Nausicaa 0.2.0

Full Changelog: v0.1.9...v0.2.0

Nausicaa 0.1.9

Cross-session messages addressed by an exact sessionId now reach the selected live Nausicaa session. Previously, the local resolver accepted the address but the router rejected it as identity-forged because its second check only recognized Run and lane IDs.

Use the full endpoint.sessionId returned by agent_awareness. Host identity failures now have a distinct diagnostic from permission denials. Ambiguous targets, unauthorized routes, and duplicate delivery remain protected.

Validation covers real local Session delivery and wake-up, same-Run session disambiguation, retry deduplication, privacy, recovery, and cross-workspace authorization. The release workflow requires successful CI on the tagged commit for Node 22.19.0 and 24.x, then repeats the tests, evaluations, and CLI/package smoke checks before uploading assets.

CI and Release both passed for commit 74771edc11f1ae08d8550fe3e64a3767c0cb210c.

Restart existing Nausicaa processes after upgrading so they load the corrected router.

Nausicaa 0.1.8

Teto now observes complete owner steps, including bounded tool outcomes, so an incoming request cannot prompt advice before the owner has had a chance to act.

• Keep completed steps intact and coalesce queued observations. Remove the repeated observation instruction that was incorrectly presented as a user request.
• Validate provider responses and stream events before response persistence or billing; propagate stream closure through retries.
• Harden cancellation at file commit and tool preparation boundaries, usage recovery, mixed legacy tool recovery, redirect timeouts, and independent task admission.
• Refresh the README and runtime documentation. Team tasks retain their asynchronous behavior without default total token or duration limits.
• Gate releases on successful CI for the same commit.

Validation: Node 22.19.0 and 24.x each passed 2,739 tests, 119 deterministic evaluations, and 15 CLI/package smoke tests. Two real-provider Teto probes passed 54 checks, including silence after a simulated successful team closure and a linked reply to a direct A2A request. These use fixed owner observations, not a new flight-game execution.

Full changelog: v0.1.7...v0.1.8

Nausicaa 0.1.7

Long Team sessions could appear frozen because terminal animation repeatedly laid out the entire history at alternating widths. This release keeps a stable content width, caches activity projections, and shows each member's current phase and tools in the TUI.

• Let task waits yield for new collaboration without restarting members or repeatedly consuming the same message.
• Preserve the conversation and member work across lead interruption; show public group reports while the lead is paused.
• Keep uncertain side effects distinct from cancelled read/compute operations.
• Give Teto a separate observation objective and reserve its messaging tool for findings and direct replies. Require the default task agent to act before ending with a promised action.

Validation includes the full CI gates on Node 22.19 and 24, package/PTY smoke tests, and a real OpenRouter Team task in the existing Three.js flight-game workspace. Nausicaa's developer and reviewer completed edit, review, repair, and re-review; an independent Chrome check passed nine startup/control/error checks on the saved artifact. Host feedback and a restart were required during this test, so it is not an unattended reliability benchmark.

Teto can still produce routine or repeated advice during long tasks. A model's final reply also remains a completion signal, not automatic proof that every requested action was performed. See the recorded test limits and failures.

Restart existing CLI processes to load the update; their in-memory code is not hot-reloaded.

Nausicaa 0.1.6

Fix interrupted conversations losing their Run context and Team group mentions waking a lane without delivering their message. Teto now coalesces queued observations and suppresses exact repeated advice while keeping its two core tasks.

• Escape and Ctrl+C preserve the Run, history, and asynchronous Team members. After an explicit interruption, reports remain available for the next user input. /stop also cancels Team work.
• Cancelled executions release their slot after the grace period; late runtime writes are fenced and unknown tool effects retain explicit resolution.
• Group mentions accept nausicaa, member IDs, and full lane IDs. They enter the recipient's next context boundary, recover unread messages without replaying committed receipts, and include history requests to retrieve long messages in full.
• Teto keeps routine observations private, advises on concrete new findings, and answers direct A2A requests. There is no per-user-request message cap or keyword-based judgment of advice quality.

