Codex Capability Router 1.1.3

V1.1.3 正式發布，整合 V1.1.0–V1.1.2 本機修正版的累積修改。此版本以 V1.0.0 為基線，不包含 V1.0.1 的 Skill Preference Memory。

修改內容

• 測試資料改用 SKILL.fixture.md，相容測試只在暫存目錄建立 SKILL.md，避免 fixture 被當成正式技能。
• 結構化路由要求最終選用 Skill／Provider 具有同任務的 selected 批次證據；缺少證據時拒絕定案並列出 IDs。
• 分開呈現已選 Provider、候選判斷覆蓋與 Host 執行回報；每筆 ExecutionAttempt 核對 Receipt、Provider 及原始需求。
• 補齊已安裝 runtime 的 Host API 指引，明確區分相容呼叫、正式路由與實際執行。
• 更新中英文說明、V1.1.3 安裝指令與版本歷程；V1.1.0–V1.1.2 未獨立發布，V1.0.1 已停用。

驗證

• 最終文件修改後，完整測試：317／317 通過。
• compileall、版本一致性、文件連結、歷史資產與公開文字靜態檢查：通過。
• 繁體中文首頁與相容頁一致；本次修改檔案保留 UTF-8、無 BOM 與 CRLF。
• Router runtime 與版本 metadata 共 22 個檔案維持原有 V1.1.3 雜湊；SKILL.md 維持 180 詞。

已知限制

Skill 指引不能保證 Host 每次完全遵從。批次資料檢查驗證提交資料的結構與綁定，不能證明模型已逐項閱讀或獨立判斷所有候選。

缺少判斷或仍為 needs_detail 的候選保留 PARTIAL。FINALIZED 僅代表選用資料定案，不代表完整覆蓋、語意選用正確或下游執行成功。

Router 不執行或安裝工具，也不替 Host 授權。實際技能套用、工具呼叫與工作成果仍需各自的執行證據。

Codex Capability Router 1.0.0

Codex Capability Router 1.0.0

正式穩定版 / Stable release.

This release promotes the beta.10 runtime to 1.0.0 without changing Router functionality. It refreshes the public documentation and adds release-verification coverage.

Highlights

• Traditional Chinese primary README and an equivalent English edition.
• Original, consistent animal characters across the hero, an eight-panel tutorial numbered 1-8, and a four-panel policy comic numbered 1-4.
• Version-controlled SVG and Mermaid architecture, plus an editable FigJam flow.
• Short Windows PowerShell and macOS/Linux installation instructions.
• Current contracts: Host semantic decisions, deterministic Python validation/evidence, batch coverage, Skill handoff, execution needs, Provider decisions, and separate execution evidence.
• CRLF/LF/mixed-line-ending regression coverage that rejects mismatched or missing installation blocks.

Verified release gates

After the final source/documentation/asset mutation:

• Focused tests: 13/13 PASS.
• Full unittest suite: 288/288 PASS.
• compileall and git diff --check: PASS.
• Strict UTF-8, U+FFFD, private-path and known-credential-pattern checks: PASS across 104 public text files.
• README local links/assets, SVG/XML parsing, and ZH/EN version/core-fact parity: PASS.
• Visual inspection at approximately 830px README content width and a 390px viewport: PASS.
• Fresh candidate clone and then a fresh public v1.0.0 tag clone: PASS, including repeated focused/full tests.
• Runtime scope: only version changed; historical beta tags preserved.

Boundaries retained

FINALIZED is not execution success. Staged candidates are not automatically semantically considered; unresolved or missing decisions remain PARTIAL. Plugin is not a Provider. There is no fixed top-k or fixed Skill maximum, no persistent preference learning claim, and no claim that automatic Host integration works in every environment.

The existing beta.10 diagnostic limitation remains: supporting_metrics.provider_selected_total can retain its preparation value. Use the final selected Provider list and selected_count for the finalized selection; see the release notes below.

