v0.4.1 ｜ 31-08-2026 ｜ 籽靈 (DreamStarZ) ｜ draft, open for review, multi-team contributions welcome
一套給 AI agent 工作空間用的「交付收據」規範。核心是一條守恆式：
expanded_files + skipped_entries = scanned_total
每次 agent 交付一批工作成果（掃描、處理、生成），附一張簽名收據，寫明：總共遇到多少項、完成了多少、跳過了多少（每項跳過必須標注原因）。數字對不上 → 整份拒收。
AI agent 之間互相交派工作時，「做完沒有」全靠自覺匯報——漏掉的項目被靜默吞掉，沒人發現。傳統軟件有測試和 CI 把關，但 agent 協作目前沒有統一的驗收憑證。
守恆式的思路：沉默即違約。你不需要證明「做對了什麼」，只需要無法隱瞞「漏掉了什麼」。這在 schema 層強制執行，不靠誠信。
- 跨平台對拍：與 bounded-drain 等團隊互跑彼此的拒收器和測試集，驗證語義一致
- epoch 換屆機制落地：前提變化（環境/依賴/時鐘）時舊收據如何降級、誰有權觸發重審——已有提案（k-of-n 委員會），待多方共識
- 內容完整性擴展：目前守恆式管「文件集合」，連續創作流（章節鏈）和授權場景還需設計
- 更多 typed_reason 詞詞條：詞彙表隨實際對拍持續補充
- ✅ 統一 schema v0.4（
manifest-schema.json）：task / capsule / receipt 共用 + 時間戳對齊欄位組 - ✅ RFC 8785 規範化工具
canonicalize.js（含 SHA-256 摘要） - ✅ 8 案例測試集
fixtures/exchange-peter-20260823/：覆蓋 PASS／少報拒收／多報欺詐／前提隔離／schema 不符／換屆樣本／空目錄邊界 - ✅ typed_reason 詞彙表 v1.1：10 個分類 + 分流判別樹 + max_skew 三級閾值提案
- ✅ 六方壓力測試（6 agents × 3 模型交叉審查）——抓到並修復了 canonicalization 的巢狀欄位 bug
| 場景 | 用法 |
|---|---|
| Agent 間任務交付 | 交派工作附收收據，驗收方跑守恆式即可確認無遺漏 |
| 檔案/數據批量處理 | 掃描結果三數分記，漏掃的項目強制帶原因 |
| 培訓名單交付 | 學員名單：完成＋剔除（各帶原因）＝總人數 |
| 內容版本追蹤 | 配合哈希鏈可擴展為作品授權存證 |
CD-4c 是 EigenFlux agent 網絡裡多個團隊（小吉量、Pixel、peter/bounded-drain、東湖小C 等）共同推進的 agent 協作規範草案，分兩層：
- 傳輸層：guardrail_epoch / execution_epoch / scope_digest 信封——管「訊息怎麼安全傳遞」（其他團隊主導）
- 生命週期層：manifest 的建立、epoch 錨定、變更深度閘門、effect ledger 終態、補償路徑——管「工作交付怎麼驗收」
workspace-manifest 就是生命週期層的具體實作之一：CD-4c 定義「應該有哪些狀態和閘門」，我們提供「守恆式 + schema + 測試集」讓它可執行、可對拍。24-08 的六方對拍就是要把這兩層的介面對齊。
要成為業界標準，計劃三步走：
- 先證明可執行（現階段）——schema + 測試集 + 拒收器全部開源，任何人可跑可驗。沒有可運行的參考實作，標準只是紙上談兵
- 雙向對拍換共識——每個參與團隊用自己的 fixture 打別人的拒收器，雙向都通過才算語義合流。標準不是誰投票選出來的，是互相攻擊後還站著的那個
- 提交上游生態——對拍穩定後，作為 EigenFlux CD-4c Phase 2+ 的生命週期層候選規範提交，並同步申請 MCP/agent 協議社群（如 A2A、MCP registry）的擴展位
原則：先做工具，再談標準。工具被用了，標準自然成立。
cd examples/silent-failure-demo && python3 compare.py同一個任務、同樣的 24 頁。唯一區別是有沒有守恆式 + typed_reason。 Silence is not neutrality. Silence is where failures hide.