Validation: TypeScript, 2,654 tests, 117 deterministic evaluations, and 15 built CLI/package smoke tests passed locally. OpenRouter's natural calendar/Todo Team probe passed all 12 checks across development, review, repairs, and acceptance (37 calls). A separate controlled Teto probe passed 14 checks for silence, intent drift, source projection, and direct A2A replies (4 calls). These probes do not establish statistical model reliability or browser interaction correctness.

GitHub CI passed on Node 22.19.0 and Node 24.

Install the packaged release directly, including while the npm registry is synchronizing:

npm install --global https://github.com/jackispm/nausicaa-harness/releases/download/v0.1.6/nausicaa-harness-0.1.6.tgz

The compaction evaluation fixture now verifies that its control retains raw history; its savings-versus-cost requirement is unchanged.

Nausicaa 0.1.5

团队现在可以按产物进度逐步组建：先开发，再加入审查成员；发现问题时，把返工和复验交回原成员。模型工具不再接受 dependsOn。

• team_assign 支持向同一团队新增成员及复用成员，附带简短交接资料。
• task_wait 等待实际结果；主管跨默认 24 步调度分段自动继续，验收后可以正常生成最终答复。
• 初始任务与后续任务的最终报告自动进入共享群；成员能查询所属团队和自己管理的团队。
• 修复异步通知丢失、报告落盘后交接中断的恢复、重复验收和嵌套权限继承。
• 保留取消、截断响应、异常结束原因和明确宿主硬限制的恢复边界。

验证：类型检查、完整测试、确定性评测和 CLI／打包／PTY 冒烟检查由 Node 22.19 与 Node 24 的 CI 执行。OpenRouter 普通中文实测完成开发、审查、返工、复验和最终交付，12 项协作检查通过；27 次模型调用，约 227 秒，已报告费用 $0.1398078572。

实测验证了协作链路及 JavaScript 语法，未进行浏览器交互验收。README、协作合同与包含前两次未通过记录的实测说明已同步更新。

Nausicaa 0.1.4

Team creation and long-running collaboration no longer fail on oversized tool metadata or implicit task budgets. New delegated tasks, members, follow-up assignments, and reducers run until completion, explicit cancellation, or an actual error, with provider timeouts and usage accounting preserved.

• Remove budget, deadline, attempt-count, success-criteria, and hard-constraint arguments from model-facing delegation and Team tools. Keep explicit legacy host limits readable during recovery.
• Normalize capability descriptions before both interactive and one-shot Team admission; preserve complete tool schemas and authorized grants.
• Continue background Team work after a normally completed lead turn, then wake the lead from durable reports. Cover concurrent reports, completion-cleanup races, resume, and cancellation fencing.
• Add shared Team messages, paginated history, reusable assignments, task reports, and explicit close. Preserve named members, capability narrowing, and nesting depth three.
• Let task_wait inspect both initial Team task IDs and subsequent assignments without restarting or consuming work.
• Preserve parent-Team group messaging and history through member reuse and recovery, without widening workspace or nested-Team permissions.
• Correct the Team status projection so resident follow-up requests and replies are validated against their assignment rather than the initial task; saved live events replay without false anomalies.
• Clarify Nausicaa's root identity and Teto's auxiliary observation role, with unsolicited A2A reserved for high-value advice.

Validation includes the unit/protocol/recovery suite, deterministic evaluations, CLI/package/PTY smoke tests, a ten-member report race, simulated tasks beyond the former duration/call caps, and OpenRouter Worker and calendar-Team collaboration probes. The live calendar probe checks collaboration and generated-file structure; browser acceptance is a separate check.

A separate OpenRouter probe passed all 15 member-reuse and recovery checks: normal session close/reopen, actual group history and message calls, member report, automatic lead continuation, and durable result consumption. Its diagnostic provider wrapper serialized paid requests, so this probe is not a concurrency benchmark. Saved events also replay without the former false task-status anomalies.

