fix: make the v0.2.2 release gate evidence-state independent

Author: zxhwolfe

CHANGELOG.md

@@ -6,14 +6,14 @@ The project follows Semantic Versioning for distributable Skill releases.
 
 ## [0.2.2] - 2026-08-11
 
-Release candidate prepared for validation. This version is not tagged or published.
+v0.2.2 introduces a persistent live Host Eval evidence gate for release validation.
 
 ### Changed
 
 - Added stricter public client origin and freshness contracts, target-platform CLI filtering, and real Host Eval failure exit codes.
 - Documented the official Radar origin rule and the strict Host Eval observability/manual-review boundary.
 - Kept packaged Skill runtime copies synchronized after the nested response-contract checks.
-- Added a persistent post-v0.2.1 Host Eval evidence gate; v0.2.2 remains blocked until a completed live RC and manual semantic review are committed.
+- Added a persistent post-v0.2.1 Host Eval evidence gate: publication requires a complete live suite, structured manual semantic review, and a passing release verifier.
 - Hardened the release evidence verifier against fabricated grades, wrong suites/versions, incomplete or truncated runs, case-contract drift, stale code revisions, and unstructured manual attestations; fixed grading for the v0.2.1 suite.
 - Required successful `feed`/`sources`/`history` helper execution without an explicit custom origin and with contract-valid Radar JSON before Host Eval records runtime workflow evidence; source reads, help, failed calls, and composed shell commands no longer qualify.
 - Synced authenticated Premium identity forwarding guidance and preserved the published v0.2.1 release record below.
@@ -61,7 +61,7 @@ No unreleased changes.
 
 ### Distribution
 
-- `v0.2.1` is the current public immutable release; `v0.2.0` remains immutable and retains the last manual ChatGPT web-upload evidence.
+- `v0.2.2` is the current package line; `v0.2.0` remains immutable and retains the last manual ChatGPT web-upload evidence, while v0.2.1 and v0.2.2 Host Eval results do not claim ChatGPT Web UI validation.
 - No Hosted MCP transport is required for public Radar connectivity because standalone ChatGPT Skills already reach the live public read endpoints.
 - A future authenticated App/Plugin/OAuth transport may be introduced only for Premium account-bound capabilities such as server Topic Insight.
 

README.md

@@ -4,9 +4,9 @@
 
 [简体中文](README.zh-CN.md)
 
-Latest public release: **v0.2.1**
+Current package version: **v0.2.2**
 
-v0.2.0 proved that the standalone Skills can upload/run in ChatGPT, execute their bundled runtime, reach the live AI Workstation Topic Radar, and compose Creator → Brief. v0.2.1 keeps that workflow but adds stricter query/provenance rules and a critical cost boundary: **normal public Skill usage must not consume AI Workstation server-side LLM quota.**
+v0.2.0 proved that the standalone Skills can upload/run in ChatGPT, execute their bundled runtime, reach the live AI Workstation Topic Radar, and compose Creator → Brief. v0.2.2 keeps that workflow and adds deterministic `python3` helper invocation, bounded Radar scans, strict runtime/evidence grading, isolated Skill fixtures and neutral Host workspaces, and persistent release evidence. **Normal public Skill usage must not consume AI Workstation server-side LLM quota.**
 
 Topic Intelligence sits on top of the existing AI Workstation Global Topic Radar. It does not build another crawler, score, database, or persistence layer.
 
@@ -52,7 +52,7 @@ The Creator Skill selects one finalist and preserves the exact live `id`. The Br
 
 If only the Brief Skill is installed, it can run one bounded live selection pass (normally no more than 5 candidates), select at most one, optionally inspect its history, and then use the **current ChatGPT/Codex/agent model** to produce the brief.
 
-## Public Skill cost boundary — v0.2.1
+## Public Skill cost boundary — v0.2.2
 
 The distributable public Skills are designed to spread without silently spending the publisher's model budget.
 
@@ -125,7 +125,7 @@ In normal public mode these editorial fields are produced by the **host model**,
 
 ## Query-quality rules
 
-v0.2.1 also hardens real issues found in ChatGPT smoke testing:
+v0.2.2 also hardens real issues found in ChatGPT smoke testing and Host Eval:
 
 - content format/duration/language/audience are **not** Radar platform/source filters;
 - explicit topic scope must be preserved from the first query (`AI` should not first expand to generic technology);
@@ -187,7 +187,7 @@ See [`docs/distribution.md`](docs/distribution.md).
 
 The published v0.2.0 ZIPs were manually tested in ChatGPT web in Creator-only, Brief-only, and both-Skills shapes. Package upload, Skill discovery, bundled runtime execution, live Radar access, and behavioral composition passed.
 
-The v0.2.1 release candidate was separately validated without a ChatGPT login through three isolated fresh-agent runs (Creator-only, Brief-only, and both-Skills), extracted-package execution, live `feed`/`sources`/`history`, blocked-network behavior, and zero-`/insight` checks. This is host/runtime acceptance, not a claim that the v0.2.1 ZIP upload UI was re-tested.
+The v0.2.1 and v0.2.2 Host/Codex Eval results are separate host/runtime acceptance, not claims that the corresponding ChatGPT Web ZIP upload UI was re-tested. The latest v0.2.2 evidence additionally persists the complete live suite, manual review, and verifier output.
 
