deepscientist-mode

# DeepScientist Mode This skill is the compact operating manual for the Hermes-native DeepScientist plugin. The full manual lives in `docs/USAGE.md` inside the plugin source. ## Core identity Hermes is the DeepScientist working agent. Use native `ds_*` tools for durable quest state, memory, artifacts, baseline gates, ideas, experiments, paper bundles, and quest-local bash execution. Hard rules: - 不要调用全局 npm `ds` for normal operation. - 不要暴露 raw MCP. - Do not open Web UI or TUI surfaces. - Do not use social/browser/messaging connector surfaces from DeepScientist. - Use native tools only; the former background-service layer is not part of this plugin. - Keep only active stage context in the working prompt; do not load all DeepScientist stage skills at once. - Plugin skills are namespaced resources. Always load them with `deepscientist:<skill>` (for example `deepscientist:scout`, `deepscientist:strict-research`, `deepscientist:paper-fetch`); bare names such as `scout` or `paper-fetch` may not appear in ordinary `skills_list` and should not be used with `skill_view`. - Keep DeepScientist project memory separate from Hermes long-term memory. Use Hermes `memory` only for user preferences or cross-project stable facts. ## 项目本地存储 Default storage follows upstream `ds --here` semantics. If Hermes is launched from a project directory, the DeepScientist runtime home is: ```text <project>/DeepScientist/ ``` This tree stores: - `memory/` - `quests/` - quest-local `memory/` - `config/` - `runtime/` - `logs/` - `cache/` - `runtime/hermes-session-map.json` Do not assume a shared DeepScientist home. If state appears missing, first check the Hermes launch directory and then inspect `<project>/DeepScientist/`. ## Agent runbook ### 首次接手任务 1. Confirm the current Hermes working directory is the intended project directory. 2. Run `ds_doctor`. 3. Run `ds_list_quests` to discover existing project quests. 4. If continuing, call `ds_set_active_quest`; if starting new work, call `ds_new_quest`. 5. Call `ds_get_quest_state` before deciding next actions. 6. Select exactly one relevant stage skill, such as `deepscientist:scout` or `deepscientist:experiment`. 7. Use Hermes tools for reading files, editing code, web research, and ordinary tests. 8. Persist research state back into DeepScientist with `ds_memory_write`, `ds_artifact_record`, or specialized artifact tools. 9. Before final reply, re-check quest state if important changes were made. ### Continuing an existing quest 1. `ds_list_quests`. 2. Select the likely quest by title/goal/recent events. 3. If ambiguous, ask the user to choose; do not merge unrelated quests. 4. `ds_set_active_quest`. 5. `ds_get_quest_state`. 6. Use `ds_memory_search` for prior conclusions, constraints, baseline choices, and experimental setup. 7. If the same research project moves from user-gated planning to autonomous execution, or back to review, call `ds_update_quest_mode` on the existing quest. Do not create a new quest just to change modes. 8. Continue work and write durable updates. ### Creating a new quest Use `ds_new_quest` when the user gives a new research goal and no existing quest clearly matches. Hermes chooses the mode contract; do not make the user type a mode command for normal use. Defaults and overrides: - Default is `workspace_mode="copilot"`, `decision_policy="user_gated"`, `need_research_paper=false`, `final_goal="open_ended"`. - Choose `workspace_mode="autonomous"` only when the task should be owned across multiple steps without stopping after the current request unit. - Keep `workspace_mode` separate from `final_goal`: autonomous does not imply paper writing. - For autonomous non-paper tasks, explicitly set `final_goal` such as `idea_optimization`, `literature_scout`, `baseline_reproduction`, `analysis_report`, or `quality_result`, plus `delivery_mode`, `completion_criteria`, and `mode_rationale`. - Set `final_goal="paper"` or `need_research_paper=true` only for explicit paper/posting/paper-bundle goals. After creation: 1. `ds_get_quest_state`. 2. Set or infer the starting stage, usually `scout`, `baseline`, or `idea`. 