Self-Evolution v2.1.0-rc.4

Self-Evolution v2.1.0-rc.3

This release candidate adds guarded knowledge edits and portable task
continuation, improves lifecycle/source review, and introduces an executable
engineering evaluation suite. It is published for evaluation; stable-release
model-effectiveness evidence remains pending.

Changes

• Resolve archived Decision IDs while keeping current retrieval limited to
  active Guides and accepted Decisions. Detect invalid supersession and compare
  Guide routes by scope, purpose and consumer conditions.
• Report affected source paths, baselines and review limits in text/JSON output.
• Add kb write <document> <proposal> <expected-sha256|absent> with participating
  Windows/Linux process locks, conflict detection and interrupted-write recovery.
• Provide optional continuation/1 guidance and a template, including actual
  patch transfer, branch/base checks and stale verification handling.
• Focus standing Skill instructions on project engineering work. Initialization
  remains three files; adapters remain disabled by default.
• Add 24 executable A/B engineering scenarios, six baseline preparation paths,
  three-session refactor chains, eight curated public history pairs, subprocess
  transfer tests and raw evidence/cost derivation.
• Cover complete smoke workspaces, actual executable version/hash and valid
  knowledge schemas. Reject receipt replay, arm relabeling and missing controls.

Validation and evidence boundary

Publication requires clean-install CI, type/format checks, CLI tests, engineering
catalog/runner tests, legacy migration/evidence checks, and distribution bundle
verification on the released commit. GitHub Actions records Linux and Windows
results for that commit.

The offline suite uses deterministic fake CLI processes and synthetic fixtures.
It demonstrates implementation and orchestration behavior, not model quality or
automatic Capture benefit. Native-memory controls, fresh paired model campaigns,
independent review and real cross-harness engineering effects remain pending.
The stable standard, private and continuity gates retain their requirements.

Install and verify

npx skills add https://github.com/D1ChangGeng/self-evolution/tree/v2.1.0-rc.3 --skill self-evolution -g -y

Alternatively, verify SHA256SUMS and extract the attached ZIP. The
self-evolution/ directory contains the Skill, bundled CLI, references, license,
usage and migration documentation. Node.js 22+ is required for CLI operations:

node <skill-dir>/references/bin/kb.mjs init --project-root <project>
node <skill-dir>/references/bin/kb.mjs check --project-root <project>

Back up the current installation before replacing it. Restore that installation
to roll back the Skill; project knowledge is ordinary repository content and
retains its normal Git history. Existing v1-to-v2 migration/rollback support
remains part of the 2.x contract. Direct editors still use Git conflict handling;
guarded writes cover participating CLI entry points only.

Self-Evolution v2.1.0-rc.3

Self-Evolution v2.1.0-rc.3

This release candidate adds guarded knowledge edits and portable task
continuation, improves lifecycle/source review, and introduces an executable
engineering evaluation suite. It is published for evaluation; stable-release
model-effectiveness evidence remains pending.

Changes

• Resolve archived Decision IDs while keeping current retrieval limited to
  active Guides and accepted Decisions. Detect invalid supersession and compare
  Guide routes by scope, purpose and consumer conditions.
• Report affected source paths, baselines and review limits in text/JSON output.
• Add kb write <document> <proposal> <expected-sha256|absent> with participating
  Windows/Linux process locks, conflict detection and interrupted-write recovery.
• Provide optional continuation/1 guidance and a template, including actual
  patch transfer, branch/base checks and stale verification handling.
• Focus standing Skill instructions on project engineering work. Initialization
  remains three files; adapters remain disabled by default.
• Add 24 executable A/B engineering scenarios, six baseline preparation paths,
  three-session refactor chains, eight curated public history pairs, subprocess
  transfer tests and raw evidence/cost derivation.
• Cover complete smoke workspaces, actual executable version/hash and valid
  knowledge schemas. Reject receipt replay, arm relabeling and missing controls.

Validation and evidence boundary

Publication requires clean-install CI, type/format checks, CLI tests, engineering
catalog/runner tests, legacy migration/evidence checks, and distribution bundle
verification on the released commit. GitHub Actions records Linux and Windows
results for that commit.

