ai-tierforge

Route LLM calls to the right model. Track what each task actually costs. Catch when your "cheap" tier isn't saving you money.

You set up a cheap model for simple tasks and an expensive one for hard tasks. Sounds smart. But nobody tracks what happens when the cheap model fails 3 times and escalates. That "cheap" task just cost you 5 retries plus the expensive call.

Most tools track per-call cost. ai-tierforge tracks per-task cost — the retries, the escalations, the failed attempts — so you can see what your tier routing actually costs.

pip install ai-tierforge

Requires Python 3.11+. Only two runtime deps: pyyaml and requests.

Write a config:

# tiers.yaml
tiers:
  architect:
    model: gpt-5-nano
    max_tokens: 8000
    use_for: [spec, review]
    provider: openai-compatible
    endpoint: https://opencode.ai/zen/v1
    api_key_env: OPENCODE_API_KEY

  workhorse:
    model: deepseek-v4-flash
    max_tokens: 4000
    use_for: [code, chat]
    provider: openai-compatible
    endpoint: https://opencode.ai/zen/v1
    api_key_env: OPENCODE_API_KEY

  utility:
    model: deepseek-v4-flash
    max_tokens: 2000
    use_for: [tickets, summaries]
    provider: openai-compatible
    endpoint: https://opencode.ai/zen/v1
    api_key_env: OPENCODE_API_KEY

pricing:
  deepseek-v4-flash:
    input: 0.00000014
    output: 0.00000028
  gpt-5-nano:
    input: 0.00000005
    output: 0.0000004

budgets:
  per_task:
    limit: 0.10
    on_exceed: downgrade

Use it:

from ai_tierforge import TierRouter

router = TierRouter.from_yaml("tiers.yaml")

result = router.route(task_type="code", prompt="Write a unit test for auth")

print(result.tier)        # workhorse
print(result.model)       # deepseek-v4-flash
print(result.response)    # the model's output

# What did this task actually cost? (including retries + escalations)
report = router.cost_report()
print(report.cost_per_task("code"))

Or from the command line:

ai-tierforge validate tiers.yaml
ai-tierforge --config tiers.yaml route code "Write a unit test for auth"
ai-tierforge --config tiers.yaml report --type code
ai-tierforge --config tiers.yaml budget check --scope team:payments

1. Routes by task type — code goes to workhorse, spec goes to architect, tickets goes to utility. You define the mapping in YAML.

2. Tracks real cost per task — not just the successful call, but every retry, every escalation, every failed attempt. Uses Decimal so you don't get float rounding errors on micropayments.

3. Escalates on failure — if workhorse times out or hits a rate limit, the router tries architect automatically. The escalation trace is logged so you can see exactly what happened.

4. Enforces budgets — set a per-task, per-day, or per-project limit. When exceeded: warn, downgrade to a cheaper tier, or hard stop. You pick the action per scope.

5. Separates routing from failover in logs — "routed to DeepSeek because it's cheap" is a different event than "fell back to GPT because DeepSeek was down". You can grep for one without the other.

6. Works with any OpenAI-compatible endpoint — OpenCode Zen, OpenAI, DeepSeek, vLLM, LiteLLM, Portkey.

If 80% of your tasks escalate from the cheap tier to the expensive tier, you're paying for both on every task. That's worse than just using the expensive model from the start.

ai-tierforge tracks escalation rate per task type and per tier, and alerts when it crosses your threshold:

escalation:
  default_threshold: 0.30
  per_tier:
    workhorse: 0.25
    utility: 0.40
  max_retries: 3

Routed real GitHub issues, PRs, and code from FastAPI, Pydantic, httpx, and Rich through the tier system. All 8 test scenarios passed.

What happened:

• code tasks routed to DeepSeek (workhorse) at ~$0.0007/call
• spec tasks routed to gpt-5-nano (architect) at ~$0.001/call — it found a real version mismatch in fastapi PR #16018
• tickets and summaries routed to DeepSeek (utility) at ~$0.0001/call
• Workhorse timeout triggered automatic escalation to architect — call succeeded
• Budget exceeded → HARD_STOP blocked the next call
• Budget exceeded → DOWNGRADE switched to workhorse, saving 60% on that call
• CLI round-trip worked via YAML config with custom endpoint and pricing

Full reports:

• docs/pass3-real-data.md — final run, 8/8 pass, real FastAPI data
• docs/pass2-synthetic-data.md — synthetic post-fix, 6/8 pass, 47.6% savings
• docs/1st-test-pass-synthetic-data.md — first run, 5 bugs found

agent calls router.route("code", prompt)
  → matches "code" to workhorse tier
  → checks budget (allowed? blocked? downgrade?)
  → calls the provider adapter
  → records cost in the ledger
  → success? return the result
  → failure? retry or escalate to architect
  → all retries exhausted? raise RouterExhaustedError

tiers:
  architect:
    model: gpt-5-nano
    max_tokens: 8000
    use_for: [spec, review]
    provider: openai-compatible
    endpoint: https://api.example.com/v1   # optional
    api_key_env: MY_API_KEY                # optional

First tier in the YAML is highest priority (architect). Last is lowest (utility). This order determines escalation direction.

Per-token costs. If you skip this, the adapter falls back to built-in pricing for common models:

pricing:
  deepseek-v4-flash:
    input: 0.00000014
    output: 0.00000028

Three scopes, each independent:

budgets:
  per_task:
    limit: 0.10
    on_exceed: downgrade    # warn | downgrade | hard_stop
  per_day:
    limit: 5.00
    on_exceed: hard_stop
  per_project:
    limit: 50.00
    on_exceed: warn

logging:
  routing: true       # log routing decisions
  failover: true      # log failover events
  level: debug        # info or debug (debug includes prompt/response)
  output: stdout      # stdout or a file path

git clone https://github.com/deghosal-2026/ai-tierforge.git
cd ai-tierforge
pip install -e ".[dev]"

pytest                              # 133 tests
ruff check src/ tests/              # lint
mypy src/                           # type check

# Field tests with real data
OPENCODE_API_KEY=oc_zen_... python tests/field/run_field_test.py \
  --data-dir tests/field/realdata --count 2 --fresh

src/ai_tierforge/
  types.py             # dataclasses + enums
  exceptions.py        # 7 custom exceptions
  config.py            # YAML loader + validator
  router.py            # routing, retries, escalation
  cost.py              # per-task cost ledger
  slo.py               # escalation tracker + routing logger
  budget.py            # budget enforcer
  omlx.py              # local model adapter
  cli.py               # CLI
  adapters/
    base.py            # ProviderAdapter protocol
    openai_compat.py   # OpenAI-compatible adapter

tests/
  test_*.py            # unit tests
  field/               # field test runner + real data

ai-tierforge sits above your existing gateway. It doesn't replace LiteLLM or Portkey — it adds tier routing and task-level cost tracking on top.

• loopguard — detects stuck agent loops. loopguard decides when to escalate, ai-tierforge decides where and tracks the cost.
• Agent Spend Protocol (ASP) — pre-call budget enforcement. ASP asks "can this tenant afford this call?" ai-tierforge asks "was this task worth what it cost?"

MIT

• Issues: github.com/deghosal-2026/ai-tierforge/issues — bug reports and feature requests
• Security: SECURITY.md — report vulnerabilities privately
• Contributing: CONTRIBUTING.md — coding standards and PR process
• Changelog: CHANGELOG.md — release history