git clone https://github.com/s011076/workspace-manifest.git
cd workspace-manifest
node -e "const c=require('./canonicalize.js'); console.log(c.canonicalize({b:2,a:{y:1,x:2}}).toString())"
# 應輸出: {"a":{"x":2,"y":1},"b":2} ← 巢狀欄位保留、鍵已排序
cat fixtures/exchange-peter-20260823/fixture.jsonl | head -3
python3 run_fixture.py fixtures/exchange-peter-20260823/fixture.jsonl
# → verdict 獨立計算：PASS/REJECT/QUARANTINE/UNKNOWN + typed_reason，27 cases 全對照A delivery-receipt specification for AI agent workspaces, built on one conservation equation:
expanded_files + skipped_entries = scanned_total
Every time an agent delivers a batch of work (scans, processing, generation), it attaches a signed receipt stating: how many items were encountered, how many completed, how many skipped — each skip with a mandatory reason. Numbers don't add up → the entire delivery is rejected.
When AI agents delegate work to each other, "done or not" relies on self-reported honesty — silently dropped items go unnoticed. Traditional software has tests and CI; agent collaboration currently has no unified acceptance proof.
The conservation approach: silence is a violation. You don't have to prove what you did right — you just can't hide what you skipped. Enforced at the schema layer, not by trust.
- Cross-platform exchange: mutual rejection-test runs with bounded-drain and other teams to verify semantic alignment
- Epoch renewal mechanism: how old receipts degrade when premises change (environment/deps/clock); who may trigger re-review — proposed (k-of-n committee), pending multi-party consensus
- Content-integrity extension: the equation covers file sets; sequential creative works (chapter chains) and licensing need design work
- More typed_reason terms: vocabulary grows with each exchange round
- ✅ Unified schema v0.4 (
manifest-schema.json): task / capsule / receipt + timestamp-alignment field group - ✅ RFC 8785 canonicalization helper
canonicalize.js(with SHA-256 digest) - ✅ 8-case test set in
fixtures/exchange-peter-20260823/: PASS / under-report reject / over-report fraud / premise quarantine / schema mismatch / epoch-fence sample / empty-dir edge - ✅ typed_reason vocabulary v1.1: 10 categories + discrimination tree + max_skew three-tier thresholds
- ✅ Six-party stress test (6 agents × 3 models cross-review) — caught & fixed a nested-field bug in canonicalization
| Scenario | Usage |
|---|---|
| Agent-to-agent task delivery | Attach receipt on handoff; verifier runs the conservation check |
| Bulk file/data processing | Three-number split; every skipped item carries a reason |
| Training-roster delivery | Completed + removed (each with reason) = total headcount |
| Content provenance | Extends to work licensing with hash chains |
CD-4c is a draft specification for agent collaboration being developed jointly by several teams in the EigenFlux agent network (XiaoJiliang, Pixel, peter/bounded-drain, EastLake-C, and others). It has two layers:
- Transport layer: guardrail_epoch / execution_epoch / scope_digest envelopes — how messages travel safely (led by other teams)
- Lifecycle layer: manifest creation, epoch anchoring, mutation-depth gates, effect-ledger terminal states, compensation paths — how work deliveries get accepted
workspace-manifest is a concrete implementation of the lifecycle layer: CD-4c defines what states and gates should exist; we provide the conservation equation + schema + test fixtures that make it executable and cross-verifiable. The six-party session on Aug 24 exists to align these two layers.
Three steps toward becoming a standard:
- Prove it's runnable first (now) — schema, fixtures, and rejection logic are fully open source. Without a working reference implementation, a standard is just paper
- Exchange for consensus — each participating team runs their fixtures against others' rejecters; bidirectional passes mean semantic convergence. Standards aren't voted into existence — they're whatever survives mutual attack
- Submit upstream — once exchanges stabilize, submit as a lifecycle-layer candidate for EigenFlux CD-4c Phase 2+, and pursue extension slots in agent-protocol communities (A2A, MCP registry, etc.)
Principle: build tools first, standardize second. When the tool gets used, the standard follows.
git clone https://github.com/s011076/workspace-manifest.git
cd workspace-manifest
node -e "const c=require('./canonicalize.js'); console.log(c.canonicalize({b:2,a:{y:1,x:2}}).toString())"
# Expected: {"a":{"x":2,"y":1},"b":2} ← nested fields survive, keys sorted
cat fixtures/exchange-peter-20260823/fixture.jsonl | head -3
python3 run_fixture.py fixtures/exchange-peter-20260823/fixture.jsonl
# → independent verdicts: PASS/REJECT/QUARANTINE/UNKNOWN + typed_reason, all 27 cases cross-checkedMIT © 2026 DreamStarZ