Todos, reflection, shortlisting
Enable todos, a reflection pass, and the LLM tool shortlister (find_tools).
Todos, reflection, and shortlisting
These are Orchestrator knobs — they change how CugaLite plans and retries, not which tools you attach.
Architecture: Reflection and todos run inside the agent loop. The configured shortlister selects tools for find_tools and provider binding when the tool catalog is large. See How CUGA works.
All three live under [advanced_features] in settings.toml.
Todos
Off by default. When on, CugaLite gets a create_update_todos tool and prompt instructions so it can keep a checklist for multi-step work (and reset it between tasks).
[advanced_features]
enable_todos = trueUse it when a task has several independent steps you want the model to track. Leave it off for short, single-tool runs.
Reflection
Off by default. After a sandbox/code step, CUGA checks whether the result actually moved the task forward. If not, it can retry instead of answering from a bad state.
[advanced_features]
reflection_enabled = trueTurn it on when wrong-but-confident answers are costly. Leave it off when latency matters more than a second look.
Pre-execute VERIFY
VERIFY inspects a proposed mutating code block before the sandbox runs it. It is independent of post-execution reflection and defaults to off:
[advanced_features]
pre_execute_verify_enabled = true
pre_execute_verify_timeout = 60Read-only blocks skip the check. A revise decision skips the proposed block so the agent can revise it; ok or unknown allows execution. Errors/timeouts fail open, and revision caps prevent endless verification loops. VERIFY is an advisory LLM check, not human approval or a deterministic authorization boundary. Use Tool Approval and ToolGuard for their respective controls.
Tool shortlister (find_tools)
When an app has more tools than the threshold, CugaLite does not dump the whole catalog into the prompt. It exposes find_tools: the model describes what it needs in natural language, and the configured shortlister returns the relevant subset.
[advanced_features]
shortlisting_tool_threshold = 35 # find_tools when total tools > this
cuga_lite_enable_few_shots = true # few-shots for the shortlister pathShortlisting strategies
| Strategy | Behavior |
|---|---|
llm | Ask the model to choose tools; default. |
embedding | Rank tools with local cosine similarity without an LLM call. |
hybrid | Prefilter with cosine similarity, then ask the LLM for the final selection. |
| Dotted class path | Load a custom shortlister implementation. |
[shortlister]
strategy = "hybrid"
fallback_strategy = "llm"
threshold = 128
top_k = 128
max_results = 10
min_score = 0.15
embedding_provider = "local"
embedding_model = "BAAI/bge-small-en-v1.5"advanced_features.shortlisting_tool_threshold decides when the prompt hides tools behind find_tools. [shortlister].threshold decides when the cosine stage engages. Shortlister embeddings are local by default and do not inherit [storage.embedding]. Use [shortlister.discovery] and [shortlister.bind_cap] for separate runtime-discovery and provider-cap overrides.
See the shortlister design and extension guide for custom implementations.
Over-cap on native bind_tools (providers that reject > ~128 tools) uses the configured shortlister:
[advanced_features]
cuga_lite_bind_tools_max_count = 128 # 0 disables the cap
cuga_lite_bind_tools_pad_to_cap = false
cuga_lite_bind_tools_include_find_tools = false