fix: validate and size governed patch outputs

packages/casan-harness/scripts/bash/goal-orchestrator.py

@@ -267,6 +267,23 @@ def extract_patch(text: str) -> str:
     return patch
 
 
+def validate_patch_check(patch: str) -> None:
+    """Require a candidate diff to apply before it can pass H2 or H3."""
+    with tempfile.NamedTemporaryFile("w", encoding="utf-8", suffix=".patch") as handle:
+        handle.write(patch)
+        handle.flush()
+        try:
+            check = subprocess.run(
+                ["git", "apply", "--check", "--whitespace=error", handle.name],
+                cwd=ROOT, capture_output=True, text=True, timeout=30,
+            )
+        except FileNotFoundError as error:
+            raise ValueError("goal_patch_validator_unavailable") from error
+    if check.returncode != 0:
+        detail = (check.stderr or check.stdout or "git_apply_check_failed").strip().replace("\n", " ")[:160]
+        raise ValueError(f"goal_patch_check_failed:{detail}")
+
+
 def validate_write_output(text: str) -> str:
     """Fail at the producing harness when a write-intent reply is not a diff.
 
@@ -274,7 +291,9 @@ def validate_write_output(text: str) -> str:
     reviewer has completed, while ensuring H2/H3 accurately identify a model
     that violated the required output contract.
     """
-    return extract_patch(text)
+    patch = extract_patch(text)
+    validate_patch_check(patch)
+    return patch
 
 
 def patch_repair_attempts() -> int:
@@ -286,6 +305,15 @@ def patch_repair_attempts() -> int:
     return min(max(configured, 0), 1)
 
 
+def patch_output_tokens() -> int:
+    """Keep a multi-hunk patch from being cut at the generic chat limit."""
+    try:
+        configured = int(os.environ.get("CASAN_GOAL_PATCH_MAX_OUTPUT_TOKENS", "6000"))
+    except ValueError:
+        configured = 6000
+    return min(max(configured, 512), 8192)
+
+
 def patch_repair_models(primary_model: str):
     """Use direct cloud -> gateway -> local order for one bounded H2 recovery."""
     configured = [item.strip() for item in os.environ.get("CASAN_GOAL_PATCH_REPAIR_MODELS", "").split(",") if item.strip()]
@@ -309,7 +337,7 @@ def provider_for_model(model: str) -> str:
     return "model"
 
 
-def repair_write_output(model: str, original_prompt: str, invalid_output: str):
+def repair_write_output(model: str, original_prompt: str, invalid_output: str, contract_error="goal_patch_missing"):
     """Ask the producing model to repair format only, without relaxing H2.
 
     A write-intent goal may never advance to approval without a checked unified
@@ -321,14 +349,14 @@ def repair_write_output(model: str, original_prompt: str, invalid_output: str):
     if patch_repair_attempts() == 0:
         return False, "", {}, "goal_patch_missing"
     repair_prompt = (
-        "Your previous response violated the required write-output contract because it did not contain a complete unified git diff. "
+        f"Your previous response violated the required write-output contract: {contract_error}. "
-        "Return ONLY one complete patch inside a ```diff fence. The first non-empty line inside the fence MUST be `diff --git `. "
+        "Return ONLY one complete, applicable unified git diff inside a ```diff fence. The first non-empty line inside the fence MUST be `diff --git `. "
-        "Do not explain, plan, summarize, use placeholders, or perform side effects. Preserve the original objective and workspace restrictions.\n\n"
+        "Every hunk must be complete and the patch must pass `git apply --check`. Do not emit `index` lines, placeholder hashes, commentary, plans, summaries, or side effects. Preserve the original objective and workspace restrictions.\n\n"
         f"ORIGINAL CONTRACT:\n{original_prompt}\n\nPREVIOUS INVALID RESPONSE:\n{invalid_output[:8000]}"
     )
     candidates = patch_repair_models(model)
     candidate = candidates[0] if candidates else model
-    ok, output, metadata, reason = call_model(candidate, repair_prompt, candidate.startswith(("openai:", "anthropic:", "openai-compatible:")))
+    ok, output, metadata, reason = call_model(candidate, repair_prompt, candidate.startswith(("openai:", "anthropic:", "openai-compatible:")), max_output_tokens=patch_output_tokens())
     metadata = dict(metadata)
     metadata["repair_model"] = candidate
     if not ok:
@@ -438,7 +466,7 @@ def scan(text: str, mode: str):
         return result.returncode == 0, safe
 
