Servotab 0.7.0

Servotab 0.7.0 — strategy, continuity and incorporation

Servotab 0.7.0 brings Plan strategy, comparative Design/Debug judgments and the
Incorporate method into one release. It preserves one implicit router, twelve
explicit-only leaves and a 70-file plugin payload. No runtime service,
dependency, hook, mandatory report or new approval round is added.

Publication receipt

Servotab 0.7.0
was published on 2026-10-14 Singapore time as GitHub Latest, non-prerelease.
Tag v0.7.0 binds PR #54's
merge commit 0c79fac18628f97a5b22db932e69f7aa3859b373 and tree
ce8d088368ab760dfef52ea193284e9618244843. Main Validate run
38564454579
passed all seven jobs. Local exact-commit build, main-CI artifacts, draft
download and fresh public download passed the complete four-file release check.
The 70-file plugin ZIP is 339,822 bytes with SHA-256
b6f4a5a185ebf1bb8d41357894c8fa2d8dcf3bc3d6991cb46fa4ece66bcf90f1.
Later documentation receipts do not replace the immutable release tag or assets.

Changes

• Plan strategy and task continuity. A settled goal can still need decisions
  about sequencing, effort, evidence and recovery. plan chooses and revises
  the route, preserves the complete accepted outcome, and keeps a durable task
  record when continuity needs it. Resume from an explicitly selected task entry
  and reconcile it against current code and evidence. Router, Execute, Spec
  Chain, Delegate and Finish keep the same goal, record and responsibility owner;
  small clear work can remain inline. Planning does not switch the host into
  Plan Mode or grant implementation or publication authority.
• Incorporate specified material. incorporate replaces review-feedback
  and handles the appropriate effect of discussions, feedback, research and
  references on the target work. Cover important contributions before narrowing
  implementation priorities, interpret corrections in context, preserve
  explicitly selected forms and judge applicability. Open ideas can improve
  design, examples or checks without automatically becoming features. Carry
  adopted effects through the authorized implementation and verification.
  Technical review retains code-based verification, scoped finding dispositions
  and correction across shared-contract consumers. Design retains open choices;
  Delegate retains responsibility allocation. A reading lane can establish
  coverage, or one worker can own a complete bounded incorporation task.
• Comparative judgments. debug constrains selective-failure explanations
  with the closest verified working case. design names the condition that
  makes a recommendation preferable to its nearest viable alternative, and can
  separate conflicting properties by time, location, condition or scale when
  that helps the decision. Ordinary work requires no extra exercise.
• Source and evidence alignment. Current skill metadata, generated paths,
  icons, documentation and the Methods catalog use incorporate. The existing
  glyph is reused without geometry or raster changes; its SVG accessible title
  and description now name Incorporate. The source pack adds
  three Plan cases and two Incorporate cases, and supports named Plan studies
  in the read-only acceptance checker while retaining existing receipt checks.

Upgrade and compatibility

Change explicit $review-feedback calls to $incorporate, including saved
prompts and method-path references. No compatibility alias or parallel leaf is
shipped. The method still handles verified technical feedback; the rename
broadens the entry to other specified material without dropping that guarantee.
Finding dispositions remain judgments; the old fixed category labels,
implementation-order ladder and scripted response wording are no longer required.
The existing servotab@personal marketplace selector and installation route are
unchanged. Published v0.6.5 archives retain their original method set.

The source evidence pack has 30 cases, 26 with semantic-review requirements,
63 activation seeds and 19 adversarial controls. Deterministic package and
fixture checks establish source consistency and reject known wrong fixture
outputs; they do not establish general live-model effectiveness.

The earlier bounded Plan comparison accepted six candidate attempts and five
of six baseline attempts; one baseline recovery omitted unknown historical
publication state. That small synthetic comparison does not establish a
general effectiveness or efficiency gain. Full Plan behavioral acceptance
remains open. Supplied Incorporate trial reports lack independently inspected
raw traces; receipt-bound semantic acceptance and ordinary-language activation
are not established by this release. See Plan coverage
and Incorporate evidence.