 See:
 
@@ -271,5 +271,6 @@ Useful evidence:
 
 ## Status
 
-- **v0.2.1:** latest immutable public release; public mode uses host reasoning and does not spend AI Workstation server-side LLM quota.
-- **v0.2.0:** previous immutable public release and the most recent version manually uploaded in ChatGPT web.
+- **v0.2.2:** current package version; public mode uses host reasoning, deterministic bounded helper calls, isolated Host execution, and persistent live evidence without spending AI Workstation server-side LLM quota.
+- **v0.2.1:** prior immutable Host-validated package line; its Codex/Host Eval does not represent ChatGPT Web UI validation.
+- **v0.2.0:** previous immutable release and the most recent version manually uploaded in ChatGPT web.

README.zh-CN.md

@@ -4,9 +4,9 @@
 
 [English](README.md)
 
-最新公开版本：**v0.2.1**
+当前包版本：**v0.2.2**
 
-v0.2.0 已经证明：两个 standalone Skill 可以在 ChatGPT 上传/发现/执行，能直接访问实时 AI Workstation Topic Radar，也能完成 Creator → Brief 组合流程。
+v0.2.0 已经证明：两个 standalone Skill 可以在 ChatGPT 上传/发现/执行，能直接访问实时 AI Workstation Topic Radar，也能完成 Creator → Brief 组合流程。v0.2.2 在此基础上固定了 `python3` 和确定性的 helper 调用、bounded Radar 扫描、严格 runtime/evidence grading、隔离 Skill fixture 与中性的 Host workspace，并引入持久化 release evidence。
 
 v0.2.1 继续优化真实 ChatGPT 使用中发现的问题，并新增一个非常重要的商业/成本边界：
 
@@ -56,7 +56,7 @@ Creator 只选一个 finalist，并把同一个 live `id` 交给 Brief。Brief 
 
 如果只装 Brief，它会做一次 bounded live selection（通常最多 5 个候选），选最多一个，再按需要查 history，然后由**当前宿主模型**直接生成研究就绪的内容简报。
 
-## v0.2.1：公开 Skill 的成本边界
+## v0.2.2：公开 Skill 的成本边界
 
 公开 Skill 应该可以放心传播，而不是每有一个人调用，就消耗你网站服务器的大模型额度。
 
@@ -232,7 +232,7 @@ v0.2.0 已经在 ChatGPT Web 做过真实 Creator-only / Brief-only / Both-Skill
 - live Radar：PASS
 - 双 Skill 组合：行为验证 PASS
 
-v0.2.1 候选包另外完成了不依赖 ChatGPT 登录的三条隔离新代理验收（Creator-only / Brief-only / Both-Skills）、解压包执行、实时 `feed` / `sources` / `history`、断网无本地回退和零 `/insight` 检查。这属于宿主/运行时验收，不声称重新测试了 v0.2.1 的 ChatGPT ZIP 上传界面。
+v0.2.1 和 v0.2.2 的 Codex/Host Eval 属于宿主/运行时验收，不冒充对应版本的 ChatGPT Web UI ZIP 上传验证。v0.2.2 另外持久化了完整 live suite、人工审核和 verifier evidence。
 
 详见：
 
@@ -297,5 +297,6 @@ python3 -m unittest discover -s tests -v
 
 ## 当前状态
 
-- **v0.2.1**：当前最新公开不可变版本；公开模式由宿主模型完成 Brief，不消耗 AI Workstation 服务器端 LLM Token。
-- **v0.2.0**：上一公开不可变版本，也是最近一次在 ChatGPT Web 手工上传验证的版本。
+- **v0.2.2**：当前包版本；公开模式由宿主模型完成 Brief，并通过确定性 helper、隔离 Host 执行和持久化 live evidence，不消耗 AI Workstation 服务器端 LLM Token。
+- **v0.2.1**：上一条 Host/Codex 验证版本；其 Host Eval 不代表 ChatGPT Web UI 验证。
+- **v0.2.0**：上一条公开不可变版本，也是最近一次在 ChatGPT Web 手工上传验证的版本。

docs/chatgpt-install.md

@@ -7,7 +7,7 @@ Topic Intelligence ships as two Agent Skills:
 - `creator-topic-opportunity-research`
 - `evidence-backed-content-brief`
 
-Latest public release: **v0.2.1**.
+Current package version: **v0.2.2**.
 
 ## Current ChatGPT availability
 
@@ -44,7 +44,7 @@ references/quality-contract.md
 LICENSE
 ```
 
-v0.2.1 changes the bundled helper into a **public no-cost read transport**. It exposes only:
+v0.2.2 changes the bundled helper into a **public no-cost read transport** with deterministic `python3` invocation and bounded scans. It exposes only:
 
 ```text
 feed
@@ -94,7 +94,7 @@ v0.2.1 additionally addresses issues found in that smoke:
 - no second selection after a valid handoff;
 - public Brief uses host reasoning with zero AI Workstation server-model spend.
 
