UNRESOLVED", f"parent_claims '{parent}' not in dossier set")
+            )
+
+    if constitution.mandatory_tags and not c.tag:
+        out.append(Violation(c.claim_id, "MISSING_TAG", "tag required by constitution"))
+
+    if constitution.banned_domains and c.source_url:
+        src_lower = c.source_url.lower()
+        for dom in constitution.banned_domains:
+            if dom.lower() in src_lower:
+                out.append(
+                    Violation(c.claim_id, "BANNED_DOMAIN", f"source_url on banned domain '{dom}'")
+                )
+
+
 def check_invariants(
     claims: Iterable[ClaimWitness],
     hasher: Any = None,
@@ -165,54 +219,14 @@ def check_invariants(
     VERIFIED claims; without it, only structural rules are enforced.
     """
     constitution = constitution or LEGACY_CONSTITUTION
-    claims = list(claims)
-    known_ids = {c.claim_id for c in claims}
+    claim_list = list(claims)
+    known_ids = {c.claim_id for c in claim_list}
     out: List[Violation] = []
 
-    for c in claims:
-        # --- tag/kind-specific structural rules ---
-        if c.tag == TAG_VERIFIED or c.kind == KIND_CLAIM:
-            if not c.source_hash:
-                out.append(Violation(c.claim_id, "VERIFIED_REQUIRES_HASH", "source_hash missing"))
-            if not c.verbatim_quote:
-                out.append(Violation(c.claim_id, "VERIFIED_REQUIRES_QUOTE", "verbatim_quote missing"))
-            if hasher is not None and c.source_hash and c.verbatim_quote:
-                passed, _conf, msg = witness_check(c, hasher)
-                if not passed:
-                    out.append(Violation(c.claim_id, "WITNESS_CHECK_FAILED", msg or "quote not in source"))
-
-        if c.tag == TAG_INFERRED or c.kind == KIND_INFERENCE:
-            if not c.parent_claims:
-                out.append(Violation(c.claim_id, "INFERRED_REQUIRES_PARENTS", "parent_claims empty"))
-            if not c.deductive_logic:
-                out.append(Violation(c.claim_id, "INFERRED_REQUIRES_LOGIC", "deductive_logic missing"))
-
-        if c.tag == TAG_HYPOTHESIS or c.kind == KIND_HYPOTHESIS:
-            if not c.falsification:
-                out.append(Violation(c.claim_id, "HYPOTHESIS_REQUIRES_FALSIFICATION", "falsification missing"))
-
-        if c.tag == TAG_NEGATIVE_KNOWLEDGE or c.kind == KIND_NEGATIVE_KNOWLEDGE:
-            if not c.query:
-                out.append(Violation(c.claim_id, "NEG_KNOWLEDGE_REQUIRES_QUERY", "query missing"))
-            if not c.finding:
-                out.append(Violation(c.claim_id, "NEG_KNOWLEDGE_REQUIRES_FINDING", "finding missing"))
-
-        # --- cross-reference resolution ---
-        for parent in c.parent_claims:
-            if parent not in known_ids:
-                out.append(
-                    Violation(c.claim_id, "PARENT_UNRESOLVED", f"parent_claims '{parent}' not in dossier set")
-                )
-
-        # --- domain-pack rules ---
-        if constitution.mandatory_tags and not c.tag:
-            out.append(Violation(c.claim_id, "MISSING_TAG", "tag required by constitution"))
-        if constitution.banned_domains and c.source_url:
-            for dom in constitution.banned_domains:
-                if dom.lower() in c.source_url.lower():
-                    out.append(
-                        Violation(c.claim_id, "BANNED_DOMAIN", f"source_url on banned domain '{dom}'")
-                    )
+    for c in claim_list:
+        _check_verified_and_inferred(c, hasher, out)
+        _check_hypotheses_and_neg_knowledge(c, out)
+        _check_constitution_and_parents(c, known_ids, constitution, out)
 
     return out
 
@@ -294,7 +308,7 @@ def compute_epistemic_score_from_claims(
     }
     weight_sum = 0.0
     tier_weighted = any(
-        w != 1.0 for w in constitution.tier_weights.values()
+        not math.isclose(w, 1.0, rel_tol=1e-7) for w in constitution.tier_weights.values()
     )
 
     for c in claims:

package/runner/research_swarm.py

@@ -26,6 +26,47 @@ from runner.auditor_engine import EpistemicAuditorEngine
 from skills.research_cache.hasher import SourceHasher
 from skills.swarm_config.configure import load_config
 