3. Write initial constraints and known facts with `ds_memory_write`. ### Switching mode inside the same quest A quest is the research-project boundary, not a single question and not a fixed execution mode. For follow-up requests inside the same research project, keep the same quest and switch its mode when the phase changes. Use `ds_update_quest_mode` when: - Copilot planning/doc editing/experiment design has converged and Hermes should start autonomous experiment execution. - Autonomous execution should pause and return to user-gated review or plan revision. - The terminal goal for the next phase changes while the project identity stays the same. Rules: - Do not create a new quest merely to switch between `copilot` and `autonomous`. - Switching to autonomous requires `mode_rationale`; also provide `final_goal`, `delivery_mode`, and concrete `completion_criteria` whenever possible. - Autonomous still does not imply paper writing. For experiment execution use `final_goal="quality_result"`, `delivery_mode="experiment_execution"`, and `need_research_paper=false` unless the user explicitly asks for a paper. - Switching back to copilot defaults to `decision_policy="user_gated"` and should preserve the same quest id. ## 工具选择规则 Use DeepScientist tools when the result should be part of the durable research record: - `ds_doctor`: verify native runtime. - `ds_list_quests`: find existing quests. - `ds_new_quest`: create a quest. - `ds_set_active_quest`: bind this Hermes session to a quest. - `ds_get_quest_state`: inspect current state. - `ds_update_quest_mode`: switch an existing quest between copilot/autonomous without changing quest identity; use for phase transitions inside the same research project. - `ds_add_user_message`: append user instructions to a quest. Use `record_only=true` when the instruction should be durable context but should not wake/queue a pending user message. - `ds_record_user_requirement`: record durable user requirements into quest conversation and `active-user-requirements.md` without leaving `pending_user_message_count > 0`. - `ds_memory_search`: search project/quest memory. - `ds_memory_read`: read a memory card. - `ds_memory_write`: store durable research facts, constraints, conclusions. - `ds_artifact_record`: record generic evidence or structured outputs. - `ds_confirm_baseline`, `ds_waive_baseline`, `ds_attach_baseline`, `ds_create_local_baseline`: manage baseline gates. Prefer `ds_create_local_baseline` for local stubs under `baselines/local/<baseline_id>/baseline.md` before calling `ds_confirm_baseline`. - `ds_submit_idea`: record a research idea candidate. - `ds_record_main_experiment`: record a main experiment run. - `ds_create_analysis_campaign`, `ds_get_analysis_campaign`, `ds_record_analysis_slice`: record and diagnose systematic analysis. Before writing/paper bundle work, inspect active analysis campaign state and finish or close pending slices. - `ds_submit_paper_outline`, `ds_submit_paper_bundle`: record writing outputs. Use `candidate -> select` for outline selection; `selected` is only a compatibility alias for `select`. Markdown-only bundles are supported; the Hermes wrapper counts `##` sections and aligns returned guidance with the latest quest anchor. - `ds_bash_exec`: run/list/read/wait/stop quest-local execution that should be logged by DeepScientist. Set `allow_project_root=true` only when project-root workdir is required; set `summary_mode=true` for compact provenance output; complex Python should be written to a `.py` file before execution instead of large inline heredoc. - `ds_workflow_smoke_report`: produce a non-mutating Hermes-only checklist for dataset inspection, baseline, experiment, analysis, paper bundle, and final report handoff. - Strict research tools: when the user asks for careful/strict literature investigation or a survey, the router may recommend strict research but must not force it. The Hermes agent decides from the full user intent whether to enable strict research; if yes, use `ds_strict_research_prepare`, `ds_strict_research_record_candidate`, `ds_paper_reliability_verify`, and `ds_strict_research_init_bibliography` to enforce broad candidate scouting, reliability verification, conservative filtering, quest-local PDF/reference storage, and bibliography updates before writing. - `ds_pause_quest`, `ds_resume_quest`, `ds_stop_quest`: update quest lifecycle. Load `deepscientist:paper-fetch` when a DeepScientist quest needs arXiv/OpenReview/PDF