Configuration reference
All configuration is read from HEAL_* environment variables (the nearest .env is auto-loaded and overrides the process environment). This page is generated from the settings schema, so it always matches the installed version.
Feature switches
Master toggles and opt-in behaviours.
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_ENABLED |
boolean |
True |
— |
Master switch for the healing engine. |
HEAL_LOCATOR_TIERS |
string |
selection |
matches ^(selection|generation)$ |
Locator pipeline: 'selection' (candidates + index pick, default) or 'generation' (DOM prompt). |
HEAL_HEAL_ASSERTIONS |
boolean |
False |
— |
Enable assertion-drift healing (opt-in). |
HEAL_FORM_FILL |
boolean |
False |
— |
Allow form-diagnosis to fill fields (opt-in; diagnose-only by default). |
HEAL_FIX_TIER |
report / patch / in-place |
report |
— |
Highest fix-application tier allowed. |
HEAL_WARM_START |
boolean |
True |
— |
Preload known broken->healed locator mappings from the healing history. |
Default model
The model used for every agent role unless overridden below. Point HEAL_BASE_URL at any OpenAI-compatible endpoint or use a pydantic-ai provider string in HEAL_MODEL (e.g. openai:gpt-4.1-mini).
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_MODEL |
string |
— |
— |
Default model name (e.g. 'MiniMax-M2.5' or 'openai:gpt-4.1-mini'). |
HEAL_BASE_URL |
string |
— |
— |
Default OpenAI-compatible endpoint base URL. |
HEAL_API_KEY |
string |
— |
— |
Default API key for the endpoint. |
HEAL_OUTPUT_MODE |
auto / tool / native / prompted |
auto |
— |
Default structured-output mode. |
Per-role overrides
Each agent role (triage, locator, vision, rca) can override the default model, endpoint, key and output mode. Unset values fall back to the defaults above — so HEAL_MODEL alone is enough for simple setups.
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_TRIAGE_MODEL |
string |
— |
— |
Model for the triage agent (falls back to HEAL_MODEL). |
HEAL_TRIAGE_BASE_URL |
string |
— |
— |
Endpoint for the triage agent (falls back to HEAL_BASE_URL). |
HEAL_TRIAGE_API_KEY |
string |
— |
— |
API key for the triage agent (falls back to HEAL_API_KEY). |
HEAL_TRIAGE_OUTPUT_MODE |
auto / tool / native / prompted |
— |
— |
Output mode for the triage agent (falls back to HEAL_OUTPUT_MODE). |
HEAL_LOCATOR_MODEL |
string |
— |
— |
Model for the locator agent (falls back to HEAL_MODEL). |
HEAL_LOCATOR_BASE_URL |
string |
— |
— |
Endpoint for the locator agent (falls back to HEAL_BASE_URL). |
HEAL_LOCATOR_API_KEY |
string |
— |
— |
API key for the locator agent (falls back to HEAL_API_KEY). |
HEAL_LOCATOR_OUTPUT_MODE |
auto / tool / native / prompted |
— |
— |
Output mode for the locator agent (falls back to HEAL_OUTPUT_MODE). |
HEAL_VISION_MODEL |
string |
— |
— |
Vision model for screenshot analysis (falls back to HEAL_MODEL). |
HEAL_VISION_BASE_URL |
string |
— |
— |
Endpoint for the vision agent (falls back to HEAL_BASE_URL). |
HEAL_VISION_API_KEY |
string |
— |
— |
API key for the vision agent (falls back to HEAL_API_KEY). |
HEAL_VISION_OUTPUT_MODE |
auto / tool / native / prompted |
— |
— |
Output mode for the vision agent (falls back to HEAL_OUTPUT_MODE). |
HEAL_RCA_MODEL |
string |
— |
— |
Model for the root-cause-analysis agent (falls back to HEAL_MODEL). |
HEAL_RCA_BASE_URL |
string |
— |
— |
Endpoint for the RCA agent (falls back to HEAL_BASE_URL). |
HEAL_RCA_API_KEY |
string |
— |
— |
API key for the RCA agent (falls back to HEAL_API_KEY). |
HEAL_RCA_OUTPUT_MODE |
auto / tool / native / prompted |
— |
— |
Output mode for the RCA agent (falls back to HEAL_OUTPUT_MODE). |
Budgets
Caps that keep healing bounded in CI. Breaching the run cap degrades to RCA-only instead of failing the test run.
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_MAX_FAILURE_SECONDS |
number |
60.0 |
> 0 |
Wall-clock cap per healing transaction. |
HEAL_MAX_FAILURE_TOKENS |
integer |
50000 |
> 0 |
Token cap per healing transaction. |
HEAL_MAX_RUN_TOKENS |
integer |
2000000 |
> 0 |
Token cap per test run; breach degrades to RCA-only. |
HEAL_REQUEST_LIMIT |
integer |
8 |
> 0 |
Max LLM requests per agent run within one transaction. |
HEAL_AGENT_RETRIES |
integer |
3 |
≥ 0 |
Output-validator retries per agent run. |
Timing recovery
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_READY_TIMEOUT_SECONDS |
number |
20.0 |
> 0 |
Max wait for page-ready in timing recovery. |
Reporting
Where artifacts and cross-run history are written.
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_REPORT_DIR |
string |
— |
— |
Report directory; defaults to /heal. |
HEAL_HISTORY_DB |
string |
— |
— |
Path to the cross-run healing history SQLite db. |
Other
| Variable |
Type |
Default |
Constraints |
Description |
HEAL_PROBE_CAPABILITIES |
boolean |
True |
— |
Verify the resolved output mode works before healing with it, and fall back to a mode that does. One tiny call per endpoint, cached for the run. |