Post-Experiment Knowledge

0.8.0-beta.1 · 实验后知识管理插件 · Windows x64 一键安装测试版

把已完成实验整理为“专题主报告 → 实验索引 → 完整证据与追溯档案”，保留原始证据，持续修订结论，避免报告反复堆叠。不执行实验。

1. 在 GitHub Releases 下载 Windows x64 ZIP 并解压。
2. 双击 install.cmd，安装器会校验文件并注册到当前用户的 Codex。无需安装 Python 或 pip，无需管理员权限。
3. 新开 Codex 对话，直接说明要归档、分析、保存报告或全部完成；Agent 根据场景选择技能。也可显式输入 $post-experiment-knowledge。每次新请求先用原生选项卡问清需求、确认模板和具体保存位置，再开始处理；已有偏好仅预填，不自动执行。

需要已有支持插件的 Codex。飞书功能另外需要自己的 lark-cli 和账号授权；本地 Markdown 无需飞书。请先阅读 Beta 已知限制。本包不会携带作者的实验资料、私有地址或凭据。

Record, Synthesize, Sync for already-produced experiment results. The Windows x64 release includes its own CPython 3.14.10 runtime: users do not install Python or pip. The core has no third-party dependencies. No training execution, scheduling, monitoring, GPU management, model APIs or telemetry.

Designed for Codex and similar local development Agents. Distribute the local plugin/Skill directly; no public plugin-directory submission, hosted application or extra MCP server is required. Feishu is optional and uses each installer's own lark-cli installation and existing account authorization. No account, tenant or experiment data is bundled.

An explicitly confirmed personal-space Wiki parent may be kept in private application-data preferences. Setup exposes that default to the Agent; it never becomes a public plugin default, silently selects an upload scope or falls back to Drive root. Topic-specific report/doc bindings remain separate.

Start without a platform

From an extracted Windows x64 release directory:

skills\post-experiment-knowledge\scripts\expkm.cmd --help
skills\post-experiment-knowledge\scripts\expkm.cmd doctor
skills\post-experiment-knowledge\python\python.exe -I -B examples\synthetic30.py --output PRIVATE_DEMO_DIRECTORY

The synthetic example produces 30 independent archived records, a single maintained conclusion, an experiment index, history and a Markdown sync preview. It does not run an ML model. Choose a destination outside this source tree. A source checkout or another platform can use python skills/post-experiment-knowledge/scripts/expkm.py and python examples/synthetic30.py with an existing Python 3.10+ installation; source-only ZIPs do not include an interpreter.

Optional core installation: python -m pip install --no-build-isolation --no-deps .. For an Agent Skill or Plugin package, use the packaging helpers described below; do not copy private data into the repository.

Four skills, one runtime

All four allow implicit invocation and share one core, private store/config and isolated bundled interpreter. A full request completes every required stage; a narrower confirmed request cannot silently add others. Host semantic matching is not a deterministic watcher.

Research topic
├── Main report: current conclusions, key evidence, limits and decisions
├── Experiment index: attempts, differences, states and archive links
└── Complete archive
    ├── Raw evidence: results, configuration, logs and artifact references
    └── Trace: record IDs, revisions, hashes and claim relationships

Every request starts with confirmed intake

Before formal experiment processing, clarify this request's goal, source scope, requirements and expected outputs, require a confirmed template and exact save locations, then obtain confirmation of the complete summary. This also applies to Record-only and analysis-only. Prior topic choices may prefill fields; they do not authorize a new request.

Use native choice cards for goal/scope, template source, save method, upload scope and final confirmation. Send exactly one question per native call and wait for its actual answer before advancing. The private cards start|next|answer|append|status|close queue retains unanswered items across interruptions; intake confirmation refuses to skip them. Re-present the pending card if it was not visible. Paths, links and special requirements use the same tool's free-text fields. A preselection or silence is not confirmation. Without native cards, use the same one-at-a-time numbered fallback.

The template card offers:

1. Provide an existing template file/link/structure.
2. Explicitly reuse a previously supplied template.
3. Ask Codex to recommend a template from actual experiment content/results, then review it.

The recommendation branch first confirms requirements, bounded source scope and exact positions, then uses a restricted template-recommendation intake for necessary read-only context. It cannot archive, synthesize research conclusions or publish. Show the recommendation/reasons and obtain template confirmation before a separate final execution intake. No recommendation is automatically adopted.

Commands and boundaries

Use skills/post-experiment-knowledge/scripts/expkm.cmd on Windows. Global --store, --config and --intake options precede the command. Real data/settings/receipts stay outside the source and installed plugin.