No private local capability inventory is included in this release. No PyPI publication or hardware execution is claimed.

繁體中文 README | English README | Release notes | Validation procedure

Codex Capability Router v0.2.0-beta.10

This release fixes the beta.9 audit defects in remediation phase A and updates the README and current routing contract.

Fixes

• Preserve explicitly configured deep Skill roots during root compression, including Plugin provenance and known-system boundaries.
• Handle missing or unreadable roots without crashing or attributing diagnostics to a previous source.
• Recover the specific changed selected Skill once, refresh authoritative metadata, and require revalidation when identity or eligibility changes.
• Separate inventory staging from Host semantic decisions. Validate task-bound batch responses and report incomplete or unresolved coverage as PARTIAL.
• Align the current Skill entrypoint, routing references, user guide, and Chinese/English READMEs; label older contracts as historical.

Verification

• Main and the annotated v0.2.0-beta.10 tag resolve to commit 563a962.
• Local checkout and a new clone from GitHub: 283 unittest tests passed, with 0 failures, 0 errors, and 0 skipped.
• Both checkouts passed compileall; the fresh clone has a clean Git status.
• Package and project versions both equal 0.2.0-beta.10; no runtime dependencies were added.
• Local checks passed for 91 UTF-8 files, 56 document/image references, README parity, and git diff --check.
• Changed files, regressions, checks, and evidence boundaries

Scope

Phases B/C are not implemented. Automatic Host integration, real LLM semantic accuracy, real Provider invocation, production, hardware, and browser rendering remain NOT VERIFIED. Global Skill/runtime installations are not updated by this source release. The original supplied audit and remediation documents remain local and are excluded from the release.

Codex Capability Router v0.2.0-beta.9

Codex Capability Router v0.2.0-beta.9

This pre-release consolidates the beta3-beta9 routing architecture work.

Highlights

High-recall Skill routing

• Every resolved Skill receives semantic consideration.
• Any Skill with plausible task-relevant value may be selected.
• Weak, redundant, and overlapping relevant Skills may be selected together.
• No fixed Skill maximum or top-k semantic truncation.

Authoritative discovery

• Uses bounded authoritative Skill roots.
• Resolves logical Plugins to deterministic exact package roots.
• Does not perform whole-disk or unrestricted Plugin-cache discovery.
• Package declarations participate in existence-first routing.

Fast Skill inventory reuse

• Uses a minimal fixed Skill root plan.
• $CODEX_HOME/skills/.system is handled as a known SYSTEM subtree, not a third global root.
• RootPlanSnapshot and SkillInventorySnapshot avoid repeated filesystem discovery.
• Ordinary routes reuse cached inventory without reopening Plugin manifests or all SKILL.md files.

Host and Provider routing

• Formal Provider kinds: app, mcp, builtin_tool, host_tool.
• Unknown Host hierarchy remains selectable as host_tool.
• Readiness is diagnostic and does not hide an existing identity-resolved capability from semantic consideration.
• Plugin remains package/provenance information, not a formal Provider.

Safe Skill source binding

• A canonical Skill may keep multiple authoritative provenance sources.
• One deterministic physical source drives profile, fingerprint, handoff path, and instructions.
• Selected-Skill freshness mismatch uses targeted refresh instead of a full inventory rescan.
• Handoff retries at most once.
• Semantic digest changes return SELECTION_REVALIDATION_REQUIRED to the Host controller.

Safety boundaries

• Codex / Host LLM owns semantic TaskAnalysis and selection.
• Python owns deterministic discovery, validation, fingerprints, handoff safety, and Receipt finalization.
• FINALIZED Receipt remains separate from immutable ExecutionAttempt evidence.