Artifacts and publication boundary

The clean-commit release builder produces four exact-source assets:
servotab-0.7.0-plugin.zip, servotab-0.7.0-source.zip,
release-receipt.json and SHA256SUMS. The plugin ZIP contains only the
manifest-owned payload under one servotab/ root; the source ZIP includes the
marketplace and maintainer tooling. Follow Releasing for
the deterministic gate, exact-commit checks and downloaded-asset verification.
Current publication receipts belong in current state.

GitHub publication, OpenAI Directory publication, local installation and
activation, and website deployment have separate receipts. The last separately
observed public Directory version is 0.6.4. This release supplies the ZIP for
an owner-controlled manual upload; it does not perform that upload. No new
live-model evaluation is part of release preparation.

中文说明

0.7.0 汇总这几轮 Plan strategy、Design/Debug 判断与 Incorporate 更新，保持一个
implicit router、十二个 explicit leaves 和 70-file plugin payload，不新增 runtime
service、dependency、hook、强制报告或审批轮次。

Plan 在目标已清楚但顺序、投入、证据或恢复仍需判断时形成并调整 execution strategy，
需要连续性时维护同一个明确 task entry；恢复先核对当前代码与证据。Router、Execute、
Spec Chain、Delegate 与 Finish 延续同一目标、记录和责任，小任务可以保持 inline。
它不切换 host Plan Mode，也不增加执行或发布权限。

incorporate 接替 review-feedback，负责让指定讨论、反馈、研究和参考对目标工作
产生恰当影响：先覆盖重要贡献，再选择实施重点；结合上下文处理纠正，保留明确选定的
形式，判断适用性，并把采纳后的影响带到获授权的实现与验证。技术 review 的代码核实、
有界 finding disposition 与 shared-contract correction 均保留。开放想法也可以改善
设计、例子或检查，无须自动变成功能。Design 保留开放决策，Delegate 保留责任分配。

Debug 用最近已验证的正常对照约束选择性失败解释；Design 指明推荐方案胜过最近可行
替代项的决定条件，必要时按时间、位置、条件或尺度分离冲突属性，不给普通任务增加
固定练习。

升级时把保存的 $review-feedback 调用和相关 method-path references 改为
$incorporate；新包不保留 alias 或平行 leaf。servotab@personal 与安装路线不变。
Source pack 现有 30 cases、26 semantic-review cases、63 activation seeds 和
19 adversarial controls。此前 Plan 有界对照为候选 6/6、基线 5/6 accepted，不能外推
普遍效果或效率收益；完整 Plan 行为验收与 Incorporate receipt-bound semantic
acceptance 仍未关闭。本次不新增 live-model 调用。GitHub release、手动 Directory
上传、安装生效与网站部署继续分别记录。

Servotab 0.6.5

Servotab 0.6.5 — connected execution across methods

Servotab 0.6.5 connects the complete-outcome method across planning, execution, review correction, and cross-method continuity. It adds no skill, runtime service, dependency, required report, or evaluation runner.

Servotab 0.6.5 was published on 2026-10-01 Singapore time as the public Latest, non-prerelease GitHub release. The observed public OpenAI directory listing remains 0.6.4; GitHub publication does not establish a directory update, website deployment, installation or activation.

Publication receipt

Tag v0.6.5 resolves exactly to merge commit c9fe5e6e773adb6ea99f349244f0fe3f5dadffdb and tree a3afab72613ed40ae9f9eaf50009865e220623de from PR #50. Main Validate run 36769375703 passed all seven jobs, including exact-source release artifacts. The four public custom assets are servotab-0.6.5-plugin.zip, servotab-0.6.5-source.zip, release-receipt.json and SHA256SUMS. The main-CI artifact, downloaded draft and fresh public download passed the complete release-builder check against the local exact-commit build. The 70-file plugin ZIP has SHA-256 21d22612b4cd45cac69d176a612232dd47e065f762636c1d047d13b7d491c58d.

The immutable release tag and assets remain bound to that source. Later documentation receipts do not replace them. Current installation, website and directory observations are recorded separately in current state.