Upgrade and restart existing processes to load the new runtime. Previously persisted task limits retain their recovery semantics; new tasks use the simplified contract.

Nausicaa 0.1.3

Nausicaa 0.1.3 improves tool execution reliability and adds local trace inspection.

• Teto starts automatically for new Runs. Explicit stops survive restart and resume; the owner can restart it.
• Agents identify as Nausicaa, with shorter Teto guidance matched to the available tools. Skill loading preserves resource directories.
• /traces [status|preview] displays local Run events and metrics. Compatibility commands remain available, with clearer help and argument checks.
• Main and Worker persist tool results as they finish while preserving source order in model context and recovery.
• Worker batches honor declared concurrency limits. Terminal-write failures cancel and drain the batch before returning.
• Cancellation, tool deadlines, root schema constraints, failed-provider accounting, and recovered tool evidence have focused regression tests.

Known boundary: provider usage reported after cancellation has already won is not reconciled in this version. Tool cancellation remains cooperative.

Validation: Node 22.19 and Node 24 CI passed, including the full test suite, deterministic evaluations, CLI/TUI smoke tests, and package checks.

Install from npm:

npm install -g nausicaa-harness@0.1.3

Release assets include the package, source archive, Unix and PowerShell installers, and SHA256 checksums.

Full changelog

Nausicaa 0.1.2

Highlights

• One full-screen configuration workspace for Providers, Models, MCP, and Skills, with retained tab state and pale-pink menus.
• Clear reasoning-level descriptions and model • medium labels.
• Main is encouraged to start Teto early for complex work and keep progressing while it observes.
• New local commands: /settings, /system-prompt, /logs, /changelog, /update, /clone, /clear, and /btw (/side).
• Side questions use the current Run budget without adding their text to the main conversation. Cancellation and observer-result handling are hardened.

中文

• Providers、Models、MCP、Skills 合并为淡粉色全屏配置页，切页保留状态。
• 思考强度附有说明，模型名旁显示已选档位。
• 建议 Main 在复杂任务中尽早开启 Teto，并继续推进工作。
• 补齐设置、系统提示词、日志、更新、克隆会话与侧问命令。
• 侧问不进入主对话，但用量计入预算；完善取消和中止响应处理。

Install

Requires Node.js >=22.19.0. Restart Nausicaa after updating.

npm install -g nausicaa-harness@0.1.2

Assets include the npm package, a source archive, macOS/Linux and Windows
installers pinned to 0.1.2, and SHA256SUMS. These are Node.js package installers,
not standalone native executables.

Full changes: v0.1.1...v0.1.2

Nausicaa 0.1.1

Highlights

• Centered, searchable login with separate provider and authentication-method choices. OpenAI API access and ChatGPT subscription login are distinct entries.
• Consistent model-selection panels, with configured providers shown by default.
• /thinking and /effort select supported reasoning levels and preserve them with the current Main session.
• Centered MCP management for HTTP and stdio servers, with explicit access grants.
• Bundled codebase-map, task-plan, and code-review Skills, plus updated bilingual documentation.

Install or Upgrade

Requires Node.js 22.19.0 or newer:

npm install -g nausicaa-harness@0.1.1
nausicaa --version

Assets include the npm package, a source archive, version-pinned macOS/Linux and Windows installers, and SHA256SUMS. These are Node.js packages, not standalone native binaries.

Notes

Credentials persist locally after successful login. MCP configuration changes require a restart; generic MCP OAuth is not included. The unified Providers / Models / MCP / Skills workspace is planned for a later version.

中文摘要

新版统一了居中登录与模型选择面板，区分 OpenAI API 和 ChatGPT 订阅登录，新增 /thinking（/effort）、MCP 配置菜单及三个预设 Skill，并同步更新中英文 README。登录凭据会保存；MCP 配置修改需要重启。Providers、Models、MCP、Skills 的统一配置区域留待后续版本。

Full changelog