Verification

Development:

• full regression: 270/270 PASS
• compile/static/privacy: PASS
• real duplicate-source pdf Skill handoff: PASS

Installed beta9:

• runtime manifest: 25/25 SHA match
• stale runtime files: 0
• neutral import: PASS
• targeted selected-Skill refresh: PASS
• static verification: PASS

Status

This is a pre-release.
Stable v0.2.0 is not published yet.

v0.2.0-beta.1 — LLM Task Analysis + Supporting Capability Selection

v0.2.0-beta.1

Pre-release for the LLM TaskAnalysis and Supporting Capability Selection
architecture. This release keeps the beta.4 route, Receipt, lifecycle, and
Skill semantics while adding a bounded lazy Supporting Provider path.

Added

• Immutable strict LLM TaskAnalysis with task_summary, work_items,
  deliverables, constraints, and quality_expectations.
• Skill-only prepare_route_context() with deterministic fingerprints and
  validated validated_decision_payloads.
• Execution Needs and true-lazy prepare_supporting_context().
• Provider-level Supporting decision contracts with
  ReadinessEvidenceCertificate, deterministic Provider digests, and one
  bounded request_detail / final_selection protocol.
• v0.2 Receipt fields for TaskAnalysis, Execution Needs, Supporting status,
  Provider readiness, fingerprints, metrics, and bounded detail references.
• Formal Supporting scope for the certified instances MCP node_repl and
  builtin-equivalent functions.exec_command.

Changed

• The v0.2 production path requires TaskAnalysis; legacy task_summary is a
  compatibility projection and must match it.
• route() finalizes Skill and Supporting decisions together; it remains the
  only formal Receipt and FINALIZED entry point.
• Supporting Provider discovery is skipped entirely when
  execution_needs=[].
• The existing 12 Skill routing scenarios remain exactly 6 zh-TW and 6
  en, migrated to the mandatory TaskAnalysis input contract.

Fixed

• Bounded legacy frontmatter normalization for explain-code.
• Compatibility parsing for scalar/block metadata, simple allowed-tools
  lists, limited scalar metadata, and one safe scalar-leaf level under
  metadata.source_frontmatter.
• Malformed, sensitive, unavailable, controller, routing-support, and unsafe
  records remain rejected.

Verified

• Phase 1 focused: 30/30 PASS; Phase 1 full: 102/102 PASS.
• Phase 2 focused: 43/43 PASS; Phase 2 full: 111/111 PASS.
• Phase 3 focused: 36/36 PASS; Phase 3 full: 123/123 PASS.
• Phase 4 focused: 14/14 PASS; Phase 4 full: 137/137 PASS.
• Phase 5 deterministic full: 137/137 PASS.
• Phase 5 focused Phase 1–4: 45/45 PASS.
• Codex Live Acceptance A–E: PASS.
• Positive Provider selection: mcp:node_repl and
  builtin_tool:functions.exec_command.
• Lazy not_required path: PASS; no-match path: PASS;
  Receipt/privacy: PASS.

Known limitations

• The current Host did not expose a Router-trusted Skill availability
  declaration during v0.2 Live Acceptance.
• Formal Supporting Provider scope is limited to certified instances.
• App remains INSUFFICIENT_RUNTIME_EVIDENCE.
• Plugin remains NO_RUNTIME_SAMPLE.
• Other MCP and builtin providers are not automatically trusted.
• Supporting detail expansion was covered deterministically but was not
  naturally triggered in the five Live Acceptance cases.
• Live destructive stale mutation was intentionally not performed.

This is a pre-release preparation only. No tag, GitHub Release, push, or
release commit is created by this change.