The offline suite uses deterministic fake CLI processes and synthetic fixtures.
It demonstrates implementation and orchestration behavior, not model quality or
automatic Capture benefit. Native-memory controls, fresh paired model campaigns,
independent review and real cross-harness engineering effects remain pending.
The stable standard, private and continuity gates retain their requirements.

Install and verify

npx skills add https://github.com/D1ChangGeng/self-evolution/tree/v2.1.0-rc.3 --skill self-evolution -g -y

Alternatively, verify SHA256SUMS and extract the attached ZIP. The
self-evolution/ directory contains the Skill, bundled CLI, references, license,
usage and migration documentation. Node.js 22+ is required for CLI operations:

node <skill-dir>/references/bin/kb.mjs init --project-root <project>
node <skill-dir>/references/bin/kb.mjs check --project-root <project>

Back up the current installation before replacing it. Restore that installation
to roll back the Skill; project knowledge is ordinary repository content and
retains its normal Git history. Existing v1-to-v2 migration/rollback support
remains part of the 2.x contract. Direct editors still use Git conflict handling;
guarded writes cover participating CLI entry points only.

Self-Evolution v2.1.0-rc.2

Self-Evolution v2.1.0-rc.2

This release candidate adds guarded knowledge edits and portable task
continuation, improves lifecycle/source review, and introduces an executable
engineering evaluation suite. It is published for evaluation; stable-release
model-effectiveness evidence remains pending.

Changes

• Resolve archived Decision IDs while keeping current retrieval limited to
  active Guides and accepted Decisions. Detect invalid supersession and compare
  Guide routes by scope, purpose and consumer conditions.
• Report affected source paths, baselines and review limits in text/JSON output.
• Add kb write <document> <proposal> <expected-sha256|absent> with participating
  Windows/Linux process locks, conflict detection and interrupted-write recovery.
• Provide optional continuation/1 guidance and a template, including actual
  patch transfer, branch/base checks and stale verification handling.
• Focus standing Skill instructions on project engineering work. Initialization
  remains three files; adapters remain disabled by default.
• Add 24 executable A/B engineering scenarios, six baseline preparation paths,
  three-session refactor chains, eight curated public history pairs, subprocess
  transfer tests and raw evidence/cost derivation.
• Cover complete smoke workspaces, actual executable version/hash and valid
  knowledge schemas. Reject receipt replay, arm relabeling and missing controls.

Validation and evidence boundary

Publication requires clean-install CI, type/format checks, CLI tests, engineering
catalog/runner tests, legacy migration/evidence checks, and distribution bundle
verification on the released commit. GitHub Actions records Linux and Windows
results for that commit.

The offline suite uses deterministic fake CLI processes and synthetic fixtures.
It demonstrates implementation and orchestration behavior, not model quality or
automatic Capture benefit. Native-memory controls, fresh paired model campaigns,
independent review and real cross-harness engineering effects remain pending.
The stable standard, private and continuity gates retain their requirements.

Install and verify

npx skills add https://github.com/D1ChangGeng/self-evolution/tree/v2.1.0-rc.2 --skill self-evolution -g -y

Alternatively, verify SHA256SUMS and extract the attached ZIP. The
self-evolution/ directory contains the Skill, bundled CLI, references, license,
usage and migration documentation. Node.js 22+ is required for CLI operations:

node <skill-dir>/references/bin/kb.mjs init --project-root <project>
node <skill-dir>/references/bin/kb.mjs check --project-root <project>

Back up the current installation before replacing it. Restore that installation
to roll back the Skill; project knowledge is ordinary repository content and
retains its normal Git history. Existing v1-to-v2 migration/rollback support
remains part of the 2.x contract. Direct editors still use Git conflict handling;
guarded writes cover participating CLI entry points only.

Self-Evolution v2.1.0-rc.1

Self-Evolution v2.1.0-rc.1

This release candidate adds guarded knowledge edits and portable task
continuation, improves lifecycle/source review, and introduces an executable
engineering evaluation suite. It is published for evaluation; stable-release
model-effectiveness evidence remains pending.

