Default to DeepSeek-V4.1-Flash (deepseek-flash), on dated rate cards

DESIGN.md

@@ -115,7 +115,9 @@ eval/            harness lives with the code it gates (see EVAL.md)
   usable input / 1M advertised / 384k output; effort levels
   off|low|high|xhigh|max; budget math always uses 616k.
 - Flash-first, pro-escalation: flash is the default executor; `--model`
-  pins. A session transcript is pinned to ONE model (Round 4 rule); the
+  pins. (Since V4.1 Flash, 2026-09-10, `deepseek-flash` is the default
+  and pro is no longer the stronger tier on DeepSeek's own agent
+  benchmarks; see README "Models". Escalation stays open as a mechanism.) A session transcript is pinned to ONE model (Round 4 rule); the
   `/pro` escalation mechanism — same-transcript switch vs
   evidence-packet handoff — is an open question, eval-gated at M4.
 - Usage accounting from provider `usage` only; cost computed from our own
@@ -201,8 +203,8 @@ Small and stable — the schema is part of the cached prefix:
 - `task` tool: spawn/wait/result/cancel lifecycle (DeepSeek-TUI shape),
   parallel by default, results return as compressed reports.
 - Roles as prompt+permission presets: explorer (read-only, flash),
-  implementer (write, flash), reviewer (read-only, pro), tester (bash,
-  flash). Verified pattern: bounded sub-contexts beat one giant context.
+  implementer (write, flash), reviewer (read-only, pro until 2026-09-10,
+  flash since), tester (bash, flash). Verified pattern: bounded sub-contexts beat one giant context.
 - Every sub-agent gets a budget envelope: max tokens, max turns, max
   wall-clock. A looping child dies quietly; the parent gets a partial
   report. (Capability-token idea from research, minimum viable form.)

EVAL.md

@@ -86,7 +86,7 @@ lower hit% while improving quality and total cost).
 
 ## Workflow
 
-- `bun eval/run.ts --tasks all --adapters dsc --models deepseek-v4-flash
+- `bun eval/run.ts --tasks all --adapters dsc --models deepseek-flash
   --n 3` -> appends runs.jsonl
 - `bun eval/report.ts` -> matrix like evot's README table.
 - CI (dsc repo): quick-tag tasks on every engine PR; full matrix

README.md

@@ -1,9 +1,8 @@
 # deepseek-code (`dsc`)
 
-A DeepSeek-native coding agent for the v4 flash/pro series (pro is GA
-as of 2026-08-12, model version DeepSeek-V4-Pro-0813 — same
-`deepseek-v4-pro` model ID). TypeScript, Bun, zero runtime
-dependencies.
+A DeepSeek-native coding agent. Defaults to DeepSeek-V4.1-Flash
+(`deepseek-flash`, released 2026-09-10); `deepseek-v4-pro` stays
+selectable. TypeScript, Bun, zero runtime dependencies.
 
 Not a Claude Code clone and not a general multi-provider harness. Being
 DeepSeek-only is the design: one protocol state machine (the
@@ -83,7 +82,7 @@ printf 'read src/parse.js\nnow fix the regex\n/cost\n' | bun src/cli.ts
 Useful flags:
 
 ```
---model deepseek-v4-flash|deepseek-v4-pro
+--model deepseek-flash|deepseek-v4-pro   (default deepseek-flash)
 --cwd DIR                 working directory for the run
 --continue                interactive: resume the latest session here
 --save / --resume ID      persist the session to SQLite (~/.dsc/)
@@ -94,13 +93,33 @@ Useful flags:
 --verbose                 -p only; stream progress to stderr
 ```
 
