Configuration

For anyone who wants to change a setting. Most people never need to.

Two ways to set a value

Where	How	Notes
Settings → Tacit (UI)	flip the switch / pick the model / type the cap	writes only the keys you changed to ~/.dsh/storages/tacit/config.patch.json
Profile patch (YAML)	add an id-targeted row to ~/.dsh/profiles/web/cordis.patch.yml	any key, incl. the ones with no UI control; takes effect after restarting npx @deepseek-ai/dsh web

# ~/.dsh/profiles/web/cordis.patch.yml
- id: tacit
  config:
    model: deepseek-v4-pro
    autoDailyBudget: 50
    enrichPrompts: true

Precedence: UI patch > YAML > built-in defaults. A key set once in the UI keeps winning over YAML until you change it again in the UI (patch keys are never removed).

All keys

Key	Default	Allowed	What it does	UI
model	deepseek-v4-flash	deepseek-v4-flash, deepseek-v4-pro	model used for every Tacit call	Analysis model
autoAnalyze	true	boolean	analyze messy / corrected turns automatically	Auto-analyze messy turns
learnFromGood	true	boolean	also analyze a clean turn that follows a messy one (what you included the second time); counts against the daily cap	Also learn from a clean prompt right after a messy turn
autoDailyBudget	30	0–1000	cap on automatic analyses per local calendar day; 0 disables them	Daily cap on automatic analyses
autoMinSteps	15	1–500	a finished turn with at least this many model steps counts as messy (auto-analysis only)	—
steerAgent	true	boolean	inject the learned directives into every new conversation's system prompt	Inject learned directives into the system prompt
directiveEvery	3	1–100	new analyses between two directive distillations	—
directiveTrialTurns	10	1–500	finished turns a candidate directive stays on trial	—
directiveWorseBy	0.15	0–1	retire a candidate when its trial correction rate exceeds the baseline by more than this (or its messy-turn rate by more than twice this)	—
reviewCandidates	false	boolean	a freshly distilled directive stays queued until you press Start trial on it	Review new directives before their trial
bootstrapConcurrency	1	1–4	analyses run at once during Learn from my last 20 turns; same calls and cost, less waiting	—
costHistoryDays	30	7–365	days of detailed usage runs kept; older day files are deleted	Keep detailed usage history (days)
costWarnDailyUsd	0	0–10000	visual warning at 80 %/100 % of this daily Tacit spend; 0 disables	Warn above this daily spend (USD)
costWarnMonthlyUsd	0	0–10000	same as costWarnDailyUsd, over a calendar month; 0 disables	Warn above this monthly spend (USD)
enrichPrompts	false	boolean	opt-in: one small call before each send appends learned context (never rewrites your words); uncapped	Add learned context before each send
liveSuggestions	true	boolean	show the ✨ Improve prompt button	Enable the composer Improve button
maxPatterns	12	1–50	mistake patterns kept in the profile and offered to ✨ Improve	—
maxKeptTurns	60	1–1000	turns retained per conversation (newest kept)	—
maxPromptChars	4000	200–100000	prompt text kept per turn	—
maxAssistantChars	4000	200–100000	final answer kept per turn	—
maxToolCallChars	500	100–20000	tool-call arguments kept per call	—
maxToolCallsPerTurn	50	1–500	tool calls kept per turn	—

Good to know

• Out-of-range numbers are clamped silently to the allowed range; nothing is logged.
• 0 only means zero for autoDailyBudget, costWarnDailyUsd and costWarnMonthlyUsd. For every other numeric key 0 falls back to the default (the code uses value || default).
• A model outside the allowlist silently becomes deepseek-v4-flash when it comes from YAML, and is rejected with 400 bad-request when set through the UI/API.
• liveSuggestions: false hides the button only; the /api/tacit/improve route stays callable.
• Unknown keys are dropped, not rejected — an old learningThreshold from dsh-prompt-coach is simply ignored.
• DSH_HOME moves the whole storage root ($DSH_HOME/storages/tacit/).
• The bootstrap size (20) is an argument of the /api/tacit/bootstrap route (limit, 1–50), not a setting; the UI always sends 20.
• The retention selector only offers 7 / 14 / 30 / 90 / 180 / 365 days. A costHistoryDays set to another value inside 7–365 (e.g. via YAML) stays in effect, but the Data & privacy card's selector shows 30 until you pick one of its own options, which then overwrites it.
• Changing steerAgent or any directive affects new conversations only, with one exception. Removing a directive or switching one off reaches open conversations at their next system-prompt assembly; everything else waits for the next conversation. See How it works §7.

---

← Docs index · See also: Privacy, cost & limitations