agentledger-runtime 1.4.0__tar.gz → 1.4.2__tar.gz

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/CHANGELOG.md

@@ -8,6 +8,40 @@ This project follows semantic versioning for the stable runtime-core contract. O
 
 No unreleased changes.
 
+## 1.4.2 - 2026-06-19
+
+### Added
+
+- Added dedicated Inspector read models and HTML panels for archived model calls and model-proposed tool calls.
+- Added stronger Inspector links between model evidence, proposed tool calls, runtime events, Tool Ledger rows, and failure records.
+- Added model evidence and proposed-tool references to portable failure exports, including Langfuse, OpenTelemetry, LangSmith, Temporal-style, and CI-friendly local mappings.
+- Added a dependency-free model evidence example showing how to record externally executed provider/gateway calls without making AgentLedger a model router.
+- Hardened boundary lint for direct database calls, direct filesystem mutation, and risky tool metadata that omits idempotency, approval, or sandbox boundaries.
+- Added a shareable side-effect safety showcase that contrasts naive duplicate side effects with AgentLedger safe retry and generated Inspector HTML.
+- Added adoption content for the showcase, including a Chinese article draft, short social posts, and a deterministic GIF recording flow.
+
+### Clarified
+
+- This is a four-language 1.4.2 release train. Runtime-core semantics remain aligned across Python, Go, TypeScript, and Rust.
+- Inspector and boundary lint remain companion/read-model tooling distributed through the Python reference package; Inspector can still consume language-neutral evidence from non-Python runtimes.
+- AgentLedger continues to record model evidence and tool proposals, not provider routing, retry, fallback, key management, or pricing catalogs.
+
+## 1.4.1 - 2026-06-19
+
+### Added
+
+- Added the Runtime Model Evidence Boundary across Python, Go, TypeScript, and Rust runtime-core packages.
+- Added `agentledger.model.evidence.v1` for externally executed model-call evidence.
+- Added `model_call_requested`, `model_call_completed`, `model_call_failed`, and `tool_call_proposed` runtime events.
+- Added APIs to record archived model requests/responses, model failures, model-proposed tool calls, token usage, and USD cost evidence without making AgentLedger a model router.
+- Added `model_call_failed` to failure attribution, failure lifecycle, Inspector failure detection, adversarial review checks, and cross-language tests.
+
+### Clarified
+
+- AgentLedger records model evidence produced by user code, agent frameworks, provider SDKs, LiteLLM/new-api/one-api, or enterprise gateways.
+- AgentLedger does not own provider routing, retry, timeout, fallback, key management, provider SDK compatibility, or pricing catalogs in runtime-core.
+- This is a four-language runtime-core patch release. Inspector remains a language-neutral companion viewer distributed through Python/PyPI.
+
 ## 1.4.0 - 2026-06-13
 
 ### Added
@@ -328,3 +362,7 @@ No unreleased changes.
 - Exact optional framework packages are not implemented.
 - Rust, TypeScript, and Go implementations are not started.
 - Production sandbox hardening requires external deployment policy.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/CODE_OF_CONDUCT.md

@@ -20,3 +20,7 @@ AgentLedger is an open-source engineering project. Contributors, maintainers, an
 ## Enforcement
 
 Maintainers may edit or remove comments, close issues, reject contributions, or restrict participation when behavior harms the project. Security concerns should be reported through `SECURITY.md`.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/CONTRIBUTING.md

@@ -96,3 +96,7 @@ Good first contribution areas:
 - TypeScript worker/client targeting the runtime contract
 - Rust runtime primitive or worker experiments
 - Go infrastructure worker/client experiments
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agentledger-runtime
-Version: 1.4.0
+Version: 1.4.2
 Summary: Runtime reliability layer for Agent Harness stacks
 Project-URL: Homepage, https://github.com/yaogdu/AgentLedger
 Project-URL: Repository, https://github.com/yaogdu/AgentLedger
@@ -21,7 +21,7 @@ Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Software Development :: Libraries :: Python Modules
 Requires-Python: >=3.11
 Provides-Extra: all
-Requires-Dist: agentledger-inspector<2,>=1.4.0; extra == 'all'
+Requires-Dist: agentledger-inspector<2,>=1.4.2; extra == 'all'
 Requires-Dist: agentledger-langfuse<2,>=1.4; extra == 'all'
 Requires-Dist: agentledger-langgraph<2,>=1.4; extra == 'all'
 Requires-Dist: agentledger-mcp<2,>=1.4; extra == 'all'
@@ -33,7 +33,7 @@ Requires-Dist: agentledger-sandbox-docker<2,>=1.4; extra == 'all'
 Provides-Extra: docker
 Requires-Dist: agentledger-sandbox-docker<2,>=1.4; extra == 'docker'
 Provides-Extra: inspector
-Requires-Dist: agentledger-inspector<2,>=1.4.0; extra == 'inspector'
+Requires-Dist: agentledger-inspector<2,>=1.4.2; extra == 'inspector'
 Provides-Extra: langfuse
 Requires-Dist: agentledger-langfuse<2,>=1.4; extra == 'langfuse'
 Provides-Extra: langgraph
@@ -68,7 +68,7 @@ AgentLedger `1.4.x` is a runtime reliability layer for Agent Harness stacks. It
 
 Most agent frameworks focus on planning, reasoning, and workflow logic. AgentLedger sits underneath or beside LangChain, LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, LlamaIndex, Semantic Kernel, or custom agents to provide runtime guarantees around state, tools, evidence, replay, and recovery.
 
-In a full harness stack, systems such as LangGraph, Temporal, Langfuse, MCP, model routers, storage backends, and sandbox providers can each own the layer they are good at. AgentLedger owns the reliability substrate between them: durable execution, tool/model governance, evidence, replay, policy/sandbox boundaries, cost/failure attribution, and adapter contracts.
+In a full harness stack, systems such as LangGraph, Temporal, Langfuse, MCP, model gateways, storage backends, and sandbox providers can each own the layer they are good at. AgentLedger owns the reliability substrate between them: durable execution, tool/model governance, evidence, replay, policy/sandbox boundaries, cost/failure attribution, and adapter contracts.
 
 Python remains the reference implementation, and Go, TypeScript, and Rust now have native runtime-core baselines aligned to the same language-neutral runtime contract. Provider-specific drivers and framework-native adapters intentionally vary by ecosystem. See `docs/LANGUAGE_IMPLEMENTATION_COMPARISON.md` for the exact four-language comparison and adapter boundary.
 
