K Fleet v3.3.2

K Fleet now defaults to suitable mainstream libraries and frameworks for general-purpose capabilities and checks existing custom mechanisms during design assessment and code review, even when they have no reported failures.

Changes

• Design and implementation research ecosystem options before selecting or retaining custom mechanisms. Routine technical decisions remain autonomous and current selection evidence is reused.
• Verification and the optional read-only reviewer distinguish defects, supported engineering improvements, and unverified leads. Improvement findings include a concrete alternative, evidence, expected benefits, and adoption costs.
• Test writing applies the same policy to general-purpose test tooling. Review findings do not authorize unrelated migrations or automatically block release.
• Updated method boundaries, setup reminders, and English/Chinese documentation. No new public skill or required agent was added.

Update

After npm publication, run npx --yes k-fleet@3.3.2 update, or use npx --yes github:KeterVM/k-fleet update. Add --all to update registered projects. Explicitly request kf-setup to refresh project reminders (CLI/IDE: $kf-setup).

Skills are downloaded from the repository's default branch; pinning the CLI version does not pin skill contents. GitHub Release creation and npm publication are separate actions.

Validation

Checked metadata for five affected skills, local links, instruction consistency, reviewer TOML and read-only/model-neutral settings, setup's explicit invocation policy, both READMEs, CLI syntax, and git diff --check. Inspected npm pack --dry-run contents (eight expected files, including reviewer configuration).

These checks establish source and packaging consistency, not skill effectiveness. Actual implementation and review behavior has not been evaluated; no evaluation harness was added.

K Fleet v3.3.1

K Fleet v3.3.1 corrects installation, update, and skill-discovery guidance in the READMEs and CLI messages.

Changes

• Refresh the English and Chinese READMEs against the current skills and CLI.
• Clarify Node.js, npm/npx, Git, and network prerequisites, including when launching with Bun.
• Explain the differences between install, update, and status, and between npm's latest tag, the CLI's skill catalog, and skill contents fetched from GitHub.
• Replace mandatory restart instructions with conditional troubleshooting. Correct explicit skill invocation examples to $kf-setup in the CLI/IDE.
• Remove unsupported model and memory-plugin recommendations.

Updating

Use the current GitHub CLI:

npx --yes github:KeterVM/k-fleet update

After the maintainer publishes the npm package:

npx --yes k-fleet@3.3.1 update

Add --all to update registered projects. Codex detects skill changes automatically; restart it if an update does not appear. Explicitly request kf-setup to refresh managed project reminders. Installation and update preserve AGENTS.md.

This GitHub Release does not publish the npm package. CLI versions determine the requested skill names; skill contents still come from the repository's default branch.

Validation

• Checked bilingual command examples, catalog agreement, local documentation links, and CLI syntax for the included changes.
• Checked version 3.3.1 and expected package files with an npm packaging dry run. No archive was generated and no npm publication was performed.

This patch changes documentation and CLI messages. These checks do not establish new evidence of skill effectiveness.

K Fleet v3.3.0

K Fleet v3.3.0 extends the skills catalog across the product lifecycle and strengthens engineering decision guidance.

Changes

• Add kf-discover-product, kf-design-experience, kf-release-product, kf-operate-product, and kf-evaluate-product, with conditional references for each method.
• Extend the CLI catalog and setup reminders to include the new methods. Update English and Chinese documentation.
• Clarify unresolved user intent before dependent design, code, or tests, including fixes and refactors.
• Strengthen responsibility boundaries, naming, file placement, dependency selection, and complete migrations. Retained compatibility must protect an established contract or transition need.
• Remove fixed catalog-size wording and agent orchestration instructions. Preserve method applicability, authorization boundaries, and completion evidence.

Updating

From a target repository, use the current GitHub CLI:

npx --yes github:KeterVM/k-fleet update

After the maintainer publishes the npm package, the versioned npm command will also be available:

npx --yes k-fleet@3.3.0 update

Add --all to update registered projects. Restart Codex after updating, then explicitly run /kf-setup if you want to refresh managed project reminders. Installation and update preserve AGENTS.md.

