dominds 1.2.7 → 1.3.1

package/dist/docs/context-health.md

@@ -116,15 +116,27 @@ Levels are derived from the two thresholds:
 ### Continuation package (“接续包”)
 
 The remediation workflow centers around a _continuation package_ (a scannable, actionable bundle of
-context that survives a new course).
+context that survives a new course). When the agent is still clear-headed, this should preferably be
+kept as **one structured reminder**. But when the agent is already muddled in caution/critical,
+multiple rough reminders are acceptable as a bridge.
 
-Recommended structure (multi-line; scale by task size; focus on details not covered in Taskdoc):
+Recommended structure (multi-line but compact; focus on details not covered in Taskdoc):
 
 - First actionable step
 - Key pointers (files/symbols/search terms)
 - Run/verify (commands, ports, env vars)
 - Easy-to-lose ephemeral details (paths/ids/urls/sample inputs)
 
+Rules:
+
+- Keep normal reminders concise and few.
+- When preparing a new course, prefer compressing into one continuation-package reminder if the
+  agent is still clear-headed.
+- If the agent is already muddled, prioritize preserving volatile information even via multiple rough
+  reminders; in the next course, the first action should be to review, merge, and delete redundancy.
+- Do not duplicate Taskdoc content except for a short bridge when strictly needed.
+- Do not paste long raw logs/tool outputs into the continuation package.
+
 ### Caution (yellow)
 
 When `level === 'caution'`, the driver auto-inserts a **role=user** guidance prompt into the _next_
@@ -138,7 +150,7 @@ Current behavior:
   generations; configurable per model).
 - Each inserted prompt requires the agent to **curate reminders** (at least one call):
   - `update_reminder` (preferred) / `add_reminder`
-  - Maintain a continuation-package draft inside reminders
+  - Prefer a structured continuation-package reminder; if already muddled, rough multi-reminder carry-over is acceptable
   - Then `clear_mind` when it becomes scannable/actionable
 
 ### Critical (red)
@@ -147,7 +159,7 @@ When `level === 'critical'`, the driver enters a **countdown remediation** (max
 
 - On each turn, the driver records a **role=user prompt** (persisted as a user message) that is
   visible in the UI as a user prompt. This prompt tells the agent to:
-  - curate reminders via `update_reminder` / `add_reminder` (best-effort continuation package), and
+  - curate reminders via `update_reminder` / `add_reminder`, preferably into a continuation-package reminder but allowing rough multi-reminder carry-over when already muddled, and
   - then call `clear_mind` to start a new course.
 - The prompt includes a countdown: after **N** turns the system will automatically clear.
 - When the countdown reaches 0, the driver **automatically calls** `clear_mind` (with empty args; no
@@ -198,7 +210,7 @@ Note (zh UI copy):
   - `caution`: driver inserts a persisted role=user prompt (UI-visible user instruction).
     On entering `caution` it inserts once; while still `caution` it reinserts on a cadence (default: every
     10 generations; configurable per model). Each time, the agent must call at least one of
-    `update_reminder` / `add_reminder` and maintain a continuation-package draft, then `clear_mind` when ready.
+    `update_reminder` / `add_reminder` and preserve a continuation package. A single structured reminder is preferred, but multiple rough reminders are acceptable when already muddled; in the new course the agent should reconcile them first, then `clear_mind` when ready.
   - `critical`: driver runs a countdown remediation (max 5 turns) using **recorded role=user prompts**.
     Each prompt includes a countdown and instructs reminder curation + `clear_mind`. When the countdown
     reaches 0, the driver auto-executes `clear_mind` and starts a new course (no Q4H, no suspension).