 
-def call_model(model: str, prompt: str, cloud: bool, timeout_seconds=None):
+def call_model(model: str, prompt: str, cloud: bool, timeout_seconds=None, max_output_tokens=None):
     if not model:
         return False, "", {}, "model_unconfigured"
     with tempfile.TemporaryDirectory() as directory:
@@ -452,7 +480,10 @@ def call_model(model: str, prompt: str, cloud: bool, timeout_seconds=None):
         # ingest the brief and produce a bounded plan. The outer goal timeout
         # remains the hard ceiling; this only raises the router's 60s default.
         environment.setdefault("CASAN_MODEL_TIMEOUT_SEC", os.environ.get("CASAN_GOAL_LOCAL_TIMEOUT_SEC", "240") if not cloud else "120")
-        environment.setdefault("CASAN_MODEL_GENERATE_MAX_TOKENS", os.environ.get("CASAN_GOAL_MAX_OUTPUT_TOKENS", "1400"))
+        configured_output = os.environ.get("CASAN_GOAL_MAX_OUTPUT_TOKENS", "1400")
+        if max_output_tokens is not None:
+            configured_output = str(max_output_tokens)
+        environment["CASAN_MODEL_GENERATE_MAX_TOKENS"] = configured_output
         timeout = max(1, int(timeout_seconds or os.environ.get("CASAN_GOAL_MODEL_TIMEOUT", "300")))
         environment["CASAN_MODEL_TIMEOUT_SEC"] = str(min(
             int(environment.get("CASAN_MODEL_TIMEOUT_SEC", timeout)), timeout
@@ -728,7 +759,7 @@ def run(job_path: str) -> int:
                 repair_provider = provider_for_model(repair_model)
                 stage(job_path, "local-worker", "running", "Repairing invalid patch output contract", repair_provider, repair_model)
                 emit(goal_id, "H2-tool", "running", "Worker is repairing patch output contract", {"reason": str(first_error), "max_attempts": patch_repair_attempts(), "repair_provider": repair_provider, "repair_model": repair_model})
-                repaired, repaired_output, repaired_meta, repair_reason = repair_write_output(local_model, local_prompt, safe_local)
+                repaired, repaired_output, repaired_meta, repair_reason = repair_write_output(local_model, local_prompt, safe_local, str(first_error))
                 local_meta = merge_usage(local_meta, repaired_meta)
                 if not repaired:
                     reason = repair_reason
@@ -754,7 +785,7 @@ def run(job_path: str) -> int:
         stage(job_path, "cloud-reviewer", "running", "Independent reviewer is challenging and improving the local solution", job.get("cloud_provider", ""), cloud_model)
         emit(goal_id, "H3-eval", "running", "Cloud reviewer is evaluating the local solution", {"provider": job.get("cloud_provider", ""), "model": cloud_model})
         reviewer_contract = (
-            "Return ONLY the corrected, complete unified git diff inside a ```diff fence. Preserve valid implementation details, fix defects, and include no prose outside the diff. "
+            "Return ONLY the corrected, complete unified git diff inside a ```diff fence. Preserve valid implementation details, fix defects, and include no prose or `index` lines outside the diff. Every hunk must be complete and pass `git apply --check`. "
             if write_intent else
             "Return one final actionable solution with ordered steps and acceptance checks. Respond in the same language as the objective. "
         )

packages/casan-harness/scripts/bash/model-call.py

@@ -56,6 +56,17 @@ PROVIDER_LOG = os.path.join(REPO_ROOT, ".specify/logs/level5/provider-usage.json
 REQUEST_TIMEOUT = max(1, int(os.environ.get("CASAN_MODEL_TIMEOUT_SEC", "60")))
 
