Introduce a lightweight checkpoint/rollback system for sandbox state
during plan execute and apply flows. CheckpointManager snapshots
the sandbox working directory before each phase and can restore it
on failure, giving the execution engine a reliable undo mechanism.
Key changes:
- SandboxCheckpoint model, Checkpointable protocol, and
CheckpointManager in infrastructure/sandbox/checkpoint.py
- PlanExecutor gains optional checkpoint_manager with pre/post
execute hooks and automatic rollback on failure
- PlanApplyService gains optional checkpoint_manager with pre-apply
checkpoint and rollback helper
- 12 BDD scenarios (features/sandbox_checkpoints.feature)
- 5 Robot Framework smoke tests (robot/sandbox_checkpoint_smoke.robot)
- ASV benchmarks for creation, rollback, and listing operations
- Reference documentation in docs/reference/sandbox.md
ISSUES CLOSED: #183
- Add CHANGELOG.md entry for M6 autonomy acceptance suite
- Split m6_autonomy_acceptance_steps.py (774 lines) into
m6_facade_steps.py (398 lines) and m6_guardrails_steps.py (399 lines)
to comply with the project's 500-line guideline
Closes#211
Fix 4 errored scenarios: patch get_container instead of the lazily-
imported ContextService (which is not a module-level attribute of
cleveragents.cli.commands.context).
Fix project context show (line 112): mock _get_namespaced_project_repo
to return a repo (not a tuple), mock get_container for session_factory,
and mock _read_policy to bypass DB access.
Fix context exclusion pattern (line 152): replace fnmatch.fnmatch with
PurePosixPath.match for proper ** glob handling and use a path with a
leading directory segment.
Also fix context list empty assertion to match actual CLI output
('No files in context' rather than 'No context'), and patch
Path.exists for the add command so the mock service is reached.
Add Alembic migration m4_002_skill_flattened_tools to extend the skills
table with five new columns: flattened_tools_json, includes_json,
capability_summary_json, yaml_text, and flattening_hash (SHA-256). A
defence-in-depth uniqueness constraint (uq_skills_name) is also added.
Update SkillModel with the new column definitions and extend
SkillRepository with update_flattened_tools(), get_flattened_tools(),
needs_refresh(), recompute_flattening_hash(), and
invalidate_cached_summaries() methods. The existing update() method
now nulls all cached fields on mutation (hash-based invalidation).
All new repository methods follow the session-factory pattern with
@database_retry and flush-but-don-t-commit semantics. Structured
logging via structlog records cache updates and invalidations.
Database schema docs updated with the new skills table columns and a
persistence-field-to-domain-model mapping table.
Tests:
- 6 Behave scenarios covering create, invalidation, hash staleness,
refresh recomputation, uniqueness constraint, and namespace filtering
- 2 Robot Framework smoke tests (round-trip and invalidation)
- 3 ASV benchmarks (persist, refresh check, namespace list)
ISSUES CLOSED: #166
Remove FakeListLLM as a silent fallback in agent graph constructors
(plan_generation.py, context_analysis.py, auto_debug.py). All three now
raise ValueError when llm=None, making missing-provider errors explicit.
Add Settings.mock_providers flag and validate_provider_availability()
method. Update container.get_ai_provider() to check Settings.mock_providers
first, with env-var fallback for backward compatibility.
Add resolve_provider_by_name() helper to the provider registry and export
it from cleveragents.providers. Add structlog trace logging to
ProviderRegistry.get_default_provider_type() to record selection reasoning.
Update all existing behave step files, robot tests, and benchmarks that
relied on the implicit FakeListLLM default to pass an explicit LLM
instance instead.
Add new BDD tests (features/provider_fixes.feature with 17 scenarios),
Robot Framework integration tests (robot/provider_detection_smoke.robot),
and ASV benchmarks (benchmarks/provider_selection_bench.py).
ISSUES CLOSED: #323
Create DecisionService application-layer service that wraps
DecisionRepository with structured logging and UnitOfWork transaction
management. Wire DecisionService and PlanLifecycleService into the DI
container as Factory providers.
Inject DecisionService into PlanLifecycleService so that phase
transitions automatically record decisions: start_strategize records a
strategy_choice decision and start_execute records an
implementation_choice decision. Decision recording is optional and
never blocks lifecycle transitions.
Add Behave feature (5 scenarios), Robot Framework smoke tests (2 test
cases), ASV benchmarks (3 benchmark classes), and DI reference
documentation.
ISSUES CLOSED: #173
Tighten ToolRuntime._enforce_capabilities() to block ANY tool with
writes=True when plan_read_only is set, removing the not-cap.read_only
loophole that allowed certain write tools through. Tool name is now
always included in the ToolAccessDeniedError message.