package/dist/llm/defaults.yaml

@@ -68,6 +68,15 @@ providers:
           min: 1
           description: Provider-specific max tokens override.
     models:
+      gpt-5.4:
+        name: GPT-5.4
+        optimal_max_tokens: 200000
+        # Caution remediation reinjection cadence in generation turns (default: 10).
+        caution_remediation_cadence_generations: 10
+        context_length: 272000
+        input_length: 272000
+        output_length: 32768
+        context_window: '272K'
       gpt-5.3-codex:
         name: GPT-5.3 Codex
         optimal_max_tokens: 200000

package/dist/minds/system-prompt-parts.js

@@ -144,7 +144,7 @@ function getMemoryPromptCopy(ctx) {
             progressLine: '- 其中 `progress` 是全队共享公告牌：用于“阶段性进度快照”（关键决策/当前状态/下一步），不是流水账。',
             injectedTaskdocLine: '- 重要：差遣牒内容会被系统以内联形式注入到上下文中（本轮生成视角下即为最新；注入内容不包括全局约束）。需要回顾时请直接基于上下文里的差遣牒内容回顾与决策，不要试图用通用文件工具读取 `*.tsk/` 下的文件（会被拒绝）。',
             constraintsLine: '- 约定：`constraints` 只写任务特有的硬要求，不得写入系统提示/工具文档里已明确且由系统强制执行的通用规则（例如 `*.tsk/` 封装禁止通用文件工具）。一经发现重复，必须删除并告知用户。',
-            remindersLine: '- 提醒项（即编号提醒，工作集）：当前对话的高频工作记录/关键细节（偏私有，不作为全队公告）；保持少量（常见 1–3 条），优先 `update_reminder` 压缩/合并，不再需要就 `delete_reminder`。',
+            remindersLine: '- 提醒项（即编号提醒，工作集）：当前对话的高频工作记录/关键细节（偏私有，不作为全队公告）；默认保持少量（常见 1–3 条），优先 `update_reminder` 压缩/合并，不再需要就 `delete_reminder`。准备 `clear_mind` 开启新一程对话时，最好整理成“结构化接续包提醒项”；其中只保留差遣牒未覆盖、但恢复工作容易丢的细节（如第一步、关键定位、运行/验证信息、临时 ids/路径）。但若你已进入吃紧/告急、头脑不够清楚，允许先保留多条粗略提醒项把信息带过桥，新一程开始后第一时间复核、合并、删除冗余。',
             teamMemoryLine: '- 团队记忆：稳定的团队约定/工程规约（跨任务共享）。',
             personalMemoryLine: '- 个人记忆：稳定的个人习惯/偏好与职责域知识；记忆会在每次生成时自动注入上下文，应保持少量且准确（关键文档/代码的精确路径 + 最小必要事实）。不要记录具体任务状态。',
             subdialogDutyLine: `你当前处于支线对话：此处不允许 \`change_mind\`。当你判断需要更新差遣牒（尤其是 progress 公告牌）时，请在合适时机直接诉请差遣牒维护人 \`@${ctx.taskdocMaintainerId}\` 执行更新，并给出你已合并好的“新全文/替换稿”（用于替换对应章节全文）。不要声称已更新，除非看到回执。`,
@@ -155,7 +155,7 @@ function getMemoryPromptCopy(ctx) {
             subdialogWorkflowLine: `工作流：先做事 → 再提炼（\`update_reminder\`；必要时整理差遣牒更新提案并诉请 \`@${ctx.taskdocMaintainerId}\` 合并写入）→ 然后 \`clear_mind\` 清空噪音。`,
             mainlineWorkflowLine: '工作流：先做事 → 再提炼（`update_reminder` + `change_mind(progress)`）→ 然后 `clear_mind` 清空噪音。',
             contextHealthLine: '系统会自动监控并提示上下文健康度：进入"吃紧"或"告急"状态时会插入用户可见提示。当收到此类提示时，立刻停止继续大实现/大阅读；先提炼，再 clear。',
-            taskdocLogLine: '不要把长日志/大段 tool output 直接塞进差遣牒；差遣牒只写结论+下一步；细节只保留必要摘录放提醒项。',
+            taskdocLogLine: '不要把长日志/大段 tool output 直接塞进差遣牒；差遣牒只写结论+下一步；提醒项也只留可扫读摘录。若头脑清楚，接续包提醒项应尽量结构化、便于快速恢复；若已吃紧/告急，允许先保留多条粗略提醒项求稳，但进入新一程后必须尽快收敛。',
         };
     }
     return {
@@ -167,7 +167,7 @@ function getMemoryPromptCopy(ctx) {
         progressLine: '- Taskdoc `progress` is the team’s shared bulletin board: distilled milestone snapshots (key decisions/current status/next steps), not raw logs.',
         injectedTaskdocLine: '- Important: the Taskdoc content is injected inline into the context (the latest as of this generation; injected content excludes global constraints). Review the injected Taskdoc instead of trying to read files under `*.tsk/` via general file tools (they will be rejected).',