The CLI version determines the requested skill names; skill contents are downloaded from the repository's default branch. This GitHub Release does not publish the npm package.

Validation

• Skill frontmatter, names, internal references, local documentation links, setup invocation policy, and reviewer TOML checked during the included changes.
• Actual CLI installation and update checked against local source in an isolated temporary target, including skill contents, lock entries, reviewer configuration, registration, and preservation of existing project guidance.
• Release CLI syntax and help checked; npm packaging dry run confirmed version 3.3.0 and the expected package files. No package archive was generated and no npm publication was performed.

These checks establish structural and installation behavior. Effectiveness on real product tasks remains unproven.

K Fleet v3.2.0

K Fleet 3.2.0 removes the external memory requirement from its engineering methods and setup reminders. You can select other skills and memory plugins for your project.

Changes

• Removed memory-provider names, configuration, connection checks, and memory operations from public skills and managed reminders.
• Refined skill selection and reference loading. The seven public skill names remain unchanged, and setup still requires an explicit user request.
• Added a Chinese README and simplified both READMEs. They explain the core ideas, helper skill sources, and recommendations for other skills and plugins.
• Added a repository-local research skill for maintainers. Consumer installation does not include it.
• Corrected outdated security documentation and added release maintenance instructions.
• Included the Chinese README and changelog in the npm package file list.

Update

From the target repository, use the CLI from this release:

npx --yes 'github:KeterVM/k-fleet#v3.2.0' update

Restart Codex. Run /kf-setup explicitly to replace old managed reminders that require a memory provider. Setup preserves separate project instructions.

The CLI downloads skills from the repository's default branch. Selecting a CLI version does not pin the downloaded skill revision. This GitHub Release does not publish the npm package.

Checks

• Checked all seven public skill names and frontmatter, 18 internal references, local documentation links, and setup's explicit-only policy.
• Checked CLI syntax, package contents, and matching README commands and link targets across languages.
• These checks cover structure and packaging. They do not establish better engineering outcomes or memory-plugin interoperability.

Full changes: v3.1.0...v3.2.0

K Fleet v3.1.0

K Fleet 3.1.0 clarifies the roles, expected results, engineering judgments, and completion boundaries of its existing methods. Detailed guidance now lives in 18 references selected by the task's open decisions.

• Requirements connect scope and acceptance conditions to the complete user outcome.
• Codebase design covers domain modeling, contracts, cross-module behavior, operational concerns, and evidence for tradeoffs.
• Implementation emphasizes complete integrated behavior, maintainability, and proportionate self-checks.
• Test writing covers meaningful assertions, dependency fidelity, isolation, concurrency, and regression protection.
• Verification selects environments and tools for the affected deliverable, distinguishes defects from uncertainty, and scopes correction checks to the affected behavior.
• Skill evolution uses conditional discovery, authoring, and actual-use assessment paths. Setup permits delegation when useful and authorized.

The seven public skill names, explicit-only setup, Supermemory authority and isolation rules, and CLI installation contract remain unchanged.

All seven skills passed structural validation, and all 18 internal references resolve within their owning skills. The revised instructions have not yet been assessed in representative software tasks; these checks do not establish improved behavior.

Update an existing project with npx --yes k-fleet@latest update, then restart Codex. Run /kf-setup explicitly if you want to refresh the project's managed reminders.

K Fleet v2.1.1

K Fleet now loads deeper engineering guidance when a task has unresolved or cross-layer risks. Routine local changes reuse established context, while implementation and design retain module ownership and necessary supporting behavior.

This release also includes the orchestration prompt and shared engineering-method improvements committed since v2.1.0. The package retains one public orchestrator and the optional read-only reviewer.

Validation: five CLI tests, 22 fixture tests, structure checks, 42-case corpus lint, skill validation, and evidence-binding checks passed. Four targeted candidate observations and one original-source presentation control completed.

Known validation limitation: the module-placement fixture was blocked by automatic approval review before any writes. The maintainer explicitly authorized this release with that observation incomplete. The recorded result remains blocked, not passing; the release waiver is bound to the exact source and corpus. See evals/PROGRESSIVE_DISCLOSURE_FORWARD_TEST_REPORT.md for evidence and limits. No general efficiency or full behavioral-pass claim is made.

