harn-vm 0.9.11

Async bytecode virtual machine for the Harn programming language
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31

# Adaptive rate/concurrency governor limits (Layer 1 of the LLM rate governor).
#
# `[provider_limits.<provider>]` rows configure the per-(provider, org_key)
# governor in `crate::llm::rate_governor`, consumed only when the
# `llm.rate_governor` flag (`HARN_LLM_RATE_GOVERNOR=1`) is on. Keeping limits in
# the catalog (not at call sites) makes them provider-decoupled and overridable
# by the same catalog/flags delivery path as every other capability fact.
#
# Fields (all optional; unset axes fall back to the conservative built-in
# default in rate_governor.rs):
#   max_concurrency  ceiling the AIMD limiter climbs toward on sustained success
#   min_concurrency  floor the AIMD limiter halves toward on a throttle signal
#   rpm              requests-per-minute token bucket (omit to disable)
#   tpm              tokens-per-minute token bucket (omit to disable)
#   adaptive         AIMD on/off; false pins the limit at max_concurrency
#   [provider_limits.<p>.backoff]  base_ms / max_ms / multiplier / jitter for the
#                    circuit-breaker OPEN window (exp backoff + full jitter,
#                    Retry-After always wins)
#
# Providers without an explicit row use the built-in default
# (max_concurrency=8, min=1, adaptive, 1s→60s x2 backoff with jitter).
[provider_limits.anthropic]
max_concurrency = 12
min_concurrency = 1
adaptive = true

[provider_limits.anthropic.backoff]
base_ms = 1000
max_ms = 60000
multiplier = 2.0
jitter = true