Changelog

One entry per release: what changed, and what it means for how you work. Full commit-level history lives on GitHub.

0.15.0 — 2026-09-12

The plan-work release: replace separate planning entry points with one adaptive planner that owns a coherent mission from outcome through coverage, decisions, feasibility, and verification.

• Coherent delivery units and epics — shape-work now turns a settled mission into independently deliverable units or an explicit epic contract with shared acceptance, dependencies, aggregate verification, and ownership.
• Continuation, tracker, and scope boundaries — approved authority continues across workflow boundaries without repeating settled decisions; tracker maintenance remains mandate-bound, and scope handling pauses only affected branches while independent authorized work continues.
• Executable behavior contracts — behavior fixtures and scorecards make planning depth, continuation, tracker non-mutation, epic readiness, and aggregate completion observable and testable.

0.14.0 — 2026-09-06

The project-controls release: quality-ratchet can run explicit repository-owned quality commands and reuse successful static evidence only when its declared inputs remain unchanged.

• Repository-owned controls — configure existing architecture, lint, type, or test commands in .agent-os/quality.json; Agent OS does not install or infer analyzers.
• Explicit delivery gates — controls remain advisory by default, while project-selected required checks block on failures, unavailable tools, and timeouts.
• Conservative evidence reuse — opt-in static caching fingerprints repository state, declared tool inputs, configuration, runtime, executable, and environment before reusing a pass.
• Auditable policy changes — changed quality policy must be restored or acknowledged with a task-grounded reason before checks continue.
• Exception deltas — new type escapes, lint suppressions, and skipped or focused tests are reported relative to the entry state and gate only when the project explicitly requires them.

0.13.1 — 2026-09-06

Public release verification now installs the exact npm package into a temporary isolated prefix and invokes its installed CLI directly through Node, avoiding npm 11's silent npm exec no-op.

0.13.0 — 2026-09-06

The proportional-delivery release: make implementation structure, review, continuation, prototyping, and validation decisions explicit and evidence-calibrated.

• Structure and contract ownership — guidance before implementation across TypeScript/Zod, React, and Java, keeping contracts and ownership explicit.
• Review and continuation — evidence-calibrated findings, targeted re-review, and next-action recommendations; sequential continuation only under existing explicit authority, with batch-work still opt-in.
• Prototype decisions — bounded prototype and experiment guidance for deciding when to learn before committing to implementation.
• Cross-boundary verification — proportional testing across boundaries and clarified quality-ratchet evidence limits.

0.12.1 — 2026-08-27

Public release verification now uses explicit npm package and command selection, compatible with npm 11.

0.12.0 — 2026-08-27

The check-work release: review completed work independently without confusing review, testing, or delivery authority.

• New workflow check-work — triggers for diffs, commits, branches, pull requests, and completed implementations, while excluding test-only, status-only, initial-design, and repository-wide audit requests.
• Report and fix modes — report mode is read-only; fix mode applies only supported in-scope findings, verifies the result, and obtains targeted independent re-review.
• Findings-first contract — findings include priority, exact location, evidence and impact, and required change, with explicit APPROVED, CHANGES_REQUESTED, or BLOCKED outcomes.

0.11.0 — 2026-08-07

The proportional-testing release: protect changed behavior without turning every small fix into a large test-generation exercise.

• New discipline proportional-testing — inspects existing coverage and requires a realistic, previously unprotected regression before adding another test.
• Behavior-first test selection — prefers existing and table-driven tests, observable contracts, and the cheapest test level that provides sufficient confidence.
• Focused bug-fix loop — establishes one reproducing regression check when practical, then runs relevant affected tests instead of generating unrelated cases.
• Delivery integration — deliver-work applies proportional testing during implementation while preserving every required repository, CI, and release check.

0.10.0 — 2026-08-06

The simplification release: remove unnecessary solutions without turning readable code into code golf.

• New workflow simplifier — reduces code, dependencies, abstractions, and solution layers while preserving required behavior, readability, safety boundaries, and useful tests.
• New workflow simplifier-audit — scans a repository for high-confidence simplification opportunities without applying changes.
• New discipline simplifier-review — reviews the current diff for unnecessary complexity and is now part of deliver-work before a candidate is frozen.
• New reference skill list-skills — reports the installed Agent OS skills and their invocation boundaries without depending on a hardcoded inventory.
• Codex experiment documentation — records the tested custom-instruction setup, mutation and rollback procedure, observed results, and remaining limitations.

