llm-cost-watchdog（LLM cost watchdog 搜索）

Cost Watchdog 💰

Real-time cost tracking layer for LLM-based agents. Prices every call live, detects runaway loops in code, enforces budget ceilings mid-execution.

1. Identity

Observes LLM spend without disturbing the agent. Prevents $2,400-overnight-loop disasters by making cost a first-class concern: priced at write time, budgeted at check time, surfaced in reports.

2. Triggers

Activate when:

• User mentions cost, budget, tokens, billing.
• Code contains LLM API calls (Anthropic, OpenAI, OpenRouter, Google, Groq, ...).
• Agent loops or recursive workflows.
• Batch / streaming processing with unclear bounds.
• /cost-watchdog [command] is invoked.

3. Commands

Run via python3 scripts/cost_watchdog.py <cmd> (or hook into your own CLI).

4. Pricing layer

Source chain

openrouter/*            → OpenRouter API (live)           → static fallback
anything else           → LiteLLM JSON (live, cached 24h) → OpenRouter (permissive) → static fallback

• 2600+ models indexed across chat, completion, embedding, image, audio,

video, rerank, OCR, search modes.

• 30+ providers in the static fallback: Anthropic, OpenAI, Google,

Groq, Mistral, Cohere, DeepSeek, Perplexity, xAI, Bedrock, Azure, and more.

• Unit-aware: token, image, second, query, page, character,

pixel. Alternatives never compare across units.

• Circuit breaker opens after 3 consecutive network failures for a host;

falls through to cache/static until the cool-down ends (60s).

Tuning

Refresh static pricing

python3 scripts/refresh_pricing.py

Regenerates references/pricing.md from the live sources so the offline fallback is fresh. Aborts if fewer than 100 rows came back (protects against clobbering on a network outage).

5. Tracking layer — how we know what was spent

Four independent paths, all write to ~/.cost-watchdog/usage.jsonl:

Usage log rotates daily: usage.YYYY-MM-DD.jsonl. session_total(since=...) skips files outside the window before scanning.

Aggregation uses canonical_family() so claude-haiku-4-5-20251001, claude-haiku-4-5, and claude-haiku-4.5 are one row in reports.

6. Budget enforcement

Two mechanisms:

1. Write-time check (race-safe): append_usage(entry, budget_ceiling=X)

takes an fcntl.flock on a sidecar, sums the current session, and refuses the write (raises BudgetExceeded) if the call would cross X.

1. Post-write check: wrappers compare cumulative spend to CW_BUDGET_USD

after logging and raise if over. Used when the wrapper doesn't know the ceiling at call time.

Either path stops the agent mid-loop; the LLM call still returns to the caller, but the next one blocks.

7. Code audit (AST)

python3 scripts/cost_watchdog.py audit path/to/agent.py

Walks the AST and reports:

• CRITICAL — while True with an LLM call and no max_iterations-style bound.
• CRITICAL — function that recurses and calls an LLM API with no depth argument.
• HIGH — plain while that calls an API with no retry/iteration counter.
• MEDIUM — LLM call missing max_tokens / max_completion_tokens.
• MEDIUM — function with ≥5 sequential LLM calls (batching candidate).

Every finding has a file line number. No more count('def ') > 3 and count('self.') > 5 → "recursion detected" false positives.

8. Detection — "what model is the agent using?"

python3 scripts/cost_watchdog.py detect

Five probe layers, ranked by confidence:

Emits a table or --json.

9. Files

10. Quality checklist

• [x] Live pricing from LiteLLM + OpenRouter, 24h-cached, with static fallback.
• [x] Exact-match model lookup (no substring conflation).
• [x] Multi-modal (token / image / second / query / page / character).
• [x] Unit-aware alternatives (never compares tokens to images).
• [x] AST-based code audit with line numbers.
• [x] Provider-aware tokenization (no more tiktoken-for-Claude).
• [x] Variance-based confidence (no += 0.05 theater).
• [x] Atomic writes to all shared state files.
• [x] fcntl.flock-guarded budget check-and-log (no race).
• [x] Circuit breaker on flaky networks (no 5s hang per call).
• [x] Streaming capture via SDK wrappers; streams flagged in errors.jsonl via HTTP capture.
• [x] Daily log rotation + date-scoped aggregation.
• [x] Canonical model families (variants collapse in reports).
• [x] errors.jsonl surfaces silent failures; cost_watchdog errors shows them.
• [x] Cassette tests for LiteLLM + OpenRouter parse paths (schema-drift safety net).
• [x] 73 logic tests passing.

11. Known limits (be honest)

• Tokenizer heuristics for Claude/Gemini/etc. are calibrated from docs,

not measured. Run cost_watchdog validate-tokens <model> to check drift against the provider's authoritative count when you have an API key.

• install_global_capture() can't see streaming responses — httpx exposes

an empty body until the user reads the stream. Use track_openai / track_anthropic for stream coverage; http_capture logs skipped streams to errors.jsonl so the gap is visible.

• Non-httpx SDKs (older Cohere, boto3 with custom transport) need the

per-SDK wrappers — HTTP capture won't see them.

• LiteLLM community data can lag 24-48h on brand-new models. OpenRouter's

API is truly live for anything it routes.

12. Testing

python3 -m unittest tests.test_cost_watchdog     # 73 tests
python3 scripts/code_audit.py test_risky_code.py # sample risks
python3 scripts/cost_watchdog.py report          # current spend summary