-Both v4 models are 1M context / 384K max output and can be picked
-anywhere a model is chosen (`--model`, `/model` in interactive,
-`dsc job add --model`). Flash is the default executor; pro is the
-stronger tier and the right pick for planning and review passes — the
-built-in `reviewer` sub-agent role already runs on it. V4-Pro GA ships
-with thinking on by default upstream, and reasoning tokens count
-against `max_tokens`, which dsc's budgets already account for.
+## Models
+
+| name | model | off-peak USD / 1M (hit / miss / out) |
+| --- | --- | --- |
+| `deepseek-flash` (default) | DeepSeek-V4.1-Flash | 0.003 / 0.15 / 0.60 |
+| `deepseek-v4-pro` | DeepSeek-V4-Pro-0813 | 0.022 / 0.66 / 1.98 |
+
+Peak (01:00-04:00 and 06:00-10:00 UTC, Monday to Friday) costs double;
+`/cost` and the turn footer price usage on the card and period in force
+when they print. Both are 1M context / 384K max output and can be
+picked anywhere a model is chosen (`--model`, `/model` in interactive,
+`dsc job add --model`).
+
+Flash is the default and every built-in sub-agent role runs on it,
+including `reviewer`, which ran pro until V4.1. The reason is
+DeepSeek's own changelog: on the benchmarks both its 2026-09-10 (V4.1
+Flash) and 2026-08-13 (V4 Pro GA) entries report, Flash scores
+Terminal Bench 2.1 90.6 vs 87.9, NL2Repo 65.4 vs 61.5, CyberGym 88.1 vs
+83.3 and HLE with tools 63.9 vs 60.0; HLE without tools goes the other
+way, 36.8 vs 42.7. We have not re-run our own eval on V4.1 yet.
+
+The retired names `deepseek-v4-flash` and `deepseek-v4-flash-vision-exp`
+are still accepted, as the API still accepts them: they are served by
+V4.1 Flash and billed at the Flash price. Thinking is on by default
+upstream, and reasoning tokens count against `max_tokens` (V4.1 Flash
+was measured spending a whole 600-token budget on reasoning), which
+dsc's budgets already account for.
 
 ## What is actually built
 
@@ -161,7 +180,7 @@ pinned benchmark suite. Numbers decide; taste does not. See
 the raw rows behind them.
 
 ```bash
-bun eval/run.ts --tasks all --adapters dsc --models deepseek-v4-flash --n 3
+bun eval/run.ts --tasks all --adapters dsc --models deepseek-flash --n 3
 bun eval/report.ts
 ```
 

eval/pricing.json

@@ -1,9 +1,36 @@
 {
-  "source": "thevibeworks/deepseek-docs content/en/quick_start/pricing.md (fetched 2026-08-02)",
-  "unit": "USD per 1M tokens, regular (off-peak) rates",
-  "note": "DeepSeek announced a coming peak-hour policy: 2x all items during 9:00-12:00 and 14:00-18:00 Beijing time. Not yet in effect; runs.jsonl rows record regular-rate cost only. Re-check the pricing doc before trusting absolute dollar figures.",
-  "models": {
-    "deepseek-v4-flash": { "input_hit": 0.0028, "input_miss": 0.14, "output": 0.28 },
-    "deepseek-v4-pro": { "input_hit": 0.003625, "input_miss": 0.435, "output": 0.87 }
-  }
+  "source": "thevibeworks/deepseek-docs content/en/quick_start/pricing.md (re-read 2026-09-18) and content/en/updates.md",
+  "unit": "USD per 1M tokens. Cards from 2026-08-16 on are OFF-PEAK rates; peak (01:00-04:00 and 06:00-10:00 UTC, Mon-Fri) is 2x. runs.jsonl cost_usd is always computed on the off-peak rate of the card in force when the row was written, so rows from different hours compare.",
+  "tiers": {
+    "deepseek-flash": "flash",
+    "deepseek-v4-pro": "pro",
+    "deepseek-v4-flash": "flash",
+    "deepseek-v4-flash-vision-exp": "flash"
+  },
+  "cards": [
+    {
+      "since": "1970-01-01T00:00:00Z",
+      "label": "flat card published 2026-08-02, no peak",
+      "rates": {
+        "flash": { "input_hit": 0.0028, "input_miss": 0.14, "output": 0.28 },
+        "pro": { "input_hit": 0.003625, "input_miss": 0.435, "output": 0.87 }
+      }
+    },
+    {
+      "since": "2026-08-16T16:00:00Z",
+      "label": "V4 GA repricing (changelog 2026-08-13)",
+      "rates": {
+        "flash": { "input_hit": 0.007, "input_miss": 0.22, "output": 0.66 },
+        "pro": { "input_hit": 0.022, "input_miss": 0.66, "output": 1.98 }
+      }
+    },
+    {
+      "since": "2026-09-10T11:00:00Z",
+      "label": "V4.1 Flash (changelog 2026-09-10); instant inferred, see src/provider/catalog.ts",
+      "rates": {
+        "flash": { "input_hit": 0.003, "input_miss": 0.15, "output": 0.6 },
+        "pro": { "input_hit": 0.022, "input_miss": 0.66, "output": 1.98 }
+      }
+    }
+  ]
 }

