Skip to content

Releases

DCA v1.0.4

Choose a tag to compare

@ykw-9826 ykw-9826 released this 28 Sep 07:15
7b49d4a

DCA v1.0.4

中文

本版本修复 DCA 对实际 Codex exec --json 事件流的兼容性。

主要更新

  • 接受首个 turn 前结构化的非致命 diagnostic,仅限符合协议的 completed error item。
  • 其它 pre-turn item 继续 fail closed,顶层 stream error 继续保持 fatal。
  • runtime activity contract 更新为 dca-activity/3。
  • historical revalidation 生成派生记录,原始事件和历史 verdict 保持不可变。
  • 不将缺少结构化关联的 stderr 文本单独视为已确认的 tool rejection。

验证

961 项全量自动化测试、GitHub Hosted CI、tag-triggered Release Validation、package build 和 fresh-wheel 非 editable 安装均已通过。


English

This release fixes DCA compatibility with real Codex exec --json event streams.

Highlights

  • Accepts structured, non-fatal diagnostics before the first turn, limited to protocol-valid completed error items.
  • Keeps other pre-turn items fail-closed and top-level stream errors fatal.
  • Versions the runtime activity contract as dca-activity/3.
  • Historical revalidation produces derived records without changing raw events or historical verdicts.
  • Does not treat uncorrelated stderr text alone as confirmed tool rejection.

Validation

All 961 automated tests passed, along with GitHub Hosted CI, tag-triggered Release Validation, package builds and fresh-wheel non-editable installation.

DCA v1.0.3

Choose a tag to compare

@ykw-9826 ykw-9826 released this 23 Sep 12:47
d4a8e36

DCA v1.0.3

中文

本版本主要增强 Codex hook 通知系统的可靠性。

主要更新

  • 增加多通道通知 fan-out,最多三个 sink 顺序发送,单路失败不阻断其余通道。
  • 支持 Server酱³ / SC3 和 ServerChan Turbo,兼容厂商实际 SendKey 格式,并安全编码 URL 路径。
  • 增加 PermissionRequest 通知策略:默认 OFF;REQUEST_OBSERVED 仅用于诊断,不表示用户正在等待人工审批。
  • 支持旧 hook 的安全迁移,包含预览、SHA256 校验、备份和回滚记录;更新后的 hook 仍需人工 review/trust。
  • 使用按协议划分的有限传输预算:Turbo 的 socket timeout / 发送总期限为 2 / 2.5 秒,SC3 为 4 / 5 秒。生成的 hook timeout 为 17 秒,覆盖最多三路顺序发送;这些预算不是送达 SLA。
  • 不进行盲目通知 retry;不确定的发送结果保持 DELIVERY_UNKNOWN。

验证

869 项离线自动化测试、fresh-wheel 安装和三通道真实通知验收均已通过。服务端 ACK 与用户确认的手机接收分别核验,不相互推断。v1.0.3 发布时尚未配置 GitHub Actions CI。

现有安装和 hooks 不会自动更新。迁移需显式授权,修改后的 hook definition 需人工 review/trust;已有 delivery history 应继续保留。


English

This release improves notification reliability for Codex hooks.

Highlights

  • Multi-channel fan-out supports up to three sequential sinks, with failures isolated per sink.
  • Server酱³ / SC3 and ServerChan Turbo are supported, including vendor SendKey formats and safe URL path encoding.
  • PermissionRequest notifications default to OFF. REQUEST_OBSERVED is diagnostic and does not establish that a user is waiting for human approval.
  • Legacy hook migration includes a preview, SHA256 checks, backups and rollback receipts. Changed hooks still require manual review and trust.
  • Transport budgets are finite and protocol-specific: Turbo uses a 2-second socket timeout and a 2.5-second send deadline; SC3 uses 4 and 5 seconds. Generated hooks have a 17-second timeout for up to three sequential sends. These budgets are not a delivery SLA.
  • There are no blind notification retries. Ambiguous outcomes remain DELIVERY_UNKNOWN.

Validation

All 869 offline automated tests passed, along with fresh-wheel installation and live three-channel notification acceptance. Service acknowledgements and user-confirmed phone receipt were checked separately. GitHub Actions CI was not configured when v1.0.3 was published.

Existing installations and hooks do not update automatically. Migration requires explicit authorization; review and trust changed hook definitions manually, and preserve existing delivery history.

DCA v1.0.2

Choose a tag to compare

@ykw-9826 ykw-9826 released this 22 Sep 01:59

DCA v1.0.2 — Dispatcher for Codex Agents

Release candidate notes. Publication requires separate approval.

English

This update makes result interpretation explicit for noninteractive external
agents. Existing tasks need no new field: AgentTask.runtime_contract defaults
to structured_output="strict_json" and rejected_user_input="fail". It does
not grant file access, tool permissions or approval authority.

