Diagnostics Rules & levers Pipelines Executions & journal Trace table Metrics ModelsMonths: 2026-11 2026-10

Rules and levers

Which setting or prompt governs the pattern each diagnostic rule reports (LEVERS and NO_LEVER in scripts/lib/meta_levers.mjs, ADR-0047): this page is generated from that table. A lever says where to look, not what caused a finding. A Diagnostics card shows it only for an active warning or critical finding of a rule below; an info finding (a guardrail working as designed) or a resolved one has none. When the rule fires and what counts as an anomaly: Rules and thresholds.

Rules with a lever

Rule	Stage	What it governs	Where to look, and what a change requires	Decisions
`review-withheld`	review	Reviews are withheld on code the prompt budget or the CI evidence leaves unseen	`config/models.yaml#review_max_input_tokens` · run the live evals.yml (suite review) `config/models.yaml#anthropic_review_max_diff_chars` · run the live evals.yml (suite review) `AI_PROVIDER_REVIEW` · run the live evals.yml (suite review) `config/review-evidence.yaml#checks` · run the live evals.yml (suite review) `prompts/pr-review-system.md` · run the live evals.yml (suite review)	ADR-0026, ADR-0038, ADR-0049
`review-coverage`	review	Reviews are cut or withheld on a prompt budget that does not fit the diffs	`config/models.yaml#review_max_input_tokens` · run the live evals.yml (suite review) `config/models.yaml#anthropic_review_max_diff_chars` · run the live evals.yml (suite review) `AI_PROVIDER_REVIEW` · run the live evals.yml (suite review)	ADR-0028, ADR-0038
`autofix-ineffective`	autofix	Auto-fix attempts end without a delivered push	`config/models.yaml#autofix_max_input_tokens` · no eval suite measures it: verify by hand and say so in the PR `config/models.yaml#autofix_diff_ratio` · no eval suite measures it: verify by hand and say so in the PR `config/models.yaml#autofix_feedback_ratio` · no eval suite measures it: verify by hand and say so in the PR `prompts/auto-fix-system.md` · no eval suite measures it: verify by hand and say so in the PR	ADR-0017, ADR-0029
`conflicts-unresolved`	conflicts	Conflicts-fix attempts end without a merge commit	`config/models.yaml#conflicts_max_input_tokens` · no eval suite measures it: verify by hand and say so in the PR `AI_PROVIDER_CONFLICTS` · no eval suite measures it: verify by hand and say so in the PR `prompts/conflicts-system.md` · no eval suite measures it: verify by hand and say so in the PR	ADR-0017, ADR-0059
`codegen-blocked`	generation	Code generation is blocked before it proposes a patch	`config/models.yaml#generation_max_input_tokens` · no eval suite measures it: verify by hand and say so in the PR `prompts/generation-system.md` · no eval suite measures it: verify by hand and say so in the PR `prompts/validation-system.md` · run the live evals.yml (suite validation)	ADR-0029
`llm-fallbacks`	provider	The primary provider fails often enough that the fallback answers	`AI_PROVIDER_<STAGE>` · no eval suite measures it: verify by hand and say so in the PR `retry policy (ADR-0022)` · no eval suite measures it: verify by hand and say so in the PR `provider tier` · no eval suite measures it: verify by hand and say so in the PR	ADR-0022, ADR-0025

Every change also needs docs and tests in the same pull request, and is applied by a person: the dashboard and /loop-insights create nothing on GitHub.

Rules without a lever

Rule	Why there is none
`review-duplicates`	fixed in .github/ workflow triggers, which no lever may name (ADR-0046 §5); the ADR-0028 marker is code, not a setting
`runs-failed`	a failed run needs a root cause (docs/runbook.md), not a setting
`stage-failures-*`	a repeated stage failure needs a root cause (docs/runbook.md), not a setting
`review-stuck`	depends on the content of the pull request
`escalations`	a person is the lever (ADR-0001)
`external-review-churn`	a practice (AGENTS.md, PR Review Comment Workflow) and the size of the pull request, not a setting; the reviewer is external
`rate-limit-waits`	a symptom of the provider tier; revisit with data
`duration-*`	a symptom of the provider tier; revisit with data

A lever never points at

a workflow or token scope (.github/)
the protected-paths floor (ADR-0044)
an eval threshold, dataset or floor (ADR-0031)
the eval suites (ADR-0031)
the auto-fix attempt cap (ADR-0034)
a secret or credential

ADR-0046 §5: no lever lowers a threshold, removes a dataset case or weakens a gate.