Implement StrategyCoordinator and FusionEngine as named facades over
the existing ACMS pipeline components, providing clean public APIs
for parallel strategy execution with proportional budget allocation
and fragment fusion with dedup/conflict resolution/knapsack packing.
Key changes:
- Add StrategyCoordinator with parallel execution and confidence-based budget allocation
- Add FusionEngine with URI+hash dedup, max-depth conflict resolution, greedy knapsack packing
- Add budget overage guard with lowest-relevance fragment dropping
- Add per-strategy max caps enforcement
- Wire into existing ContextAssemblyPipeline
- Add Behave BDD tests, Robot integration tests, ASV benchmarks
- Add docs/reference/acms_fusion.md
ISSUES CLOSED: #192
Implemented lazy container activation for devcontainer-instance resources
with ContainerLifecycleState enum tracking six states (inactive, starting,
active, stopping, stopped, error) with validated transitions. Extended
DevcontainerHandler with devcontainer up CLI integration and JSON output
parsing for container start. Added periodic health checking via
devcontainer exec ping with configurable interval. Added agents resource
stop and agents resource rebuild CLI commands for manual lifecycle
control. Wired session close and plan completion hooks to automatic
container cleanup. Includes lifecycle state persistence in resource
registry with timestamped transitions. Added Behave BDD tests, Robot
integration tests, and ASV activation latency benchmarks.
- Added remoteWorkspaceFolder absolute-path validation
- Aligned spec: handler name, rebuild types, --yes flag on stop/rebuild
- Added registry re-read in stop_container success path for consistency
- Added session_id field to ContainerLifecycleTracker for scoped cleanup
- Scoped stop_all_active_containers to session_id when provided
- Wired _cleanup_devcontainers into fail_apply and fail_execute
- Wired start_health_check into activate_container success path
- Restructured facade session close to always run container cleanup
even without session service (F4)
- Re-read tracker from registry in activate_container success path
- Added evict_terminal_trackers to cap registry growth
- Updated devcontainer_resources.md: health check auto-start, scoped
cleanup hooks, known limitations for eviction and sandbox_strategy
- Wired evict_terminal_trackers into stop_all_active_containers so
terminal-state trackers are actually evicted in production
- Made stop_container idempotent: returns early when container is
already in a terminal state instead of raising ValueError
- Fixed benchmark health check thread leak in TimeActivationLatency
by clearing registry after each timing loop
- Added rebuild pass-through (--reset-container flag to devcontainer up)
- Added host_workspace_path field on ContainerLifecycleTracker so
health probes use the host-side path for devcontainer exec
- Wired lazy activation into DevcontainerHandler.resolve() for
devcontainer-instance resources in non-running states
- Changed _default_strategy from SNAPSHOT to NONE (container
itself provides isolation; SandboxFactory raises NotImplementedError
for snapshot)
- Restricted _STOPPABLE_TYPES to devcontainer-instance only
(container-instance is not directly stoppable via CLI)
ISSUES CLOSED: #514
Add RepoIndexingService with incremental refresh, language detection,
SHA-256 hashing, and policy enforcement. Includes domain models,
DB persistence, DI wiring, Behave/Robot/ASV tests, and reference docs.
ISSUES CLOSED: #195
## Summary
Fix `agents project show` not finding a project immediately after creation. Extends the `session.commit()` fix from #589 to also cover `update()` and `delete()` in `NamespacedProjectRepository`.
## Changes
**Production fix** (`src/cleveragents/infrastructure/database/repositories.py`):
- Add `session.commit()` to `create()`, `update()`, and `delete()` methods
- Add `finally: session.close()` guard to all three methods
- Update class docstring to reflect commit-per-method pattern
**Tests & benchmarks**:
- 3 Behave BDD regression scenarios (`features/project_show_after_create.feature`)
- Robot Framework integration smoke tests with "not found" assertion (`robot/project_show_after_create.robot`)
- ASV benchmarks for create-then-show round-trip (`benchmarks/project_show_after_create_bench.py`)
## Review feedback addressed
- **F1**: Removed unrelated em-dash CHANGELOG edits — wrote clean entry from scratch
- **F2**: Kept Suite Setup/Teardown (required for `${PYTHON}` variable); updated stale docs
- **F3**: Added "not found" string assertion to Robot negative test case
- **F4**: Removed redundant `Base.metadata.create_all()` from `_make_fresh_repo()` helper
- Updated all stale TDD "expected to fail" comments — this PR includes the fix
## Process
- Single squashed commit, rebased onto `master` (no merge commits)
- Prescribed commit message from issue #590 metadata
ISSUES CLOSED: #590
Reviewed-on: cleveragents/cleveragents-core#593
Reviewed-by: Rui Hu <rui.hu@cleverthis.com>
Co-authored-by: Brent E. Edwards <brent.edwards@cleverthis.com>
Co-committed-by: Brent E. Edwards <brent.edwards@cleverthis.com>
## Summary
Fix `agents project create` not persisting projects to the database. `NamespacedProjectRepository.create()` called `session.flush()` but never `session.commit()`, so projects were invisible to subsequent `agents project list` invocations that open a separate session.