  • json_or_single_fence optionally accepts one exact outer JSON fence after raw
    JSON parsing fails. It removes only the specified wrapper bytes: no trimming,
    JSON repair, field insertion or reserialization. Raw and normalized bytes,
    SHA256 values, removed ranges and validation-input provenance are retained.
  • Activity classifier dca-activity/2 distinguishes observable execution,
    pre-execution rejection, no tool activity, and unknown/incomplete evidence.
    It checks thread, turn, item and terminal ordering. Invalid lifecycle evidence
    cannot acquire a rejection exemption or become success through valid JSON.
  • warn_if_runtime_rejected requires call-correlated, lifecycle-consistent
    evidence that request_user_input was rejected before execution. Required
    input or unfinished work still blocks acceptance; DCA never answers on behalf
    of the user. The supported CLI mappings do not establish such a real rejection
    event; the positive offline demonstration uses an explicitly synthetic protocol.
  • dca revalidate creates an independent, hash-pinned interpretation of approved
    historical artifacts. It neither runs old commands nor changes source attempts,
    and adds no model calls or token usage.
  • batch collect --selection explicitly selects original or qualified derived
    results. Exact-once coverage and plan-owned roles remain mandatory; original
    plus derived double counting, automatic promotion and ledger rewrites are refused.

This is not a general Markdown or damaged-JSON repair feature. Stderr text, model
claims and nonzero exits do not by themselves prove tool rejection. Incomplete,
unknown or lifecycle-invalid evidence cannot authorize the opt-in acceptance
paths. Not every historical shard is recoverable: both supplied real-history
cases remained fail-closed and were not collected, even though their structured
payloads passed schema and application coverage checks.

Validation used offline tests, test executables and a clean wheel installation;
no new model/provider or notification requests were made. File/tool authority,
runtime protection and the notification boundary remain unchanged. See the
runtime contract in the source ZIP for exact rules and evidence limits.

简体中文

本次更新让非交互外部 Agent 的结果解释方式由任务明确选择。旧任务无需新增字段:
AgentTask.runtime_contract 默认仍为 structured_output="strict_json" 和
rejected_user_input="fail",不授予文件、工具或审批权限。

  • 可选的 json_or_single_fence 只在原始 JSON 解析失败后,按精确规则移除一层
    外围 JSON fence。不 trim、不修复 JSON、不补字段,也不重新序列化。原始与
    归一化字节、SHA256、移除区间和实际校验输入的 provenance 分别保存。
  • dca-activity/2 区分可观察的工具执行、执行前拒绝、无工具活动和未知/不完整
    证据,并检查 thread、turn、item 与 terminal 的顺序。生命周期错误不能获得
    拒绝豁免,也不能因为 JSON 合法就变成成功。
  • warn_if_runtime_rejected 要求可按调用关联、生命周期一致的执行前拒绝证据。
    仍需用户输入或工作未完成时不能接受,DCA 不替用户回答。当前支持的 CLI 映射
    尚不能确认这种真实拒绝事件;离线正向 Demo 使用明确标记的合成协议。
  • dca revalidate 对获准读取的历史 artifacts 生成独立、hash-pinned 的重新解释
    记录。不执行历史命令、不改原始 attempt,也不增加模型调用或 token 用量。
  • batch collect --selection 显式选择原始或合格的 derived result,仍要求
    exact-once 覆盖和计划确定的角色,不重复计数、不自动提升角色、不改旧 ledger。

这不是通用 Markdown 或损坏 JSON 修复器。stderr 文本、模型自述和非零退出码,
都不能单独证明工具在执行前遭到拒绝。证据未知、不完整或生命周期错序时,不能
通过兼容选项放行。也不是所有旧 shard 都能恢复:本次提供的两个真实历史 case,
虽然结构化内容通过了 schema 和业务 ID 覆盖检查,最终仍失败关闭,均未收集。

本版使用离线测试、测试 executable 和全新 wheel 安装验证,没有新增模型/provider
或通知请求。文件与工具授权、runtime 保护及通知边界保持不变。精确规则和证据限制
见源码 ZIP 内的运行合同。

DCA v1.0.1

Choose a tag to compare

@ykw-9826 ykw-9826 released this 21 Sep 03:04

DCA v1.0.1 — Dispatcher for Codex Agents

Release candidate draft; no tag, GitHub Release or package upload is implied.

English

This update adds optional task-scoped file and tool permissions to external agent
execution. Without a policy, the existing read-only, no-tool behavior remains.

  • Explicit read/write roots and tool allowlists fail closed. Supported grants are
    shell or unified exec, web search, and exact configured HTTPS MCP tools.
  • Standalone Codex profile sidecars supply model/provider selection. Reasoning
    effort and verbosity are optional; missing values are not inferred. Invalid
    configuration and I/O failures preserve complete immutable failure artifacts.
  • Restricted Linux self-exec compatibility adds only the selected canonical
    Codex executable as an internal read dependency (related upstream issue:
    openai/codex#29049). It never grants an entire release or Codex home.
  • A separate runtime deny boundary rejects user read/write grants overlapping
    the selected standalone release, including installations outside Codex home.
    Unknown layouts fail closed for explicit capability tasks before launch.
    Requested permissions, effective permissions and protection metadata remain
    separately auditable. See capability policy in the source ZIP.

