Cost accounting and cheaper code paths

Cost accounting and cheaper code paths in tobi-techy/MIMI-AGENT.

Draft

tobi-techywants to merge intofrom

5 commits34 files changed+1969-100

Draft. Cuts one recorded e2e run (43 agent turns) from ₦967 as main counted it to ₦706, measured offline at ₦0 on the recorded run. Details in COST.md.

	Agent cost	Per turn	Model calls
Before, as main counted it (cached input at the full rate)	₦967.48	₦22.50	117
Before, cached input at Meta's cached rate	₦769.13	₦17.89	117
After (defaults)	₦705.93	₦16.42	111

Of the ₦261.55 drop, ₦198.35 is accounting (cached input priced at $0.15 per 1M, which Meta already bills) and ₦63.20 is real spend cut by the plain-code paths.

What is in it

One metering seam: router, steer, binding, readers, SAM and speech to text booked to the daily naira cap and the run budget; prices in mimi/models/prices.py (MIMI_CACHED_RATE=0 restores the old accounting).
Plain-code paths (MIMI_PLAIN_CODE): balance, promo and quiet approvals on by default; delivery receipt and stock paths built but off until a live e2e re-record. Same tools, tiers, Go policy, approvals and claims gate; any surprise hands the step to the model.
prompt_cache_key per shared prefix (body only, recordings replay unchanged).
Shared-prefix context layout behind MIMI_CTX_SHARED_PREFIX=1, off (needs a live re-record).
Demo kill switch also refuses new and reset sessions.

Measure: META_API_KEY= python -m mimi.api.e2e --no-write (after) and the same with MIMI_PLAIN_CODE=0 (before). No live model calls, no deploy.

Activity

Demo kill switch also refuses new and reset sessions (503, the refused-turn body); reads keep working. Demo caps approved as set

68316d1Oct 29

One metering seam for every model call: router, steer, binding, readers, SAM and speech to text are booked to the daily naira cap and folded into the run budget; cached input priced at Meta's cached rate; prices in one place; e2e report counts steer and router calls

72fd0b8Oct 29

Cheaper code paths: a prompt_cache_key per shared prefix, a shared-prefix context layout behind MIMI_CTX_SHARED_PREFIX (off), and plain-code paths for deterministic turns (MIMI_PLAIN_CODE; balance, promo and quiet approvals on by default)

78ba6afOct 30

COST.md: cost per turn on the recorded e2e run (N22.50 counted at the full rate, N17.89 billed, N16.42 after), what was cut and what was only accounting, how to measure it again

bff5ed5Oct 30

coderabbitai[bot]1d ago

IMPORTANT

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:
reviews:
  auto_review:
    drafts: true

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

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

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

Worker-kill test pins MIMI_PLAIN_CODE=0: its scripted model must drive every step, and the default balance path answered the test's balance question in code

dda61cbOct 30

This pull request is a draft

Merging is unavailable until it is marked ready for review.