eval/run.ts

@@ -2,7 +2,7 @@
 // Matrix runner: (task, adapter, model) x N -> runs.jsonl.
 // Usage:
 //   bun run.ts --tasks bugfix-slugify,feature-tdd --adapters dsc \
-//     --models deepseek-v4-flash --n 3
+//     --models deepseek-flash --n 3
 // Design: deepseek-code/EVAL.md. One run = fresh fixture copy -> headless
 // invocation -> dumb bash verifier -> metrics row. Sequential on purpose:
 // DeepSeek's prefix cache is time-sensitive and parallel runs would blur
@@ -51,13 +51,17 @@ const taskNames =
         .sort()
     : taskArg.split(",");
 const adapterNames = arg("adapters", "dsc").split(",");
-const models = arg("models", "deepseek-v4-flash").split(",");
+const models = arg("models", "deepseek-flash").split(",");
 const n = parseInt(arg("n", "1"), 10);
 
 const pricing = await Bun.file(join(EVAL_DIR, "pricing.json")).json();
+/** Off-peak cost on the card in force now (see pricing.json "unit"). */
 function costUsd(model: string, u: Usage | null): number | null {
-  const p = pricing.models[model];
-  if (!p || !u) return null;
+  const tier = pricing.tiers[model];
+  if (!tier || !u) return null;
+  const now = Date.now();
+  const card = pricing.cards.filter((c: { since: string }) => Date.parse(c.since) <= now).at(-1);
+  const p = card.rates[tier];
   return (u.inputFresh * p.input_miss + u.cacheRead * p.input_hit + u.output * p.output) / 1e6;
 }
 

package.json

@@ -3,7 +3,7 @@
   "version": "0.7.0",
   "private": false,
   "type": "module",
-  "description": "DeepSeek-native coding agent for the v4 flash/pro series. TypeScript/Bun, zero runtime dependencies.",
+  "description": "DeepSeek-native coding agent, defaulting to DeepSeek-V4.1-Flash (deepseek-flash). TypeScript/Bun, zero runtime dependencies.",
   "scripts": {
     "test": "bun test tests",
     "eval": "bun eval/run.ts",

src/cli.ts

@@ -25,7 +25,7 @@
 // Key resolution: $DEEPSEEK_API_KEY, else ~/.dsc/key.
 
 import { resolve } from "node:path";
-import { DEFAULT_BASE_URL, DEFAULT_MODEL, MODELS } from "./provider/catalog";
+import { CURRENT_MODELS, DEFAULT_BASE_URL, DEFAULT_MODEL, MODELS } from "./provider/catalog";
 import { resolveApiKey } from "./config";
 import { schedulerCli } from "./scheduler/cli";
 import { addUsage } from "./provider/types";
@@ -57,7 +57,7 @@ const USAGE = `usage:
   dsc job|ps|serve             scheduled jobs (dsc job for details)
   dsc skills                   discovered SKILL.md skills and their sources
 
