Luis Mendes
1ffb0fddbe
Enhanced the CorrectionService to properly isolate correction scope:
1. analyze_impact() now populates excluded_decisions by computing
the set difference between all plan decisions and the affected
subtree, ensuring root and sibling decisions are tracked.
2. Added rollback_tier computation that counts parent hops from
the target decision to the root, enabling depth-aware rollback
strategies (tier 0 = root targeted, tier N = N levels deep).
3. Enhanced dry-run report with excluded decisions list, rollback
tier, and tier-0 warning when entire tree is affected.
4. Added subtree isolation validation confirming root exclusion
and sibling non-contamination for non-root corrections.
5. Fixed status state-machine regression in execute_revert() where
analyze_impact() overwrote status back to ANALYZING from EXECUTING.
Guard now only transitions to ANALYZING when status is PENDING.
execute_revert() now transitions through ANALYZING before EXECUTING
for correct lifecycle ordering.
6. Fixed validate_subtree_isolation() to use structural-only BFS for
the sibling invariant check, so that influence-DAG-caused sibling
reachability is not misreported as an isolation violation (per spec
§ Affected Subtree Computation).
7. Fixed false-positive cycle-detection warnings from convergent
(diamond) topologies by replacing per-node seen_this_round with a
global enqueued set in BFS, preventing duplicate queue insertions
from different parents.
8. Added dry_run enforcement guard in _assert_executable() to prevent
execution of dry-run-only corrections per spec (§ plan correct
--dry-run: "Show impact without executing").
9. Fixed generate_dry_run_report() to preserve request status so that
generating a preview does not advance the correction lifecycle
(dry-run is non-mutating per spec). Status restoration now uses
try/finally to guarantee recovery even when analyze_impact() raises
after transitioning the status.
10. Improved cycle-detection log message accuracy to cover both
structural tree and influence DAG sources.
11. Extracted cost/time estimation constants (_COST_PER_DECISION,
_RECOMPUTE_SECONDS_PER_DECISION) from magic numbers.
12. Added terminal-state guard in analyze_impact() to reject
re-analysis after execution (APPLIED/FAILED/CANCELLED/REJECTED),
preventing audit-data corruption. Promoted terminal-status set
to a module-level _TERMINAL_STATUSES frozenset constant.
13. Added mode validation in execute_revert()/execute_append() to
prevent mode-mismatched execution (e.g. calling execute_revert
on an APPEND correction).
14. Fixed _collect_all_decisions() to always include the target
decision in the universe, preventing broken partition invariant
for isolated single-node plans.
15. Fixed tier-0 dry-run warning to only trigger when the target
is genuinely in the structural tree (avoiding false warnings for
nodes not present in the tree).
16. Removed duplicate as_cli_dict regression scenarios in
resource_type_deferred_physical.feature.
ISSUES CLOSED: #845
CleverAgents Core
CleverAgents is a Python-first automation platform. It provides a unified agents CLI, embedded runtime, and service orchestration tools while embracing modern Python tooling.
Highlights
- Unified CLI entry points:
cleveragentsandagents - Fast Typer/Click-based interface with parity for help/version behavior
- Behavior-driven coverage via Behave and Robot Framework
- Nox automation for linting, typing, testing, docs, builds, and benchmarks
- MkDocs-powered documentation with CleverAgents branding
Quick Start
# clone the CleverAgents core repository
git clone https://git.cleverthis.com/cleveragents/core.git
cd core
# install dependencies
python -m venv .venv
source .venv/bin/activate
pip install -e ".[dev,tests,docs]"
# set up pre-commit hooks and verify tooling
bash scripts/setup-dev.sh
# verify the CLI
agents --help
agents --version
Developing
Pre-commit hooks run automatically on every git commit (formatting, linting, type
checking, security scanning). To run checks manually:
# core validation
nox -s format # ruff auto-formatting
nox -s lint # ruff linting
nox -s typecheck # pyright type checking
nox -s unit_tests # behave unit tests
nox -s integration_tests # robot integration tests
# quality & security
nox -s security_scan # bandit security scanning
nox -s dead_code # vulture dead code detection
nox -s complexity # radon complexity analysis
nox -s pre_commit # run all pre-commit hooks
nox -s adr_compliance # verify ADR compliance
For the full quality automation guide, see docs/development/quality-automation.md.
Documentation
nox -s docs
nox -s serve_docs
Tests
Behave feature scenarios live under features/ and Robot suites under robot/. Use the Nox sessions above to execute them in parity with the implementation plan.
Observability
LangSmith tracing is optional and off by default. Enable it by exporting CLEVERAGENTS_LANGSMITH_ENABLED=true along with a project name and API key (CLEVERAGENTS_LANGSMITH_PROJECT, CLEVERAGENTS_LANGSMITH_API_KEY). The settings module automatically mirrors these values to LANGCHAIN_TRACING_V2, LANGCHAIN_PROJECT, and LANGCHAIN_API_KEY, so LangChain/LangGraph agents emit traces without extra wiring. Additional knobs such as CLEVERAGENTS_LANGSMITH_ENDPOINT, CLEVERAGENTS_LANGSMITH_USER_ID, and CLEVERAGENTS_LANGSMITH_TAGS are documented in docs/observability.md.
LLM provider configuration
CleverAgents ships with a LangChain/LangGraph powered provider registry that discovers whichever API keys you export and automatically selects the best available provider. The CLI now uses actors: select an actor with --actor <name> (or set a default via agents actor set-default). Actors embed provider/model choices; built-in actors are seeded from CLEVERAGENTS_DEFAULT_PROVIDER / CLEVERAGENTS_DEFAULT_MODEL, then fall back to the built-in order (openai → anthropic → google → azure → openrouter → groq → together → cohere → gemini).
Required environment variables
| Provider | Primary variables |
|---|---|
| OpenAI | OPENAI_API_KEY |
| Anthropic | ANTHROPIC_API_KEY |
GOOGLE_API_KEY or GOOGLE_GENAI_API_KEY |
|
| Azure OpenAI | AZURE_OPENAI_API_KEY plus AZURE_OPENAI_ENDPOINT/AZURE_OPENAI_DEPLOYMENT |
| OpenRouter | OPENROUTER_API_KEY (+ optional CLEVERAGENTS_OPENROUTER_ORGANIZATION for sanitized headers) |
| Gemini | GEMINI_API_KEY or GOOGLE_GEMINI_API_KEY |
| Cohere | COHERE_API_KEY |
| Groq | GROQ_API_KEY |
| Together | TOGETHER_API_KEY |
Set CLEVERAGENTS_DEFAULT_PROVIDER to pin the global provider (for example export CLEVERAGENTS_DEFAULT_PROVIDER=openai) and CLEVERAGENTS_DEFAULT_MODEL to lock in a model ID. When unset, the registry picks the first configured provider and uses its published default model such as gpt-4o for OpenAI or claude-sonnet-4-20250514 for Anthropic.
Diagnostics and testing shortcuts
agents diagnosticsprints whether the registry can see your credentials and which actor/provider is selected.agents tellandagents buildrequire--actor <name>unless a default actor is set; useagents actor set-default <name>to configure one.- Built-in actors (
<provider>/<model>) are immutable, custom actors must be namedlocal/<id>, and the default actor cannot be removed. Use--unsafewhen adding/updating configs marked unsafe; runtime only warns when invoking unsafe actors. CLEVERAGENTS_TESTING_USE_MOCK_AI=trueforces the in-repo mock provider so Behave/Robot suites never hit external APIs.- The full capability matrix (streaming, tool calls, JSON mode, etc.) is documented in
docs/reference/providers.md.