 
+def generation_max_tokens(role):
+    """Return a bounded generation budget shared by every model provider."""
+    if role in ("classify", "judge"):
+        return 16
+    try:
+        configured = int(os.environ.get("CASAN_MODEL_GENERATE_MAX_TOKENS", "1400"))
+    except ValueError:
+        configured = 1400
+    return min(max(configured, 64), 8192)
+
+
 def enforce_call_budget():
     """Fail closed when a run exceeds CASAN_MODEL_MAX_CALLS. The count is tracked in
     CASAN_MODEL_CALL_COUNTER_FILE so it spans the many per-step model-call processes
@@ -205,7 +216,7 @@ def call_openai(model_name, prompt, role):
         "model": model_name,
         "messages": [{"role": "user", "content": prompt}],
         "temperature": 0 if role in ("classify", "judge") else 0.2,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
     }
     data = json.dumps(body).encode()
     req = urllib.request.Request(
@@ -272,7 +283,7 @@ def call_openai_compatible(model_name, prompt, role):
         "model": model_name,
         "messages": [{"role": "user", "content": prompt}],
         "temperature": 0 if role in ("classify", "judge") else 0.2,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
     }
     data = json.dumps(body).encode()
     req = urllib.request.Request(
@@ -335,7 +346,7 @@ def call_anthropic(model_name, prompt, role):
     url = f"https://{host}/v1/messages"
     body = {
         "model": model_name,
-        "max_tokens": 16 if role in ("classify", "judge") else 512,
+        "max_tokens": generation_max_tokens(role),
         "messages": [{"role": "user", "content": prompt}],
     }
     data = json.dumps(body).encode()

packages/casan-harness/tests/goal-patch-workflow-tests.py

@@ -38,6 +38,11 @@ class GoalPatchWorkflowTests(unittest.TestCase):
         value = ORCHESTRATOR.extract_patch("```diff\ndiff --git a/apps/okr/frontend/a.ts b/apps/okr/frontend/a.ts\n--- a/apps/okr/frontend/a.ts\n+++ b/apps/okr/frontend/a.ts\n@@ -1 +1 @@\n-a\n+b\n```")
         self.assertTrue(value.startswith("diff --git"))
 
+    def test_write_contract_rejects_a_truncated_diff_before_h3(self):
+        truncated = "diff --git a/apps/okr/frontend/a.ts b/apps/okr/frontend/a.ts\n--- a/apps/okr/frontend/a.ts\n+++ b/apps/okr/frontend/a.ts\n@@ -1 +1 @@\n-a\n"
+        with self.assertRaisesRegex(ValueError, "goal_patch_check_failed"):
+            ORCHESTRATOR.validate_write_output(truncated)
+
     def test_invalid_write_output_gets_one_bounded_repair_attempt(self):
         repaired = "```diff\ndiff --git a/apps/okr/frontend/a.ts b/apps/okr/frontend/a.ts\n--- a/apps/okr/frontend/a.ts\n+++ b/apps/okr/frontend/a.ts\n@@ -1 +1 @@\n-a\n+b\n```"
         with patch.dict(os.environ, {"CASAN_GOAL_PATCH_REPAIR_ATTEMPTS": "1"}, clear=False), \
@@ -68,6 +73,7 @@ class GoalPatchWorkflowTests(unittest.TestCase):
         self.assertEqual(reason, "ok")
         self.assertEqual(usage["repair_model"], "openai:gpt-4o-mini")
         self.assertEqual(call.call_args.args[0], "openai:gpt-4o-mini")
+        self.assertEqual(call.call_args.kwargs["max_output_tokens"], ORCHESTRATOR.patch_output_tokens())
 
     def test_patch_outside_workspace_is_denied(self):
         job = {"workspace": {"context_roots": ["apps/okr/frontend"]}}