ContextLean v0.3.2 — Privacy disclosure

ContextLean v0.3.2 — Privacy disclosure

This patch adds the privacy notice and its public URL to the Codex and
Claude plugin manifests for directory submission. The deterministic distribution
package includes the notice, and the repository and package READMEs link to it.

The notice documents local project access and output, the absence of a hosted
ContextLean collection service, and the provider data flow of explicitly requested
live benchmarks. Contact: andyst-dev@proton.me.

All four Skill behaviors and v0.3.0 Balanced guidance remain unchanged. The optional
measurement helper changes only its reported version. No model benchmark was run,
and no historical evidence was changed. Published v0.3.1 remains untouched.

Directory review and publication are separate from this GitHub release. This
release does not claim Anthropic or OpenAI approval or new performance results.

Validation and artifact

• 139 offline tests passed with no skips, including native Codex/Claude lifecycle
  checks. No model session or live benchmark was run.
• Ruff 0.15.7 lint and format checks, four Skill validators, Claude manifest/catalog
  validation and whitespace checks passed. Source-root CLAUDE.md and shared catalog
  policy warnings remain expected.
• The four Skill bodies, references and UI metadata are byte-identical to v0.3.1.
  Published historical evidence and the original Bootstrap specification are unchanged.

Release asset: contextlean-0.3.2.zip — 23 files, 34,882 bytes.
SHA-256 (also recorded in contextlean-0.3.2.sha256):

41a23e53dfa06b7d4576f87f874368c4d0faf0d0739c2ab81adb242b60374772

ContextLean v0.3.1 — Packaging and distribution readiness

ContextLean v0.3.1 — Packaging and distribution readiness

This patch release delivers the completed repository cleanup and Codex/Claude Code
distribution work. The product remains dependency-free and skills-only.

• Cleaner active repository structure, with historical documentation separated
  from current installation and developer documentation.
• Completed Old-vs-Balanced comparison tooling removed from active maintenance;
  legacy Bootstrap compatibility machinery retained as historical test support,
  outside the packaged Skill.
• Simpler contributor guidance and clearer installation, update and uninstall flows.
• Native Codex and Claude Code packaging and four-Skill discovery support.
• Deterministic distribution ZIP export from the canonical shared Skills.
• Offline validation of fresh installation, discovery, relocation, updates and
  uninstall safety, including preservation of customized project guidance.
• Listing metadata and a shared icon prepared for separate platform submissions.

Balanced runtime guidance is unchanged. There are no new Skills, no Skill behavior
changes and no benchmark rerun. The optional measurement helper changes only its
reported product version. Published benchmark evidence is unchanged; historical
v0.2.1 performance results are not v0.3.1 performance claims.

Update through the existing native plugin commands in the
installation guide.
The version bump enables normal updates from 0.3.0 without modifying generated
project guidance. Public-directory submissions are not part of this release.

Validation and artifact

• 139 offline tests passed, including native Codex/Claude installation lifecycle
  tests; no model sessions were run.
• Plugin/Skill validators, Ruff lint/format, Markdown links/anchors, version and
  package hygiene checks passed. Claude retains the documented catalog-policy warning.
• All 14 historical checksum ledgers (3,363 entries) and eight historical
  fingerprints passed. Published evidence changed files: zero.
• All four SKILL.md files and Balanced runtime guidance are byte-identical to
  distribution commit b107a96c679f517e5ef0724fcaa879d8934ab85b. Representative
  generated guidance remains 4,752 bytes.

Release asset: contextlean-0.3.1.zip — 22 files, 33,470 bytes.
Its companion contextlean-0.3.1.sha256 records SHA-256:

28cf9489e9fe7c93d22e7b34a993b3badfb1bdd67767268be2e471483e66eff3

ContextLean v0.3.0 — Balanced

ContextLean v0.3.0 — Balanced

Bootstrap once. Read less. Change locally.

ContextLean v0.3.0 makes Balanced the current product contract: compact maintained
project guidance that helps coding agents find the right code, avoid unnecessary
exploration, make focused changes and verify them proportionally.

Balanced runtime contract

The ordinary coding guidance retains owner-first/search-first navigation, the reuse
ladder, focused implementation and safety constraints, root-cause bug fixing,
proportional verification, reuse of equivalent passed checks, and targeted map
correction when work invalidates mapped information or encounters stale entries.

Representative automatic context falls from 5,486 to 4,752 bytes (−13.38%).
This measures the representative generated AGENTS.md plus its Claude wrapper;
project-specific maps vary. It is not a model-token or performance guarantee.

