Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 3 additions & 0 deletions docs/evaluation-protocol.md
Original file line number Diff line number Diff line change
Expand Up @@ -169,6 +169,9 @@ The same gate runs after **real stdio MCP** calls when you pass
never call `local_*`. Accept / rewrite / reject then run on the scored
candidate. Security-sensitive delegated jobs first call `local_review`
and attach those notes to the premium packet; that still cannot approve.
`MCP_JOBS` includes `rename_exception_across_files` so the apply gate
sees a harder multi-file case; premium accept of an unproven local
result is blocked (`accept_unproven_local`).

The MCP server also refuses secret filenames, private-key / token
blobs, oversized file sets, and `max_tokens` above 4096 **before**
Expand Down
12 changes: 12 additions & 0 deletions src/local_coding_slm/eval/jobs.py
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,7 @@
from __future__ import annotations

from local_coding_slm.eval.cases import CASES_BY_ID
from local_coding_slm.eval.cases_harder import RENAME_EXCEPTION_PARTIAL
from local_coding_slm.eval.orchestrate import OrchestrationJob
from local_coding_slm.eval.review import accept, reject, rewrite
from local_coding_slm.eval.routing import RouteSignals, mechanical_signals
Expand Down Expand Up @@ -96,6 +97,17 @@ def clamp(value: int, lo: int, hi: int) -> int:
eval_case=CASES_BY_ID["implement_clamp"],
review=reject(notes="do not apply until bounds are reviewed"),
),
OrchestrationJob(
id="mcp_rename_exception_unproven",
signals=mechanical_signals(),
eval_case=CASES_BY_ID["rename_exception_across_files"],
local_replies=(
RENAME_EXCEPTION_PARTIAL,
RENAME_EXCEPTION_PARTIAL,
RENAME_EXCEPTION_PARTIAL,
),
review=accept(notes="do not apply unproven harder local"),
),
)


Expand Down
25 changes: 25 additions & 0 deletions tests/test_eval_mcp_orchestrate.py
Original file line number Diff line number Diff line change
Expand Up @@ -4,8 +4,10 @@

import unittest

from local_coding_slm.eval.cases_harder import HARDER_CASE_IDS
from local_coding_slm.eval.harness import run_campaign, run_orchestrated_campaign
from local_coding_slm.eval.jobs import MCP_JOBS
from local_coding_slm.eval.orchestrate import run_job
from local_coding_slm.eval.record import summarize


Expand Down Expand Up @@ -108,6 +110,29 @@ def test_job_table_covers_keep_and_delegate(self) -> None:
self.assertTrue(any(job.eval_case is None for job in MCP_JOBS))
self.assertTrue(any(job.eval_case is not None for job in MCP_JOBS))

def test_mcp_jobs_intersect_harder_cases(self) -> None:
case_ids = {
job.eval_case.id for job in MCP_JOBS if job.eval_case is not None
}
overlap = case_ids & set(HARDER_CASE_IDS)
self.assertIn("rename_exception_across_files", overlap)
self.assertTrue(overlap)

def test_harder_rename_exception_job_blocks_accept_unproven_local(self) -> None:
job = next(
item
for item in MCP_JOBS
if item.eval_case is not None
and item.eval_case.id == "rename_exception_across_files"
)
result = run_job(job)
self.assertEqual(result.case_id, "rename_exception_across_files")
self.assertTrue(result.delegated)
self.assertFalse(result.local_passed)
self.assertFalse(result.applied)
self.assertEqual(result.outcome, "blocked")
self.assertEqual(result.blocked, "accept_unproven_local")


if __name__ == "__main__":
unittest.main()
Loading