post-execution tokens/api usage estimation.

[hackmajoris]

Could be that this idea is unfeasible, but it's worth to ask your opinion. Do you think such feature is feasible/makes sense?

The business guys are very hot on this

[pySilver]

I guess there is no problem to accumulate token counters across responses. But for what? Tasks are very different and you can't say that "ok on average development with ralphex costs me N". I think business guys needs to change the way how they think instead.

[hackmajoris]

I agree with your point. The issue they face is not specific to Ralphex. It’s a general problem that arises when a client requests a project implementation and wants to estimate the cost of AI usage.

[umputun]

short answer: parts of it are feasible, but I'm not adding it.

this was already proposed and implemented - #253 with PR #254, 928 lines across 13 files. I closed both, and closed the narrower opt-in usage-hook variant proposed in the same thread. Nothing here changes what that turned on: who needs the number and what decision it drives.

on feasibility, so the question gets a straight answer. Claude's stream-json result event does carry usage and a cost field, and codex writes typed token_count records into its session rollout that ralphex already tails for streaming output. Parsing is not the blocker. These two are:

• every alternative-provider wrapper ends its stream with an empty result and forwards no usage: codex-as-claude.sh:105, gemini-as-claude.sh:115, agy-as-claude.sh:125, pi-as-claude.sh:227, opencode-as-claude.sh:202, copilot-as-claude.sh:736. Custom external review scripts are plain text by design. A run under any of those prints zero, and a zero you can't tell apart from a cheap run is worse than no number at all
• closer to what you actually asked for: on a subscription plan there is no dollar figure to report. Codex reports plan_type: pro and used_percent of a rolling weekly window. That's a percentage, not money. Max/Pro subscribers, who are most of the reason --codex exists, spend no marginal dollars per run. The number your business guys want only exists on per-token API billing

what's available today: claude /usage, the Anthropic console, codex's own tracking, and ccusage if you want to read claude's transcripts directly. For per-task granularity, customize ~/.config/ralphex/prompts/task.txt to have the agent report its own session usage at the end of each task - each task runs in a fresh session, so that lines up.

and pySilver's point stands, as you agreed: this isn't a ralphex-shaped problem.