Previous authorized exact-read, default-restricted and exact-write live checks
passed on Ubuntu 24.04 / Codex 0.155.1 through production codex exec. The later
deny-boundary fix is checked offline; it does not broaden normal permission
compilation. No live calls were repeated for this candidate. The same Codex
version's codex sandbox linux diagnostic dispatch was broken; that diagnostic
failure is separate from the verified production path.

Only the canonical standalone/releases/<release-id>/bin/codex layout is
recognized for explicit capability tasks. Launcher symlinks and ambiguous paths
are refused. Other installations/platforms are not certified. Identity/hash
rechecks are not atomic protection against privileged concurrent replacement.
Event audits are not OS enforcement. Missing served-model metadata remains
NOT_REPORTED; provider compatibility warnings do not prove model identity.
DCA validates execution and result integrity, not external-agent answer quality.

Deferred to v1.0.2, not implemented here: structured-output JSON fence
normalization and raw/normalized hashes, tool-attempt taxonomy, noninteractive
request_user_input handling, and historical shard revalidation.

简体中文

本次更新为外部 Agent 增加可选的单任务文件与工具权限。没有显式 policy 时,仍按
原有只读、禁用工具的方式运行。

  • 文件读写范围与工具分别授权,非法配置失败关闭。支持 shell 或 unified exec、
    web search,以及已配置 HTTPS MCP server 的精确工具白名单。
  • 支持 Codex standalone profile sidecar。model/provider 必填,reasoning effort
    与 verbosity 可选,缺失时不推断默认值;配置或 I/O 错误保留完整不可变失败记录。
  • Linux 受限文件系统的 self-exec 兼容处理仅添加所选 canonical Codex binary 的
    内部 read 依赖,不开放整个 release 或 Codex home;相关上游问题为
    openai/codex#29049。
  • 独立的 runtime 拒绝范围阻止用户读写所选 standalone release 及重叠路径,安装
    位于 Codex home 之外时同样生效。显式 capability 任务遇到未知布局会在启动前
    失败关闭。用户请求、实际权限和保护元数据分开记录,详见源码 ZIP 内的权限文档。

此前获授权的精确读取、默认受限及限定写入 live 检查,已在 Ubuntu 24.04 /
Codex 0.155.1 的 production codex exec 路径通过。本次 deny-boundary 修复
只做离线验证,不扩大正常权限编译结果,没有重跑 live 调用。同版本
codex sandbox linux 诊断命令的 dispatch 故障,不等于 production 路径失败。

显式 capability 目前只识别 canonical
standalone/releases/<release-id>/bin/codex 布局;拒绝 launcher symlink 与
有歧义路径,不宣称其他安装或平台均已认证。identity/hash 复检不构成对特权并发
替换的原子防护,event audit 也不是 OS enforcement。served model 缺失保持
NOT_REPORTED,provider warning 不能证明实际服务身份。DCA 保证执行与结果
完整性检查,不保证外部答案正确。

继续留待 v1.0.2,未在本版实现:JSON 外围 fence 归一化与 raw/normalized hash、
tool-attempt taxonomy、非交互 request_user_input 处理、历史 shard 重新核验。

DCA v1.0.0

Choose a tag to compare

@ykw-9826 ykw-9826 released this 08 Sep 11:41

English

DCA v1.0.0 is the first public release.

DCA lets Codex hand tasks to external model agents such as GLM and DeepSeek, then collect their results back into the main workflow.

This release includes:

  • Run tasks with independently configured external model agents
  • Batch task execution and result collection
  • Compare outputs from different models
  • Keep execution records, failures and result history
  • Resume completed workflows without repeatedly polling from the main GPT model
  • Retry failed tasks without rerunning successful ones
  • Optional ServerChan and HTTPS webhook notifications
  • Local test demos that do not call real models

DCA is currently focused on external-agent execution and long-running task management. More external runtimes and practical workflows will be added in later versions.

简体中文

DCA v1.0.0 是第一个正式发布版本。

DCA 可以让 Codex 把任务交给 GLM、DeepSeek 等外部模型 Agent,再把结果收回来继续主流程,省掉手工在不同模型之间搬 Prompt 和反复查看任务进度的过程。

这个版本主要包括:

  • 调用独立配置的外部模型 Agent
  • 批量执行和收集任务结果
  • 对比不同模型的输出
  • 保存执行记录、失败记录和结果历史
  • 外部任务完成后自动回到主流程,不需要 GPT 反复查询进度
  • 重试失败任务时跳过已经成功的部分
  • 可选 Server酱和 HTTPS webhook 通知
  • 不调用真实模型的本地测试 Demo

目前 DCA 主要解决外部 Agent 调用和长任务管理。后续版本会继续增加新的外部运行环境和实际工作流支持。