gitea-admin/CASAN
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix: strengthen goal patch repair routing

Browse Source
This commit is contained in:
thanhnv
2026-07-18 11:36:03 +07:00
parent f23fb98953
commit 4f8a53bda9
5 changed files with 43 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
packages/casan-harness/scripts/bash/goal-orchestrator.py
+23 -13
View File
@@ -299,10 +299,10 @@ def validate_write_output(text: str) -> str:
def patch_repair_attempts() -> int:
"""Return the bounded number of chances to repair an invalid model diff."""
try:
configured = int(os.environ.get("CASAN_GOAL_PATCH_REPAIR_ATTEMPTS", "1"))
configured = int(os.environ.get("CASAN_GOAL_PATCH_REPAIR_ATTEMPTS", "2"))
except ValueError:
configured = 1
return min(max(configured, 0), 1)
configured = 2
return min(max(configured, 0), 3)
def patch_output_tokens() -> int:
@@ -354,14 +354,23 @@ def repair_write_output(model: str, original_prompt: str, invalid_output: str, c
"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:")), max_output_tokens=patch_output_tokens())
metadata = dict(metadata)
metadata["repair_model"] = candidate
if not ok:
return False, "", metadata, f"goal_patch_repair_failed:{reason}"
return True, output, metadata, "ok"
candidates = patch_repair_models(model)[:patch_repair_attempts()]
last_metadata, last_reason = {}, "goal_patch_missing"
for candidate in candidates:
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
last_metadata = metadata
if not ok:
last_reason = f"goal_patch_repair_failed:{reason}"
continue
try:
validate_write_output(output)
except ValueError as error:
last_reason = str(error)
continue
return True, output, metadata, "ok"
return False, "", last_metadata, f"goal_patch_repair_invalid:{last_reason}"
def merge_usage(primary: dict, additional: dict) -> dict:
@@ -755,10 +764,11 @@ def run(job_path: str) -> int:
try:
validate_write_output(safe_local)
except ValueError as first_error:
repair_model = patch_repair_models(local_model)[0] if patch_repair_models(local_model) else local_model
repair_candidates = patch_repair_models(local_model)[:patch_repair_attempts()]
repair_model = repair_candidates[0] if repair_candidates else local_model
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})
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, "repair_candidates": repair_candidates})
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:
packages/casan-harness/tests/goal-patch-workflow-tests.py
+17 -1
View File
@@ -46,7 +46,8 @@ class GoalPatchWorkflowTests(unittest.TestCase):
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), \
patch.object(ORCHESTRATOR, "call_model", return_value=(True, repaired, {"output_tokens": 9}, "ok")) as call:
patch.object(ORCHESTRATOR, "call_model", return_value=(True, repaired, {"output_tokens": 9}, "ok")) as call, \
patch.object(ORCHESTRATOR, "validate_write_output", return_value=repaired):
ok, value, usage, reason = ORCHESTRATOR.repair_write_output("ollama:test", "original prompt", "implementation plan")
self.assertTrue(ok)
self.assertEqual(reason, "ok")
@@ -67,6 +68,7 @@ class GoalPatchWorkflowTests(unittest.TestCase):
with (
patch.dict(os.environ, {"CASAN_GOAL_PATCH_REPAIR_ATTEMPTS": "1", "CASAN_GOAL_PATCH_REPAIR_MODELS": "openai:gpt-4o-mini,anthropic:claude,openai-compatible:auto/coding,ollama:ornith"}, clear=False),
patch.object(ORCHESTRATOR, "call_model", return_value=(True, repaired, {}, "ok")) as call,
patch.object(ORCHESTRATOR, "validate_write_output", return_value=repaired),
):
ok, _, usage, reason = ORCHESTRATOR.repair_write_output("ollama:ornith", "original", "plan")
self.assertTrue(ok)
@@ -75,6 +77,20 @@ class GoalPatchWorkflowTests(unittest.TestCase):
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_repair_tries_next_stronger_candidate_when_first_patch_is_corrupt(self):
repaired = "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+b\n"
with (
patch.dict(os.environ, {"CASAN_GOAL_PATCH_REPAIR_ATTEMPTS": "2", "CASAN_GOAL_PATCH_REPAIR_MODELS": "openai:gpt-4.1,openai-compatible:aug/claude-sonnet"}, clear=False),
patch.object(ORCHESTRATOR, "call_model", side_effect=[(True, "corrupt", {}, "ok"), (True, repaired, {}, "ok")]) as call,
patch.object(ORCHESTRATOR, "validate_write_output", side_effect=[ValueError("goal_patch_check_failed:corrupt"), repaired]),
):
ok, output, usage, reason = ORCHESTRATOR.repair_write_output("ollama:ornith", "original", "invalid")
self.assertTrue(ok)
self.assertEqual(output, repaired)
self.assertEqual(reason, "ok")
self.assertEqual(usage["repair_model"], "openai-compatible:aug/claude-sonnet")
self.assertEqual([item.args[0] for item in call.call_args_list], ["openai:gpt-4.1", "openai-compatible:aug/claude-sonnet"])
def test_patch_outside_workspace_is_denied(self):
job = {"workspace": {"context_roots": ["apps/okr/frontend"]}}
content = "diff --git a/package.json b/package.json\n--- a/package.json\n+++ b/package.json\n@@ -1 +1 @@\n-a\n+b\n"