Configuration and companion CLI reference#
Generated by python scripts/check_docs.py --write-reference; checked by the same script.
Run python -m localagentbridge --config PATH companion ACTION --help for action flags.
Configuration file selection: explicit --config, then LAB_CONFIG, then current-directory config.toml.
Relative data_dir/source_root/codex_home paths resolve against the selected TOML file.
Existing LAB_MODEL/LAB_ENDPOINT/LAB_DATA_DIR and role overrides remain supported.
| Top-level key | Default / source |
|---|---|
model |
'qwen3:4b' |
endpoint |
'http://localhost:11434' |
allow_remote |
False |
max_input_chars |
10000 |
max_output_tokens |
600 |
max_output_bytes |
3000 |
num_ctx |
8192 |
timeout_seconds |
180 |
no_inference_deadline |
False |
thinking |
False |
data_dir |
factory: <lambda> |
poll_seconds |
2 |
discover_seconds |
30 |
collapse_seconds |
10 |
stale_seconds |
300 |
otel_port |
4318 |
codex_home |
factory: <lambda> |
source_root |
None |
roles |
factory: dict |
routes |
factory: <lambda> |
consensus_roles |
factory: <lambda> |
search |
factory: SearchConfig |
delegation |
factory: DelegationConfig |
model_registry |
factory: list |
adaptive |
factory: AdaptiveConfig |
agentic |
factory: AgenticConfig |
selective |
factory: dict |
Companion actions#
| Action | Flags/arguments |
|---|---|
contract-usage-import |
--rollout |
contract-usage-bind |
--kind, --session, --logical-id, --run, --task, --category, --evidence |
contract-usage-report |
--since, --cutoff, --session |
contract-route |
--request |
contract-submit |
--request, --run |
contract-review |
--event, --candidate, --outcome, --validation, --coverage |
contract-correct |
--job, --note |
contract-resource-release |
--job, --evidence |
contract-qualify-start |
--backend, --model, --activate, --context-tokens |
contract-qualify-record |
--job |
contract-metric |
--job, --request |
contract-integrate |
--job, --candidate |
inventory |
|
select |
model |
provider-status |
--backend, --model |
create |
|
submit |
--run, --request, --model, --execution, --no-deadline |
workflow-submit |
--run, --request |
workflow-next |
--job |
workflow-checkpoint |
--job, --request |
workflow-status |
--job |
controller |
--codex-exe, --once |
start |
--codex-exe |
status |
--run |
pause |
|
resume |
|
recover |
|
shutdown |
|
cancel |
--run, --job |
init-review |
--run, --codex-exe |
reconcile-session |
--run, --evidence |
reconcile |
--event, --disposition, --evidence |
decide |
--event, --outcome, --validation, --local |
finalize |
--run, --summary, --validation-passed |
dock |
The companion uses data_dir/companion/companion.sqlite3 (schema 3), with private
callback artifacts under reviews/<run>/<delivery>/. No control endpoint is exposed.
The CLI/store interface is the dock contract; state names appear in USER-GUIDE.md.
Existing nested sections: roles, routes, search, delegation, model_registry, adaptive, agentic.
See the existing profiles and workflow guides for those bounded interfaces.
Optional selective.defaults selects a model/backend; selective.model_profiles binds exact identity,
limits and provider controls. selective.project_root restricts that personal profile to one checkout.
Explicit job model overrides the persistent selector, which overrides the configured default.
Profile output/attempt/retry limits may be narrowed; changed context requires separate qualification.
selective.host_limits requires whole-host RAM admission, hard/sustained headroom and GPU ceilings.
See SELECTIVE-SETUP.md for runtime bindings and SELECTIVE-PILOT.md for accounting semantics.
A direct job with no_inference_deadline keeps the HTTP connect timeout at five seconds;
it disables generation/read deadlines, not workflow budgets or callback watchdogs.
Analysis task JSON requires task_key, tool, objective and content. Routed execution
also accepts role/routing/consensus/roles. Tools: analyze_code, summarize_context,
review_diff, analyze_test_output and second_opinion. Workflow submission uses
task_key and the existing typed task instead. See COMPANION-WORKFLOWS.md.
Configured roles select routed execution by default; --model or --execution exact
explicitly selects one exact model. Workflow jobs retain the full v0.5 policy.