[Programcode]

Cost fix 2 of 7 from the cost analysis.

What changes

When a request does web research, the server makes two small model calls before the answer: one decides whether an Ask question needs a web search, and one writes follow-up search queries. Until now both ran on the model the user picked. A Pro user on Claude Opus 5.5 at high effort paid Opus prices for two short JSON replies on every researched request.

Both calls now run on Claude Haiku 5.5 at low effort. The model is one value, BACKGROUND_MODEL in supabase/functions/_shared/model-catalog.ts, with its effort next to it. The answer the user sees still comes from their chosen model. The helpers' token usage is still added to the parent request's ai_requests row, as before.

Affected features: Generate, Attack, Brainstorm, Factoids (follow-up queries) and Ask with web search or citations (search decision plus follow-up queries). Crossfire, Sparring and streaming Ask make no research calls and are unchanged. Jev routing and ranking calls already use their own fixed model.

Things to know

• The usage row still carries the user's model in its model column while its token totals now mix Haiku and the user's model. The dollar cost comes from OpenRouter per call, so the total charge stays correct. Splitting it per model belongs to the cost-logging fix.
• Haiku writes the follow-up queries, so search quality for Pro users could shift slightly. These are 2 or 3 short search strings, so I don't expect a visible difference.

Verification

• npm run verify passed locally at this head: docs check, lint, typecheck, unit tests, 356 Deno edge tests, build, service worker test, and 30 desktop Chromium smoke tests. GitHub Actions is broken repo-wide, so this is a local pass only, not a CI pass.
• Extended the existing Ask handler test to assert that the helper calls use BACKGROUND_MODEL and the answer uses the requested model. It fails against the old code and passes with this change.
• Real-app check with verify-debatelab (fake provider): on an Opus account, every helper call in Generate, Attack, Brainstorm, Factoids and Ask (web search and citation mode, desktop and mobile) asked for Haiku 5.5, and every answer asked for Opus 5.5. Details in the PR comment.
• Codex review of bd692f7: no findings.

Deploying needs the ai Edge Function redeployed; this merge alone doesn't change production behaviour.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
debate-navigator	Ready	Preview	Oct 22, 2026 11:00am UTC
debatelab-staging	Ready	Preview	Oct 22, 2026 11:00am UTC

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

ℹ️ Recent review info

⚙️ Run configuration

• Source mode: Auto
• Repository: Programcode/DebateLab
• Branch: cost/haiku-background-steps
• File: .coderabbit.yaml
• Commit: bd692f7
• Layers: Repository YAML (base), Organization UI (inherited)
• Review profile: CHILL
• Plan: Free
• Run ID: b799a31b-7d2e-4ef6-af18-47decab79c9b

📥 Commits

Reviewing files that changed from the base of the PR and between 4f9e0a0 and bd692f7.

📒 Files selected for processing: 4

• docs/architecture.md
• supabase/functions/_shared/model-catalog.ts
• supabase/functions/ai/handler-research.test.ts
• supabase/functions/ai/index.ts

Included review availability: This review used your included allowance. Your plan provides up to 1 included review per hour; 0 remain after this review.

📝 Walkthrough

Research helper calls now use a fixed background model and reasoning effort. Answer generation continues to use the user-selected model. The Ask test checks the model used for each call, and the architecture documentation describes the change.

Changes: Background research model

Layer / File(s)	Summary
Background model constants supabase/functions/_shared/model-catalog.ts	Adds exported constants for the background model, anthropic/claude-haiku-5.5, and its "low" reasoning effort.
Research call wiring and validation supabase/functions/ai/index.ts, supabase/functions/ai/handler-research.test.ts, docs/architecture.md	Research calls use the fixed background settings. The Ask test checks that two helper calls use the background model and the answer call uses the requested model. The architecture documentation describes the model assignments and usage accounting.

Estimated code review effort: 2 (Simple) | ~10 minutes

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Note

🎁 Summarized by CodeRabbit Free

Your organization is on the Free plan. CodeRabbit will generate a high-level summary and a walkthrough for each pull request. For a comprehensive line-by-line review, please upgrade your subscription to CodeRabbit Essentials by visiting https://app.coderabbit.ai/settings/billing.

_{Comment @coderabbitai help to get the list of available commands.}

[Programcode]

Real-app check with verify-debatelab (fake provider, local Supabase) at bd692f7. The demo user is on Claude Opus 5.5. The fake provider records the model each call asked for:

Recipe	Helper calls	Answer
attack-run	research-queries → Haiku 5.5	Opus 5.5
brainstorm-run	research-queries → Haiku 5.5	Opus 5.5
factoids-run	research-queries → Haiku 5.5	Opus 5.5
gen-research	research-queries → Haiku 5.5	Opus 5.5
ask-web (desktop, mobile)	research-decision + research-queries → Haiku 5.5	Opus 5.5
ask-cited (desktop, mobile)	research-decision + research-queries → Haiku 5.5	Opus 5.5
ask-stream	none (streaming Ask does no research)	Opus 5.5

All passed. ask-cited failed once on desktop during the first run: the click on the Options button timed out before any request went out. It passed when rerun alone on desktop and on mobile. The fake mode doesn't prove OpenRouter accepts Haiku with these parameters live. Haiku is already the Basic default and goes through the same callOpenRouter path, though.

Codex review (gpt-6.1-sol, high) of bd692f7: no findings.