## Changes
**Production fix** (`src/cleveragents/infrastructure/database/repositories.py`):
- Replace `session.flush()` with `session.commit()` in `NamespacedProjectRepository.create()`
- Add `finally: session.close()` guard for proper session lifecycle
**Tests & benchmarks**:
- 4 Behave BDD regression scenarios (`features/project_create_persist.feature`)
- Robot Framework integration smoke tests (`robot/project_create_persist.robot`)
- ASV benchmarks for create-then-list round-trip (`benchmarks/project_create_persist_bench.py`)
## Review feedback addressed
- **H3**: Added `finally: session.close()` to `create()` method
- **M1**: Removed redundant `session.flush()` before `session.commit()`
- **M2**: Updated stale TDD "expected to fail" comments — this PR includes the fix
- **M4**: Rewrote CHANGELOG entry to describe the fix, not just tests
- **F1**: Updated Robot documentation (Suite Setup/Teardown kept — needed for `${PYTHON}`)
- **F2**: Fixed bare assertion `"my-app"` to `"local/my-app"` in namespace scenario
- **F3**: Removed redundant `Base.metadata.create_all()` from `_make_fresh_repo()`
## Process
- Single squashed commit, rebased onto `master` (no merge commits)
- Prescribed commit message from issue #589 metadata
ISSUES CLOSED: #589
Reviewed-on: cleveragents/cleveragents-core#591
Reviewed-by: Rui Hu <rui.hu@cleverthis.com>
Co-authored-by: Brent E. Edwards <brent.edwards@cleverthis.com>
Co-committed-by: Brent E. Edwards <brent.edwards@cleverthis.com>
## Summary
Fixes the `plan correct` CLI handler which never passes the decision tree or influence edges to `CorrectionService`, producing single-node impact analysis regardless of the plan's actual decision structure.
Closes#606
## Root Cause
The `correct_decision()` handler in `plan.py` created a bare `CorrectionService()` and called `analyze_impact()` and `execute_correction()` without passing `decision_tree` or `influence_edges`. Both parameters default to `None` -> empty dicts, causing `_compute_affected_subtree()` BFS to return only the single target decision, ignoring all descendants and influence-DAG dependents.
## Changes
### Production Fix (`src/cleveragents/cli/commands/plan.py`)
- Resolve `DecisionService` via `get_container()` (following the pattern used by `plan explain` and `plan tree`)
- Build structural tree adjacency list from `decision_svc.list_decisions(plan_id)` using `parent_decision_id` relationships
- Fetch influence edges from `decision_svc.get_influence_edges(plan_id)`
- Pass both `decision_tree` and `influence_edges` to `svc.analyze_impact()` and `svc.execute_correction()`
### Existing Test Fixups
- Updated 4 step definition files that mock `CorrectionService` to also mock the new `DecisionService` resolution path
### New Tests
- **Behave BDD**: 3 scenarios verifying tree/edge forwarding (dry-run subtree, execution subtree, leaf node)
- **Robot Framework**: 3 integration smoke tests
- **ASV Benchmark**: Tree building and analyze_impact overhead benchmarks
## Quality Gates
- `nox -s lint` — PASSED
- `nox -s typecheck` — 0 errors
- `nox -s unit_tests` — 9,109 scenarios, 0 failures
- `nox -s coverage_report` — 97%
ISSUES CLOSED: #606
Reviewed-on: cleveragents/cleveragents-core#639
Co-authored-by: Jeffrey Phillips Freeman <jeffrey.freeman@cleverthis.com>
Co-committed-by: Jeffrey Phillips Freeman <jeffrey.freeman@cleverthis.com>
Implement spec-mandated pre-flight guardrail checks that validate plan
readiness before entering the Strategize phase:
1. Action schema validation — verifies action exists and is well-formed.
2. Actor availability — confirms all 4 actor roles registered.
3. Skill/tool existence — transitively resolves all tools/skills.
4. Automation policy — verifies profile permits execution.
5. Rollback feasibility — ensures all tools checkpointable if required.
6. Resource accessibility — shallow connectivity check on linked resources.
7. Validation attachment resolution — pre-resolves all applicable validations.
PlanPreflightGuardrail service runs all 7 checks via run_all_checks().
On first failure, raises PreflightRejection with check name and message.
Wired into plan_lifecycle_service before Strategize phase.
Behave BDD: 18 scenarios covering all 7 checks (positive + negative).
Robot Framework: 3 integration smoke tests.
ASV benchmarks: pre-flight check execution time.
ISSUES CLOSED: #582
ToolRegistryRepository.create(), .update(), and .delete() called
session.flush() but never session.commit(). The CLI factory creates a
raw sessionmaker without a UnitOfWork wrapper, so the transaction was
never committed and SQLAlchemy performed an implicit rollback when the
session was garbage-collected.
The same bug existed in ValidationAttachmentRepository.attach() and
.detach().
Changes:
- Add session.commit() after session.flush() in all five mutating
methods across ToolRegistryRepository and
ValidationAttachmentRepository.
- Add finally: session.close() to guarantee session cleanup regardless
of success or failure.
- Update class docstrings to reflect the new commit-on-write semantics.
- Add Behave BDD feature (tool_add_persist.feature) with scenarios for
single-tool round-trip, multi-tool persistence, and duplicate
rejection, using file-based SQLite to reproduce the cross-session
issue.
- Add Robot Framework integration test (tool_add_persist.robot) with
add-then-list and fresh-list-empty scenarios.
- Add ASV benchmark (tool_add_persist_bench.py) with
track_list_after_add_count metric.
Key decisions:
- File-based SQLite (not in-memory) is used in tests because the bug
only manifests when the session/engine is fully disposed between add
and list, simulating separate CLI invocations.
- Step patterns are prefixed with "tool-persist" to avoid AmbiguousStep
collisions with existing tool_registry_steps.py.
- The commit-in-repository approach was chosen over adding a UnitOfWork
to the CLI factory because the CLI commands are simple CRUD operations
that should auto-persist without requiring callers to remember to
commit.
ISSUES CLOSED: #621
Wire SkillService to use SkillRepository for database persistence,
fixing the bug where `agents skill add` stored skills only in an
in-memory OrderedDict that was lost when the CLI process exited.
Changes:
- SkillService now accepts optional skill_repo and session_factory
parameters. When provided, add_skill() persists to the database
and the constructor pre-loads existing skills from DB rows.
- _get_skill_service() in the CLI now creates a DB-backed service
following the same engine/sessionmaker/repository pattern used by
the tool CLI (tool.py).
- _reset_skill_service() now installs a fresh in-memory SkillService
(instead of setting None) to avoid DB side-effects during unit
testing with parallel runners.
- remove_skill() also persists the deletion to the database.
Test coverage:
- Behave BDD: features/skill_add_persist.feature (4 scenarios)
- Robot Framework: robot/skill_add_persist.robot (3 smoke tests)
- ASV benchmark: benchmarks/skill_add_persist_bench.py
ISSUES CLOSED: #620
Implemented the remaining ACMS pipeline components and advanced context
strategies:
Pipeline Phase 3:
- FragmentOrdererProtocol + RelevanceCoherenceOrderer: orders fragments
by relevance while maintaining narrative coherence via UKO node prefix
grouping. Groups related fragments together, sorts groups by max
relevance, and within groups orders by relevance desc / depth asc.
- PreambleGeneratorProtocol + ProvenancePreambleGenerator: generates
provenance preamble with strategy contributions (fragment counts and
token percentages), confidence indicators (avg/min/max), tier and
depth distribution, UKO node coverage, and coverage gap detection.
Advanced Strategies:
- ArceStrategy (quality 0.95): adaptive recursive context expansion with
iterative multi-backend refinement and configurable iteration limit
(default 5) to prevent unbounded refinement. Uses composite scoring
(relevance + depth + diversity) with contextual boosting for fragments
related to the current top-ranked anchor set.
