dsh-diary

dsh-diary · Diary Plugin

A diary plugin for dsh (DeepSeek Harness): write your diary on a paper-letter-styled web page where the raw text is saved to disk first and the AI comment is generated after — a failed summary never loses your writing.

中文版 · Built-in guide: press ? at the top right of the page (or open /diary/guide) after installing.

Preview

Division of labor

Quick start

git clone https://github.com/zhm20001/dsh-diary.git
cd dsh-diary
pnpm install            # compiled lib/ is committed — no build needed
dsh plugin --profile web add <path-to-this-repo>
# restart dsh web, then open http://127.0.0.1:<port>/diary

On first open the page shows a setup card: enter a diary directory (e.g. ~/Documents/Diary; ~ expands to your home) and save — it is written to the plugin's config.json and takes effect immediately, no restart. Then write your first entry → "Save & summarize".

Daily flow: open the page → write → submit. The raw text hits disk immediately; if the summary fails, your text is safe — press "Retry summary only". Drafts autosave to browser localStorage (keyed per date). The dsh web sidebar also gets a "Diary" entry at its footer (src/client.js, label follows your browser language). Each submit also quietly maintains a user profile (see "Long-term memory"); a failure there only shows a non-blocking note that saving is unaffected.

Long-term memory (user profile)

Alongside the summary, every save maintains a user profile — plain markdown, stored at .diary-meta/profile.md inside the diary directory, next to (but conceptually separate from) the emoji meta table, and readable by you at any time.

• Injection: before generating the day's summary, the profile plus a digest of the last week's recorded days (each day's keyword + one-liner, extracted from past comment blocks — zero extra LLM calls) is appended to the summary request as reference data; the day's full text remains the only subject. A missing or empty profile simply omits that section and the summary proceeds as before.
• Update: right after the comment lands, one independent profile-update call runs (old profile + the day's full text → new profile), synchronous and written to disk before the response returns. The profile is layered as core self / current phase / recent events; the update rules are built in and fixed (the model adjudicates what recent events to drop once they go unmentioned past the retention period; there is a hard size cap) and are not user-editable — you can change the comment prompt (how the context is used), not the update rules (how the profile changes).
• Failure semantics: a failed profile update never affects saving or summarizing — the API still returns 200 and the page shows one "Profile update failed, saving unaffected" note; there is no retry button (the next submit self-heals). A failed summary still takes the existing 502 path and does not trigger a profile update.
• Full rebuild: the "Generate profile from history" button in the settings card is only clickable when no profile file is detected (disabled when one exists — it never overwrites a profile under maintenance); after a cost warning and a second confirmation it rebuilds an initial profile from your entire history in chronological chunks. Moving machines = copy .diary-meta/profile.md.

There is no switch and no new config key: daily LLM usage roughly doubles (one extra profile-update call plus the profile riding along with the summary), which is negligible for a single-user setup. Your diary text is never modified by "forgetting" — the profile is recomputable derived state.

Configuration

Three layers, highest priority first:

Customizing the comment prompt: the 📝 button next to the "comment model" selector opens the prompt card — read the system prompt that drives the AI comment, then edit, save, or reset to default. It is stored as summaryPrompt in config.json and takes effect immediately (precedence as in the table above). Saving validates the output contract: the three markers 【今日关键词】 (keyword), 【一句话总结】 (one-line summary) and 【评注】 (comment) must stay (the 【当日 emoji】 line is optional — without it that day's heatmap cell falls back to green); a custom prompt only shapes future summaries, never rewrites comment blocks already in your diary. A broken prompt injected via a cordis patch fails loudly at summary time (no silent fallback).

Other plugin options (cordis patch layer): provider / model (summary model, defaults deepseek-official / deepseek-chat), temperature (0.6), timeoutMs (120000), nightCutoff (6), pagePath (/diary; if changed, update the href in src/client.js too), summaryPrompt (comment prompt override).

Privacy & your data

• Your full diary text is stored only on your machine. The outbound calls are summary generation and the profile update: the day's full text (plus the profile and last week's digest for the summary) goes to the configured (or selected) LLM provider. Heatmap data, meta files and drafts never leave your machine (drafts live in browser localStorage).
• Data is self-contained in the diary directory: entries are YYYY-MM-DD.md; the day's emoji lives in the hidden .diary-meta/YYYY.json and the user profile in .diary-meta/profile.md inside the same directory. Backup = copy the whole directory (don't forget .diary-meta/ when migrating, or emoji silently degrade to plain green cells and the profile falls back to a cold start).
• "Has a diary" is always derived from filenames in the directory at request time (suffixed files like 2026-04-28-note.md count as 04-28; multiple files per day are not double-counted) — manual create/delete/rename shows up on the heatmap immediately; orphaned emoji in meta are ignored harmlessly.

Semantics

• 30-hour clock: anything written between midnight and nightCutoff (default 6 am) counts as the previous day, with timestamps shown as 25:xx–29:xx — late-night entries never split across two files.
• Resubmitting the same day appends the new text and regenerates the summary for the whole day: the old AI comment block is stripped and rewritten; everything you wrote above it is untouched. Convention: don't hand-edit below the ## AI评注 heading.
• File contract: <diaryDir>/YYYY-MM-DD.md, front matter from the template (empty date: field or {{date}} placeholder supported).
• On summary failure the API returns stage:'summary' + hint (HTTP 502); a save failure returns stage:'save'.
• Heatmap: one frame per half-year, a fixed 27-week × 7-day grid starting Monday; frames start on the Monday of/before the half-year's first day, and adjacent frames intentionally share an edge week. Written = the day's emoji (green cell when no emoji); past unwritten = faded; future = grey; today gets a ring. The denominator is the actual number of days in that half-year. In rare years (July 1 falling on a Sunday, e.g. 2018/2029) the H2 window spans 190 days — the Dec-31 cell is not drawn (counts unaffected).
• Backfilling comments for old entries / moving directories / deleting a day: see the "Your data" section of the built-in guide.

Development

pnpm install
pnpm typecheck
pnpm test               # vitest: core pure functions + path resolution + atomic writes
pnpm build              # regenerates lib/ (CI verifies it matches src/)

# Smoke: a dev overlay points the plugin at /tmp/diary-sandbox — never touches your real diary
dsh web --patch dev.patch.yml   # then browse /diary-dev

Structure: src/core.ts pure functions (no IO) → src/paths.ts path vars (config.json read/write) → src/service.ts routing & orchestration (webServer/httpServer dual-name compatible) → src/summary.ts summary LLM pipeline → src/profile.ts long-term memory (profile read/write, injection assembly, update loop, full rebuild) → src/page.ts page + src/guide.ts guide. lib/ is compiled output, committed with the repo (clone-and-use, no build); changes to src/ must be committed together with lib/.

Host requirement: a dsh web host (the internal harness build exposes ctx.webServer, the npm @deepseek-ai/dsh-host-webserver rc line exposes ctx.httpServer — both are supported).

License

MIT