Skip to content

roadmap.md

Latest commit

 

History

History
95 lines (72 loc) · 4.36 KB

roadmap.md

File metadata and controls

95 lines (72 loc) · 4.36 KB
Raw

Roadmap

Entropy Loop Core aims to stay small and dependable while sharpening one idea: compiling agent failures into reusable assets. This roadmap is conservative and directional, not a commitment. It contains no private commercial detail.

v0.1.0 — Failure Compiler foundations

  • Typed data contract: Task, AgentOutput, VerificationResult, FailureTrace, Lesson, RetryContext, LoopResult, RegressionCase.
  • In-memory MemoryStore with failure history and relevant-lesson recall.
  • Deterministic, fluent Verifier.
  • Deterministic LessonGenerator (no LLM, no network).
  • generate_regression_case for test-like cases from failures.
  • EntropyLoop orchestrating verify → trace → learn → retry.
  • entropy-loop demo CLI, a worked example, tests, and docs.

v0.2.0 — Failure classification + policy + reproducible regression (current)

  • FailureCategory and richer VerificationResult (category, details).
  • VerificationPolicy + Verifier.from_policy — declarative rule configuration.
  • Deterministic, public-safe failure fingerprint on every FailureTrace.
  • Category-based LessonGenerator.
  • EvaluationSummary and summarize for run rollups.
  • Regression export (export_regression_case / export_regression_cases).
  • Improved entropy-loop demo and a new entropy-loop doctor.
  • docs/reliability-model.md and expanded docs.

v0.3.0 — Replay (current)

Theme: failures should not only be remembered — they should be replayed. Generated regression cases become a runnable suite so a remembered failure can be re-checked, not just stored. Shipped: RegressionSuite, RegressionRunner, RegressionRunResult, RegressionReport, suite import/export and JSON save/load, and entropy-loop replay-demo.

v0.4.0 — Memory policy

Decide what to remember, group, and forget. MemoryPolicy, LessonCompactor, LessonMemory, CompactionResult, memory import/export, and entropy-loop memory-demo — deterministic lesson deduplication (by fingerprint or category), minimum-occurrence and count caps, and drop-empty hygiene.

v0.5.0 — Regression pack + CI gate

Turn agent failures into portable CI checks. RegressionPack, RegressionPackRunner, JSON/JUnit reports, and entropy-loop run-pack with stable exit codes (0/1/2) so a reappearing regression fails the build.

v0.6.0 — Agent adapter + live pack refresh

Refresh a pack from a live local agent, explicitly. AgentCommand, CommandAgentAdapter, RegressionPackRefresher, and entropy-loop refresh-pack ... -- <command> — run your agent per case, capture fresh outputs, then gate with run-pack. Local-only, no shell by default, no secret injection.

v0.7.0 — Regression triage + baseline diff

Explain what changed between two runs. CaseTransition, RegressionTriage, TriagePolicy, RegressionTriageEngine, JSON/Markdown triage reports, and entropy-loop compare-reports BASELINE CURRENT --fail-on new-failures — classify each case as newly failing, fixed, persistent, skipped, or missing, and fail CI only on new regressions. Deterministic and local-only; no network, no GitHub API. The run-pack JSON report now carries a backward-compatible case_results list.

v0.8.0 — GitHub Action + CI evidence bundle (in progress)

Make adoption a few lines of YAML. CIEvidenceBundle, CIEvidenceWriter, GitHub Actions step-summary helpers, entropy-loop ci-demo, entropy-loop write-ci-evidence, and a first-party composite Action (action.yml) — compare reports, write a deterministic local evidence bundle, and append a step summary. No GitHub API, no PR comments, no artifact uploads, no GITHUB_TOKEN. No version bump until a release is cut.

Later / not now

Deferred until a real, public use case justifies the added surface:

  • async support,
  • a custom verifier / rule registry,
  • persistence adapters (file, then others) as a simple optional interface,
  • richer evaluation reports,
  • framework and CI integrations.

Non-goals

  • Proprietary or vendor-specific agent logic.
  • LLM or network calls inside the core.
  • Databases, vector stores, or embeddings in the core.
  • Heavyweight orchestration, web servers, dashboards, or UI frameworks.
  • Private commercial roadmap details of any kind.

Commercial products may build private policies, datasets, dashboards, and deployment workflows on top of this open-source core — those live elsewhere, not in this repository.