Changes

• Resolve archived Decision IDs while keeping current retrieval limited to
  active Guides and accepted Decisions. Detect invalid supersession and compare
  Guide routes by scope, purpose and consumer conditions.
• Report affected source paths, baselines and review limits in text/JSON output.
• Add kb write <document> <proposal> <expected-sha256|absent> with participating
  Windows/Linux process locks, conflict detection and interrupted-write recovery.
• Provide optional continuation/1 guidance and a template, including actual
  patch transfer, branch/base checks and stale verification handling.
• Focus standing Skill instructions on project engineering work. Initialization
  remains three files; adapters remain disabled by default.
• Add 24 executable A/B engineering scenarios, six baseline preparation paths,
  three-session refactor chains, eight curated public history pairs, subprocess
  transfer tests and raw evidence/cost derivation.
• Cover complete smoke workspaces, actual executable version/hash and valid
  knowledge schemas. Reject receipt replay, arm relabeling and missing controls.

Validation and evidence boundary

Publication requires clean-install CI, type/format checks, CLI tests, engineering
catalog/runner tests, legacy migration/evidence checks, and distribution bundle
verification on the released commit. GitHub Actions records Linux and Windows
results for that commit.

The offline suite uses deterministic fake CLI processes and synthetic fixtures.
It demonstrates implementation and orchestration behavior, not model quality or
automatic Capture benefit. Native-memory controls, fresh paired model campaigns,
independent review and real cross-harness engineering effects remain pending.
The stable standard, private and continuity gates retain their requirements.

Install and verify

npx skills add https://github.com/D1ChangGeng/self-evolution/tree/v2.1.0-rc.1 --skill self-evolution -g -y

Alternatively, verify SHA256SUMS and extract the attached ZIP. The
self-evolution/ directory contains the Skill, bundled CLI, references, license,
usage and migration documentation. Node.js 22+ is required for CLI operations:

node <skill-dir>/references/bin/kb.mjs init --project-root <project>
node <skill-dir>/references/bin/kb.mjs check --project-root <project>

Back up the current installation before replacing it. Restore that installation
to roll back the Skill; project knowledge is ordinary repository content and
retains its normal Git history. Existing v1-to-v2 migration/rollback support
remains part of the 2.x contract. Direct editors still use Git conflict handling;
guarded writes cover participating CLI entry points only.

Self-Evolution v2.0.0

Self-Evolution 2.0.0 turns the skill into a lightweight, maintained project wiki for agentic engineering. The host harness owns task state, planning, tools, and permissions; the skill routes and verifies durable project knowledge, resolves stale or conflicting claims, and turns useful experience into scoped future actions. Codex, Claude Code, and OpenCode integration guidance is included, with optional adapters kept separate from the core skill.

The standard/core release gate passed against the frozen previous-v2 baseline and three paired public runs on 1302-1. LongMemEval cleaned full scored 90.8% for this release versus 88.2% for the baseline (+2.6 points). LongMemEval-V2 small scored 7.32% versus 9.31% (-1.996 points), within the declared 2-point overall tolerance; its low absolute score and narrow margin remain explicit limitations. All six read-only Codex, Claude Code, and OpenCode engineering execution/review samples passed. The 17,215-file evidence bundle can be validated offline from the repository.

Validation: npm run ci, npm run eval:external:test (127 tests), kb check, skill validation, and npm audit passed. GitHub CI passed on Ubuntu and Windows: https://github.com/D1ChangGeng/self-evolution/actions/runs/34706933619.

The private integrated campaign remains pending under the optional private profile. The V2 pilot uses text state evidence and bounded lexical trajectory selection; it does not establish screenshot-dependent CodexMemory compatibility or broad end-to-end engineering quality. The model endpoint exposed an alias rather than an immutable build, so the release gate uses three complete paired runs and per-question majority results.

Commit: f56bff3c24cafa1ca1b976d24f3b512d3bf0e5f4
Skill subject SHA-256: 315cb30f1bf7dbb0446a8ab5436764dbfb0ff830e5b95762b8300306b225f952
Bundled CLI SHA-256: 96ad0fde0c34535f29115199f24440e16fa74506dea47c120e3df37bb9b84daf
Public evidence archive SHA-256: 3b95f0a23a3342e60a435a33a78eaffc454019f1cf9e96062c62945c4c4a75e0

For v1 projects, start with the migration guide. New installations require Node.js 22 or newer for the bundled kb CLI.