1. setup --topic quality --stage record|synthesize|sync (Sync --sync-kind report|archive|both) returns required questions. The coordinator's default workflow mode clarifies an ambiguous outcome.
2. Confirm the chosen template with template preview/confirm and the exact private result directory with record-location preview/confirm. Requested reports additionally select exact destination targets; requested uploads select archive-destination scope/parent. No arbitrary default platform or path.
3. Author the private request metadata file, run intake preview --file PRIVATE_REQUEST, show the concrete summary, await the user's response and intake confirm --proposal PRIVATE_PROPOSAL --approve VERSION.
4. Pass --intake VERSION to Record/Synthesize/new Sync and export commands. The CLI checks scope, stages, authorized roots and current template/location versions before payload reads. New report/archive plans bind the intake; apply cannot replay a plan under another intake. Original plan review, allowlists and write/upload flags still apply.
5. Keep the coordinator's workflow manifest current and verify every requested output separately. Both report targets require two receipts; archives require exact parent and downloaded size/SHA-256. Archive freshness audits reuse the sealed size budget, including explicit large export limits.
6. Close the request with intake close --version VERSION after delivery. A closed request cannot be reused. New requests use new IDs and confirmation, even with identical preferences.

See the intake schema and recommendation procedure. Current synthesis/export operates on all existing records of the selected topic. Do not silently widen a subset request; use an explicitly separate topic/store or explain the unsupported scope.

The mandatory result location for local-only work is the exact private store. No unrequested report document is required, but the mandatory template still is. Prepared reports can go directly to Sync after intake; don't force re-recording or re-analysis.

Templates preserve headings/order and general data/environment/metric tables while allowing different result formats. No built-in template is silently chosen. Historical records/plans remain verifiable; new applies need current intake. Core Python primitives are trusted application APIs and remain platform independent; the CLI gate cannot prevent an authorized developer from calling them directly.

The runtime proves explicit confirmation parameters and bindings, not that a human saw a card or who clicked. It cannot force a host to invoke this plugin. Semantic interpretation and readability still need Agent/user review.

Record defaults to --source-mode auto: finite known sources extract metadata; code, binaries, unknown encodings, malformed JSON and NaN/Infinity inputs are preserved as inert exact bytes with a raw-only descriptor and no inferred metrics. No Agent library bypass, code execution, unpickling or original-file conversion is required. Use --source-mode structured for strict finite parsing or --source-mode raw to skip extraction. Metadata sidecars remain strict finite JSON. Local archival does not grant upload permission or bypass export credential checks.

Feishu uses each user's own optional lark-cli, account and exact target authorization. Node creation/relocation uses available external tools only after the concrete parent/names/setup scope are confirmed; no experimental content is populated before final intake. No automatic login, permission escalation, model service, telemetry or sharing change.

Upload private evidence and trace archives

Only for requested archival Sync, choose no upload, trace only, or raw evidence plus trace; report-only work does not ask about uploads. archive plan --topic quality builds a deterministic ZIP with an experiment index, immutable record revisions, knowledge history and SHA-256 manifest; full scope additionally includes archived raw objects. Configs, credentials and journals are excluded by construction. External weights/resources are references, not implicitly downloaded. Review the exact inventory and credential guard before upload.

archive apply --plan PRIVATE_PLAN --approve PLAN_DIGEST --allow-feishu-upload writes only to the selected folder, which must also appear in the private feishu.allowed_archive_folders list. It uses the existing user's lark-cli identity and does not change permissions. Verification downloads the uploaded ZIP and checks folder membership, size and SHA-256. A known upload is not repeated after a failed readback; an unknown result stops for reconciliation. Report and archive receipts are separate. See archive operations for recovery and size limits. These ZIPs are private evidence artifacts, never plugin release packages.

For Wiki archives, select --archive-wiki WIKI_PARENT_TOKEN instead of --archive-folder and allowlist it under private feishu.allowed_archive_wiki_nodes. The adapter uploads directly under that node and verifies exact Wiki parent/object/name plus downloaded size/SHA-256. Folder and Wiki targets cannot be mixed. After an explicitly authorized relocation, archive adopt can verify and bind the existing file to a new reviewed location plan without uploading it again; original receipts remain history.

Before a new upload, the adapter checks the actual folder-list read capability. A valid login or successful upload does not prove that permission. Missing scopes return needs_feishu_authorization with the exact scope names, before any upload or pending-write journal; current lark-cli directory listing can require space:document:retrieve separately from upload/download scopes. No permission is silently added. UTF-8 .log inputs are supported as plain text.

