ROADMAP-12 — P1 — No runs.model column or cost attribution; cached-token detail ignored #104

Series: ROADMAP — filed from the 2026-10-27 whole-project assessment · Backlog task: M15-T05 · Priority: P1 · Effort: medium

Problem

Runs cannot be attributed by model or cost:

• runs has token totals but no model column, despite LLM_PROVIDER/LLM_MODEL being configurable per deployment and per client
• Cached-token accounting is ignored: OpenAI prompt_tokens_details and Gemini cached_token_count are not read (only prompt_tokens / candidates_token_count), so cost models cannot distinguish cached vs fresh input
• The unrestricted provider payload is persisted into raw_metadata with no field filtering or size cap

Token placement correctness is ROADMAP-06; this issue is the data model + attribution.

Evidence

• agentkit/openai extraction agentkit/llm/openai.py:199–200, metadata :202–207
• agentkit/llm/gemini.py:252–255 usage extraction, :270–277 metadata dump
• agentkit/db/models.py:38–47 — runs columns

Suggested fix / acceptance criteria

• runs.model persisted at run creation (and recorded in traces)
• Cached-token counts captured in TokenUsage and surfaces in metadata
• raw_metadata filtered to a bounded, documented field list
• Docs describe cost reading; (optional) aggregated /metrics-style sums arrive with ROADMAP-25

Suggested branch / commit

feat/run-model-cost · feat(db): record run model and cached-token/cost attribution