@askalf/dario 6.19.2 → 6.20.1

package/docs/keys.md

@@ -34,12 +34,45 @@ on its next request, no restart.
 | `--expires=30d` | Refused after this, like a revoked key. `12h`, `2w`, or an ISO date. |
 | `--budget=$5/day` | A daily cap on the API-equivalent price of the key's traffic. See **Budgets**. |
 | `--budget-tokens=2M/day` | A daily cap on tokens, every bucket counted. See **Budgets**. |
+| `--effort=low` | The effort this key's requests run at when they name none themselves (no model-name suffix, no effort field). `low`, `medium`, `high`, `xhigh` or `max`. See **Default effort**. |
 
 `dario keys revoke <name>` refuses a key from now on and keeps it in the list;
 `dario keys rotate <name>` prints a new secret under the same name, seat,
 models and expiry, and the old secret stops at once; `dario keys remove
 <name>` forgets it.
 
+## Default effort
+
+`dario keys create alice --effort=low`, or `dario keys effort alice low` on an
+existing key (`--clear` to remove it), sets the effort the key's requests run
+at when they name none themselves. Resolution per request, first match wins:
+
+1. an effort suffix on the model name (`claude-opus-5-5:low`);
+2. the proxy's `--effort` / `DARIO_EFFORT` pin, when set to a level (`client`
+   is not a level: under it the key's default still fills in);
+3. the client's own effort field (`output_config.effort`, `reasoning_effort`, `reasoning.effort`);
+4. the key's default;
+5. `high`, the proxy's long-standing fallback.
+
+So a key default never overrules a client that chose, and the operator's pin
+still overrules everything, as it always has. On the Claude path the level
+goes out as `output_config.effort`.
+
+On the Codex path it goes out as `reasoning.effort`, on both routes: the
+Messages-shape request dario translates, and the Responses body Codex CLI
+sends, which dario otherwise forwards as written (the level is added to its
+`reasoning` object; `summary` and the rest stay). Two differences there. The
+proxy pin has never reached the Codex backend, so it is not in the order. And
+a Messages-shape client's `thinking.budget_tokens` is that client's choice on
+Codex (it becomes `reasoning.effort`), so it outranks the key's default the
+way the client's own field does. With none of those, the backend's own
+default applies, not `high`.
+Why the knob is per key: a reasoning model at its default effort can spend
+most of its output thinking on work that did not need it (the `Thinking:` line
+in `dario usage` shows the share per key), and the lane that needs the cap is
+usually one consumer, not the whole proxy. `dario keys list` shows an EFFORT
+column; `GET /admin/keys` and `dario keys --json` carry `effort`.
+
 ## Budgets
 
 `dario keys create alice --budget=$5/day --budget-tokens=2M/day` — or
@@ -91,7 +124,9 @@ fields (an empty body clears), audited as `key_budget`.
   to the key; the header cannot overrule a credential.
 - **`dario usage --by-key`** — the lifetime API-equivalent number split per
   consumer: named keys, header names, and the `u_…` hash of a client's user
-  id, whichever named the request.
+  id, whichever named the request. The tokens line under each key includes
+  its thinking share: how many of the key's output tokens were the model's
+  reasoning, and what they cost at the output rate.
 - **The log file** — `consumer` on every request line; `reject: "key-model"`
   on a model refusal; `event: "admin.key_create|key_rotate|key_revoke"` with
   `key: <name>` when the admin API changed something.

package/docs/pool-keys-and-analytics.md

@@ -69,8 +69,11 @@ The rolling window forgets on every restart; the **ledger** does not. Since 6.6
       Claude       $388   1,204 reqs   (Opus 5 $301 · Sonnet 5 $86.68)
       ChatGPT    $24.75     311 reqs   (gpt-5.6-terra $24.75)
     Today $48.20 · Last 7d $413 · Last 30d $413
+    Thinking: 2.1M of 3.4M output tokens (62%) · $156 of the number above
 ```
 
+The `Thinking:` line is the share of output tokens that were the model's reasoning, read from the usage block upstream returns (Claude's `output_tokens_details.thinking_tokens`, and on the ChatGPT-subscription engine `output_tokens_details.reasoning_tokens`, both counted inside output), and what that reasoning would have cost at each model's output rate. It is the first number to look at when a bill surprises you: a reasoning model at its default effort can spend most of its output thinking on work that did not need it, and `output_config.effort` (or a `:low` model-name suffix, see [configuration](configuration.md#per-request-effort-by-model-name)) is a cheaper fix than a smaller model. `dario usage --by-key` shows the share per key, so the lane that is thinking the most is one line away.
+
 Only served requests count. Traffic that was metered anyway — an API key upstream, or Anthropic's paid `extra_usage` overage — is kept in its own column and reported as spent, not saved. `dario usage --card` writes the headline as a 640×320 SVG you can drop in a README or a post, and `--donut` writes the same number as three rings — by model, by key, subscription vs metered; `--no-ledger` / `DARIO_LEDGER=0` turns the file off, `DARIO_LEDGER_PATH` moves it, and `GET /analytics/ledger` is the per-day table behind the number. Details: [api-equivalent-spend.md](api-equivalent-spend.md).
 
 A request served on a Max plan's API credits (the `credits:<model>` rung of the failover chain, key in `DARIO_CREDITS_API_KEY`, [failover](two-plans.md#failover-between-subscriptions)) is counted as spent the same way in the ledger; `/analytics`, `/metrics` and the dashboard show it under its own bucket, `credits`.

package/docs/system-prompt.md

@@ -13,6 +13,8 @@ The empirical basis for this feature lives in [`docs/research/system-prompt-clas
 | `aggressive` | Partial + remove the prompt-level RLHF restatements (`IMPORTANT: Assist with authorized security testing…`, `IMPORTANT: You must NEVER generate or guess URLs…`) and the `# Executing actions with care` section. | <3% above partial |
 | `<file path>` | Replace the slot entirely with the contents of a file you control. The escape hatch for users running well-defined agent workflows with their own system prompt. | depends on your prompt |
 
+What `partial` finds to strip depends on the prompt Claude Code sends for the model. The Sonnet 5 prompt has the `# Tone and style` and `# Text output` sections and the Doing-tasks bullets. The Fable prompt has a one-line rule about code comments, which `partial` swaps for a "be thorough" instruction. The base prompt (Claude Code 2.1.316) and the Opus 5 prompt carry none of these, so `partial` sends them unchanged and only `aggressive` changes them.
+
 ## Aggressive vs partial — what's the actual difference?
 
 Aggressive is provided for completeness, not because it does meaningful work. The added removals are *prompt-level restatements* of refusal categories — reminders the prompt makes about RLHF behavior that's already trained into the model's weights. Removing the reminder doesn't remove the trained behavior. We measured this: 9 trials (3 prompts × 3 strip levels), aggressive vs partial added <3% practical change on benign tasks.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@askalf/dario",
-  "version": "6.19.2",
+  "version": "6.20.1",
   "description": "Use your Claude and ChatGPT subscriptions in Cursor, Cline, Aider, Claude Code and the Agent SDK at subscription pricing, not per-token API bills. One local Anthropic + OpenAI-compatible endpoint: either plan answers either wire shape, opt-in failover when one hits its limit, and a ledger of what each plan absorbed.",
   "type": "module",
   "bin": {
@@ -91,7 +91,7 @@
     "node": ">=18.0.0"
   },
   "devDependencies": {
-    "@types/node": "^26.6.5",
+    "@types/node": "^26.6.6",
     "tsx": "^4.23.15",
     "typescript": "^5.7.0"
   }