Changes

The release connects four existing decision points that the earlier releases already owned:

• plan selects an early executable path through the real entry point and material state boundary before it sequences the rest of the work, so the first slice closes a real observable outcome instead of only the easiest isolated piece.
• execute connects and observes that entry/state path before expanding peer features. It keeps risk-appropriate red-green work and the remaining accepted scope intact, so connecting slices never becomes a reason to drop tests, shrink the request, or defer the full outcome.
• review-feedback carries a confirmed correction to every consumer of the same shared contract while protecting legitimate exceptions and explicit write limits. It verifies the correction against the real contract rather than applying it to the reported symptom alone.
• The router and the explicit execute leaf keep one goal, plan, responsibility owner, evidence set, and remaining scope continuous across the methods that participate in a task, and withdraw investigation or ceremony whose purpose has ended instead of restarting the commission at each method boundary.

The router's implicit activation topology is unchanged. Clear changes still stay direct, and stronger method still appears only when risk, scope, or uncertainty warrants it.

Evidence cases

The source pack now contains twenty-five cases; twenty-one declare non-empty human_review_requirements, and the twelve existing adversarial controls are unchanged. The package still contains one implicit router, twelve explicit-only leaves, and 70 manifest-owned files.

feedback-correction-radius is the new case. A confirmed review correction must reach every consumer of the same user-date contract while preserving the UTC audit exception and the exact write boundary. The prompt and the visible regression expose only the chart symptom and do not name the sibling consumers, so the repair radius must be derived from the implementation. An independent command oracle checks the complete user-date contract: chart, detail, and share output must move together, UTC audit semantics must remain unchanged, exact minute offsets must be preserved, and only days.py may change. Semantic review requires trace evidence that the agent derived that repair radius from the code rather than from the prompt alone.

The existing complete-notes case now reserves semantic review for whether the selected CLI and persistence path was actually exercised before broad migration and error-path expansion, early enough to inform later implementation rather than being checked only at final verification. Its deterministic complete-delivery oracle is unchanged.

New deterministic fixture regressions check the correction-radius control; expected reports remain reference controls, never fabricated attempt receipts. Semantic requirements must be judged from real attempt evidence, and deterministic success establishes fixture mechanics only.

Validation and release boundary

Run the repository's complete deterministic source/package gate from AGENTS.md and the Python regressions:

python3 -m unittest discover -s scripts -p 'test_*.py' -q

When the standalone fieldlab CLI is available, the no-target-model subject checks are:

fieldlab validate fieldlab-pack.json
fieldlab selftest fieldlab-pack.json
fieldlab list fieldlab-pack.json

These checks establish source, generated-package, and deterministic fixture consistency for this source revision. They do not execute a target model, close declared semantic-review requirements, install or activate the package, deploy the website, or update the OpenAI directory. Exact publication state belongs in current state.

No paid model calls, live router changes, or network operations are required by this release. Any future live synthetic attempt needs its own plan and explicit invocation budget under the existing evaluation contract.

中文说明

0.6.5 把 complete-outcome 主线接到 plan、execute、review-feedback 与跨 method 连续性四个既有决策点上，不新增 skill、runtime service、dependency、强制报告或 evaluation runner。

plan 在选择实施顺序前，先识别穿过真实 entry point 与关键 state boundary 的早期可执行路径；execute 在批量扩展 peer features 前先接通并观察该路径，同时保留 risk-appropriate 的 red-green 工作与剩余已接受范围；review-feedback 把已经核实的修正带到同一 shared contract 的全部 consumers，同时保护合法例外与明确写入限制；router 与显式 execute leaf 让 goal、plan、responsibility owner、evidence 与 remaining scope 在多个 methods 之间保持连续，并及时撤回目的已结束的调查或仪式。Router 的 implicit 激活拓扑不变。

Source pack 现有 25 个 case，其中 21 个声明非空 human_review_requirements，12 个既有 adversarial controls 不变；package 仍是一个 implicit router、十二个 explicit leaves、70 个文件。新增的 feedback-correction-radius 要求一条已核实的 review 修正覆盖同一个 user-date contract 的全部 consumers，同时保留 UTC audit 例外与精确写入边界：prompt 与可见回归只暴露 chart 症状，不点名 sibling consumers，修复半径必须来自实现；独立 command oracle 检查完整 contract，semantic review 要求 trace 证据。既有 complete-notes 保留原有 deterministic oracle，只把 semantic review 用于判断所选 CLI 与持久化路径是否在广泛迁移与错误路径扩展之前真正被执行。