@@ -76,7 +76,9 @@ Python remains the reference implementation, and Go, TypeScript, and Rust now ha
 
 | Need | Go to |
 | --- | --- |
+| See the core value in 3 minutes | [examples/showcase/duplicate_side_effect_crash](examples/showcase/duplicate_side_effect_crash/README.md) |
 | Install and run the first example | [docs/GETTING_STARTED.md](docs/GETTING_STARTED.md) |
+| Decide whether AgentLedger fits your use case | [docs/USE_CASES.md](docs/USE_CASES.md) |
 | Choose Python / Go / TypeScript / Rust | [docs/LANGUAGE_QUICKSTART.md](docs/LANGUAGE_QUICKSTART.md) |
 | Find runnable examples | [examples/README.md](examples/README.md), [go/examples/README.md](go/examples/README.md), [typescript/examples/README.md](typescript/examples/README.md), [rust/examples/README.md](rust/examples/README.md) |
 | Query runtime tables | [docs/QUERY_EXAMPLES.md](docs/QUERY_EXAMPLES.md) |
@@ -108,6 +110,12 @@ Inspector exports self-contained, read-only HTML for local or internal debugging
 | What is not in core? | Planning engines, full eval systems, RAG/vector memory, trace stores, application administration backends, and sandbox infrastructure providers. |
 | How should other languages work? | This repo is contract-first. Python is the reference runtime; Go, Node/TypeScript, and Rust now have native runtime baselines under `go/`, `typescript/`, and `rust/`. Runtime-ready requires `contracts/agentledger.runtime.v1.json`, the shared semantic manifest `contracts/conformance/runtime_semantics.v1.json`, shared conformance fixtures, and per-language conformance commands. |
 
+## When to use it
+
+Use AgentLedger when an agent starts doing work you may need to recover, prove, or replay later: creating tickets, sending emails, changing files, calling internal APIs, querying sensitive systems, waiting for human approval, or running long enough that a worker crash is realistic.
+
+The fastest way to validate the value is the [3-minute side-effect safety showcase](examples/showcase/duplicate_side_effect_crash/README.md). It contrasts a naive retry that sends an external email twice with an AgentLedger-managed retry that records the side effect once, resumes safely, and exports Inspector HTML evidence. For concrete production scenarios, guarantee boundaries, and adoption guidance, see [docs/USE_CASES.md](docs/USE_CASES.md).
+
 ## Scope principle
 
 AgentLedger keeps the runtime thin but hard to replace: core only owns guarantees that cannot be reliably enforced outside the runtime boundary. Everything else should integrate through adapters, contracts, conformance tests, and examples.
@@ -153,7 +161,7 @@ For concrete stack patterns, from a minimal local harness to a Temporal + LangGr
 | Durable orchestration | Temporal, Ray, Kubernetes workers | provide agent-specific leases, fencing, checkpoints, cancellation, cost/failure attribution, and replay semantics inside the worker step |
 | Observability / eval UI | Langfuse, LangSmith, OpenTelemetry, custom dashboards | export structured runtime events, evidence bundles, trace/cost/failure data, and correlation IDs |
 | Tool and context protocols | MCP, internal tool servers, provider SDK tools | enforce schema, permissions, approval, sandbox, idempotency, and audit before side effects happen |
-| Model providers / routers | OpenAI, Anthropic, Gemini, Bedrock, Ollama, LiteLLM, enterprise gateways | provide the runtime model-call contract, archived responses, budget/fallback/replay semantics, and optional provider adapters |
+| Model providers / gateways | OpenAI, Anthropic, Gemini, Bedrock, Ollama, LiteLLM/new-api/one-api, enterprise gateways | execute or route model calls externally while AgentLedger records model evidence, archived responses, proposed tool calls, cost/failure attribution, and replay semantics |
 | Storage / artifacts | SQLite, Postgres, MySQL, S3/MinIO, internal stores | keep runtime metadata, state versions, migrations, blob refs, and evidence refs durable and conformance-tested |
 
 The intended production shape is therefore not `AgentLedger instead of LangGraph/Temporal/Langfuse`. It is `AgentLedger with LangGraph/Temporal/Langfuse` where AgentLedger governs the model/tool/state boundary that those systems otherwise cannot enforce by themselves.
@@ -185,6 +193,7 @@ The repository includes cross-language 3-minute side-effect safety demos, MCP go
 
 | Goal | Demo | Run |
 | --- | --- | --- |
+| Shareable crash/retry showcase | Python | `PYTHONPATH=src python3 examples/showcase/duplicate_side_effect_crash/demo.py` |
 | 3-minute side-effect safety | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/three_minute_demo/demo.py`; `cd go && go run ./examples/three_minute_demo`; `cd rust && cargo run --example three_minute_demo`; `cd typescript && node examples/three_minute_demo/three_minute_demo.js` |
 | MCP tool governance | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/mcp_governance/demo.py`; `cd go && go run ./examples/mcp_governance`; `cd rust && cargo run --example mcp_governance`; `cd typescript && node examples/mcp_governance/mcp_governance.js` |
 | Python | `examples/travel_assistant/demo.py` | `python3 examples/travel_assistant/demo.py` |
@@ -341,7 +350,7 @@ AgentLedger is also not a new LLM SDK, not a workflow engine, not a general obse
 
 ## Current maturity
 
-AgentLedger 1.4.x is a stable runtime-core line with Python as the reference implementation and Go, TypeScript, and Rust covered by shared runtime-core parity gates. The four language runtimes share the same runtime-core failure semantics: failed steps, retry scheduling, cancellation, lease recovery, Tool Ledger unknown-state handling, cost/failure attribution, evidence, replay, normalized failure envelopes, failure lifecycle, causal graph, replay plan, regression report, alerts, and failure export mappings. The current runtime-core release is 1.4.0 across Python, Go, TypeScript, and Rust; Inspector remains a language-neutral companion viewer distributed through Python/PyPI. It is suitable for local use, framework adapter integration, reliability semantics validation, and production pilot preparation with explicit adapter boundaries.
+AgentLedger 1.4.x is a stable runtime-core line with Python as the reference implementation and Go, TypeScript, and Rust covered by shared runtime-core parity gates. The four language runtimes share the same runtime-core failure semantics and model evidence boundary: failed steps, retry scheduling, cancellation, lease recovery, Tool Ledger unknown-state handling, cost/failure attribution, evidence, replay, normalized failure envelopes, failure lifecycle, causal graph, replay plan, regression report, alerts, failure export mappings, archived model request/response/failure evidence, and model-proposed tool-call records. The current runtime-core release is 1.4.2 across Python, Go, TypeScript, and Rust; Inspector remains a language-neutral companion viewer distributed through Python/PyPI. It is suitable for local use, framework adapter integration, reliability semantics validation, and production pilot preparation with explicit adapter boundaries.
 