Historical reports

organize --topic quality --input SNAPSHOT --out PRIVATE_INVENTORY inventories full DocxXML or a lark-cli fetch envelope without requiring report settings. This is structural diagnosis, not automatic semantic migration. The Agent creates an explicit block mapping, a concise main report, an experiment index and a complete archive. Re-run with --mapping and --main to validate coverage and produce a private migration preview. Plain Markdown history can be recorded as evidence; use Markdown sync for maintained sections.

Original bytes and resource references are retained. External resource binaries are not implicitly downloaded. Live Feishu writes support atomic replacements, not automatic archival reorganization; full historical relocation remains a separately approved operation.

Package or install

For local plugin installation, extract a release ZIP and run the following from its post-experiment-knowledge directory on a host with Codex plugin CLI support:

codex plugin marketplace add .
codex plugin add post-experiment-knowledge@post-experiment-knowledge-beta

The packaged local marketplace points at the extracted plugin. Keep that directory available for refreshes. Start a new chat or refresh/restart the host if discovery has not updated, then invoke $post-experiment-knowledge. Follow the first-use introduction and selection cards. This uses a local marketplace, not public submission. Refer to the official local marketplace instructions for host-specific UI differences. Other Agents can adapt the Skill format; their native plugin installation and implicit routing are not claimed to be tested.

For the Windows x64 release, the Agent invokes scripts\expkm.cmd inside the installed Skill. It always uses that Skill's bundled interpreter with isolated imports, ignores PYTHONHOME/PYTHONPATH, and does not modify registry or PATH. Runtime files and licenses are included at skills/post-experiment-knowledge/python/. It runs from arbitrary working directories and paths with spaces/non-ASCII names. It does not install packages or download a runtime on first use. If python/python.exe is missing, the launcher reports an incomplete package rather than using system Python. Other platform bundles are not yet provided; their source/Skill installation still requires Python 3.10+.

Feishu additionally requires lark-cli and each user's own authorization/private allowlist; local Markdown works without them. Missing dependencies are explained during setup rather than silently installed or configured.

from pathlib import Path
from expkm.packaging import build_plugin, install_skill
official_zip = Path('OFFICIAL_DOWNLOAD_DIRECTORY') / 'python-3.14.10-embed-amd64.zip'
build_plugin(Path('.'), Path('PRIVATE_BUILD_DIRECTORY') / 'post-experiment-knowledge-windows-x64.zip', windows_runtime=official_zip)
install_skill(Path('.'), Path('PRIVATE_SKILLS_DIRECTORY') / 'post-experiment-knowledge', windows_runtime=official_zip)

The legacy install_skill helper installs one discoverable coordinator with companion instructions and one runtime; use the plugin for four independently discoverable skills. The standalone entry includes generic core and, with windows_runtime, the same interpreter/licenses. Existing Skill directories are never overwritten. Omitting that argument deliberately builds a source-only package. The ZIP has portable and compatibility manifests, a local marketplace catalog and an onboarding entrypoint; no MCP server or hooks are required. Host UI rendering and account setup remain host-specific.

Builders obtain the embedded ZIP from the official Python 3.14.10 release. runtime/windows-x86_64.json pins the official archive SHA-256 and every member name, size and SHA-256. A different archive, extra member or modified member fails packaging. Binary runtime files bypass text scanning only after exact lock verification; all generic code/doc files still use the distribution allowlist and scan. The final ZIP is checked again. CPython and bundled third-party notices are retained in its original LICENSE.txt. The runtime is upgraded only through a reviewed lock/package update, never silently from a user's environment. See Python embedded-distribution guidance.

Validate and troubleshoot

Install pytest in a development environment, then run python -m pytest -q. It is not a runtime dependency. Use verify --record RECORD_ID to check archived evidence and reindex to rebuild the disposable SQLite index. CLI failures deliberately suppress raw exception payloads; inspect the private input and config in the authorized local session.

Conflicts: regenerate the preview from the current source. Unknown write outcome: re-read/verify the same plan before any further apply. Never rerun an experiment to fix a record or upload problem.

See ARCHITECTURE.md, RESEARCH.md, VALIDATION.md, BETA.md and ROADMAP.md. Real research and real acceptance artifacts are private and excluded from this package. Project code is provided under MIT License; the bundled CPython runtime retains its original PSF and third-party notices.