0.9.1 — 2026-08-02

The portable-installer correction: the npm CLI now owns skill installation instead of requiring Codex or Claude Code to install themselves.

• Direct by default — installs the packaged skills into the selected user or project skill directories with no host CLI requirement.
• Managed updates — an Agent OS manifest records exactly which skill directories the installer owns; updates replace those directories and preserve unrelated skills.
• Portable policy sync — Node now applies the managed policy block on every supported platform, with malformed-marker preflight before any mutation.
• Native plugin option — --method plugin retains marketplace installation when explicitly selected and skips Git refresh for local Codex marketplaces.

0.9.0 — 2026-08-02

The guided installer release: agent-os can now be installed and refreshed through npm without a repository checkout.

• New npm CLI — npx @sockulags/agent-os install guides Codex and Claude Code setup, including Claude install scope and optional shared-policy synchronization.
• Updates — npx @sockulags/agent-os update refreshes the selected host plugin and can resync the shared policy; explicit flags support repeatable automation.
• Validation — the package manifest, CLI tests, package contents, plugin manifests, and documentation version copies are checked together.

0.8.0 — 2026-07-31

The planning-entry release: agent-os gained an entry point for work whose outcome is not yet clear.

• Earlier planning entry — from a vague desire, through questioning, to a plain-language summary that could guide the next planning step. The historical entry continued into decision mapping or shaping, carrying the summary as a ## TLDR at the top of the artifact it created.
• New workflows understand-work and explain-work — the questioning loop and the plain-language gate as standalone skills, composable by the planning entry and invocable on their own against any decision, plan, diff, or pull request.
• New discipline notice-lesson — a developer interruption or correction is treated as a misunderstanding signal: named in one sentence, course-corrected, and offered to record-lesson when the lesson is durable. It never edits policy itself.
• New workflow record-lesson — records one durable lesson at the narrowest level where it holds: the repo's project policy, or the global policy.md behind the managed blocks.
• Policy: one continuation exception — a workflow may continue into the workflow its exit contract names once the developer approves that exit.
• Docs: verified Codex install commands (codex plugin marketplace add, codex plugin add), this changelog, a troubleshooting page, a landing-page session transcript, transcripts on more skill pages, and a planning prologue in the worked example.
• Validation: documentation prose may no longer hardcode skill counts — a validator check with a red case replaces the reminder that failed.

0.7.0 — 2026-07-31

Delivery-ready issues. Shape-work must materialize implementation-ready issues before a branch counts as delivery-ready; the ready frontier belongs to the developer, and issue count never selects an execution workflow.

0.6.2 — 2026-07-30

Independent review gate in deliver-work: one adversarial reviewer by default for material change, a precisely qualified small-fix exception, and a hard stop before mutation when the host exposes no real reviewer launch tool.

0.6.1 — 2026-07-30

Evidence-backed documentation: session transcripts on discipline pages and honest eval reporting that excludes unmeasurable cases from the denominator instead of hiding them.

0.6.0 — 2026-07-30

Contract-driven workflows. Approval checkpoints and receipt ceremony were removed; authority now follows the request, and workflows are governed by outcome, boundaries, and ground truth.

0.5.0 — 2026-07-29

Batch orchestration: the batch-work workflow with a coordinator-owned manifest, hash-based definition-drift detection behind a deterministic script, isolated worker worktrees, and aggregate verification that worker-local checks can never substitute for.

0.4.x — 2026-07-28 → 2026-07-29

Chart-to-shape handoffs with stable identity and idempotent retries (0.4.0–0.4.1), prototype evidence records, and the structural validator with its red-case suite (0.4.2).

0.3.0 — 2026-07-28

The earlier decision-map workflow — broad efforts became parallel graphs of evidence-backed decision tickets — and this documentation site.

0.2.0 — 2026-07-22

Initial public release: the workflow and discipline skill set shared between Claude Code and Codex, dual plugin manifests, and the managed global policy block.