adept-at/Adept Triage Agent

AI-powered test failure triage to determine if failures are test or product issues

View on GitHub

Trust Signals

Scorecard Score: not yet scored
Maintenance Recency: Activelast commit Jul 10, 2026
License: MIT

Pinned Snippet

workflow.ymlSHA-pinned

uses: adept-at/adept-triage-agent@1d71ed283909ca56ef0e2e045048ab1011eec1c2 # v1.54.3

tags can be moved; commit SHAs can't. why a SHA?

Inputs

name	description	required	default
GITHUB_TOKEN	GitHub token for API access. Use a PAT or app token when REPOSITORY or AUTO_FIX_TARGET_REPO points to a different repo.	no	${{ github.token }}
OPENAI_API_KEY	OpenAI API key for AI analysis	yes	—
ERROR_MESSAGE	Error message to analyze (optional if using workflow artifacts)	no	—
ERROR_FILE	Test file path when using ERROR_MESSAGE directly (e.g. cypress/e2e/login.cy.ts). Required for grounded auto-fix in direct-error mode.	no	—
ERROR_TEST_NAME	Test case title when using ERROR_MESSAGE directly (e.g. should login successfully). Improves skill matching and validation scoping.	no	—
PERSIST_RESULTS	When false, skip DynamoDB skill/failure/outcome writes for this run. Outputs and artifacts are still emitted. Use for synthetic canary runs.	no	true
WORKFLOW_RUN_ID	Workflow run ID to fetch logs from	no	—
JOB_NAME	Specific job name to analyze	no	—
CONFIDENCE_THRESHOLD	Minimum confidence level for verdict (0-100)	no	70
PR_NUMBER	Pull request number to fetch diff from	no	—
COMMIT_SHA	Commit SHA associated with the test failure	no	—
BRANCH	Branch being tested (used to fetch branch diff when no PR number available)	no	—
REPOSITORY	App/source repository in owner/repo format for PR, branch, or commit diff lookup. Workflow runs and artifacts are still read from the repo where this action executes.	no	${{ github.repository }}
TEST_FRAMEWORKS	Test framework: "cypress" or "webdriverio".	no	cypress
ENABLE_AUTO_FIX	Enable automatic branch creation with fix. Consumer opts in by setting this to true.	no	false
AUTO_FIX_BASE_BRANCH	Base branch to create fix branch from	no	main
AUTO_FIX_MIN_CONFIDENCE	Minimum fix confidence required to apply auto-fix (0-100)	no	70
AUTO_FIX_TARGET_REPO	Repository where repair source files are fetched and fix branches are created (owner/repo format). Use this when test code lives in a different repo than REPOSITORY.	no	${{ github.repository }}
ENABLE_VALIDATION	Enable validation of auto-fixes before accepting them. Validation runs remotely by default (dispatches VALIDATION_WORKFLOW with VALIDATION_TEST_COMMAND). Set ENABLE_LOCAL_VALIDATION=true to run validation in-container instead.	no	false
ENABLE_LOCAL_VALIDATION	When true with ENABLE_AUTO_FIX, ENABLE_VALIDATION, and VALIDATION_TEST_COMMAND, runs the validation test locally in-container (clone, apply, test, push + PR only on pass, up to 3 iterations). When false (default), validation is dispatched to VALIDATION_WORKFLOW remotely.	no	false
VALIDATION_WORKFLOW	Name of the validation workflow file (e.g., validate-fix.yml). Used by the remote validation path when ENABLE_LOCAL_VALIDATION is false.	no	validate-fix.yml
VALIDATION_PREVIEW_URL	Preview URL for validation tests (if different from original). Replaces {url} in VALIDATION_TEST_COMMAND.	no	—
VALIDATION_SPEC	Spec file for validation tests (if different from detected). Replaces {spec} in VALIDATION_TEST_COMMAND.	no	—
VALIDATION_TEST_COMMAND	Test command template with {spec} and {url} placeholders. Used by BOTH validation paths — passed to the remote VALIDATION_WORKFLOW via workflow_dispatch, or evaluated locally when ENABLE_LOCAL_VALIDATION=true. Example: npx wdio run wdio.conf.ts --spec {spec} -t sauce	no	—
NPM_TOKEN	Token for npm registry auth (e.g., GitHub Packages read token). Falls back to GITHUB_TOKEN when not set.	no	—
PRODUCT_REPO	Product repository (owner/repo). All browser-test repos target the same frontend. Used for fetching recent product commits and PR diffs.	no	adept-at/learn-webapp
PRODUCT_DIFF_COMMITS	Number of recent product commits to include in the diff context (default: 5)	no	5
TRIAGE_AWS_REGION	AWS region for DynamoDB skill store.	no	us-east-1
TRIAGE_DYNAMO_TABLE	DynamoDB table name for skill store.	no	triage-skills-v1-live
MODEL_OVERRIDE_FIX_GEN	Optional override for the fix-generation agent model. When set (e.g. "gpt-5.5-pro"), fix-gen reverts to this model without requiring a new release. Reasoning effort automatically resets to 'none' when the override is not a gpt-5.5 model.	no	""
MODEL_OVERRIDE_REVIEW	Optional override for the review agent model. Same rollback mechanism as MODEL_OVERRIDE_FIX_GEN, independent. Effort auto-resets the same way.	no	""