Documentation

• Illustrated v0.2 User Guide
• Installed Skill Live Test Report
• v0.2 Architecture / Design

Post-release documentation was added after the v0.2.0-beta.1 tag.
The release tag itself remains unchanged.

v0.1.0-beta.4

Integration Hardening

v0.1.0-beta.4 不取代 beta.3 的語意 Skill Selection 設計，專注強化 production integration boundary。

• Production Route Enforcement：正式 Router 結果必須由 production routing.route(SelectionRouteInput(...)) 產生；外層不能偽造正式 Selection Receipt。
• Automatic Selection Receipt：自動記錄 candidates、preliminary selection、full handoff、final selected Skills、selection_status、expanded retrieval、correction 與 finalized state；不保存 private chain-of-thought 或敏感資料。
• Canonical Skill ID Enforcement：machine path 全程使用 canonical Skill ID；display name 僅供顯示，不猜測 display name 到 Skill ID。
• Route Finalization：lifecycle 為 OPEN 到 FINALIZED；FINALIZED 後不得 correction、expanded retrieval 或改寫 selected Skills；新工作建立新的 Route。
• beta.3 semantic Skill Selection 核心保持不變。

Validation

• 91/91 automated tests PASS
• Integration Live Acceptance 4/4 PASS
• compileall、UTF-8/U+FFFD、git diff --check、production source scan PASS

Known Observation

Existing runtime inventory may report malformed Skill metadata diagnostics. This observation predates beta.4 and was not introduced or changed by this release。

v0.1.0-beta.3

摘要

v0.1.0-beta.3 將 Final Skill Selection 交由 Codex 主模型依工作語意決定，並完成 Phase 0～5 驗收。本版本為 Pre-release，不建立 v0.1.0 stable。

主要變更

• Final Skill Selection 由 Codex 主模型根據工作語意決定。
• keyword、category、provides 不再主導 final selection，也不再存在 keyword → Skill ID production mapping。
• 新增 Skill inventory、cache、fingerprint，以及 Basic / Enriched Profile。
• 新增 recall-first Candidate Retrieval 與 eligible Explicit Skill inclusion。
• 對 Codex 初選 Skill 完整讀取 SKILL.md，執行 Full SKILL.md Applicability Check。
• Expanded Retrieval 最多一次；Correction 最多一次。
• selection output 使用 selected / no_matching_skill。
• 移除 PRIMARY / OPTIONAL 與 3+2 selection semantics。
• 無 legacy fallback、無 silent fallback。
• 新增或更新 Skill 不需修改 Router production mapping。

驗證

• Full automated tests：81/81 PASS。
• Phase 5 Full Live Acceptance：5/5 PASS。
• compileall：PASS。
• UTF-8/U+FFFD：PASS。
• git diff --check：PASS。
• Production source scan：PASS。

Known Observations

• Existing runtime inventory may report malformed Skill metadata diagnostics.
• This was observed before beta.3 and was not introduced by this release.
• It did not cause missed or incorrect Skill selection in beta.3 acceptance；本觀察不屬於 beta.3 regression。

Release boundary

本版本為 v0.1.0-beta.3 Pre-release。既有 beta.1 與 beta.2 tags 保持不變；本次未建立 v0.1.0 stable。

Codex Capability Router v0.1.0-beta.2

Codex Capability Router v0.1.0-beta.2

Status: Beta / Pre-release

v0.1.0-beta.2 is the second public beta of Codex Capability Router, a local-first Codex skill for discovering, classifying, explaining, and routing available capabilities with deterministic and privacy-conscious behavior.

What’s New in Beta 2

Improved route-only behavior

Route-only mode now correctly separates capability selection from capability execution.

When a user asks Codex to only determine which skills should be used:

• the router still selects capabilities for the actual target task;
• selected downstream capabilities are not executed;
• execution suppression no longer changes capability ranking or selection.

Controller self-selection protection

codex-capability-router is permanently excluded from downstream capability selection.

The router acts as the controller and cannot select itself as a task capability.

Internal routing support is no longer exposed as a task capability

Internal discovery/support primitives such as exec_command, when used only for read-only capability inspection, are excluded from:

• Selected Skills
• Selected Capabilities
• PRIMARY
• OPTIONAL

This keeps the result focused on capabilities that actually serve the user's target task.

