fix(commandcode): char-safe model canonicalize, MiniMax pricing, per-turn input

[junhoyeo]

Problem

Three confirmed bugs in crates/tokscale-core/src/sessions/commandcode.rs (PR #717):

• (a) PANIC (high): canonicalize_model byte-sliced base[base.len() - PROMO_SUFFIX.len()..] guarded only by a length check, not a char boundary. A non-ASCII model id from the untrusted ~/.commandcode/config.json could place that byte index mid-codepoint and panic the parse.
• (b) PRICING (low): Stripping the org prefix turned MiniMaxAI/MiniMax-M3-Free into bare MiniMax-M3, but the client's own command-code provider is not a pricing provider, so the resolver never reached the minimax/... pricing key and cost resolved to 0.
• (c) O(N^2) INPUT (low): Per-assistant-turn input was estimated from context_chars, a cumulative running sum of all prior messages that grows across the session, inflating reported input versus other clients.

Fix

• (a) Char-safe, case-insensitive -free strip via str::get(..), which returns None (instead of panicking) when the tail is not on a char boundary.
• (b) Recover the real provider from the gateway id with provider_identity::inferred_provider_from_model (e.g. minimax) and use it as the message provider_id, matching the claudecode convention. Falls back to command-code when no provider is inferred.
• (c) Estimate each turn's input from only the new context introduced since the previous response (user prompt + tool results), not the cumulative total. Summed over the session this charges each message's content exactly once.

Tests

• test_canonicalize_model_does_not_panic_on_non_ascii — non-ASCII/emoji model ids do not panic (fails before fix a).
• test_minimax_model_resolves_nonzero_pricing — a MiniMax session resolves non-zero pricing against a minimax/... key (fails before fix b).
• test_input_is_per_turn_delta_not_cumulative + rewritten pinning test test_commandcode_input_is_per_turn_delta — per-turn input does not accumulate (fails before fix c).
• Updated the module doc-comment to describe per-turn (not cumulative) input estimation.

cargo test -p tokscale-core (985 lib + integration tests) and cargo clippy -p tokscale-core --tests both pass clean.

Residual concern

provider_id for MiniMax sessions is now minimax rather than command-code (intentional, so pricing resolves and matches how claudecode attributes multi-provider models). Gateway ids that infer no known provider keep the prior command-code provider unchanged.

🤖 Generated with Claude Code

---

Summary by cubic

Fixes three issues in the Command Code session parser: safe model canonicalization to prevent panics, correct MiniMax pricing via provider inference, and per-turn input accounting to stop inflated usage. Improves stability, pricing accuracy, and consistent usage metrics.

• Bug Fixes
  • Model id: safely strip -free with char-boundary checks; non-ASCII ids no longer panic.
  • Pricing: infer the real provider from gateway ids (e.g., MiniMaxAI/MiniMax-M3-Free -> provider minimax, model MiniMax-M3), falling back to command-code, so pricing hits minimax/....
  • Input: estimate each assistant turn’s input from only new context (user prompt + tool results) and reset after the response; avoids O(N^2) growth and over-reporting.

^{Written for commit 2bdbd61. Summary will update on new commits.}

[vercel]

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

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
tokscale	Ignored		Jun 22, 2026 7:30am

[cubic-dev-ai]

No issues found across 1 file

_{Re-trigger cubic}