+EXAMPLE_RUST_RAFT_URL = "https://github.com/example/rust-raft"
+# Common backends we can resolve without touching PATH. This is a convenience
+# fast-path, NOT an allowlist: backend resolution is a documented extension
+# point (CLI > IUMBTEMS_BACKEND_* > config), so any executable the OS can
+# resolve must work. The security property is argv-list + shell=False, which
+# run_claude_process already enforces; injection is not possible here.
+KNOWN_BACKEND_BINARIES = {"claude", "opencode", "python", "python3", "node"}
+
+
+def _validate_backend(backend: List[str]) -> List[str]:
+    """Resolve-check a backend argv. Raises ValueError only if nothing can run it.
+
+    Resolution mirrors what subprocess does with shell=False: an absolute/relative
+    path is taken as-is (checked against the spawn cwd, PROJECT_ROOT), otherwise
+    the name is looked up on PATH. Previously this rejected anything outside a
+    five-name allowlist, which broke `python3.11`, `aider`, `codex exec`,
+    `npx -y ...` and any config-supplied backend.
+    """
+    if not backend:
+        return ["claude", "-p"]
+    candidate = backend[0]
+    bin_name = Path(candidate).name
+    if bin_name in KNOWN_BACKEND_BINARIES:
+        return list(backend)
+    # Path form: resolve the same way the spawn will (cwd=PROJECT_ROOT).
+    if os.sep in candidate or (os.altsep and os.altsep in candidate):
+        resolved = (PROJECT_ROOT / candidate).resolve() if not os.path.isabs(candidate) else Path(candidate)
+        if resolved.is_file():
+            return list(backend)
+        raise ValueError(
+            f"Backend binary not found at {candidate!r} (resolved {resolved}); "
+            f"expected an executable file relative to {PROJECT_ROOT}"
+        )
+    # Bare name: ask PATH, exactly as subprocess will.
+    if shutil.which(candidate) is None:
+        raise ValueError(
+            f"Backend binary {candidate!r} is not on PATH and is not a known backend "
+            f"({', '.join(sorted(KNOWN_BACKEND_BINARIES))})"
+        )
+    return list(backend)
+
 
 class SwarmRunner:
     def __init__(
@@ -115,13 +156,26 @@ class SwarmRunner:
         cmd = self.build_agent_cmd(prompt, system_prompt_file, tools, role=role)
 
         try:
+            if cmd:
+                _validate_backend([cmd[0]])
             res = subprocess.run(
-                cmd, capture_output=True, text=True, check=True, cwd=str(PROJECT_ROOT)
+                cmd,
+                capture_output=True,
+                text=True,
+                check=True,
+                cwd=str(PROJECT_ROOT),
+                shell=False,
             )
             return res.stdout.strip()
         except subprocess.CalledProcessError as e:
             print(f"[ERROR] Claude process failed: {e.stderr}", file=sys.stderr)
             raise RuntimeError(f"Claude execution failed: {e.stderr}")
+        except ValueError as e:
+            # Bad backend config: report it as a run failure, not an abort of the
+            # whole swarm. Matches the pre-allowlist behavior of surfacing the
+            # error from inside the try.
+            print(f"[ERROR] Invalid agent backend: {e}", file=sys.stderr)
+            raise RuntimeError(f"Invalid agent backend: {e}")
 
     def _mock_claude_response(self, prompt: str) -> str:
         """Mock response generator for unit testing without live API keys."""
@@ -243,7 +297,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
             elif self.mode == "scout":
                 sample_content = "# High Performance Raft in Rust\nZero-dependency Raft implementation with 150k ops/sec throughput under Apache-2.0."
                 shash = self.hasher.store_source(
-                    "https://github.com/example/rust-raft", sample_content, "Rust Raft"
+                    EXAMPLE_RUST_RAFT_URL, sample_content, "Rust Raft"
                 )
                 dossier = {
                     "agent": "Agent Alpha (OSS Scout)",
@@ -256,7 +310,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
                             "tag": "VERIFIED",
                             "statement": "Rust-Raft achieves 150k ops/sec with zero external dependencies.",
                             "source_hash": shash,
-                            "source_url": "https://github.com/example/rust-raft",
+                            "source_url": EXAMPLE_RUST_RAFT_URL,
                             "verbatim_quote": "Zero-dependency Raft implementation with 150k ops/sec throughput under Apache-2.0.",
                         }
                     ],
@@ -390,7 +444,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
             elif self.mode == "scout":
                 sample_content = "# High Performance Raft in Rust\nZero-dependency Raft implementation with 150k ops/sec throughput under Apache-2.0."
                 shash = self.hasher.store_source(
-                    "https://github.com/example/rust-raft", sample_content, "Rust Raft"
+                    EXAMPLE_RUST_RAFT_URL, sample_content, "Rust Raft"
                 )
                 dossier = {
                     "agent": "Agent Beta (OSS Red Team)",
@@ -403,7 +457,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
                             "tag": "VERIFIED",
                             "statement": "Candidate repository has single maintainer with 9-month lull in commit history.",
                             "source_hash": shash,
-                            "source_url": "https://github.com/example/rust-raft",
+                            "source_url": EXAMPLE_RUST_RAFT_URL,
                             "verbatim_quote": "Zero-dependency Raft implementation with 150k ops/sec throughput under Apache-2.0.",
                             "severity": "LOW",
                         }
@@ -550,7 +604,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
         print("=" * 70)
 