 The runtime-core contract is stable; optional production adapters and external infrastructure hardening remain separately tracked. See [docs/MATURITY_MODEL.md](docs/MATURITY_MODEL.md), [docs/IMPLEMENTATION_STATUS.md](docs/IMPLEMENTATION_STATUS.md), and [docs/ROADMAP.md](docs/ROADMAP.md).
 
@@ -396,3 +405,7 @@ See [docs/RELEASE_CHECKLIST.md](docs/RELEASE_CHECKLIST.md) for the complete rele
 ## License
 
 Apache-2.0. See [LICENSE](LICENSE).
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/README.md

@@ -16,7 +16,7 @@ AgentLedger `1.4.x` is a runtime reliability layer for Agent Harness stacks. It
 
 Most agent frameworks focus on planning, reasoning, and workflow logic. AgentLedger sits underneath or beside LangChain, LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, LlamaIndex, Semantic Kernel, or custom agents to provide runtime guarantees around state, tools, evidence, replay, and recovery.
 
-In a full harness stack, systems such as LangGraph, Temporal, Langfuse, MCP, model routers, storage backends, and sandbox providers can each own the layer they are good at. AgentLedger owns the reliability substrate between them: durable execution, tool/model governance, evidence, replay, policy/sandbox boundaries, cost/failure attribution, and adapter contracts.
+In a full harness stack, systems such as LangGraph, Temporal, Langfuse, MCP, model gateways, storage backends, and sandbox providers can each own the layer they are good at. AgentLedger owns the reliability substrate between them: durable execution, tool/model governance, evidence, replay, policy/sandbox boundaries, cost/failure attribution, and adapter contracts.
 
 Python remains the reference implementation, and Go, TypeScript, and Rust now have native runtime-core baselines aligned to the same language-neutral runtime contract. Provider-specific drivers and framework-native adapters intentionally vary by ecosystem. See `docs/LANGUAGE_IMPLEMENTATION_COMPARISON.md` for the exact four-language comparison and adapter boundary.
 
@@ -24,7 +24,9 @@ Python remains the reference implementation, and Go, TypeScript, and Rust now ha
 
 | Need | Go to |
 | --- | --- |
+| See the core value in 3 minutes | [examples/showcase/duplicate_side_effect_crash](examples/showcase/duplicate_side_effect_crash/README.md) |
 | Install and run the first example | [docs/GETTING_STARTED.md](docs/GETTING_STARTED.md) |
+| Decide whether AgentLedger fits your use case | [docs/USE_CASES.md](docs/USE_CASES.md) |
 | Choose Python / Go / TypeScript / Rust | [docs/LANGUAGE_QUICKSTART.md](docs/LANGUAGE_QUICKSTART.md) |
 | Find runnable examples | [examples/README.md](examples/README.md), [go/examples/README.md](go/examples/README.md), [typescript/examples/README.md](typescript/examples/README.md), [rust/examples/README.md](rust/examples/README.md) |
 | Query runtime tables | [docs/QUERY_EXAMPLES.md](docs/QUERY_EXAMPLES.md) |
@@ -56,6 +58,12 @@ Inspector exports self-contained, read-only HTML for local or internal debugging
 | What is not in core? | Planning engines, full eval systems, RAG/vector memory, trace stores, application administration backends, and sandbox infrastructure providers. |
 | How should other languages work? | This repo is contract-first. Python is the reference runtime; Go, Node/TypeScript, and Rust now have native runtime baselines under `go/`, `typescript/`, and `rust/`. Runtime-ready requires `contracts/agentledger.runtime.v1.json`, the shared semantic manifest `contracts/conformance/runtime_semantics.v1.json`, shared conformance fixtures, and per-language conformance commands. |
 
+## When to use it
+
+Use AgentLedger when an agent starts doing work you may need to recover, prove, or replay later: creating tickets, sending emails, changing files, calling internal APIs, querying sensitive systems, waiting for human approval, or running long enough that a worker crash is realistic.
+
+The fastest way to validate the value is the [3-minute side-effect safety showcase](examples/showcase/duplicate_side_effect_crash/README.md). It contrasts a naive retry that sends an external email twice with an AgentLedger-managed retry that records the side effect once, resumes safely, and exports Inspector HTML evidence. For concrete production scenarios, guarantee boundaries, and adoption guidance, see [docs/USE_CASES.md](docs/USE_CASES.md).
+
 ## Scope principle
 
 AgentLedger keeps the runtime thin but hard to replace: core only owns guarantees that cannot be reliably enforced outside the runtime boundary. Everything else should integrate through adapters, contracts, conformance tests, and examples.
@@ -101,7 +109,7 @@ For concrete stack patterns, from a minimal local harness to a Temporal + LangGr
 | Durable orchestration | Temporal, Ray, Kubernetes workers | provide agent-specific leases, fencing, checkpoints, cancellation, cost/failure attribution, and replay semantics inside the worker step |
 | Observability / eval UI | Langfuse, LangSmith, OpenTelemetry, custom dashboards | export structured runtime events, evidence bundles, trace/cost/failure data, and correlation IDs |
 | Tool and context protocols | MCP, internal tool servers, provider SDK tools | enforce schema, permissions, approval, sandbox, idempotency, and audit before side effects happen |
-| Model providers / routers | OpenAI, Anthropic, Gemini, Bedrock, Ollama, LiteLLM, enterprise gateways | provide the runtime model-call contract, archived responses, budget/fallback/replay semantics, and optional provider adapters |
+| Model providers / gateways | OpenAI, Anthropic, Gemini, Bedrock, Ollama, LiteLLM/new-api/one-api, enterprise gateways | execute or route model calls externally while AgentLedger records model evidence, archived responses, proposed tool calls, cost/failure attribution, and replay semantics |
 | Storage / artifacts | SQLite, Postgres, MySQL, S3/MinIO, internal stores | keep runtime metadata, state versions, migrations, blob refs, and evidence refs durable and conformance-tested |
 