- TemporalArchaeologyStrategy (quality 0.5): historical context retrieval
from graph+cold backends. Prioritises cold-tier fragments using a
temporal scoring model (tier bonus + relevance + depth).
- PlanDecisionContextStrategy (quality 0.7): decision history retrieval
from warm/cold backends. Prioritises warm then cold tier fragments
for correction and retry scenarios.
All strategies registered in strategy registry with correct quality scores
and backend requirements. All components implement their respective
Protocol interfaces and can be injected into the ContextAssemblyPipeline
via constructor dependency injection.
Tests:
- 33 BDD scenarios in features/acms_pipeline_phase3.feature
- Robot Framework integration tests in robot/acms_pipeline_phase3.robot
- ASV performance benchmarks in benchmarks/acms_pipeline_phase3_bench.py
ISSUES CLOSED: #545
Added minimal LSP server entrypoint supporting initialize/shutdown/exit
handshake over JSON-RPC stdin/stdout transport with Content-Length
framing. Unsupported methods return MethodNotFound error with descriptive
message. Wired LSP requests through ACP facade in local mode. Added
agents lsp serve CLI command with --log-level flag, PID output, and
startup banner. Created reference documentation for the stub server.
Includes Behave BDD tests for protocol handshake, Robot smoke test, and
ASV startup latency benchmark.
ISSUES CLOSED: #203
Core domain types (FragmentProvenance, ContextFragment, ContextBudget,
ContextPayload) now extend their CRP counterparts via Pydantic v2
inheritance, ensuring isinstance compatibility across the model
hierarchy.
Key changes:
- CRP base types made frozen=True (no consumer mutates them)
- CRP AssembledContext fields changed from list to tuple (frozen consistency)
- Core types extend CRP bases: FragmentProvenance(CRPFragmentProvenance),
ContextFragment(CRPContextFragment), ContextBudget(CRPContextBudget),
ContextPayload(CRPAssembledContext)
- Removed duplicate ContextFragment dataclass from skeleton_compressor
- Updated project_context.py to pass tuples to frozen AssembledContext
- Added Behave tests (10 scenarios), Robot integration tests (3 cases),
and ASV benchmarks for the unified hierarchy
- Updated Known Limitations table in docs/reference/acms.md
ISSUES CLOSED: #569
Implemented the analyzer plugin framework with AnalyzerProtocol,
AnalyzerRegistry for registration/discovery by file extension,
PythonAnalyzer (AST-based extraction of modules, classes, functions,
imports, docstrings), and MarkdownAnalyzer (section, code block, and
link extraction). Both analyzers produce well-formed UKO triples with
proper URI schemes.
ISSUES CLOSED: #551
Add the Depth/Breadth Projection System and Skeleton Context
Propagation as specified in docs/specification.md §25265-25340
and §43057-43128:
- ProjectionSpec: frozen Pydantic model capturing a projection
request (focus, breadth, depth, gradient, domain)
- ProjectedNode: frozen model for materialized graph nodes with
resolved depth and distance
- DepthBreadthProjector: stateless BFS projector over UKO graph
adjacency with depth gradient (linear reduction by distance)
- PlanContextInheritance: service computing child plan context
from parent assembled context with skeleton injection
- ChildContextResult: frozen result model with request and skeleton
- InheritanceConfig: frozen config for skeleton_ratio (default 0.2)
- Built-in DetailLevelMap presets for code, docs, and database
Includes 27 Behave BDD scenarios, 9 Robot Framework integration
tests, and ASV benchmarks for all components.
ISSUES CLOSED: #544
Add production-grade Phase 2 (Fragment Fusion) components for the ACMS
context assembly pipeline, replacing the no-op defaults:
- ContentHashDeduplicator: Groups fragments by UKO node URI, hashes
content to detect duplicates, retains highest relevance_score.
- MaxDepthResolver: Resolves depth conflicts by keeping the highest
detail depth per UKO node, with relevance tiebreaking.
- WeightedCompositeScorer: Computes composite score from configurable
weighted factors (relevance=0.4, hierarchy=0.3, quality=0.2,
recency=0.1). Stores component breakdown in metadata.
- GreedyKnapsackPacker: Greedy knapsack selection with depth fallback
(tries depths [9,4,2,0] for oversized fragments) and minimum
fragment token threshold (10).