Add read_only flag to ChangeSetCapture with ReadOnlyViolationError
raised when write-capable tools are wrapped on a read-only plan.
Add CLI fail-fast guards on plan execute and plan apply commands that
abort before calling the service layer if plan.read_only is True.
SkillContext.enforce_write_guard() already included tool name
correctly and required no changes.
Includes 18 Behave scenarios (90 steps), Robot integration tests,
ASV benchmarks, and docs/reference/read_only_actions.md.
ISSUES CLOSED: #322
P0: reject register() after close_all() with RuntimeError.
P1: catch CancelledError in close_all(), use WeakKeyDictionary for
cancellation_reasons to prevent memory leak, guard StateManager
update_state/reset/load_checkpoint/time_travel after close().
P2: contextlib.suppress in __del__ for partial construction, re-cancel
pending tasks in cleanup_tasks_async, handle late tasks added during
await window, guard AcpEventQueue.publish() after close with _is_closed
flag and is_closed property, fix ASV TimeRegisterBatch crash.
Tests: 5 new Behave scenarios (T1-T4 + is_closed), log handler and
event loop cleanup in after_scenario (T5-T6).
Docs: async_safety.md updated for register-after-close and state
mutation guards.
ISSUES CLOSED: #321
Add AsyncResourceTracker (core/async_cleanup.py) providing a central
registry for async resources with timeout-bounded close_all(), async
context manager support, and a __del__ finalizer that logs leaked
resources by name.
Enhance LangGraphBridge with cleanup_tasks_async() that awaits
in-flight tasks with a deadline instead of fire-and-forget cancel().
Add cancellation_reasons dict to trace why tasks were cancelled.
Add StateManager.close() to properly release checkpoint file handles
and complete the RxPY BehaviorSubject. Add AcpEventQueue.close() to
dispose all subscriptions.
Includes 14 Behave scenarios (67 steps), Robot integration tests,
ASV benchmarks, and docs/reference/async_safety.md.
ISSUES CLOSED: #321
Add decision persistence layer with DecisionRepository, DecisionModel,
and Alembic migration. Includes tree queries (BFS traversal via deque,
path-to-root), superseded lookup, ordered decision path retrieval,
concrete Decision type annotations (via TYPE_CHECKING), and comprehensive
test coverage (Behave BDD, Robot Framework, ASV benchmarks). Updated
database_schema.md, CHANGELOG.md, and CONTRIBUTORS.md.
ISSUES CLOSED: #171
- Patch CorrectionService at its module path, not the local-import site
- Use request_correction/execute_correction/analyze_impact (real API)
- Replace invalid 27-char subplan_id values with valid 26-char ULIDs
- Add required --guidance flag to dry-run CLI invocation
- Fix feature file assertions to match actual CLI output
Resolves CI failures in unit_tests (job 4) and coverage (job 6) for
run 645 on PR #441.
Replace hard-coded '11 decision types' with dynamic len(DecisionType) in
the assertion and remove the literal from scenario names, step patterns,
docstrings, and docs so the tests stay correct when new DecisionType
values are added.
- Replace dead after_scenario() hook with context.add_cleanup() calls
so patchers are properly stopped (Behave only runs hooks from
environment.py)
- Fix invalid ULID constants: pad DECISION/CORRECTION to 26 chars,
replace invalid 'I' with 'J' in INVARIANT_ULID
- Add temp file cleanup for NamedTemporaryFile(delete=False) via
context.add_cleanup() in behave steps and try/finally in robot helper
- Remove decision-persistence doc sections from testing.md that belong
in PR #438, not this PR
The jitter-spread scenario used millisecond-bucketed wall-clock
timestamps to assert that retried operations did not cluster. On busy
CI runners — especially when a preceding scenario's time.sleep patch
leaked through a late cleanup — tenacity's retry waits became no-ops
and all five operations landed in the same millisecond bucket, tripping
the max_in_bucket <= 3 assertion.
Three hardening changes:
1. Jitter test: switch from time.time() to time.monotonic_ns() and
replace the fragile bucket assertion with a unique-timestamp count
(>= 2 distinct readings among 5 sequential operations).
2. Timeout test: eagerly restore time.sleep in a try/finally block so
subsequent scenarios never observe the patched no-op, regardless of
behave's cleanup ordering.
3. Async circuit breaker: lower recovery_timeout from 0.2 s to 0.1 s
(the sleep step already waits 0.2 s) to give a wider safety margin
on slow CI machines.
The CLI plain format output does not include the literal text "Dry Run"
(that string only appears in the Rich panel title). Assert against
"risk_level" which is present in the dry-run impact output instead.
freemo
brent.edwards
aditya
khyari hamza