CI Retry Gate v1.2.0 — State-Bound Reruns with Evidence-Gated Safety

CI Retry Gate v1.2.0

CI Retry Gate now binds rerun authorization to the exact workflow state that produced the evidence.

This release strengthens the core principle:

Evidence before action.

Highlights

• Added EASL-compatible subject-state binding for pre-mutation validation.
• Full-workflow reruns now revalidate the exact:
  • repository
  • run ID
  • run attempt
  • head SHA
  • workflow ID
  • failed-job set
• Selective reruns now bind authorization to:
  • run attempt
  • head SHA
  • workflow ID
  • workflow lifecycle
  • exact failed-job execution state
• Selective mode performs at most one mutation per evaluated workflow-state epoch.
• Any remaining candidate must be re-evaluated after the first mutation changes workflow state.
• Added canonical EASL subject-state conformance vectors for the Python compatibility layer.
• Added external consumer E2E coverage for the state-bound selective-rerun path.

Safety

• Workflow-state drift invalidates old rerun justification before mutation.
• Invalid or unavailable state bindings fail closed.
• Mutating GitHub API POST requests are not blindly replayed after ambiguous transport outcomes.
• Automatic reruns remain disabled by default.
• Causal Dominance remains research-only and does not participate in production retry authority.

External verification

A real external GitHub Actions consumer test demonstrated:

1. Two failed jobs qualified as selective-safe candidates.
2. Exactly one rerun mutation was issued from the evaluated state epoch.
3. GitHub advanced the workflow to a new attempt.
4. Exactly one job received a new execution.
5. The other job was carried forward without a new execution timestamp.
6. The new attempt was re-evaluated before any further write.
7. No second mutation was authorized from the old state epoch.

The published @v1 line was also exercised from an external consumer against both ALLOW and fail-closed UNKNOWN evidence contracts.

Install

Use the stable major release:

- uses: achirothmane/workflow-failure-lab@v1

Or pin this exact version:

- uses: achirothmane/workflow-failure-lab@v1.2.0

Automatic reruns remain explicit opt-in.

CI Retry Gate v1.1.0

CI Retry Gate v1.1.0

This release introduces the Evidence Producer and a verifiable runtime evidence boundary for CI retry authorization.

Highlights:

• Policy-free Evidence Producer
• Canonical evidence-producer.ci.v1 EvidenceBundle
• Canonical evidence.json runtime artifact
• SHA-256 integrity verification before authorization
• Separate-process Evidence Gate
• New outputs:
  • evidence-bundle-path
  • evidence-bundle-sha256
• Missing, malformed, conflicting, or tampered evidence fails closed to BLOCK
• Retry policy remains separate from evidence
• Recovery and recurrence checks remain conservative outer guards

Verification:

• 353 Python tests passed
• CI passed
• Compatibility Matrix passed
• Remote v1 Consumer E2E passed
• Clean-install verification passed
• Release Readiness passed

Automatic reruns remain disabled by default.

CI Retry Gate v1.0.2

CI Retry Gate v1.0.2

Evidence-gated CI retry decisions are now backed by a broader real-world regression contract and external consumer verification.

Highlights

• Added recovery and recurrence evidence across workflow attempts.
• Kept failure-cause evidence separate from retry-outcome evidence.
• Added cutoff-safe historical reliability evidence that remains supporting-only, never authorizing by itself.
• Preserved fail-closed behavior when public GitHub job logs are unavailable.
• Added bounded evidence acquisition with explicit rate-limit and evidence-budget states.
• Strengthened regression coverage using real-world CI failure patterns.
• Improved zero-install Public Run Analysis for evaluating failed public GitHub Actions runs.
• Updated the README around a faster zero-install proof path.
• Hardened Node/Jest/Vitest compatibility with isolated locked dependency graphs and a verified npm resolver.
• Fixed GitHub Action Marketplace metadata and stable @v1 consumer validation.

Release verification

Before promotion to the stable v1 line:

• the full producer test suite passed;
• remote @v1 compatibility checks passed;
• the release candidate was exercised from an independent consumer repository;
• root Action, pytest, Jest, Vitest, Setup Doctor, and ownership/JUnit checks passed externally.

Automatic reruns remain disabled by default.

Install

uses: achirothmane/workflow-failure-lab@v1

CI Retry Gate v1.0.1

CI Retry Gate v1.0.1

Marketplace-ready patch release for CI Retry Gate.

Changed

• Shortened the GitHub Action description to satisfy the GitHub Marketplace metadata limit.
• Added a CI guard to prevent future Marketplace description regressions.
• Updated stable installation guidance for the @v1 release line.

CI Retry Gate

CI Retry Gate analyzes failed GitHub Actions jobs before a rerun is allowed.

It helps prevent blind retries by using:

• failure classification
• execution provenance
• failure-step provenance
• side-effect detection
• retry-attempt limits
• selective rerun safety gates
• recurring-failure fingerprints
• CI waste analysis

Safety by default

Automatic reruns remain disabled by default.

auto-rerun: 'false'
selective-rerun: 'false'

Users must explicitly opt into rerun behavior.

Installation

- uses: othy19904-eng/workflow-failure-lab@v1
  with:
    github-token: ${{ github.token }}
    auto-rerun: 'false'
    selective-rerun: 'false'

Marketplace

Primary category: Continuous integration

Secondary category: Utilities

Research features

Causal Dominance remains research-only and is not enabled in production retry authority.

Current independent real positive-control evidence: 2/3.

License

MIT License.

CI Retry Gate v1.0.0

CI Retry Gate v1.0.0

First production-ready release of CI Retry Gate.

CI Retry Gate analyzes failed GitHub Actions jobs before a rerun is allowed, helping teams avoid blindly rerunning deterministic failures or workflows with side effects.

Highlights

• Conservative CI failure classification
• Safe retry gating for transient infrastructure and dependency/network failures
• Selective rerun of individually safe failed jobs
• Execution Provenance to bind transient evidence to the actual failed step
• Recovery Ground Truth based on real step re-execution
• Failure fingerprints and recurring-failure analysis
• CI waste measurement
• Conservative retry-policy learning
• Policy Shadow Mode
• Cross-repository Benchmark Mode
• Coverage Attribution and Rejection Intelligence
• UNKNOWN failure research and promotion gates
• Side-effect detection and retry-attempt limits

Safety by default

Automatic reruns are disabled by default.

auto-rerun: 'false'
selective-rerun: 'false'

Users must explicitly opt into rerun behavior.

Deploy, publish, migration, release, and other side-effect signals remain blocked by the safety gate.

Causal Dominance research

The Causal Dominance work remains research-only and is not enabled in production retry authority.

Current independent real positive-control evidence: 2/3.

This does not block the v1 release because the experimental override remains disabled.

Installation

- uses: othy19904-eng/workflow-failure-lab@v1
  with:
    github-token: ${{ github.token }}
    auto-rerun: 'false'
    selective-rerun: 'false'

License

MIT License.

workflow-failure-lab v0.1.0

First public release of Workflow Failure Lab.

Highlights:

• Deterministic Markdown incident reports from local workflow execution logs
• Three-state certainty model: SUCCEEDED, FAILED_CONFIRMED, INDETERMINATE
• Fail-closed retry and reconciliation guidance
• Stable CLI and exit-code contract
• Dependency-direction and import-cycle enforcement
• Clean-install release verification gate
• GitHub Actions CI