-The v0.2.1 release candidate was accepted without a ChatGPT login through isolated Creator-only, Brief-only, and both-Skills fresh-agent runs plus extracted-package/live-runtime checks. The v0.2.1 ChatGPT upload UI itself was not re-tested, so its status must not be inferred from those non-UI results. See [`v0.2.1-non-ui-host-acceptance-2026-08-10.md`](v0.2.1-non-ui-host-acceptance-2026-08-10.md).
+The v0.2.1 and v0.2.2 candidates were accepted through strict isolated Codex Host Eval, including live Radar evidence, neutral workspaces, Skill fixture isolation, manual review, and persistent verifier checks. This is Host/runtime evidence; it does **not** claim that the v0.2.2 ZIP upload UI was re-tested in ChatGPT. The last real ChatGPT Web upload evidence remains v0.2.0. See [`v0.2.1-non-ui-host-acceptance-2026-08-10.md`](v0.2.1-non-ui-host-acceptance-2026-08-10.md).
 
 ## First-use prompts
 

docs/distribution.md

@@ -2,7 +2,7 @@
 
 AI Workstation Topic Intelligence is distributed as two reusable Skills over the existing AI Workstation Global Topic Radar.
 
-Latest public release: **v0.2.1**.
+Current package version: **v0.2.2**.
 
 A `VERSION` change does not publish anything by itself; only a matching pushed `v*` tag can start the release workflow.
 
@@ -113,6 +113,13 @@ Published v0.2.1 hashes:
 81d6aac45b42c27b8f24c27ac18b6a268509fb9a8a88a0813b96feab8f034d5e  aiworkstation-topic-intelligence-0.2.1-evidence-backed-content-brief.zip
 ```
 
+v0.2.2 deterministic build hashes:
+
+```text
+e0c56957c95333ae8de28a2bfb71fcbaf59ec15bc65ace2f5d379a819a7fad68  aiworkstation-topic-intelligence-0.2.2-creator-topic-opportunity-research.zip
+80a1d10cf46b25549a0abc803fef368144c166bb5356f571b452aa6f80c6332e  aiworkstation-topic-intelligence-0.2.2-evidence-backed-content-brief.zip
+```
+
 Consumers should verify release assets against `SHA256SUMS` / `release-manifest.json`.
 
 ### 3. ChatGPT Skill upload
@@ -130,7 +137,7 @@ ChatGPT host model
 
 The uploaded public Skill should not depend on a server-side AI Workstation model call.
 
-The v0.2.1 release candidate passed isolated non-UI host/runtime acceptance. The ChatGPT v0.2.1 upload UI was not re-tested; the last manual web-upload evidence remains the separately recorded v0.2.0 smoke.
+The v0.2.1 and v0.2.2 candidates passed isolated non-UI Codex/Host acceptance. The v0.2.1 and v0.2.2 ChatGPT upload UIs were not re-tested; the last manual web-upload evidence remains the separately recorded v0.2.0 smoke.
 
 If a future ChatGPT App/Plugin/OAuth integration provides an authenticated AI Workstation account connection, that native connection may expose separate Premium capabilities. Do not put that authentication responsibility inside the portable Skill ZIP.
 
@@ -145,9 +152,10 @@ vX.Y.Z
 Pushing a matching tag triggers `.github/workflows/release.yml`, which:
 
 1. verifies tag == `VERSION`;
-2. runs the offline suite;
-3. builds deterministic archives;
-4. publishes ZIPs, manifest, and checksums.
+2. runs the offline full test suite, runtime sync, and eval dry-run;
+3. verifies persistent live Host Eval evidence and manual review;
+4. builds deterministic archives;
+5. publishes ZIPs, manifest, and `SHA256SUMS`.
 
 Existing tags such as `v0.1.0` and `v0.2.0` are immutable.
 

docs/v0.2.2-host-eval-blocked-2026-08-11.md

@@ -1,6 +1,10 @@
-# v0.2.2 Host Eval release blocker — 2026-08-11
+# v0.2.2 Host Eval release blocker — historical record, 2026-08-11
 
-Status: **BLOCKED — do not tag or publish v0.2.2**.
+Status: **HISTORICAL EARLY-RC FAILURE — not the final release status.**
+
+This document records failed early release-candidate attempts and their diagnostic facts. It must not be read as the current v0.2.2 publication state. Current release eligibility is determined by the final persistent evidence and `scripts/verify_release_evidence.py`.
+
+At the time of this incident, status was **BLOCKED — do not tag or publish v0.2.2**.
 
 The live `v0.2.1` strict Host Eval collector was run with fresh Codex processes. The run produced 7 cases: 5 timeouts, 1 non-zero exit, and 2 unobservable route observations. The second-stage grader also initially treated `v0.2.1` as an unsupported quality suite and later proved too permissive about helper-path mentions; both tooling defects are now fixed. Regrading this run with the hardened rules produces 7 unobservable cases because it contains no successful contract-valid Radar helper response. This is an honest failed RC, not release evidence.