0.6.5 已于 2026-10-01 新加坡时间发布为公开 Latest、非 prerelease 的 GitHub Release；tag、四份 assets、main CI 与重新下载的文件已按上面的 exact-source receipt 核验。公开 Directory listing 仍为 0.6.4；GitHub publication 不等于 Directory 更新、网站部署、安装或生效。确定性检查与 Field Lab validate/selftest/list 只建立 source、generated-package 与 fixture 一致性，不建立 live model behavior。任何后续 live attempt 都需要单独计划与显式 invocation budget。

Servotab 0.6.4

Servotab 0.6.4 — evidence-scoped debugging

Servotab 0.6.4 is the evidence-scoped debugging release. GitHub publication, directory submission, installation, activation, website deployment and live model evaluation remain separately evidenced states; publishing this release does not establish the later ones.

The tagged GitHub Release was published on 2026-09-25 from merge commit 877496f460c673c7468d2120e4b6a18fe90fbec0. It provides the 70-file plugin archive, source archive, release receipt and checksums. A fresh public download reproduced and verified all four assets; the plugin ZIP SHA-256 is de75cca764ae93ce51a6361e70eef6b1ee343ddffa011b3fe37501daf92f70c3. The OpenAI Plugins Directory was still observed at 0.6.3 when this release was recorded, and a later logged-out public readback the same day reported version 0.6.4, developer Yifei Fang, and all thirteen public skills, including Worktree. That readback establishes public listing metadata and inventory only: the listing exposes no package digest, so it does not establish byte identity with the GitHub plugin ZIP, and it does not prove local installation, activation or live model behavior.

Changes

This release builds the merged discussion-intake rehearsal into a 24-case source pack and sharpens existing debug and verify decisions, with a narrow alignment of the execute failure-handling handoff, without adding a skill, runtime service, dependency, required report or evaluation runner. The existing router already covers unknown causes, conflicting evidence and repeated failure; it stays unchanged.

debug separates the mechanism experiment from final acceptance. A surrogate must preserve the conditions relevant to the tested cause. Trial results can be supportive, contradictory or inconclusive; an unverified intervention or failure to restore the whole experience does not erase a contributing defect. Direct boundary violations and loss of discriminating evidence prompt an earlier diagnostic reset. Counts remain a backstop, and a reset does not automatically require an architecture search. Permanent repairs follow component and lifecycle ownership.

execute failure handling now points at the same evidence rule instead of a mechanical two-attempt reset: test one discriminating causal claim at a time, retain independently observed defects, and reset only when repeated attempts on one idea stop adding discriminating evidence.

verify checks probe entry, path and conditions, pairs disappearing errors with successful outcomes, separates mechanism repair from complete recovery, and prefers structured or stable normalized before-images for state-preservation claims.

Diagnostic cases

Three cases extend the merged 21-case source pack (20 prior cases plus the discussion-intake rehearsal) to twenty-four:

• cross-boundary-diagnosis: a local executable admission model with healthy bypass probes, a confounded earlier trial and an independent residual delay. Acceptance requires actual comparable pre/post experiments and successful delivery, not a proposed plan or an edited counter.
• host-specific-diagnosis: a desktop probe lacks a material native-host condition and cannot exclude that cause.
• diagnosis-progress: successive experiments make real causal progress despite incomplete overall recovery; a numeric failure count must not trigger a pointless restart.

The existing Field Lab schema, baseline/expected overlays and semantic-review boundary are reused. New deterministic fixture regressions check red/green delivery, residual symptoms, non-exempt paths and incorrect repairs. Expected reports are reference controls, never fabricated attempt receipts. Semantic requirements must be judged from real attempt evidence.

Validation and release boundary

Run the repository's existing source/package gate and python3 -m unittest discover -s scripts -p 'test_*.py' -q. The focused controls can be run with python3 -m unittest discover -s scripts -p 'test_diagnostic_fixtures.py' -v. Exact-source CI, release receipts and publication state are recorded separately; this note and the version string do not establish them by themselves.

