deepseekcode

Two-Model Duet — Pro Validator

deepseekcode’s second headline feature. Flash drives the main loop; Pro adjudicates the dangerous moments. Pro never runs on every turn — that would halve the cache-hit rate and double cost for no quality gain. Pro is invoked surgically, at moments where its premium is justified.

When Pro fires

Two automatic triggers:

1. Destructive-tool gate. Before any of:
   • write_file / edit_file outside cwd, in .git/, in .env*, or matching the user’s secret_path_patterns
   • bash matching destructive patterns: rm, rm -rf, git push, git reset --hard, git checkout ., git clean -f, curl -X POST|PUT|DELETE|PATCH, kubectl delete, kubectl apply -f, terraform apply, terraform destroy, psql/mysql/sqlite3 with DROP/DELETE/TRUNCATE, npm/pnpm/yarn publish, docker push
   • Plus anything in [duet].extra_destructive_patterns
2. failure_retry_with_pro. When the main loop attempts the same fix twice and both fail (same tool name + arg-hash + non-zero exit), the next single attempt is routed to Pro internally. Per-attempt only; does not persist.

What you see

◆ pro check (1.2s · 800 tok): approved
◆ pro check (3.4s · 1.2k tok): blocked — would delete uncommitted changes in pkg/auth/

Distinct color from the flash voice (magenta vs cyan). Validator output is folded by r / R along with regular reasoning blocks.

When Pro blocks

You get a choice:

◆ pro check (3.4s): blocked — <reason>
  [o]nce override · [e]dit · [c]ancel

• o — proceed anyway. Override is logged so you can review later.
• e — edit your prompt and try again.
• c — cancel; the model receives a “user cancelled after pro validation block” tool result and adapts.

Failure modes

Pro failure	Behavior
10s timeout	Skip with ◆ pro validation skipped: timeout → fall through to standard permission prompt
Network error	Same as timeout
Malformed JSON	Treat as block with override choice
Rate limited	Skip with note → fall through to permission prompt

Pro never auto-approves a destructive op because the validator failed. You always get the final say.

Edge case: /models deepseek-v4-pro

If you’ve switched the main-loop model to Pro, the destructive-tool gate becomes a silent no-op (Pro can’t meaningfully validate itself). The standard permission prompt still gates destructive paths.

Disabling

• CLI: dsc --no-duet
• Config: [duet] enabled = false

When disabled, destructive operations go through the standard permission prompt without Pro intervention.

Cost

The Cost HUD splits into two lines when the validator has fired this session:

flash · 7 steps · cache 91% · ¥0.04
pro   · 2 calls · cache 73% · ¥0.02

Pro’s cache-hit rate is naturally lower (validator prompts are constructed fresh per call), but each call is bounded to recent transcript + the proposed tool args, so total cost typically stays well under what an always-on Pro main loop would burn.

This site is open source. Improve this page.