Intentional product simplification

Historical bootstrap compatibility no longer defines the runtime contract. Setup
no longer requires a historical facet-transfer certificate or measurement report.
Ordinary work no longer requires map checking after every task, mandatory project
Skill creation/discovery, or a mandatory Lean Review handoff in generated projects.

Architecture/refactor detail remains in the explicitly requested Lean Review.
Project Skill authoring/sharing is optional documentation. Claude wrappers/imports
and safe exclusions remain Bootstrap/install responsibilities. No automatic Skill
invocation, runtime package, hooks or telemetry is introduced.

Behavioral validation

The separate Balanced product-change validation contains six Old-versus-Balanced
Refactor sessions: 6/6 passed submitted tests, original regression tests and
acceptance tests, including independent offline grading. No material retained-contract
regression was found in this task. Balanced preserved the tested Refactor behavior
after simplifying the runtime contract.

The experiment does not establish equivalence for complex architectural refactors.
Three pairs do not justify statistical-significance claims. The historical v0.2.1
30-session study and its −8.96% descriptive token aggregate remain historical
performance evidence; they are not v0.3.0 performance results.

Balanced validation ·
Historical evidence provenance ·
Contract detail ·
Release verification

ContextLean v0.2.1

ContextLean v0.2.1

Small verification-behavior clarification.

Changed

• Verification reasons about actual test coverage rather than command labels such as "targeted" or "full".
• Reuses already-passed equivalent coverage when no relevant code, test, configuration or environment change occurred.
• Still broadens verification when additional coverage is needed, after failures or unresolved results, or when project rules require it. Project-required repeated checks remain valid.

Validation

• 120 offline tests; Ruff clean
• Four Skill validators; Codex and Claude plugin validation (two existing Claude warnings)
• 74 PASS / 0 PARTIAL / 0 MISSING / 2 INTENTIONALLY REPLACED
• 71/71 permanent semantic facets preserved
• Automatic context: 5,486 bytes, below the unchanged 5,489-byte ceiling
• Audit: 9 PASS / 0 WARNING / 0 ACTION NEEDED; Lean Review: no material findings

A focused Git-backed GPT-5.6 Sol Refactor validation passed submitted, regression and acceptance tests in both conditions. ContextLean ran the six-test coverage once; the redundant equivalent rerun observed during the v0.2.0 investigation was not reproduced.

This is a focused behavioral validation, not evidence of universal performance gains, statistical confidence or causality. The historical v0.2.0 benchmark headline remains unchanged.

ContextLean v0.2.0

ContextLean v0.2.0

ContextLean gives coding agents a compact, maintained map of a repository so they can find relevant code without repeatedly exploring everything.

Highlights

• Bootstrap Repository
• Audit Context Locality
• Lean Change Review
• Benchmark Context Usage
• Codex and Claude Code support
• Compact AGENTS.md guidance with conditional detailed references
• Project-map-driven navigation
• Semantic transfer and safe-removal contracts
• Original bootstrap intent preserved

ContextLean remains dependency-free and skills-only, with no MCP server, hooks, telemetry or automatic benchmark runs.

Compact context

Representative automatic AGENTS.md guidance was reduced from 7,956 bytes → approximately 5.5 KB (5,476 bytes), while maintaining 74 PASS / 0 PARTIAL / 0 MISSING across the semantic coverage gate. Two mechanisms are intentionally replaced; all 71/71 permanent facets are preserved.

Preliminary validation

The release-aligned batch used five existing tasks and 10 fresh runs, with gpt-5.6-terra, low reasoning, and one Vanilla and one ContextLean run per task. All ten solutions passed acceptance, regression and submitted tests, including independent offline regrading.

• Total tokens: 482,024 → 341,676 (-29.12%)
• Wall time: 189.97s → 134.52s (-29.19%)
• Commands: 18 → 16 (-11.11%)
• Task success: 5/5 → 5/5

Task token differences:

• Navigation: -8.18%
• Bug Fix: -11.35%
• Feature: -17.97%
• Refactor: -24.02%
• Documentation/config: -58.77%

These results are preliminary. There was one run per condition per task. They do not establish statistical confidence, causality, or universal savings. Results apply to the tested configuration.

Unfavorable submetrics remain disclosed: Bug Fix used 35 more output tokens (+3.35%) and two more commands (3→5, +66.67%); Navigation had a higher dated credit-equivalent (+59.93%) because cache composition differed. Raw diagnostics, recoveries and previous datasets are preserved separately.

Product tested: 86043499d5a375cc6a3b295bc7c03b0999eb458c.
Full committed evidence.

Verification

114 offline tests, Ruff lint/format, all four Skill validators, Codex/Claude package validators, evidence checksums, relative links and hygiene pass. No model benchmark was rerun for release.

Installation and usage · Preserved original bootstrap.