 The intended production shape is therefore not `AgentLedger instead of LangGraph/Temporal/Langfuse`. It is `AgentLedger with LangGraph/Temporal/Langfuse` where AgentLedger governs the model/tool/state boundary that those systems otherwise cannot enforce by themselves.
@@ -133,6 +141,7 @@ The repository includes cross-language 3-minute side-effect safety demos, MCP go
 
 | Goal | Demo | Run |
 | --- | --- | --- |
+| Shareable crash/retry showcase | Python | `PYTHONPATH=src python3 examples/showcase/duplicate_side_effect_crash/demo.py` |
 | 3-minute side-effect safety | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/three_minute_demo/demo.py`; `cd go && go run ./examples/three_minute_demo`; `cd rust && cargo run --example three_minute_demo`; `cd typescript && node examples/three_minute_demo/three_minute_demo.js` |
 | MCP tool governance | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/mcp_governance/demo.py`; `cd go && go run ./examples/mcp_governance`; `cd rust && cargo run --example mcp_governance`; `cd typescript && node examples/mcp_governance/mcp_governance.js` |
 | Python | `examples/travel_assistant/demo.py` | `python3 examples/travel_assistant/demo.py` |
@@ -289,7 +298,7 @@ AgentLedger is also not a new LLM SDK, not a workflow engine, not a general obse
 
 ## Current maturity
 
-AgentLedger 1.4.x is a stable runtime-core line with Python as the reference implementation and Go, TypeScript, and Rust covered by shared runtime-core parity gates. The four language runtimes share the same runtime-core failure semantics: failed steps, retry scheduling, cancellation, lease recovery, Tool Ledger unknown-state handling, cost/failure attribution, evidence, replay, normalized failure envelopes, failure lifecycle, causal graph, replay plan, regression report, alerts, and failure export mappings. The current runtime-core release is 1.4.0 across Python, Go, TypeScript, and Rust; Inspector remains a language-neutral companion viewer distributed through Python/PyPI. It is suitable for local use, framework adapter integration, reliability semantics validation, and production pilot preparation with explicit adapter boundaries.
+AgentLedger 1.4.x is a stable runtime-core line with Python as the reference implementation and Go, TypeScript, and Rust covered by shared runtime-core parity gates. The four language runtimes share the same runtime-core failure semantics and model evidence boundary: failed steps, retry scheduling, cancellation, lease recovery, Tool Ledger unknown-state handling, cost/failure attribution, evidence, replay, normalized failure envelopes, failure lifecycle, causal graph, replay plan, regression report, alerts, failure export mappings, archived model request/response/failure evidence, and model-proposed tool-call records. The current runtime-core release is 1.4.2 across Python, Go, TypeScript, and Rust; Inspector remains a language-neutral companion viewer distributed through Python/PyPI. It is suitable for local use, framework adapter integration, reliability semantics validation, and production pilot preparation with explicit adapter boundaries.
 
 The runtime-core contract is stable; optional production adapters and external infrastructure hardening remain separately tracked. See [docs/MATURITY_MODEL.md](docs/MATURITY_MODEL.md), [docs/IMPLEMENTATION_STATUS.md](docs/IMPLEMENTATION_STATUS.md), and [docs/ROADMAP.md](docs/ROADMAP.md).
 
@@ -344,3 +353,7 @@ See [docs/RELEASE_CHECKLIST.md](docs/RELEASE_CHECKLIST.md) for the complete rele
 ## License
 
 Apache-2.0. See [LICENSE](LICENSE).
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/README.zh-CN.md

@@ -16,7 +16,7 @@ AgentLedger `1.4.x` 是面向 Agent Harness stack 的 runtime reliability layer
 
 大多数 Agent 框架关注 planning、reasoning 和 workflow logic。AgentLedger 放在 LangChain、LangGraph、CrewAI、AutoGen、OpenAI Agents SDK、LlamaIndex、Semantic Kernel 或自定义 Agent 的下方或旁边，提供 state、tool、evidence、replay、recovery 相关的 runtime guarantees。
 
-在完整 Harness stack 里，LangGraph、Temporal、Langfuse、MCP、model router、storage backend、sandbox provider 等系统可以各自负责擅长的层。AgentLedger 负责它们之间的 reliability substrate：durable execution、tool/model governance、evidence、replay、policy/sandbox boundary、cost/failure attribution 和 adapter contracts。
+在完整 Harness stack 里，LangGraph、Temporal、Langfuse、MCP、model gateway、storage backend、sandbox provider 等系统可以各自负责擅长的层。AgentLedger 负责它们之间的 reliability substrate：durable execution、tool/model governance、evidence、replay、policy/sandbox boundary、cost/failure attribution 和 adapter contracts。
 
 Python 仍然是 reference implementation；Go、TypeScript、Rust 已有 native runtime-core baseline，并对齐同一份 language-neutral runtime contract。provider-specific driver 和 framework-native adapter 会按各语言生态差异处理。四语言实现对比和 adapter 边界见 `docs/zh/LANGUAGE_IMPLEMENTATION_COMPARISON.md`。
 
@@ -24,7 +24,9 @@ Python 仍然是 reference implementation；Go、TypeScript、Rust 已有 native
 
 | 需求 | 入口 |
 | --- | --- |
+| 3 分钟看到核心价值 | [examples/showcase/duplicate_side_effect_crash](examples/showcase/duplicate_side_effect_crash/README.md) |
 | 安装并跑通第一个示例 | [docs/zh/GETTING_STARTED.md](docs/zh/GETTING_STARTED.md) |
+| 判断 AgentLedger 是否适合你的场景 | [docs/zh/USE_CASES.md](docs/zh/USE_CASES.md) |
 | 选择 Python / Go / TypeScript / Rust | [docs/zh/LANGUAGE_QUICKSTART.md](docs/zh/LANGUAGE_QUICKSTART.md) |
 | 找可运行示例 | [examples/README.md](examples/README.md)、[go/examples/README.md](go/examples/README.md)、[typescript/examples/README.md](typescript/examples/README.md)、[rust/examples/README.md](rust/examples/README.md) |
 | 查询 runtime 表 | [docs/zh/QUERY_EXAMPLES.md](docs/zh/QUERY_EXAMPLES.md) |
@@ -56,6 +58,12 @@ Inspector 会导出自包含、只读的 HTML，适合本地或内网排查问
 | 哪些不属于 core？ | Planning engine、完整 eval 系统、RAG/vector memory、trace store、托管应用产品和托管 sandbox infrastructure。 |
 | 其它语言怎么做？ | 这个 repo 是 contract-first。Python 是 reference runtime；Go、Node/TypeScript、Rust 已在 `go/`、`typescript/`、`rust/` 下有 native runtime baseline。runtime-ready 必须对齐 `contracts/agentledger.runtime.v1.json` 和共享语义清单 `contracts/conformance/runtime_semantics.v1.json`，通过共享 conformance fixtures，并提供各语言 conformance command。 |
 