Also adds:
- ScoredFragment frozen Pydantic model (spec §42825) with
composite_score, score_components, and fragment reference
- score_detailed() method on WeightedCompositeScorer returning
ScoredFragment objects for callers needing full breakdowns
- All components implement v1 Protocol signatures from acms_service.py
and can be DI-injected into ACMSPipeline constructor
Testing:
- 31 Behave BDD scenarios in acms_pipeline_phase2.feature covering
deduplication, depth resolution, scoring, packing, depth fallback,
budget constraints, pipeline integration, and ScoredFragment model
- 6 Robot Framework integration smoke tests
- ASV benchmark suites for all 4 components and ScoredFragment
Quality gates: lint, typecheck (0 errors), unit_tests (8555 scenarios),
coverage (97.0%), dead_code — all passing.
ISSUES CLOSED: #540
Add production-quality Phase 1 pipeline components for the ACMS
Context Assembly Pipeline:
- ConfidenceWeightedSelector: strategy selection with preference
boosting and confidence-based ranking
- ProportionalBudgetAllocator: proportional token budget distribution
with min_useful_budget enforcement and largest-remainder rounding
- ParallelStrategyExecutor: concurrent strategy execution via
ThreadPoolExecutor with per-strategy timeouts and circuit breaking
- CircuitBreaker: per-strategy failure tracking with configurable
threshold and explicit reset
- ContextAssemblyPipeline: extends ACMSPipeline with Phase 1
production components and per-stage timing (StageTimings)
Includes 28 Behave BDD scenarios, 9 Robot Framework integration tests,
and ASV benchmarks for all components.
ISSUES CLOSED: #539
Created ScoredFragment frozen model wrapping ContextFragment with
composite_score, score_breakdown, and rank fields. Added pipeline-
specific fragment models in domain/contexts/ with proper equality
based on uko_uri + detail_depth for deduplication support.
ISSUES CLOSED: #538
Implement the first three built-in context strategies for the ACMS v1
context assembly pipeline:
1. SimpleKeywordStrategy (quality 0.3) - Keyword matching on fragment
content with word-density fallback. Universal fallback strategy.
2. SemanticEmbeddingStrategy (quality 0.6) - Jaccard word-overlap
similarity scoring between query and fragment content.
3. BreadthDepthNavigatorStrategy (quality 0.85) - UKO node hierarchy
navigation prioritising fragments near focus nodes with higher
detail depths. Primary strategy for code projects.
All strategies implement the v1 ContextStrategy Protocol from
acms_service.py and can be registered with ACMSPipeline via
register_strategy().
Includes:
- 28 Behave BDD scenarios covering ranking, budget, capabilities,
can_handle confidence, explain, empty input, and pipeline
registration
- 9 Robot Framework integration tests
- ASV benchmarks at 10/100/1000 fragment scales for all 3 strategies
- Vulture whitelist entries for public API symbols
- 100% coverage on context_strategies.py
ISSUES CLOSED: #541
- Add AsyncJob domain model with status state machine and Pydantic validation
- Add AsyncWorker service with configurable concurrency and job store
- Add CancellationToken, WorkerHealthReport, InMemoryJobStore
- Add AsyncJobModel SQLAlchemy model and Alembic migration (m6_003)
- Add 5 async config keys to Settings (worker_id, concurrency, poll_interval, max_retries, timeout)
- Add _check_async_worker_health diagnostic check in system.py
- Add comprehensive Behave BDD tests (~60 scenarios) with full step definitions
- Add Robot Framework integration tests (6 smoke tests)
- Add ASV benchmark suite for async execution
- Add architecture documentation
- Update vulture_whitelist with new public API symbols
- All quality gates pass: lint, typecheck, unit_tests, integration_tests, coverage_report (97%)
1. Wire async job creation into PlanLifecycleService:
- Add optional job_store parameter to __init__
- Add _maybe_enqueue_async_job() helper that checks settings.async_enabled
and job store presence before creating and enqueuing an AsyncJob
- Call helper from execute_plan() (phase="execute") and apply_plan()
(phase="apply") after phase transitions
- When async is disabled or no job store is configured, behaviour is
unchanged (silent no-op)
2. Redact secrets in failed job error messages:
- Apply shared.redaction.redact_value() to the error string before
persisting to AsyncJob.error_message, preventing accidental secret
leakage (e.g. API keys in exception text) into the audit trail
Documentation:
- S1: Added specification reconciliation note (ADR-style) to
async_architecture.md addressing tension between "No Plan Queuing"
clause and the async subsystem authorised by issue #312
ISSUES CLOSED: #312
Add CrossPlanCorrectionService that implements the four child-plan-state-
