Mimir (dsh-mimir)

English | 中文

Mimir is the research-assistant plugin suite for the DeepSeek Harness: an arXiv literature surface, a persistent research wiki, a LaTeX compile tool, and an independent fresh-reviewer loop. It ports the ARIS workflow mechanisms (failed-ideas memory, reviewer independence, compile-driven writing) into one dsh plugin.

Plugin

Function plugin (name/inject/Config/apply). Injects commands, tools, subagents, storageDomain, and webServer, so a composition must also mount the storage hub, one KV backend, the domain form, and — for the web panel's PDF route — the HTTP carrier (see examples/mimir-agent/).

Web research panel

The plugin also mounts the research Remote namespace (ResearchService) plus the /research/pdf/<project id> and /research/figure/<project id>?path=<rel path> prefix routes, which back the web research workbench (dsh-client-ui-mimir): project list from the wiki's projects table, the section outline, source read/replace for the panel's editor, latexmk/tectonic compile with parsed issues, the compiled PDF stream, the literature list (listPapers), the arXiv keyword subscriptions (listArxivSubscriptions / saveArxivSubscription / deleteArxivSubscription / checkArxivSubscriptions — persisted as <workspaceDir>/arxiv-subscriptions.json outside the wiki domain, re-checked on a scheduled daily timer and on demand, new entries surfaced to the panel for one-click import), the experiment-run list (listExperiments), whitelisted markdown artifact reads (readArtifact — IDEA_REPORT.md, EXPERIMENT_PLAN.md, EXPERIMENT_LOG.md, REVIEW_REPORT.md; any other name is an invalid-artifact failure, a missing file an artifact-not-found), and the paper directory's figure scan (listFigures — .png/.jpg/.jpeg/.svg/.pdf files at the top level and one figures/ level deep), plus deleteFigure and convertFigure (host-side SVG → PDF conversion — PNG via the platform rasterizer as the fallback — for the workbench's insert-into-paper flow, probing rsvg-convert/inkscape/magick on PATH and qlmanage on macOS). Backup and migration go through exportWiki / importWiki: the export snapshots all six wiki tables into one dated JSON envelope (format: "mimir-wiki", version: 2); the import revalidates the envelope and every row against its table's zod schema before any write, merge upserts only absent primary keys (existing records are skipped, never overwritten), and replace wipes all six tables first, so it requires confirmReplace: true. The figure route serves one scanned file by workspace-relative path; a non-image extension or a .. escape is a 400, an unknown project or file a 404. Each project can point at its own paper directory: the record's optional paperDir (set via the wiki_note tool's set_project action) names a subdirectory of the workspace, and every paper call also accepts an explicit dir override (?dir= on the routes); the candidate must be a relative path confined inside the workspace — an absolute path or a .. escape is an invalid-dir failure (400 on the route). savePaperSource replaces main.tex under optimistic concurrency — the caller's base mtime is re-checked and the atomic commit runs inside the @deepseek-ai/dsh-atomic-write writer lock, so an agent writing through the file tools can never interleave a check-then-write; a displaced draft gets a conflict reply carrying the current mtime. Compile status is process memory keyed by project id; both routes 404 unknown projects. A composition without webServer (e.g. the TUI) cannot mount the suite — inject is all-or-nothing.

Config

- id: mimir
  name: 'dsh-mimir'
  config:
    workspaceDir: .research        # research workspace root, resolved against the process cwd
    reviewer:
      provider: spawn              # subagent provider route; reserved for cross-model review
      maxRounds: 3                 # per-project review-round budget
    latex:
      engine: auto                 # auto (probe latexmk then tectonic on PATH), an engine name, or an absolute binary path (basename picks the dialect)
      timeoutMs: 120000            # compile kill timeout
    arxiv:
      maxResults: 10               # default arxiv_search result cap

Tools

• arxiv_search { query, max_results? } → { results: [{ id, title, authors, summary, published, url }] } from the arXiv Atom API. Network and HTTP failures reject.
• paper_fetch { arxiv_id } → one paper record by bare id (version suffix allowed).
• wiki_note { action, ... } → the model's read/write surface over the wiki domain. Actions: add_paper, add_idea, fail_idea, add_claim, set_claim, add_experiment, set_experiment, list, get.
• latex_compile { project_dir } → { success, engine, errors, warnings, log_excerpt }; runs the resolved engine in project_dir — latexmk -pdf -interaction=nonstopmode -halt-on-error main.tex or tectonic --keep-logs --synctex main.tex — and parses the log into file/line-attributed diagnostics (tectonic's --keep-logs main.log when present, else its error:/warning: console lines).

Commands

• /research-idea <direction> — registers a project, scaffolds IDEA_REPORT.md, and instructs the model to survey arXiv, check failed ideas first, and record the idea.
• /research-plan [project id] — scaffolds EXPERIMENT_PLAN.md from the idea report; planned claims become pending wiki claims.
• /research-review <scope> <paths...> [project id] — one independent review round: a fresh reviewer subagent gets only absolute file paths (never the executor's summary) and returns a schema-validated PASS/WARN/FAIL verdict. WARN/FAIL is handed back to the agent as a revision follow-up.
• /paper-write [project id] — scaffolds paper/main.tex + references.bib and instructs the model to draft and compile-fix until clean.
• /paper-compile [dir] — compiles once (default <workspace>/paper) and reports parsed diagnostics.

Storage

The wiki is the research_wiki domain (version 2) over the configured storage-domain backend, with five tables: papers (keyed by arXiv id), ideas (failed ideas are never deleted — this is the anti-repetition memory), claims, projects (pipeline stage, artifacts, reviewRounds), and experiments (run records: name, status, scalar metrics, optional log path). The plugin closes the domain on unmount.

Templates

templates/ carries the artifact skeletons (IDEA_REPORT.md, EXPERIMENT_PLAN.md, paper/main.tex, paper/references.bib). The runtime source of these bytes is src/templates.ts; the directory exists for human reference and is not read at runtime.

Model Experience

Tool schemas and command instructions

What the model sees

Four tool schemas (arxiv_search, paper_fetch, wiki_note, latex_compile) in the tool catalog, plus structured follow-up instructions when a /research-* command runs.

Token effect

Conditional: tool schemas are present whenever the plugin is mounted; command instructions appear only after the corresponding slash command.

KV Cache effect

Append-only: follow-up instructions extend the conversation without replacing earlier request content.

Known Limitations and Deferred Work

• No real end-to-end coverage yet — the suite is verified by unit tests (log parser, wiki domain) and composition loading; a live-model example run requires DEEPSEEK_API_KEY and a TeX installation, neither of which is assumed present.
• LaTeX log parsing is single-line based — the classical 79-column log wrap is not re-joined, so a wrapped message keeps its first line only.
• Reviewer provider is deployment-selected but not cross-model — reviewer.provider picks a registered subagent route; routing a different model family to the reviewer is the provider's own configuration concern.