+## 什么时候该用
+
+当 Agent 开始执行以后可能需要恢复、证明或 replay 的工作时，就适合引入 AgentLedger：创建 ticket、发送邮件、修改文件、调用内部 API、查询敏感系统、等待人工审批，或者任务长到 worker crash 是现实风险。
+
+最快的价值验证方式是跑 [3 分钟副作用安全 showcase](examples/showcase/duplicate_side_effect_crash/README.md)。它会对比一个 naive retry 把外部邮件发送两次，以及 AgentLedger-managed retry 只记录并复用一次副作用、可安全恢复并导出 Inspector HTML evidence。更具体的生产场景、保证边界和接入判断见 [docs/zh/USE_CASES.md](docs/zh/USE_CASES.md)。
+
 ## 范围原则
 
 AgentLedger 的 runtime 要保持“薄但不可替代”：core 只内建那些不在 runtime boundary 内就无法可靠保证的能力。其它能力通过 adapter、contract、conformance test 和 example 接入成熟生态。
@@ -101,7 +109,7 @@ AgentLedger 本身不是完整 Agent Harness。它的设计目标是和 Harness
 | Durable orchestration | Temporal、Ray、Kubernetes workers | 在 worker step 内提供 agent-specific lease、fencing、checkpoint、cancellation、cost/failure attribution 和 replay semantics |
 | Observability / eval UI | Langfuse、LangSmith、OpenTelemetry、自定义 dashboard | 导出 structured runtime events、evidence bundle、trace/cost/failure data 和 correlation IDs |
 | Tool and context protocols | MCP、internal tool servers、provider SDK tools | 在副作用发生前强制执行 schema、permission、approval、sandbox、idempotency 和 audit |
-| Model providers / routers | OpenAI、Anthropic、Gemini、Bedrock、Ollama、LiteLLM、企业 gateway | 提供 runtime model-call contract、archived responses、budget/fallback/replay semantics 和 optional provider adapters |
+| Model providers / gateways | OpenAI、Anthropic、Gemini、Bedrock、Ollama、LiteLLM/new-api/one-api、企业 gateway | 在外部执行或路由 model call；AgentLedger 记录 model evidence、archived response、proposed tool call、cost/failure attribution 和 replay semantics |
 | Storage / artifacts | SQLite、Postgres、MySQL、S3/MinIO、内部存储 | 持久化 runtime metadata、state version、migration、blob ref 和 evidence ref，并通过 conformance 验证 |
 
 所以推荐的生产形态不是 `AgentLedger instead of LangGraph/Temporal/Langfuse`，而是 `AgentLedger with LangGraph/Temporal/Langfuse`：AgentLedger 负责治理这些系统自身难以强制保证的 model/tool/state boundary。
@@ -133,6 +141,7 @@ AgentLedger 本身不是完整 Agent Harness。它的设计目标是和 Harness
 
 | 目标 | Demo | 运行方式 |
 | --- | --- | --- |
+| 传播型 crash/retry showcase | Python | `PYTHONPATH=src python3 examples/showcase/duplicate_side_effect_crash/demo.py` |
 | 3 分钟副作用安全 demo | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/three_minute_demo/demo.py`; `cd go && go run ./examples/three_minute_demo`; `cd rust && cargo run --example three_minute_demo`; `cd typescript && node examples/three_minute_demo/three_minute_demo.js` |
 | MCP tool governance | Python / Go / Rust / TypeScript | `PYTHONPATH=src python3 examples/mcp_governance/demo.py`; `cd go && go run ./examples/mcp_governance`; `cd rust && cargo run --example mcp_governance`; `cd typescript && node examples/mcp_governance/mcp_governance.js` |
 | Python | `examples/travel_assistant/demo.py` | `python3 examples/travel_assistant/demo.py` |
@@ -289,7 +298,7 @@ AgentLedger 也不是新的 LLM SDK，不是 workflow engine，不是通用 obse
 
 ## 当前成熟度
 
-AgentLedger 1.4.x 是 stable runtime-core line，Python 是 reference implementation，Go、TypeScript、Rust 已由共享 runtime-core parity gate 覆盖。四种语言 runtime 共享同一组 runtime-core failure semantics：failed step、retry scheduling、cancellation、lease recovery、Tool Ledger unknown-state handling、cost/failure attribution、evidence、replay、normalized failure envelope、failure lifecycle、causal graph、replay plan、regression report、alerts 和 failure export mappings。当前 runtime-core release 是 1.4.0，并覆盖 Python、Go、TypeScript、Rust；Inspector 仍是 language-neutral companion viewer，通过 Python/PyPI 分发。适合本地使用、framework adapter integration、reliability semantics 验证，以及在明确 adapter 边界下做 production pilot 准备。
+AgentLedger 1.4.x 是 stable runtime-core line，Python 是 reference implementation，Go、TypeScript、Rust 已由共享 runtime-core parity gate 覆盖。四种语言 runtime 共享同一组 runtime-core failure semantics 和 model evidence boundary：failed step、retry scheduling、cancellation、lease recovery、Tool Ledger unknown-state handling、cost/failure attribution、evidence、replay、normalized failure envelope、failure lifecycle、causal graph、replay plan、regression report、alerts、failure export mappings、archived model request/response/failure evidence，以及 model-proposed tool-call records。当前 runtime-core release 是 1.4.2，并覆盖 Python、Go、TypeScript、Rust；Inspector 仍是 language-neutral companion viewer，通过 Python/PyPI 分发。适合本地使用、framework adapter integration、reliability semantics 验证，以及在明确 adapter 边界下做 production pilot 准备。
 
 runtime-core contract 已稳定；optional production adapter 和外部基础设施加固仍按独立阶段推进。详见 [docs/MATURITY_MODEL.md](docs/MATURITY_MODEL.md)、[docs/zh/IMPLEMENTATION_STATUS.md](docs/zh/IMPLEMENTATION_STATUS.md) 和 [docs/zh/ROADMAP.md](docs/zh/ROADMAP.md)。
 
@@ -344,3 +353,7 @@ cd ../rust && cargo run --quiet -- conformance
 ## License
 
 Apache-2.0. See [LICENSE](LICENSE)。
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/SECURITY.md