retrieval or official paper-resource verification. Do not load note-taking `clip` for this case unless the user explicitly asks to archive the paper into `llm-wiki`. Use Hermes native tools when work is local and does not itself need to become quest state: - `read_file`, `search_files`, `patch`, `write_file` for files. - `terminal` for ordinary tests/builds. - `web`/browser tools for information gathering. - `todo` for current-session task management. If Hermes native tools produce important research evidence, immediately record the evidence with the relevant DeepScientist tool. ## 阶段推进规则 Available stage skills: - `deepscientist:scout` - `deepscientist:strict-research` - `deepscientist:paper-reliability-verifier` - `deepscientist:baseline` - `deepscientist:idea` - `deepscientist:optimize` - `deepscientist:experiment` - `deepscientist:analysis-campaign` - `deepscientist:write` - `deepscientist:finalize` - `deepscientist:decision` - `deepscientist:figure-polish` - `deepscientist:intake-audit` - `deepscientist:review` - `deepscientist:rebuttal` Rules: 1. Read `ds_get_quest_state` before stage decisions. 2. Load only the active stage skill and one companion skill if truly necessary. 3. Define the current gate/output before doing work. 4. Record completion evidence as memory/artifact. 5. If switching route or stage, record the reason with `ds_artifact_record` or `ds_memory_write`. ## Deep-integrated companion/support skills The plugin bundles several formerly user-local skills as `deepscientist:*` resources. Load at most one when the current subtask needs it: - `deepscientist:experiment-execution` for formal command documents, manifest validation, `planned_not_executed` boundaries, baseline gate/comparator semantics, and real-runner backfill. - `deepscientist:quest-handoffs` for `AGENTS.md`, concise current-status handoffs, researcher packages, sync verification, and `ds_artifact_record` milestone routing. - `deepscientist:writing-plans` for implementation plans, DeepScientist experiment roadmaps, code-only passes, and formal experiment command handoffs. - `deepscientist:paper-reliability-verification` for using `ds_paper_reliability_verify` and accepted-publication evidence before citing papers as reliable support. - `deepscientist:review` for skeptical draft/report audits, claim downgrade, revision logs, experiment TODOs, and route decisions. These are plugin-owned copies. Do not depend on bare global skills with the same original names during DeepScientist work. ## Slash commands Use slash commands for quick user interaction: - `/ds help` - `/ds mode on|off|status` - `/ds doctor` - `/ds list` - `/ds active [quest_id]` - `/ds status [quest_id]` - `/ds new <goal>` - `/ds send <quest_id> <message>` - `/ds stage [stage]` - `/ds events <quest_id> [limit]` - `/ds docs <quest_id> [name ...]` Use tool calls rather than slash commands for complex agent work. ## Memory and artifact writing - Use memory for durable facts: problem framing, constraints, chosen metrics, environment assumptions, validated conclusions, and reusable lessons. - memory kind canonical values are `papers`, `ideas`, `decisions`, `episodes`, `knowledge`, `templates`; semantic aliases such as `constraint`, `context`, `observation`, `hypothesis`, `result`, and `plan` are normalized by the Hermes wrapper and retained as tags/metadata. - Use artifacts for evidence: plans, baseline decisions, experiment records, analysis results, paper bundles, review reports. - Do not write short-lived todos as memory. - Titles should be searchable and specific. - Include source paths, dates/stages, and applicability when writing memory. ## Completion checklist Before final response: 1. Confirm active quest id and stage if a quest was used. 2. Confirm important outputs were persisted via `ds_memory_write`, `ds_artifact_record`, or specialized tools. 3. Mention key `<project>/DeepScientist/...` paths if files were created/updated. 4. Mention verification results or explicitly state what was not verified. 5. Provide next-step recommendation. ## 完成任务回复 For DeepScientist tasks, final replies should include: - Active quest id. - Current stage. - DeepScientist tools used. - New/updated memory or artifacts. - Relevant project-local paths. - Verification summary. - Next step. - Completion time.