todo2code

Ticket 013 audit

Baseline

google/gemini-3.6-flash: PASS 6/6, 125,486 ms, 177,953 tokens, $0.412363, no fallback or degradation.

Candidate screening

Model Structured output Prompt / completion per 1M Context
google/gemini-3-flash-preview yes $0.50 / $3.00 1,048,576
mistralai/codestral-2508 yes $0.30 / $0.90 256,000
deepseek/deepseek-v4-pro yes $0.435 / $0.87 1,048,576

Live results

Model Result Time Tokens Cost Fallback
google/gemini-3.6-flash (fresh baseline) PASS 6/6 106,700 ms not recorded in comparison summary $0.342992 no
google/gemini-3-flash-preview PASS 6/6 64,064 ms 116,604 $0.076411 no
mistralai/codestral-2508 PASS 6/6 57,129 ms 118,920 $0.037994 no
deepseek/deepseek-v4-pro FAIL >900,000 ms no manifest unmeasured no result

Codestral was about 1.87× faster and 9.0× cheaper than the fresh Gemini 3.6 baseline. Gemini 3 Flash Preview was about 1.67× faster and 4.49× cheaper. DeepSeek was stopped at the declared run budget rather than allowed to hang.

Cross-repository result

The first real repository run exposed sequential Markdown batches. On weekly, Codestral enriched 161 records in six requests but needed 218,741 ms. Bounded concurrency of three preserved response/record audit order and reduced the same run to 53,362 ms (4.1× faster), with no degradation. The previously timeouting nlp2uri then completed 619 records in 20 requests in 194,750 ms, 176,797 tokens and $0.08588244. A large deterministic algitex scan completed 2,643 Markdown records and the full pipeline in 9.4 seconds.

Decision

Promote mistralai/codestral-2508 to the explicit default. Keep google/gemini-3-flash-preview as the first fallback/reference candidate. The selection is operational: contract adherence, latency and cost are measured; semantic quality still remains bounded by runtime validators and the offline gold suite.

The live runner now enforces its total budget by aborting provider requests; it also refuses to reuse a failed manifest older than the current attempt.