Advanced Configuration
Intentional operator-tuning controls and explicit compatibility/legacy boundaries.
Advanced Configuration
Use this page when teams need more granular control than the baseline configuration page.
Source of truth: these settings are implemented in patchpatrol/config.py and intentionally placed here when they are advanced, high-impact, or compatibility-only.
Stable Advanced Controls
These settings are part of the shipped code path and safe to use intentionally, but they are separated from the baseline contract so teams can make changes with intent.
| Variable | Default | Supported values or posture | Public meaning |
|---|---|---|---|
AI_REVIEW_TEMPERATURE | 0.2 | 0 < value <= 2 | Controls model sampling temperature. |
AI_REVIEW_TOP_P | 0.9 | 0 < value <= 1 | Controls nucleus sampling aggressiveness. |
AI_REVIEW_NUM_CTX | unset | Positive integer, Ollama-only | Sets model context window for Ollama; rejected when AI_REVIEW_PROVIDER=openai. |
AI_REVIEW_INCLUDE_GLOBS | unset | Comma-separated glob list | Include-filter allowlist for changed files before chunk planning. |
AI_REVIEW_EXCLUDE_GLOBS | unset | Comma-separated glob list | Exclude-filter blocklist for changed files before chunk planning. |
AI_REVIEW_ENABLE_PERFORMANCE_DIAGNOSTICS | false | true or false | Emits extra runtime diagnostics in meta.performance_diagnostics and adds a markdown diagnostics section. |
AI_REVIEW_CLEANUP_OWNED_MR_DISCUSSIONS | false | true or false | Enables opt-in cleanup of unresolved PatchPatrol-owned inline discussions before reposting feedback. |
AI_REVIEW_CODE_QUALITY_EXPORT | true | true or false | Controls export of gl-code-quality-report.json where applicable. |
Performance diagnostics
AI_REVIEW_ENABLE_PERFORMANCE_DIAGNOSTICS=true is the supported way to inspect
runtime behavior. It adds meta.performance_diagnostics to ai-review.json
and renders a diagnostics section in human reports. Use Generation tok/s for
Ollama decode speed, Prompt tok/s for prompt ingestion, Load time (s) for
cold-load cost, and Overhead (s) for Docker/network/wrapper overhead.
Wall total tok/s is blended prompt-plus-completion throughput and is not a
generation-speed benchmark.
Provider-native prompt/generation rates may render as n/a when an
OpenAI-compatible endpoint exposes token counts but not comparable timing
fields. Do not log raw prompts, completions, diffs, API keys, GitLab tokens, or
provider payloads for performance debugging; use artifact metadata instead.
See Artifacts Schema for exact fields and
Troubleshooting for interpretation.
Compatibility and Legacy Controls
AI_REVIEW_POST_MR_NOTEis a legacy alias preserved for backwards compatibility.- It is equivalent to
AI_REVIEW_FEEDBACK_MODE=mrwhen the explicit feedback mode is not set. - New deployments should prefer
AI_REVIEW_FEEDBACK_MODE.
- It is equivalent to
OLLAMA_TIMEOUT_SECONDSremains available as a backwards-compatible timeout alias for the Ollama path.AI_REVIEW_TIMEOUT_SECONDStakes precedence when both are set.
What is intentionally not public
- Internal defaults and implementation details beyond this page are intentionally outside the public contract.
- Any undocumented environment variable in these docs is treated as unsupported as a stable public guarantee, even if read internally for compatibility or transient implementation reasons.
When uncertain about a control, prefer the baseline configuration reference and treat absence from docs as intentional.
Return to Reference.