No paid model calls or live router changes are required by this release. Host installation, directory upload, website deployment and live model evaluation remain separate operations. Fixture success establishes fixture mechanics, not improved agent behavior. A future live attempt needs its own explicit invocation budget and review under the existing evaluation contract.

中文说明

0.6.4 把已合并的 discussion-intake rehearsal 一起纳入 24 个 case 的 source pack，并集中改进跨边界调试中的实验选择与结论范围，保留一个 router、十二个 leaves。机制复现与真实用户验收分工；surrogate 只需保留与被测因果判断相关的条件；试验无结论时保留已成立的观察；按探测实际经过的路径理解绿灯；信息停滞时及时重查诊断，不因次数机械重启或直接展开架构替换。execute 的失败处理交接同步改为同一证据规则，不再使用机械的两次尝试重置。

新增一个可执行的共享边界案例与两个判断反例。确定性检查验证案例及错误修复的拒绝能力；实际模型是否主动选对实验，仍需真实 attempt 和语义审阅。源码、完整验证、GitHub 发布、安装、生效、Directory 更新与网站部署分别记录，不把文案修改或测试通过写成行为能力已经提升。

Servotab 0.6.3

Servotab 0.6.3

0.6.3 expands worktree from isolation setup into workspace lifecycle guidance: find the right task, reuse or restore its workspace, create isolation when it solves a real conflict, and park or clean up only within the authorized scope.

Servotab 0.6.3 was published as a tagged GitHub Release on 2026-09-20 from merge commit 6a57dbb656dc2d1fcd7674d109817e358cfeffdf. The release provides the source archive, 70-file plugin archive, receipt, and checksums. Publication does not by itself establish installation, activation, website deployment, or an OpenAI directory update.

Changed behavior

• Calling $worktree no longer implies creation. Inspection and organization requests produce concrete dispositions without authorizing deletion. The ordinary-language router recognizes recovery, parking, and cleanup as well as isolation.
• Creation first checks existing task workspaces and reconciles ambiguous earlier results. Dirty state, duration, risk, or worker count alone do not demand a new checkout. Verify required starting content: ordinary Git creation at a revision does not transfer dirty prerequisites, while host tools have their own contracts.
• Parking can release an inactive, unmerged checkout with a durable named ref and a restoration route. It does not auto-commit, stash, archive, upload, merge, or delete the branch.
• Cleanup checks ignored data, tracked edits hidden by assume-unchanged or skip-worktree, detached tips, current integration targets, active ownership, locks, in-progress Git operations, submodules, and unavailable storage where relevant. Directory removal, local/remote branch deletion, data disposal, and registration pruning remain separate actions.
• Flagged tracked files that are present need a content comparison without changing the real index or its flags, or the candidate stays unresolved. Ordinary status and non-force removal do not prove that these local edits are preserved; a retained branch restores committed content only.
• Exact existing authorization is reusable for a concrete batch. Recheck each item before removal, leave changed or uncertain items pending, execute independent approved items, and verify preserved refs afterward.
• finish shares the removal contract. delegate now explains that different branch names in one checkout do not separate working files or indexes, and that worktrees still share refs and may share runtime resources.

The one-router/twelve-leaf topology, 70-file package, skills-only product boundary, and licenses remain unchanged. No runtime cleanup executor, new dependency, global registry, or unattended deletion service is added.

Evidence and limits

The source pack contains twenty cases. Three new synthetic decision rehearsals cover reuse after ambiguous creation, inspection-only organization with mixed risks, and already authorized parking with one changed candidate. Structural checks verify output existence and write scope; declared semantic review requirements judge the decisions. The explicit worktree-organize case requires actual leaf delivery/read evidence in semantic review; it does not require a redundant router-reference read. Deliberately unsafe or obstructive outputs remain adversarial controls because correct labels cannot prove sound reasoning.

scripts/test_worktree_git.py checks the underlying Git mechanisms in disposable local repositories: ignored-data deletion, non-mutating hidden-tracked-content inspection and loss on removal under both index flags, unmerged named-ref preservation/restoration, detached tips without named anchors, squash/ancestry differences, dirty prerequisites, locks, and missing/moved registrations. These tests do not execute a target model or prove host-managed cleanup behavior. The rehearsal cases do not execute real cleanup either; a future live operational case needs its own fixture, saved plan, and explicit invocation budget.