dependent behaviours from the specification when a correction's affected
subtree includes child plans:
- Not yet started → cancel the child plan
- In progress → cancel + rollback sandbox to pre-child-plan state
- Completed but not applied → cancel + rollback sandbox
- Already applied → reject the correction (CorrectionRejection)
Key additions:
- ChildPlanState enum classifying child plans into 4 states
- CorrectionRejection result type with reason and affected applied plan IDs
- CascadeAction/CascadeResult models for cascade operation tracking
- CorrectionStatus.REJECTED for rejected corrections
- Atomic cascade-or-rollback: all child plan actions succeed or the
entire cascade is rolled back
- Protocol-based dependency injection (ChildPlanLookup, ChildPlanCanceller,
SandboxRollbacker) for testability
- execute_correction_with_cascade() integrates with CorrectionService flow
Testing:
- 24 Behave BDD scenarios in cross_plan_correction.feature
- 8 Robot Framework end-to-end smoke tests
- ASV benchmarks for cascade performance with varying child plan counts
ISSUES CLOSED: #547
Add InvariantReconciliationActor that runs at the start of the Strategize
phase to reconcile invariants from four scopes (global, project, action,
plan). The actor detects conflicts, resolves them using specificity-based
precedence (plan > action > project > global), honours non_overridable
global invariants, records invariant_enforced decisions, and produces a
reconciled InvariantSet.
Changes:
- New: src/cleveragents/actor/reconciliation.py
- InvariantReconciliationActor class with collect_invariants() and run()
- reconcile_invariants() pure function
- ScopeInvariants, ConflictRecord, ReconciliationResult dataclasses
- Modified: src/cleveragents/domain/models/core/invariant.py
- Added non_overridable: bool field to Invariant model
- New: features/invariant_reconciliation_actor.feature (26 BDD scenarios)
- New: features/steps/invariant_reconciliation_actor_steps.py
- New: robot/invariant_reconciliation_actor.robot
- New: robot/helper_invariant_reconciliation.py
- New: benchmarks/invariant_reconciliation_bench.py
Closes#549
Implement the runtime execution engine for validation tool wrapping,
as specified in docs/specification.md § Tool Wrapping.
WrappedToolExecutor resolves wraps references and delegates execution
to wrapped tools, supporting composable wrapping chains with cycle
detection and depth limiting (max 10 levels).
ArgumentMapper translates arguments between wrapper and wrapped tool
schemas using the argument_mapping configuration. Supports both
forwarded parameter names and literal fixed values.
TransformExecutor runs user-supplied transform functions in a
sandboxed Python environment with restricted builtins (no imports,
no filesystem, no network access). Validates that transforms return
proper validation-format dicts with a passed boolean.
Wired into the tool package public API via tool/__init__.py exports.
All new error types (WrappedToolNotFoundError, WrappingCycleError,
WrappingDepthExceededError, TransformExecutionError) provide clear
diagnostic messages.
Tests: 20 Behave scenarios covering argument mapping, transform
execution, simple/chained delegation, error handling, and sandbox
restrictions. 8 Robot Framework integration smoke tests. ASV
benchmarks for delegation overhead measurement.
ISSUES CLOSED: #543
Implement sandbox_boundary(r) function that walks up containment edges
in the resource DAG to the nearest sandboxable ancestor, enabling
resources sharing a boundary to share one sandbox instance.
Changes:
- Add boundary.py: is_sandbox_boundary(), sandbox_boundary(),
compute_sandbox_domains(), BoundaryCache (thread-safe, per-execution)
- Update SandboxManager: resolve_sandbox_key() and
get_or_create_sandbox_for_resource() key by (plan_id, boundary_id)
instead of (plan_id, resource_id); boundary cache lifecycle methods
- Define "sandboxable" via ResourceCapabilities.sandboxable + non-none
sandbox_strategy as per specification section 24659-24674
- Export new symbols from sandbox __init__.py
- Add vulture whitelist entries for new public API
Tests:
- 26 Behave BDD scenarios (features/sandbox_boundary_algebra.feature)
- 5 Robot Framework integration tests (robot/sandbox_boundary_algebra.robot)
- ASV benchmarks for boundary walk, domain grouping, and cache performance
ISSUES CLOSED: #548
Extended _compute_affected_subtree() to BFS over both the structural tree
(parent-child plan relationships) and decision_dependencies edges (influence
DAG). The algorithm performs a single O(V+E) BFS pass that unions neighbors
from both edge sources, using a visited set for cycle detection to guard
against data corruption.