Self-Evolution v2.0.0-rc.1

Self-Evolution v2.0.0-rc.1

Pre-release status: deterministic artifact and safety checks pass, but nine model/outcome gates remain pending. The recorded state is release_ready: false. This release is not proof that v2 outperforms v1 on real-world coding tasks.

Self-Evolution v2 replaces the v1 lifecycle-heavy project-memory system with a smaller, evidence-oriented project context layer. It keeps durable knowledge only when it can change a later decision or action, retrieves the smallest relevant set, and treats current code, tests, configuration, runtime behavior, and human decisions as the final truth boundary.

Highlights

• Smaller initialization contract: v1 created 7 files; v2 creates 3 (AGENTS.md, .agents/settings.yaml, and .agents/knowledge/index.yaml) — a 57.1% reduction.
• Four clear knowledge destinations: Guides, Decisions, Observations, and Archive.
• Minimal retrieval: a thin AGENTS.md router plus Guide scope and use_when metadata select the smallest relevant context.
• Evidence before authority: stored prose routes work, but material or changed claims must be checked against current project reality.
• Deterministic Node.js 22+ CLI: initialization, indexing, checks, audit support, adapter lifecycle, and migration are implemented in the bundled kb CLI.
• Reviewable v1 → v2 migration: prepare, inspect, apply, and rollback stages are byte-accounted and reversible.
• Opt-in integrations: adapters are off by default, project-scoped, and isolated from durable project knowledge.
• Failure-driven maintenance and evaluation: proposals and gates are tied to replayable failure evidence instead of feature accumulation.
• Frozen v1 baseline: the complete 1.x implementation is retained under legacy/v1/ for migration and evaluation provenance.

Installation

npx skills add D1ChangGeng/self-evolution --skill self-evolution -g -y

This repository is the source/Skill distribution. It is not an npm package release; package.json remains private: true.

Validation

The release tag points to commit 51fcea8ac14fbe7b960c0ef6271e3998eec606a8.

• Product Vitest suite: 43/43 passed
• Maintainer evaluation contract/evidence tests: 27/27 passed
• Executable evaluation fixtures: 13/13 materialized and verified
• External campaign harness tests: 127/127 passed
• CLI bundle/source byte parity: passed
• Git-free clean snapshot: npm ci and npm run ci passed with Node.js 22.13.1 / npm 10.9.2
• GitHub Actions on main: Ubuntu and Windows passed — run 32087884462
• GitHub Actions on this tag: Ubuntu and Windows passed — run 32088089387

Recorded integrated gate state:

• 8 pass
• 0 fail
• 9 pending
• 0 blocked
• release_ready: false

See maintainer/evals/RESULTS.md for the exact gate-by-gate boundary.

Real-task comparison status

The three frozen upstream defect tasks passed their environment/oracle preflight. The formal v1/v2 campaign then failed closed at the workspace-edit gateway smoke check before any model session began.

• Completed A/B trial units: 0/18
• Completed blind reviews: 0/9
• Real-task winner: not measured / no conclusion

The public external harness and task contracts are evaluation infrastructure; “hidden” tests are hidden from the execution agent during a run, not claimed as permanently private benchmarks.

Artifact hashes

Skill SHA-256:
2f3c045501508441f3eb89431b4feea2126a9ea04b9006951605aee849b7f8e2

CLI bundle SHA-256:
6eccbda3da357dc9c4024f0bf338601c206f08bf021cc0d151390a40bc9e15ad

Evaluation contract SHA-256:
94233d2dfc6c4549f98671e97fd4d7556a2265e32e56e5abe6d8b1b15b8f0e19

Licensing

Self-Evolution is distributed under Business Source License 1.1 with the repository's Additional Use Grant and derivative-work requirements. It is not an unconditional permissive open-source grant. On 2030-04-27, the licensed work changes to Apache License 2.0. Review LICENSE for the complete terms.

Upgrade notes

• Use the staged migration flow for existing v1 projects; kb init intentionally refuses to overwrite a detected v1 knowledge base.
• Review generated migration proposals before applying them.
• Adapters remain disabled unless explicitly installed.
• Treat this RC as suitable for evaluation and controlled adoption while model/outcome gates remain pending.

Full changelog: 3b28dd2...v2.0.0-rc.1