The canonical package gate, fixture controls, and site checks validate source consistency and their stated mechanisms. No live 0.6.3 model-effectiveness result is claimed. Exact verification, publication, installation, and deployment state belongs in current state.

Operational basis

Git distinguishes per-worktree files/indexes from shared refs, documents lock/repair behavior, and defines pruning as removal of stale administrative records. Use the supported host lifecycle where it owns a workspace. See the official Git worktree documentation and branch deletion contract. The official index flag documentation and file-tag inventory explain the hidden-state check. Git's acceptance of a removal is not evidence that all local data was disposable or that a branch was integrated.

Install the tagged source package through the existing marketplace route when deliberately testing this release; no migration or legacy retirement is required. An installed 0.6.2 package keeps the earlier worktree guidance until updated to 0.6.3 and rediscovered by its host.

Servotab 0.6.2

Servotab 0.6.2

Maintenance and delegation-contract update for the quiet, risk-scaled engineering method plugin. The package retains one implicit router and twelve explicit methods while moving to a 70-file portable package layout.

Changes

• Define LF as canonical text identity so Git for Windows core.autocrlf=true checkouts and equivalent CRLF retained by older worktrees do not create false stale-package reports; binary assets remain byte-exact.
• Add the portable root plugin.json with extensions.com.openai and retain the synchronized .codex-plugin/plugin.json compatibility fallback.
• Choose responsibility before deep execution: keep clear and trivial work direct, honor explicit solo requests, reuse one active worker, and allow one coherent bounded lane when delegation actually pays.
• Preserve coupled parts within one owner, forbid recursive delegation, and sequence locally when the host exposes no subagent capability.
• Add explicit tranche-planning and cross-process complete-delivery controls plus four delegation responsibility-choice canaries, bringing the source pack to seventeen cases.
• Separate deterministic trace and workspace assertions from semantic review. One independently budgeted workspace-scoped delegation attempt passed deterministic verification and all three receipt-bound review requirements; that evidence does not establish general model effectiveness.
• Refresh public package, release, website, and publisher-identity surfaces for 0.6.2.

Package and upload

servotab-0.6.2-source.zip contains the repository marketplace and maintainer tooling. servotab-0.6.2-plugin.zip contains only the 70-file plugin payload for an owner-controlled directory upload. release-receipt.json identifies the exact source commit/tree and package manifest; SHA256SUMS checks both archives and the receipt. Verify the source provenance as well as the checksums.

The package metadata identifies Yifei Fang as publisher and maintainer. Public creator credit remains Faye & Cove.

Evidence and limits

Deterministic validation covers canonical/generated identity, exact package ownership, parsed SVGs, decoded PNGs, package/release regressions, seventeen behavior fixtures, public-tree safety, website behavior tests, and the production build. The Windows CI job checks generated payload identity under Git for Windows defaults. Check the attached receipt and the exact source revision's GitHub checks.

The tagged GitHub release establishes public source and artifact availability for the exact release revision. It does not by itself establish installation, activation, OpenAI directory submission, review, approval, or publication. The website was deployed and verified separately after the GitHub release. Publisher/creator credit, licensing, privacy boundaries, and the skills-only product model are otherwise unchanged.

Servotab 0.6.1

Servotab 0.6.1

Release hardening for the quiet, risk-scaled engineering method plugin. One implicit router, twelve explicit methods, and the 69-file skills-only package remain unchanged in topology.

Changes

• Validate real SVG XML and decoded PNG pixels, rejecting malformed assets, non-finite/undersized canvases, active content, external resources, truncated images, and source-format conversion.
• Reject ambiguous metadata, invalid manifest entries, unowned package content, and generator symlink writes.
• Clarify the ordinary test-criterion check: no mandatory mutation-testing stage or extra reviewer loop. Explicitly authorized foundational work remains possible.
• Prepare reproducible plugin/source ZIPs with exact-source receipts and checksums, gated by source/package/site CI.
• Reconcile release evidence and add two public behavior fixtures for reuse and false-green verification.