Workspace-specific specialist preference validated

When a project-specific skill is available and provides better task coverage than generic alternatives, the router can prefer the workspace-specific specialist while rejecting redundant overlapping capabilities.

Selected Capability Explanation

Routing output provides human-readable explanations for selected capabilities.

Each selected capability can include:

• Name
• Kind
• PRIMARY / OPTIONAL
• Function
• Why selected

Selection explanations are based on deterministic routing evidence such as:

• explicit request
• trigger match
• specialist match
• requirement coverage
• workspace specificity
• overlap preference
• fallback or complementary coverage

The explanation layer does not expose private chain-of-thought.

Recommendation-only capabilities remain separate

Capabilities that are only recommendations are kept separate from executable selections.

The router does not automatically:

• install plugins or skills;
• modify permissions;
• perform OAuth;
• execute recommendation-only capabilities.

---

Core Capabilities

Codex Capability Router currently supports:

• local capability discovery;
• canonical machine-readable capability registry;
• capability classification;
• deterministic routing;
• overlap and deduplication handling;
• runtime / CLI / manual-source precedence;
• provenance and conflict tracking;
• safe handling of unknown capabilities;
• PRIMARY and OPTIONAL capability selection;
• recommendation-only separation;
• English output;
• Traditional Chinese (zh-TW) output;
• bilingual human-readable capability catalogs;
• Selected Capability Explanation;
• route-only routing mode;
• controller recursion protection.

---

Real-world Validation

Beta 2 was validated in an independent STM32G0 firmware workspace.

The real-world acceptance flow verified:

• automatic router triggering;
• explicit router invocation;
• workspace-specific skill discovery;
• specialist preference;
• capability overlap and deduplication;
• controller self-exclusion;
• internal support capability exclusion;
• route-only selection semantics;
• Selected Capability Explanation;
• downstream skill execution;
• separation of software evidence, build blockers, and hardware evidence;
• preservation of HARDWARE_PENDING when hardware evidence is unavailable.

The selected project-specific firmware skill successfully identified real software-side defects while preserving unrelated build and hardware validation boundaries.

No private project source code or private workspace inventory is included in this public release.

---

Validation

Current release validation:

• 46 / 46 deterministic tests PASS

• 12 routing scenarios PASS

  • 6 Traditional Chinese
  • 6 English

• Python compile validation: PASS

• Clean-install verification: PASS

• UTF-8 / U+FFFD validation: PASS

• git diff --check: PASS

• Public safety audit: PASS

• Security/privacy audit: PASS

The v0.1.0-beta.1 tag remains unchanged.

v0.1.0-beta.2 points to the verified Beta 2 release commit.

---

Safety and Privacy

The current beta is intentionally conservative.

It does not automatically:

• install plugins;
• uninstall plugins;
• change plugin permissions;
• perform OAuth authorization;
• crawl plugin marketplaces;
• execute recursive routing;
• start background monitoring;
• synchronize registries to the cloud;
• upload telemetry;
• expose private chain-of-thought.

Capability states that cannot be reliably verified remain unknown rather than being guessed.

The public repository excludes private inventories, credentials, secrets, local backups, temporary files, and machine-specific private paths.

---

Known Limitation

Measured runtime token usage is currently unavailable.

Plugin Eval currently reports:

• Deferred static estimate: 43,789 tokens
• Total static estimate: 44,200 tokens
• Budget type: estimated-static

This estimate includes repository artifacts such as documentation, tests, fixtures, validation files, and implementation files.

It is not measured runtime token consumption.

The static estimate is retained as diagnostic information and is not treated as a Beta functional blocker.

---

Stable v0.1.0 Requirement

Promotion from Beta to stable v0.1.0 requires either:

1. empirical runtime token-usage measurement; or
2. other credible evidence demonstrating bounded runtime context usage.

Until then, this release remains a public Beta.

---

Installation

See the repository README for the current installation, update, uninstall, registry, and usage instructions.