Decision creation now supports dependency_decision_ids parameter in
record_decision() which populates the in-memory influence DAG store.
get_influence_edges() returns the adjacency list format consumed by
CorrectionService.
All public CorrectionService methods (analyze_impact, execute_revert,
execute_correction, generate_dry_run_report) accept an optional
influence_edges parameter while remaining backward-compatible (defaults
to None, preserving structural-only traversal when not provided).
Key design decisions:
- Single BFS pass over union of structural + influence edges rather than
separate traversals, ensuring O(V+E) complexity and consistent visit order
- Cycle detection via visited set with warning log (not an error) since
cycles indicate data corruption, not a programming error
- Influence edge logging: traversal emits count of influence edges processed
for observability
- Backward-compatible API: existing callers that only pass decision_tree
continue to work identically
ISSUES CLOSED: #542
Implement safety profile resolution and enforcement in the tool
execution pipeline, replacing the NotImplementedError stub with
working precedence logic and runtime safety checks.
Core changes:
- resolve_safety_profile() now resolves plan > action > project >
global precedence, returning the highest-priority non-None profile
(or DEFAULT_SAFETY_PROFILE with GLOBAL provenance when all None)
- ToolExecutionContext gains an optional safety_profile field
- ToolRuntime._enforce_capabilities() extended with three new checks:
* Unsafe tool gating: blocks tools with unsafe=True when profile
has allow_unsafe_tools=False (ToolSafetyViolationError)
* Skill category allow-list: blocks tools whose skill category
is not in allowed_skill_categories (ToolSafetyViolationError)
* Checkpoint requirement: OR-combines ctx.require_checkpoints
with safety_profile.require_checkpoints
- New ToolSafetyViolationError in tool error hierarchy
Test coverage:
- 30 updated BDD scenarios in safety_profile.feature (resolve
precedence replaces NotImplementedError stub test)
- 24 new BDD scenarios in safety_profile_enforcement.feature
- 9 Robot Framework integration smoke tests
- 4 ASV benchmark suites (construction, serialization, resolution,
provenance enum)
All nox sessions pass (typecheck 0 errors, unit_tests 7735 scenarios
0 failures, coverage 97%, integration_tests 9/9 passed, benchmarks
complete).
ISSUES CLOSED: #345
Add LLMTrace Pydantic v2 domain model with all required fields (trace_id,
plan_id, decision_id, actor, provider, model, prompt_tokens, completion_tokens,
cost_usd, latency_ms, tool_calls, context_hash, streaming, retry_count, error).
Define 14 OperationalMetricKey values (PLAN_DURATION_MS, PLAN_TOTAL_COST_USD,
PLAN_DECISION_COUNT, ACTOR_INVOCATION_COUNT, ACTOR_LATENCY_MS,
TOOL_INVOCATION_COUNT, TOOL_ERROR_RATE, CONTEXT_BUILD_TIME_MS,
CONTEXT_TOKEN_COUNT, LLM_CALL_COUNT, LLM_TOTAL_TOKENS, LLM_TOTAL_COST_USD,
LLM_AVG_LATENCY_MS, SUBPLAN_COUNT) with MetricEntry model and MetricCollector.
Add llm_traces database table (LLMTraceModel) with LLMTraceRepository for
persistence. TraceService provides recording, querying, metric computation,
plan lifecycle hooks, and optional LangSmith forwarding when
LANGCHAIN_TRACING_V2=true.
Wired into DI container as trace_service. Includes 28 Behave BDD scenarios,
6 Robot Framework smoke tests, 3 ASV benchmark suites, and reference
documentation.
Fix pre-existing cli_core server_mode test flake by mocking
resolve_server_mode in test steps to avoid stale config file interference.