Outputs

name	description
verdict	Classification result: TEST_ISSUE, PRODUCT_ISSUE, INCONCLUSIVE, PENDING, ERROR, or NO_FAILURE
confidence	Confidence score (0-100)
reasoning	Detailed explanation of the decision
summary	Brief summary for PR comments
triage_json	Complete triage analysis as JSON string
has_fix_recommendation	Boolean indicating if a fix recommendation was generated (TEST_ISSUE only)
fix_recommendation	Complete fix recommendation as JSON (when available)
fix_summary	Human-readable fix recommendation summary (when available)
fix_confidence	Confidence score for the fix recommendation (0-100)
auto_fix_applied	Whether the auto-fix completed successfully end-to-end (true/false). True only when both the fix was applied (branch created and pushed) AND validation passed in the same run. False when only a branch was created but validation failed, when validation passed but publish/apply did not complete, or when auto-fix was not attempted. The authoritative validation signal is `validation_status`.
auto_fix_branch	Name of the created branch (if auto-fix applied)
auto_fix_commit	Last commit SHA created while applying the fix (if auto-fix applied)
auto_fix_files	JSON array of modified files (if auto-fix applied)
validation_run_id	Workflow run ID of the validation workflow (legacy remote validation path only)
validation_status	Authoritative validation outcome. Set on both local and remote validation paths. "passed" means the fix was validated by the local or remote validation path independent of auto_fix_applied / push success — a "passed" status with auto_fix_applied=false indicates validation succeeded but publish/PR creation failed. Possible values: passed (local test succeeded or remote workflow concluded success with verified test evidence), failed (validation ran and reported a real failure), inconclusive (validation produced mixed/ambiguous evidence), pending (remote validation dispatched but not yet complete), or skipped (validation did not run, e.g., baseline check passed or no spec was identified).
validation_url	URL to the validation workflow run (legacy remote validation path only)
auto_fix_skipped	Whether auto-fix was intentionally withheld (true/false). Set to true when the agent decides not to apply a fix despite a TEST_ISSUE verdict — typically because the spec is chronically flaky and needs human investigation rather than another fallback fix. Verdict/confidence/reasoning still reflect the honest classification; only the auto-fix step is skipped.
auto_fix_skipped_reason	Human-readable reason when auto_fix_skipped is true (e.g., chronic flakiness message, recommendation for human follow-up).
repair_status	Repair pipeline lifecycle: not_started, skipped, in_progress, no_fix_generated, review_rejected, timed_out, cancelled, no_approved_fix, approved, applied, validated, validated_publish_failed (validation passed but publish/PR creation failed afterward — the fix was real, only delivery failed), validated_not_published (validation passed without a completed apply/publish; reserved for future ApplyResult shapes).
repair_summary	Human-readable repair outcome for Slack (orthogonal to classifier verdict)
repair_details	JSON object with iterations, lastStage, review issues, fix summary, timing
repair_iterations	Fix/review loop iterations completed in agentic repair
repair_last_stage	Last repair stage reached: analysis, code_reading, investigation, fix_generation, review
repair_review_issues	Newline-separated review issue lines when review rejected a fix (empty otherwise)