Repository:

Lzxpan/codex-capability-router

---

繁體中文摘要

Codex Capability Router v0.1.0-beta.2

這是 Codex Capability Router 的第二個公開 Beta 版本。

本次主要改善：

• 修正「只路由、不執行」模式；

• Router 不再把自己列為被選技能；

• exec_command 等內部 discovery/support 能力不再混入 Selected Capabilities；

• Workspace 專用 Skill 可以正確優先於重複的通用 Skill；

• 顯示「已選技能 / 已選能力」；

• 顯示每個能力的：

  • 名稱
  • 類型
  • PRIMARY / OPTIONAL
  • 功能
  • 選用理由；

• recommendation-only 能力與實際選用能力分離；

• 選用理由採用可稽核的 deterministic routing evidence，不揭露 private chain-of-thought。

驗證結果

• 46 / 46 tests PASS

• Routing scenarios：12

  • 繁體中文：6
  • English：6

• Clean-install：PASS

• UTF-8：PASS

• Security / Privacy audit：PASS

• 獨立 STM32G0 專案 Real-world Acceptance：PASS

已知限制

目前仍無法取得實際 Runtime token usage。

Plugin Eval 的 43,789 deferred / 44,200 total 為 estimated-static 靜態估算，不代表實際執行 Skill 時的 token 消耗。

正式升級為 stable v0.1.0 前，仍需要取得實際 Runtime token usage，或其他可信的 bounded runtime evidence。

Release status: Beta / Pre-release

v0.1.0-beta.1

Codex Capability Router v0.1.0-beta.1

First public beta release of Codex Capability Router.

Codex Capability Router is a local-first Codex skill for discovering,
classifying, explaining, and routing available capabilities using
deterministic and privacy-safe recommendations.

Highlights

• Local capability discovery
• Canonical machine-readable capability registry
• Deterministic capability routing
• Capability overlap and deduplication handling
• Runtime / CLI / manual source precedence
• Provenance and conflict tracking
• Safe handling of UNKNOWN capabilities
• English and Traditional Chinese (zh-TW) output
• Human-readable Selected Capabilities / 已選能力
• Human-readable Selected Skills / 已選技能
• PRIMARY / OPTIONAL distinction
• Capability function descriptions
• Deterministic "Why selected" explanations
• Recommendation-only capabilities separated from executable selections

Safety

• No automatic plugin installation
• No automatic plugin removal
• No automatic permission modification
• No automatic OAuth
• No marketplace crawling
• No recursive routing
• No hidden chain-of-thought exposure
• No telemetry
• Local-first registry design
• Privacy-safe public release

Validation

• 42 / 42 deterministic tests PASS
• 12 routing scenarios PASS
  • 6 Traditional Chinese
  • 6 English
• Python compile validation PASS
• UTF-8 / U+FFFD validation PASS
• Clean-install validation PASS
• Security and privacy audit PASS
• Existing React UI and firmware routing regression checks PASS

Selected Capability Explanation

Example:

Name: React UI Debugging
Kind: skill
Level: PRIMARY
Function: Diagnoses React UI regressions.
Why selected: The task matches React, UI, and debugging triggers and this capability is a specialist match for the task.

Selection explanations are derived from deterministic routing evidence
and do not expose private chain-of-thought.

Languages

• English
• Traditional Chinese (zh-TW)

Known limitation

Measured runtime token usage is currently unavailable.

Plugin Eval reports:

• Deferred static estimate: 43,789 tokens
• Total static estimate: 44,200 tokens
• Budget type: estimated-static

This estimate includes repository documentation, tests, fixtures,
implementation files, and validation artifacts. It is not measured
runtime token consumption.

Empirical runtime token measurement, or other credible bounded runtime
evidence, remains a requirement before promotion to stable v0.1.0.

Release status

Beta / Pre-release

This release is intended for public testing and feedback.

Stable v0.1.0 has not been released.