K Fleet v2.1.0

What changed

• Require a connected, repository-scoped Supermemory runtime before setup writes.
• Add a shared Engineering contract checkpoint for cross-boundary implementation,
  bug-fix, and refactor work.
• Require observable ownership, lifecycle, alternative, failure, deletion, and test
  decisions before production mutation.
• Strengthen design and verification checks so static analysis alone cannot prove
  changed behavior.
• Add a realistic cross-boundary calling regression case to the orchestration corpus.

Validation

• 30 orchestration corpus cases linted.
• 6 source-bound isolated blind release cases passed, including the new engineering
  checkpoint case.
• 22 Fleet Ledger fixture tests passed.
• Repository structure, forward-result hashes, skill validation, and GitHub CI passed.

K Fleet v2.0.0

K Fleet 2.0 is a hard cutover to one portable orchestrator.

Highlights:

• One installable kf-orchestrate-work skill with progressive workflow disclosure.
• Explicit /kf-orchestrate-work setup for safe target AGENTS.md bootstrap.
• Supermemory owns the complete scoped memory lifecycle and automatic capture.
• SkillOpt-Sleep remains an external offline optimizer with memory evolution disabled for K Fleet.
• Four source-bound Codex release smoke cases and deterministic CI validation.

Breaking change: the v1 standalone workflow, context, delegation, learning, and feedback skills are removed without compatibility shims.

Full Changelog: v1.5.0...v2.0.0

K Fleet v1.5.0

This minor release adds role-based sub-agent delegation, optional read-only companion agents, post-work skill-usage feedback, progressive entry points, and a bound cross-model routing robustness matrix while keeping eleven core workflow skills.

Highlights

• Replaces specialist-only delegation with kf-delegate-subtask, covering reviewer, verifier, reasoner, explorer, worker, and specialist roles with isolated write ownership.
• Adds the optional kf_reviewer and kf_context_auditor companion agents, with installation instructions alongside the skills.sh workflow.
• Adds kf-report-skill-usage and a sanitized, provenance-aware feedback packet contract.
• Adds progressive references and installation profiles for delegation, verification, and Context maintenance.
• Adds a 24-case multilingual naturalistic robustness corpus and an 18-run, 240-observation blind matrix with evidence binding, stopping contracts, model-specific promotion gates, and tamper checks.
• Keeps the compact catalog-description candidate unpromoted because repeated high-risk contract failures remain model-specific.

Validation

• 11 core skills, 1 feedback skill, and 2 companion agents pass repository structure validation.
• 56 core eval cases, 24 robustness cases, and 7 feedback cases pass corpus validation.
• 18 runs, 240 observations, 240 stopping judgments, and 480 invariant judgments validate with 0 excluded observations.
• All 12 skill directories pass the installed Codex skill validator.
• All 22 Fleet Ledger tests pass.
• GitHub Actions CI run #43 passed on commit dc03554.

See CHANGELOG.md and the routing robustness matrix report for details.

K Fleet v1.4.1

K Fleet v1.4.1

This patch release adds an isolated bounded-quality execution experiment without changing the eleven-skill public contract.

Highlights

• Compares the current workflow, a stronger single-owner quality contract, and an experimental sole-writing expert topology.
• Adds four artifact scenarios, hidden behavioral checks, eight routing cases, sanitized executor bundles, and a protocol-v2 runner.
• Tightens the candidate entry gate so routine local work does not create an agent tree.
• Hardens scoring against fixture leakage, baseline and arm drift, hidden-test mutation, stale identities, duplicate runs, result tampering, and structural-signal false positives.
• Records the first smoke run honestly as instruction-blind, non-promotional evidence.

Validation

• 11 public skills pass repository validation.
• 51 existing eval cases pass corpus validation.
• 8 bounded-quality cases and 4 artifact scenarios pass harness validation.
• All 22 Fleet Ledger tests pass.
• Independent final review found no actionable findings.

See CHANGELOG.md and evals/BOUNDED_QUALITY_PILOT_REPORT.md for full details.