Closes#500
Add hierarchical decomposition with 4+ levels and bounded context per
subplan. Implement decomposition heuristics (max_files_per_subplan,
max_tokens_per_subplan, language/dir clustering). Add dependency closure
computation for large graphs and DAG execution ordering. Add bounded
dependency closure with cutoff thresholds and memoization for 10K+
files. Record decomposition decisions in DecisionService
(strategy_choice + subplan_spawn entries).
New modules:
- decomposition_models.py: DecompositionConfig, DecompositionNode,
DecompositionResult, DependencyEdge, DependencyGraph
- decomposition_clustering.py: ClusteringStrategy with directory,
language, and size clustering plus deterministic sort
- decomposition_graph.py: DependencyClosureComputer with bounded
closure, topological sort, cycle detection, and memoization
- decomposition_service.py: DecompositionService orchestrating
hierarchy building and decision recording
Settings: planner_max_depth, planner_max_files_per_subplan,
planner_max_tokens_per_subplan, planner_min_files_per_subplan
Closes#205
Add domain models, application service, CLI integration, and full test
coverage for multi-project subplan orchestration.
New components:
- MultiProjectMetadata, ProjectScope, ProjectDependency domain models
- MultiProjectService for creating/managing multi-project plans
- CLI display of multi-project metadata in plan commands
- Behave BDD tests (20 scenarios), Robot Framework integration tests,
ASV benchmarks, and reference documentation
Also fixes pre-existing test flakiness in cli_core, core_cli_commands,
cli_plan_context_commands, and helper_server_stubs caused by environment
leakage and path-with-spaces issues.
ISSUES CLOSED: #199
Add ExecutionEnvironment enum (host/container) to domain models, implement
execution environment resolution with priority chain (tool > plan > project
> default), wire the tool runner to check execution_environment before
execution, and add CLI flags to plan use/execute and project context set.
When container is selected but no container resource is available, a clear
ContainerUnavailableError is raised with an actionable message.
Closes#512
Behave steps, benchmarks, vulture whitelist, and docs referenced
renamed methods (get_decisions_for_plan, get_decision_tree). Updated
to use the actual API names (list_decisions, get_tree) and the kwargs
record_decision signature.
ISSUES CLOSED: #172
Add `plan explain` and `plan tree` CLI commands that format decision
trees in json/yaml/table/rich formats. Flags control views for
superseded decisions, context snapshots, and reasoning details.
- plan explain <decision_id>: renders a single decision with optional
--show-context, --show-reasoning, --show-alternatives flags
- plan tree <plan_id>: renders full decision tree with optional
--show-superseded and --depth flags
- BFS tree building uses collections.deque (no list.pop(0))
- Behave BDD scenarios (14 scenarios, 54 steps)
- Robot Framework smoke tests with helper script
- ASV benchmarks for explain formatting and tree operations
- Updated docs/reference/plan_cli.md and CHANGELOG.md
ISSUES CLOSED: #174
Implemented SkeletonCompressorService for ACMS context inheritance, producing
compressed context representations for propagation from parent plans to child
plans. Key design decisions and implementation details:
- SkeletonMetadata (frozen Pydantic model): records ratio, original_tokens,
compressed_tokens, and source_decision_ids for full auditability of each
compression pass. Persisted on Plan.skeleton_metadata.
- SkeletonCompressorService: stateless service accepting a list of
ContextFragment objects and a skeleton_ratio in [0.0, 1.0]. Fragments are
sorted by relevance descending with a stable secondary sort on fragment_id
to guarantee deterministic output. Token budget is original_tokens*(1-ratio);
fragments are greedily selected until budget is exhausted.
- Ratio semantics: 0.0 = no compression (pass-through), 1.0 = maximum
compression (single top fragment only), None = default 0.3.
- Integration: Plan model gains optional skeleton_metadata field exposed in
as_cli_dict() under the 'skeleton' key. Service registered in DI container
as skeleton_compressor_service (Singleton, stateless).
- Tests: 22 BDD scenarios (features/skeleton_compressor.feature) covering
ratio validation, stable ordering, metadata correctness, edge cases, and
plan model integration. 6 Robot Framework smoke tests. ASV benchmark suites
at 10/100/1000 fragment scales.
- Documentation: docs/reference/skeleton_compressor.md with ratio table,
algorithm description, metadata schema, and multi-decision plan example.
ISSUES CLOSED: #194
freemo
CoreRasurae
hamza.khyari
brent.edwards
aditya
khyari hamza