The release also carries the 0.6 source improvements: dependency-ordered design decisions, bounded reuse research, risk-matched verification and review disposition, race-safe homepage motion, and per-skill SVG canvases compatible with the observed upload-scanner minimum.

Assets and installation

servotab-0.6.1-source.zip contains the repository marketplace and maintainer tooling. servotab-0.6.1-plugin.zip contains only the 69-file plugin payload for a separate owner-controlled directory upload. release-receipt.json identifies the exact source commit/tree and package manifest; SHA256SUMS checks both archives and the receipt. Verify the GitHub source as well as checksums.

From the extracted source directory, or a checkout of this release tag:

codex plugin marketplace add .
codex plugin add servotab@personal
codex plugin list --marketplace personal

Start a fresh Codex task/process and confirm the installed version. Older manifest-owned Softpowers installations must follow the migration guide; do not delete their directories by hand.

Evidence and limits

Deterministic validation covers package and asset structure, canonical/generated identity, packaging/release regressions, behavior-fixture failure/repair controls, and nine website motion tests plus the production build. Check the attached receipt and the exact revision's GitHub checks. Fixture controls do not invoke a target model; no general model-effectiveness improvement or 0.6.1 named-host acceptance is claimed.

Maintainer validation requires Python 3.10+ and pinned PyYAML/Pillow from requirements-dev.txt. Neither is shipped in the installed skill payload. A GitHub release does not update the OpenAI directory or deploy the website. Publisher/creator credit, licensing, privacy boundaries, and the skills-only product model are unchanged.

Softpowers v0.3.0-rc4

License Boundary is now a recommended standalone companion, not part of the Softpowers payload. This RC installs 14 skills, adds a pre-mutation ownership gate for projection-bearing historical layers, and makes uninstall manifest-driven across the real 12/13/14/15-skill stack. Users upgrading from an RC that bundled License Boundary should use this release uninstaller repeatedly until no active layer manages license-boundary, then install rc4 and install License Boundary v0.1.0-rc3 independently.

Softpowers v0.3.0-rc3

This release candidate integrates the standalone License Boundary skill and closes the current behavior-eval review findings.

Licensing skill

• Includes an exact offline projection of IndelibleVivi/license-boundary v0.1.0-rc2.
• Ships the complete packaged SUL-1.0 terms and project notice with the installable skill.
• Keeps the standalone repository as canonical authoring authority and rejects projection drift.

Behavior eval hardening

• Locks resume to case, prompt, fixture, runner, Codex executable/version, model, timeout, and workspace-retention identity.
• Makes fixture commits independent of global GPG-signing settings.
• Resolves relative Codex executable paths before workspace cwd changes.
• Correctly parses NUL-delimited rename records.
• Supports the full self-test on Python 3.10 and 3.13.

Pattern provenance

• Uses one schema-valid top-level decision per candidate record.
• Records exact reviewed files and review dates for all ten external-pattern samples.

Validation passed on the public-tree audit plus Ubuntu Python 3.10/3.13 and macOS Python 3.13 GitHub Actions jobs.

Softpowers v0.2.0-rc4

Softpowers 0.2.0-rc4 是一个 community-maintained、非 OpenAI 官方的 public release candidate。

这个版本包含完整的 13-skill pack：

• softpowers 是唯一具备 implicit-invocation eligibility 的 router；
• 12 个 leaf skills 是 explicit-only shortcuts，低频使用是设计的一部分；
• complete requested usable outcome 现在是 implementation default；
• Worker Lanes 使用 bounded work orders、one-writer ownership 与 explicit return packets；
• public behavior / installer feedback forms、contributor guidance 与 fail-closed public-tree audit 已加入。

安装：

git clone https://github.com/IndelibleVivi/softpowers.git
cd softpowers
./install.sh

需要 Git、Bash、Python 3.10+。CI 已覆盖 Ubuntu Python 3.10 / 3.13 与 macOS Python 3.13；native Windows 尚未验证。相同版本重复安装仍会创建新的 rollback layer；从 v0.2 rollback 到 v0.1.x 后，继续移除 legacy layer 需要使用匹配的 v0.1.x uninstaller。

• README
• Changelog
• Behavior feedback
• Installer feedback
• Validation