         # 1. Orchestrate
-        scopes = self.orchestrate_objective(objective, frontier_file)
+        self.orchestrate_objective(objective, frontier_file)
 
         # 2. Execute scopes according to DAG
         while True:
@@ -659,7 +713,7 @@ Output ONLY valid JSON representing the scope decomposition conforming to prompt
                     synthesis_lines.append("\n---\n")
 
         avg_div = round(sum(all_divergences) / max(1, len(all_divergences)), 2)
-        synthesis_lines.append(f"\n## Swarm Epistemic Audit Totals\n")
+        synthesis_lines.append("\n## Swarm Epistemic Audit Totals\n")
         synthesis_lines.append(
             f"- **Total Verified Primary Citations**: `{total_verified}`"
         )

package/runner/state_machine.py

@@ -34,12 +34,15 @@ class ScopeStatus(str, Enum):
     COMPLETE = "COMPLETE"
     FAILED = "FAILED"
 
+MANIFEST_FILENAME = "manifest.json"
+
+
 class ResearchStateMachine:
     def __init__(self, base_dir: Optional[Path] = None):
-        self.base_dir = base_dir or Path(".research")
+        self.base_dir = Path(os.path.realpath(str(base_dir or ".research")))
         self.scratchpads_dir = self.base_dir / "scratchpads"
         self.sources_dir = self.base_dir / "sources"
-        self.manifest_file = self.base_dir / "manifest.json"
+        self.manifest_file = self.base_dir / MANIFEST_FILENAME
         self._lock = threading.Lock()
         
         # Ensure directories exist
@@ -112,17 +115,17 @@ class ResearchStateMachine:
                 "audit_completed": False,
                 "created_at": datetime.now(timezone.utc).isoformat()
             }
-            temp_scope_file = (scope_dir / "manifest.json").with_suffix(".tmp")
+            temp_scope_file = (scope_dir / MANIFEST_FILENAME).with_suffix(".tmp")
             with open(temp_scope_file, "w", encoding="utf-8") as f:
                 json.dump(scope_manifest, f, indent=2)
-            os.replace(temp_scope_file, scope_dir / "manifest.json")
+            os.replace(temp_scope_file, scope_dir / MANIFEST_FILENAME)
 
     def get_scope_dir(self, scope_id: str) -> Path:
         return self.scratchpads_dir / scope_id
 
     def load_scope_manifest(self, scope_id: str) -> Dict[str, Any]:
         with self._lock:
-            scope_manifest_file = self.get_scope_dir(scope_id) / "manifest.json"
+            scope_manifest_file = self.get_scope_dir(scope_id) / MANIFEST_FILENAME
             if not scope_manifest_file.exists():
                 raise FileNotFoundError(f"Scope manifest not found for {scope_id}")
             with open(scope_manifest_file, "r", encoding="utf-8") as f:
@@ -130,7 +133,7 @@ class ResearchStateMachine:
 
     def save_scope_manifest(self, scope_id: str, manifest: Dict[str, Any]):
         with self._lock:
-            scope_manifest_file = self.get_scope_dir(scope_id) / "manifest.json"
+            scope_manifest_file = self.get_scope_dir(scope_id) / MANIFEST_FILENAME
             temp_path = scope_manifest_file.with_suffix(".tmp")
             with open(temp_path, "w", encoding="utf-8") as f:
                 json.dump(manifest, f, indent=2)

package/runner/tests/test_mcp_server.py

@@ -97,7 +97,6 @@ class TestMcpServerTools(unittest.TestCase):
         with tempfile.TemporaryDirectory() as tmp:
             tmp_path = Path(tmp)
             # Seed a source so verify_quote has something real to check.
-            sys.path.insert(0, str(PROJECT_ROOT))
             from skills.research_cache.hasher import SourceHasher
 
             hasher = SourceHasher(tmp_path)