@@ -54,3 +54,7 @@ threat modeling
 ```
 
 See `docs/SECURITY_ENTERPRISE.md` for the enterprise security model.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/contracts/agentledger.runtime.v1.json

@@ -12,6 +12,7 @@
       "video_segment"
     ],
     "media_schema_version": "agentledger.media.v0",
+    "model_evidence_schema_version": "agentledger.model.evidence.v1",
     "stream_checkpoint_fields": [
       "stream_id",
       "consumer_id",
@@ -111,7 +112,10 @@
     "agent_result_returned",
     "tool_approval_decided",
     "step_waiting_human",
+    "model_call_requested",
     "model_call_completed",
+    "model_call_failed",
+    "tool_call_proposed",
     "cost_recorded",
     "budget_check_failed",
     "failure_classified"
@@ -125,7 +129,8 @@
     "PENDING_VERIFICATION side effects are not auto-retried",
     "replay and shadow mode must not create external side effects",
     "sandbox-required tools fail closed when the executor is missing or disabled",
-    "media and stream artifacts store refs, metadata, lineage, and checkpoints instead of raw codec processing"
+    "media and stream artifacts store refs, metadata, lineage, and checkpoints instead of raw codec processing",
+    "model evidence records archived provider inputs, outputs, failures, and proposed tool calls without routing providers"
   ],
   "language_targets": [
     {

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/contracts/conformance/cost_failure_attribution.v1.json

@@ -45,6 +45,24 @@
         "failure_has_retryability",
         "failure_links_to_step_and_optional_tool_or_model"
       ]
+    },
+    {
+      "name": "runtime_model_evidence_boundary",
+      "description": "Externally executed model calls should be recorded as replayable evidence without requiring runtime-core to route provider traffic.",
+      "required_event_types": [
+        "model_call_requested",
+        "model_call_completed",
+        "model_call_failed",
+        "tool_call_proposed",
+        "cost_recorded"
+      ],
+      "required_assertions": [
+        "model_evidence_schema_version_is_agentledger_model_evidence_v1",
+        "model_call_failure_category_is_model",
+        "tool_call_proposal_is_distinct_from_tool_execution",
+        "model_cost_records_include_tokens_and_usd",
+        "runtime_core_does_not_require_provider_sdk_or_router"
+      ]
     }
   ],
   "preview_status": "implemented by Go, TypeScript, and Rust preview runtimes as runtime-core parity gates; production adapter hardening remains separate"

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/ADAPTER_CERTIFICATION.md

@@ -234,3 +234,7 @@ temporal
 The generated bundle includes package metadata, contract version, conformance commands, smoke commands, required external services, security assumptions, known limitations, and a `production_validation` block.
 
 For adapters that depend on real infrastructure, `production_validation.status` is intentionally `external-required`. That is not a failure. It means the local runtime has produced the certification contract, but a production claim still needs real credentials, service behavior, concurrency/load checks, and restore or rollback drills. P2-style production hardening must be backed by those external artifacts, not by local mocks alone.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/ADAPTER_PACKAGING.md

@@ -271,3 +271,7 @@ The adapter package boundary should not claim production hardening just because
 - media processing adapters
 - sub-agent or multi-agent runtime semantics
 - long-running UI, deployment management, or full eval system
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/ADAPTER_ROADMAP.md

@@ -56,7 +56,7 @@ These are valuable but should follow the Priority 1 adapters or remain thinner f
 | Sandbox | E2B | Good managed remote sandbox for code/tool execution. | Keep as optional remote executor adapter. |
 | Distributed execution | Ray bridge | Useful for Python distributed worker pools. | Ray should own cluster scheduling; AgentLedger owns run semantics. |
 | Deployment | Kubernetes worker recipe | Useful for pilots. | Recipe/Helm/examples first; keep deployment management outside runtime-core. |
-| Model providers/router | OpenAI, Anthropic, Gemini, Bedrock, Azure OpenAI, Ollama, LiteLLM-style bridge | Model calls affect evidence, cost, replay, fallback, and budget enforcement, but provider SDKs should not be runtime-core dependencies. | Wait for `ModelGateway`/`ModelRouter` contract before stabilizing provider packages. |
+| Model evidence endpoints | OpenAI-compatible endpoints, Anthropic-style calls, enterprise gateways | Model calls affect evidence, cost, replay, and failure attribution, but provider SDKs and routers should remain outside runtime-core. | Use the Runtime Model Evidence Boundary; do not build official router/gateway adapters unless the boundary stays evidence-only. |
 
 ## Priority 3: Experimental Or Community Adapters
 
@@ -68,7 +68,7 @@ These should not block core parity or official release claims.
 | Sandbox | Firecracker | Powerful but infra-heavy; often managed through a platform layer. |
 | Sandbox | bubblewrap | Useful on Linux local machines but less universal across macOS/Windows teams. |
 | Workflow | Airflow / Prefect / Argo | They are batch/workflow systems; useful bridges, but AgentLedger should not become a general workflow engine. |
-| Eval | LangSmith / Braintrust / custom eval platforms | Eval remains a consumer of evidence/replay contracts, not runtime-core. |
+| Eval | Langfuse, Phoenix, promptfoo, DeepEval, Ragas, OpenAI Evals, LangSmith, Braintrust, custom CI | Eval remains a consumer of evidence/replay contracts, not runtime-core. AgentLedger should provide adapters/export formats, not a standalone Eval Platform. |
 | Vector DB / RAG | Pinecone, Weaviate, Milvus, pgvector, etc. | Long-term memory/retrieval infra should be external; runtime stores refs/evidence, not knowledge retrieval logic. |
 | Routing advisor / capability router | WisePick-style route decision services | Candidate direction only, not a committed adapter. It may be evaluated as a pre-execution suggestion and post-execution feedback loop, but should remain optional and must not turn AgentLedger core into a capability router or authorization layer. |
 | Application administration backend | User, organization, billing, identity, or control-plane adapters | Out of current project scope. AgentLedger is a framework/library/runtime reliability layer. |
@@ -89,7 +89,7 @@ Python remains the reference implementation, but official adapters should conver
 | Kubernetes sandbox/backend | Recommended | Recommended | Recommended | Recommended | Prefer manifest/dry-run contract plus optional execution. |
 | Temporal bridge | Recommended | Required when Go runtime matures | Recommended | Community/optional | Match Temporal ecosystem strength per language. |
 | OpenTelemetry | Required official | Required official | Required official | Required official | Standard enterprise observability path. |
-| Model provider/router | Recommended after ModelGateway contract | Recommended after ModelGateway contract | Recommended after ModelGateway contract | Recommended after ModelGateway contract | Keep provider SDKs optional; use injected clients and conformance fixtures. |
+| Model evidence boundary | Required core | Required core | Required core | Required core | Record external model calls, failures, costs, and tool proposals without bundling provider SDKs or routers. |
 
 ## Non-Negotiable Adapter Requirements
 
@@ -113,5 +113,10 @@ Every official adapter must provide:
 6. LangChain / CrewAI / AutoGen / OpenAI Agents SDK / LlamaIndex / Semantic Kernel facades where ecosystems are stable.
 7. Kubernetes sandbox/backend recipe, then optional execution adapter.
 8. Temporal/Ray/Kubernetes scheduler/backend bridges based on real user demand.
-9. Model provider/router adapters after the runtime `ModelGateway` contract is stable.
+9. Add model evidence examples for OpenAI-compatible endpoints, Anthropic-style calls, and enterprise gateways without making AgentLedger a model router.
 10. Evaluate only, without implementation commitment, whether a Routing Advisor / Capability Router evidence boundary is useful for WisePick-style services. If later adopted, it should record externally supplied route suggestions, preserve replay by reusing stored routing evidence, and send feedback only after new real executions.
+11. Add eval adapter/export packages for mature open-source eval tools only when the boundary is evidence/replay conversion: Langfuse datasets/scores, Phoenix experiments, promptfoo test cases, DeepEval cases, Ragas rows, OpenAI Evals samples, and CI report inputs. Do not build dataset/scorer/experiment management inside AgentLedger.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/ADOPTION.md

@@ -16,8 +16,10 @@ Can I verify the reliability claim myself?
 
 | Priority | Item | Acceptance criteria |
 |---|---|---|
+| P0 | Shareable side-effect showcase | A user can run `examples/showcase/duplicate_side_effect_crash`, see naive duplicate side effects versus AgentLedger safe retry, and open generated Inspector HTML. |
 | P0 | 3-minute demo | A user can run the Python, Go, TypeScript, or Rust 3-minute demo and see one external side effect, one Tool Ledger entry, safe retry, and replay-safe evidence. |
 | P0 | README first-screen pain point | The first screen explains the tool side-effect problem before listing architecture. |
+| P0 | Shareable article and short post | A maintainer can use `docs/articles/what-happened-after-agent-tool-failure.zh-CN.md` and `docs/articles/social-posts.md` without inventing positioning copy from scratch. |
 | P1 | MCP governance example | Dependency-free MCP-style tools demonstrate schema, approval, sandbox metadata, idempotency, and audit evidence across Python, Go, TypeScript, and Rust. |
 | P1 | Public issue/discussion list | The next adoption tasks are easy to open as GitHub issues or discussions. |
 | P1 | Case study template | A real integration can be documented without exposing private data. |
@@ -28,6 +30,7 @@ These can be opened as GitHub issues when the project is ready for public tracki
 
 | Title | Type | Why it matters |
 |---|---|---|
+| Record a side-effect safety GIF from the showcase | adoption | Turns the core value into a short visual artifact for README, social posts, and issue discussions. |
 | Build a cross-language 3-minute Tool Ledger demo | example | Shows the core value in a few minutes across Python, Go, TypeScript, and Rust. |
 | Add OpenAI Agents SDK approval/replay example | example | Connects AgentLedger to a major agent SDK boundary without claiming official endorsement. |
 | Add MCP tool governance example | example | Shows how MCP-style tools should pass through policy, approval, sandbox, idempotency, and evidence. |
@@ -56,3 +59,7 @@ Case studies should be useful but conservative:
 - external demos or blog posts
 - adapter conformance reports
 - real-service hardening reports
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/ARCHITECTURE.md

@@ -150,3 +150,7 @@ evidence/replay golden fixtures
 ```
 