-  --model NAME                 ${Object.keys(MODELS).join(" | ")}
+  --model NAME                 ${CURRENT_MODELS.join(" | ")} (default ${DEFAULT_MODEL})
   --cwd DIR                    working directory for the run
   --continue                   interactive: resume the latest session here
   --resume ID                  resume a specific session
@@ -111,7 +111,7 @@ if (promptIdx >= 0 && prompt === undefined) {
 
 const model = argValue("model") ?? process.env.DSC_MODEL ?? DEFAULT_MODEL;
 if (!MODELS[model]) {
-  console.error(`dsc: unknown model "${model}" (known: ${Object.keys(MODELS).join(", ")})`);
+  console.error(`dsc: unknown model "${model}" (known: ${CURRENT_MODELS.join(", ")})`);
   process.exit(2);
 }
 const cwd = resolve(argValue("cwd") ?? process.cwd());

src/engine/compact.ts

@@ -146,6 +146,9 @@ export async function summarize(
     tools: [],
     messages: [...sanitizeForSummary(messages), ask],
     maxTokens: opts.maxTokens ?? SUMMARY_MAX_TOKENS,
+    // The cap bounds the summary; with thinking on, reasoning eats it and
+    // the call returns no text, silently degrading to the emergency summary.
+    noThinking: true,
     signal: opts.signal,
     maxAttempts: 2,
   });

src/engine/subagent.ts

@@ -7,8 +7,10 @@
 // Budget envelopes: max turns / max total tokens / max wall-clock. A
 // looping child dies quietly; the parent gets a partial report.
 // Children are model-pinned per role (Round 4 rule #1: one transcript,
-// one model) — a pro reviewer is a fresh pro context, never a mid-
-// transcript model switch.
+// one model) — a role on another model gets a fresh context, never a mid-
+// transcript model switch. Every role runs deepseek-flash since V4.1
+// Flash (2026-09-10); the reviewer ran pro before that. The benchmark
+// comparison behind the switch is in README.md, "Models".
 
 import type { Message, Usage } from "../provider/types";
 import { addUsage, zeroUsage } from "../provider/types";
@@ -44,29 +46,29 @@ const READ_ONLY_TOOLS = [readTool, bashTool];
 export const ROLES: Record<RoleName, Role> = {
   explorer: {
     name: "explorer",
-    description: "read-only investigation (fast model)",
+    description: "read-only investigation",
     tools: READ_ONLY_TOOLS,
-    model: "deepseek-v4-flash",
+    model: "deepseek-flash",
     preamble:
       "You are an explorer sub-agent: investigate the codebase to answer " +
       "the task below. Do not modify any files.",
     budget: { maxTurns: 12, maxTotalTokens: 150_000, maxWallMs: 120_000 },
   },
   implementer: {
     name: "implementer",
-    description: "makes a code change and verifies it (fast model)",
+    description: "makes a code change and verifies it",
     tools: [readTool, bashTool, editTool, writeTool],
-    model: "deepseek-v4-flash",
+    model: "deepseek-flash",
     preamble:
       "You are an implementer sub-agent: make the change described in the " +
       "task below, verify it, and report exactly what you changed.",
     budget: { maxTurns: 24, maxTotalTokens: 300_000, maxWallMs: 240_000 },
   },
   reviewer: {
     name: "reviewer",
-    description: "read-only code review (strong model)",
+    description: "read-only code review",
     tools: READ_ONLY_TOOLS,
-    model: "deepseek-v4-pro",
+    model: "deepseek-flash",
     preamble:
       "You are a reviewer sub-agent: read the code relevant to the task " +
       "below and report concrete problems (bugs, missed edge cases, broken " +
@@ -75,9 +77,9 @@ export const ROLES: Record<RoleName, Role> = {
   },
   tester: {
     name: "tester",
-    description: "runs tests/commands and reports results (fast model)",
+    description: "runs tests/commands and reports results",
     tools: READ_ONLY_TOOLS,
-    model: "deepseek-v4-flash",
+    model: "deepseek-flash",
     preamble:
       "You are a tester sub-agent: run the commands or tests named in the " +
       "task below and report the results. Do not edit source files.",