         constraintsLine: '- Convention: Taskdoc `constraints` must contain task-specific requirements only; do not include global, system-enforced rules already stated in system prompt/tool docs (e.g. `.tsk/` encapsulation bans general file tools). If duplication is found, you MUST remove it and notify the user.',
-        remindersLine: '- Reminders (i.e. numbered reminders, working set): your high-frequency per-dialog worklog + critical details (not a team bulletin board); keep it small (often 1–3 items), prefer `update_reminder` to compress/merge; delete when obsolete.',
+        remindersLine: '- Reminders (i.e. numbered reminders, working set): your high-frequency per-dialog worklog + critical details (not a team bulletin board); keep it small by default (often 1–3 items), prefer `update_reminder` to compress/merge, and delete when obsolete. When preparing `clear_mind` to start a new course, prefer a structured continuation-package reminder that keeps only details not already covered by Taskdoc but easy to lose during resume (first step, key pointers, run/verify info, volatile ids/paths). But if you are already in caution/critical and not clear-headed enough, you may carry multiple rough reminders across the course boundary first, then review/merge/delete them at the start of the new course.',
         teamMemoryLine: '- Team memory: stable shared conventions (cross-task).',
         personalMemoryLine: '- Personal memory: stable personal habits/preferences and responsibility-scope knowledge. Memory is automatically injected into context on each generation: keep it small and accurate (exact key doc/code paths + minimal key facts); do not store per-task state.',
         subdialogDutyLine: `You are currently in a subdialog: \`change_mind\` is not allowed here. When Taskdoc should be updated (especially the shared progress bulletin board), tellask the Taskdoc maintainer \`@${ctx.taskdocMaintainerId}\` with a fully merged replacement draft (full-section replacement). Do not claim it is updated until you see a receipt.`,
@@ -178,7 +178,7 @@ function getMemoryPromptCopy(ctx) {
         subdialogWorkflowLine: `Workflow: do work → distill (\`update_reminder\`; when Taskdoc needs updates, draft a merged replacement and ask \`@${ctx.taskdocMaintainerId}\`) → then \`clear_mind\` to drop noise.`,
         mainlineWorkflowLine: 'Workflow: do work → distill (`update_reminder` + `change_mind(progress)`) → then `clear_mind` to drop noise.',
         contextHealthLine: 'System will automatically monitor and alert on context health: yellow (caution/"Caution") or red (critical/"Critical") will insert a user-visible prompt. When you receive such alerts, immediately stop large implementations/reads; distill first, then clear.',
-        taskdocLogLine: 'Do not paste long logs/tool outputs into Taskdoc; Taskdoc should record decisions + next steps; keep only essential excerpts in reminders.',
+        taskdocLogLine: 'Do not paste long logs/tool outputs into Taskdoc; Taskdoc should record decisions + next steps; reminders should also keep only scannable excerpts. When clear-headed, keep continuation-package reminders structured and fast to resume from; when already in caution/critical, multiple rough reminders are acceptable as a bridge, but you must quickly reconcile them in the new course.',
     };
 }
 function buildMemorySystemPrompt(ctx) {