 This prevents each language from inventing different event names, lease semantics, idempotency behavior, or evidence bundle shapes.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/BACKUP_RESTORE.md

@@ -102,3 +102,7 @@ orchestrate cross-region restore
 ```
 
 Those operations belong to DBA, SRE, or deployment runbooks. AgentLedger provides conformance checks, evidence export, replay, and deterministic refs so the restore can be verified without running real tools.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/COMPARISONS.md

@@ -174,3 +174,7 @@ Can another language/runtime produce the same evidence semantics?
 ```
 
 That is the intended boundary: framework-neutral runtime safety and evidence, not a replacement for every framework around it.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/COMPLETE_CORE_PARITY_CHECKLIST.md

@@ -55,8 +55,8 @@ Excluded or not applicable:
 | CLI baseline | `--help`, `doctor`, `version`, `quickstart`, `conformance`, `contract validate`, `contract export` | `agentledger` CLI | `agentledger-go` | `agentledger-ts` | `agentledger-rust` | `scripts/check_complete_core_parity.py` | done |
 | Quickstart docs | Language quickstart and adapter quickstart | README + docs | `go/README.md` | `typescript/README.md` | `rust/README.md` | docs link check | done |
 | Examples | Runnable examples for core runtime and adapter API | Python examples | `go/examples/quickstart/main.go` | `typescript/examples/quickstart/quickstart.js` | `rust/examples/quickstart.rs` | `scripts/check_complete_core_parity.py` | done |
-| Package metadata | Installable package metadata in the current 1.4 release family | PyPI metadata `1.4.x` | `go.mod` tag `go/v1.4.0` for current baseline | npm package `1.4.0` for current baseline | `Cargo.toml` `1.4.0` for current baseline | `scripts/check_complete_core_parity.py` release-family check | done |
-| Package install smoke | Install/import/run from published or built package | PyPI `agentledger-runtime==1.4.0` release smoke after publish | `go/v1.4.0` clean external `go get` smoke after tag | npm `agentledger-runtime@1.4.0` release smoke after publish | crates.io `agentledger-runtime==1.4.0` release smoke after publish | publish logs + `scripts/check_complete_core_parity.py` dry-run gates | release-gated |
+| Package metadata | Installable package metadata in the current 1.4 release family | PyPI metadata `1.4.x` | `go.mod` tag `go/v1.4.2` for current baseline | npm package `1.4.2` for current baseline | `Cargo.toml` `1.4.2` for current baseline | `scripts/check_complete_core_parity.py` release-family check | done |
+| Package install smoke | Install/import/run from published or built package | PyPI `agentledger-runtime==1.4.2` release smoke after publish | `go/v1.4.2` clean external `go get` smoke after tag | npm `agentledger-runtime@1.4.2` release smoke after publish | crates.io `agentledger-runtime==1.4.2` release smoke after publish | publish logs + `scripts/check_complete_core_parity.py` dry-run gates | release-gated |
 | Release docs | Changelog, release checklist, status docs reflect exact parity boundary | docs | docs | docs | docs | docs link check | done |
 | Comparison docs | Honest overlap with LangChain/LangGraph/LangSmith/Langfuse/etc. | docs | docs | docs | docs | docs link check | done |
 
@@ -67,7 +67,7 @@ The strict core parity gate now covers the previously weak areas:
 1. CLI baseline is automatically checked for all four language CLIs.
 2. Go, TypeScript, and Rust have runnable quickstart example files.
 3. TypeScript and Rust package surfaces have package dry-run checks (`npm pack --dry-run`, `cargo package --allow-dirty --no-verify`) and post-publish install smokes are required for release.
-4. Package release families are checked across Python, TypeScript, and Rust. The current core release family is `1.4`; 1.4.0 is a four-language runtime-core feature release for Agent Failure Lifecycle.
+4. Package release families are checked across Python, TypeScript, and Rust. The current core release family is `1.4`; 1.4.2 keeps the four-language release train aligned while consolidating model evidence UX, export mappings, examples, and boundary lint around the 1.4.0/1.4.1 runtime-core features.
 5. Go external module consumption is tag/release dependent; each release should repeat the clean external `go get` smoke after pushing the `go/vX.Y.Z` tag.
 6. Real service-backed hardening remains out of scope for core parity and stays documented as optional follow-up work.
 
