Release v0.6.4

CHANGELOG.md

@@ -6,6 +6,20 @@ The format is based on Keep a Changelog and this project follows Semantic Versio
 
 ## [Unreleased]
 
+## [0.6.4] - 2026-04-02
+
+### Fixed
+- MCP-backed task completion finalization now only performs a direct provider-side status update when the active profile permits `C2_remote_bounded` transactional effects, avoiding false task failures under the default `C1_local_write` profile while preserving explicit remote close-out for C2-capable runs.
+- Corrected MCP provider documentation to reflect the actual default profile behavior for direct provider-side remote writes.
+
+## [0.6.3] - 2026-04-02
+
+### Fixed
+- Single-task scope validation now accepts compacted tasklists that move the selected completed item into the `Completed` summary section, avoiding false scope violations after compaction.
+- Inner-loop `--continue` now preserves selected-task identity so resumed builder/reviewer work cannot drift to a different task after a mid-task halt.
+- Successful MCP-backed task runs now explicitly mark the selected remote task done after commit, instead of relying only on the builder prompt to do it.
+- Automatic compaction on tracked tasklists now finalizes its own tasklist rewrite immediately, preventing compaction-only changes from leaking into the next task's diff, review, or commit.
+
 ## [0.6.2] - 2026-03-31
 
 ### Fixed

docs/providers/mcp.md

@@ -144,7 +144,9 @@ Any MCP server your agent supports works — the `mcp_server` value is passed th
 
 ## Effect policy
 
-All MCP operations are classified as `EffectClass.transactional` (C2). Under the default `DEV_IMPLEMENTATION` profile these are permitted. If you have a stricter policy configured (e.g. `C1_LOCAL_WRITE` only), MCP operations will raise `CapabilityViolation` — update `.millstone/policy.toml` to allow `C2_REMOTE_BOUNDED` effects.
+Millstone-classified MCP write operations are `EffectClass.transactional` (`C2_remote_bounded`). The default `dev_implementation` profile is `C1_local_write`, so direct provider writes that millstone issues itself are not permitted unless you configure a `C2_remote_bounded` profile with `transactional` effects allowlisted.
+
+In the default profile, prompt-driven MCP actions performed by the coding agent can still run because the agent is executing those tool calls from the rendered prompt, not through millstone's direct provider-effect path. Orchestrator-driven follow-up writes, such as explicit post-commit remote task completion, are skipped unless the active profile allows those transactional effects.
 
 ---
 

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "millstone"
-version = "0.6.2"
+version = "0.6.4"
 description = "Orchestrator for agentic coding tools"
 readme = "README.md"
 requires-python = ">=3.10"

src/millstone/artifact_providers/file.py

@@ -676,8 +676,12 @@ def get_prompt_placeholders(self) -> dict[str, str]:
                 "`- [ ]` pending, `- [x]` complete. Select the FIRST unchecked task."
             ),
             "TASKLIST_COMPLETE_INSTRUCTIONS": (
-                f"Mark exactly this one task complete by changing its `- [ ]` to "
-                f"`- [x]` in `{self.path}` and stop. Do not modify any other tasks."
+                f"Mark exactly this one task complete in `{self.path}` and stop. "
+                "Normally, change its `- [ ]` to `- [x]`. "
+                "If the file has a `Completed` summary section instead of listing older "
+                "finished tasks individually, you may instead remove the selected task from "
+                "the pending list and update only that `Completed` summary. Do not modify "
+                "any other pending tasks."
             ),
             "TASKLIST_REWRITE_INSTRUCTIONS": (
                 f"Write the entire compacted content back to `{self.path}`, "

src/millstone/artifacts/tasklist.py

@@ -414,6 +414,10 @@ def _extract_tasks(self, content: str) -> list[tuple[bool, str]]:
             tasks.append((checked.lower() == "x", task_text.strip()))
         return tasks
 
+    def _has_completed_summary_section(self, content: str) -> bool:
+        """Return True when the tasklist already uses a compacted completed summary."""
+        return bool(re.search(r"^#{2,6}\s+Completed\b", content, re.MULTILINE | re.IGNORECASE))
+
     def validate_single_task_completion(
         self,
         original_content: str,
@@ -430,7 +434,25 @@ def validate_single_task_completion(
         if not original_tasks or not new_tasks:
             return True, ""
 
+        first_unchecked = next(
+            (i for i, (checked, _) in enumerate(original_tasks) if not checked),
+            None,
+        )
+
         if len(new_tasks) < len(original_tasks):
+            if (
+                first_unchecked is not None
+                and len(new_tasks) == len(original_tasks) - 1
+                and (
+                    self._has_completed_summary_section(original_content)
+                    or self._has_completed_summary_section(new_content)
+                )
+            ):
+                expected_tasks = (
+                    original_tasks[:first_unchecked] + original_tasks[first_unchecked + 1 :]
+                )
+                if new_tasks == expected_tasks:
+                    return True, ""
             return (
                 False,
                 f"Task count decreased: {len(original_tasks)} -> {len(new_tasks)}. Deleting tasks is not allowed.",
@@ -457,10 +479,6 @@ def validate_single_task_completion(
                 f"Multiple tasks were marked complete ({len(checkoffs)}).",
             )
 
-        first_unchecked = next(
-            (i for i, (checked, _) in enumerate(original_tasks) if not checked),
-            None,
-        )
         if first_unchecked is None:
             return False, "No unchecked tasks remained, but a task was marked complete."