@@ -82,3 +82,7 @@ python3.11 scripts/check_complete_core_parity.py
 ```
 
 `check_complete_core_parity.py` must verify CLI baseline, runnable examples, package metadata, and docs links in addition to semantic conformance.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/DESIGN_AND_IMPLEMENTATION.md

@@ -244,3 +244,7 @@ The checked-in contract fixture must match:
 PYTHONPATH=src python3 -m agentledger contract export > /tmp/agentledger-contract.json
 diff -u contracts/agentledger.runtime.v1.json /tmp/agentledger-contract.json
 ```
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/DISTRIBUTED_WORKERS.md

@@ -157,3 +157,7 @@ cancelled run fences stale owner and blocks new claims
 ## Non-goals
 
 Runtime core does not create deployments, queues, clusters, connection pools, secrets, or cloud resources. It defines worker semantics and provides a reference loop plus conformance tests. Deployment adapters and operations teams own the actual orchestration environment.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/EXECUTION_BACKENDS.md

@@ -176,3 +176,7 @@ Temporal Workflow
 In this setup, Temporal owns workflow lifecycle, worker queues, retries, timers, and distributed execution. LangGraph owns graph nodes, edges, routing, and multi-agent orchestration. AgentLedger owns the model/tool boundary inside each node: Tool Ledger, idempotency, approval, sandbox, evidence, replay safety, and cost/failure attribution.
 
 This is the preferred positioning for complex production systems. AgentLedger should not claim to replace Temporal or LangGraph; it should make LangGraph nodes running inside Temporal agent-safe.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/EXTENSIBILITY.md

@@ -386,3 +386,7 @@ Blob adapters should implement `BlobStoreProtocol` and pass `BlobStoreConformanc
 Runtime core owns the contract and audit trail; adapter packages own backend-specific execution. A backend adapter should implement `SandboxExecutor.run_tool(...)`, respect `SandboxPolicy`, and return `SandboxResult` without leaking secrets into metadata. Kubernetes-based adapters should use `runtime_class` for gVisor rather than making gVisor a separate core dependency. The built-in Kubernetes adapter provides a dependency-free Job manifest dry-run and a `kubectl`-gated execution path; stronger integrations can replace it with an in-cluster controller or enterprise scheduler.
 
 Sandbox adapter packages can remain optional so users who do not want Docker/E2B/Kubernetes can use `none`, `local`, bubblewrap, or their own enterprise executor.
+
+---
+
+generated by codex cli

{agentledger_runtime-1.4.0 → agentledger_runtime-1.4.2}/docs/GETTING_STARTED.md

@@ -9,7 +9,7 @@ This is the shortest path from install to a working AgentLedger run. If you are
 | Language | Install / use | Quickstart | Examples | Package / command |
 |---|---|---|---|---|
 | Python | `pipx install agentledger-runtime` or `pip install agentledger-runtime` | `agentledger quickstart` | `../examples/README.md` | PyPI package `agentledger-runtime`, CLI `agentledger` |
-| Go | `go get github.com/yaogdu/AgentLedger/go@v1.4.0` inside a Go module | `cd go && go run ./examples/quickstart` from this repo | `../go/examples/README.md` | Go module `github.com/yaogdu/AgentLedger/go`, CLI package `.../go/cmd/agentledger-go` |
+| Go | `go get github.com/yaogdu/AgentLedger/go@v1.4.2` inside a Go module | `cd go && go run ./examples/quickstart` from this repo | `../go/examples/README.md` | Go module `github.com/yaogdu/AgentLedger/go`, CLI package `.../go/cmd/agentledger-go` |
 | TypeScript | `npm install agentledger-runtime` | `node typescript/examples/quickstart/quickstart.js` | `../typescript/examples/README.md` | npm package `agentledger-runtime`, CLI `agentledger-ts` |
 | Rust | crates.io package: `agentledger-runtime` | `cargo add agentledger-runtime` 或 `cd rust && cargo run --example quickstart` | `../rust/examples/README.md` | crate `agentledger-runtime`, binary `agentledger-rust` |
 
@@ -36,17 +36,17 @@ Use the library inside a Go module:
 
 ```bash
 go mod init your-module-name  # only if your project does not already have go.mod
-go get github.com/yaogdu/AgentLedger/go@v1.4.0
+go get github.com/yaogdu/AgentLedger/go@v1.4.2
 ```
 
 Install the optional CLI:
 
 ```bash
-go install github.com/yaogdu/AgentLedger/go/cmd/agentledger-go@v1.4.0
+go install github.com/yaogdu/AgentLedger/go/cmd/agentledger-go@v1.4.2
 agentledger-go --help
 ```
 
-Important: `go get` must run inside a Go module. `go install github.com/yaogdu/AgentLedger/go@v1.4.0` is not valid because that path is a library, not `package main`. Use `/cmd/agentledger-go` for the CLI.
+Important: `go get` must run inside a Go module. `go install github.com/yaogdu/AgentLedger/go@v1.4.2` is not valid because that path is a library, not `package main`. Use `/cmd/agentledger-go` for the CLI.
 
 ### TypeScript
 
@@ -88,6 +88,7 @@ The crate is published as `agentledger-runtime`; the library crate is imported a
 
 | Goal | Example |
 |---|---|
+| See the side-effect safety problem visually | `../examples/showcase/duplicate_side_effect_crash/README.md` |
 | Understand the core value in 3 minutes | `../examples/three_minute_demo/README.md`; Go `../go/examples/three_minute_demo`; TypeScript `../typescript/examples/three_minute_demo`; Rust `../rust/examples/three_minute_demo.rs` |
 | Smallest Python run | `../examples/hello_world/hello.py` |
 | Idempotent side effects | `../examples/side_effect_idempotency/README.md` |
@@ -149,3 +150,7 @@ python3.11 scripts/check_language_parity.py --json-report /tmp/agentledger-langu
 ```
 
 These checks verify runtime-core parity, CLI/DX baseline, examples, package metadata, docs links, and shared conformance semantics.
+
+---
+
+generated by codex cli