model/context: per-model ctx window overrides + expose window size in /api/state

This commit is contained in:
damocles 2026-05-20 15:20:07 +02:00 committed by Mara
parent 9064cd3c57
commit 770cbaccf9
3 changed files with 68 additions and 23 deletions

View file

@ -303,8 +303,30 @@ pub fn default_model() -> &'static str {
/// Overridable at runtime via `HIVE_CONTEXT_WINDOW_TOKENS` (useful for /// Overridable at runtime via `HIVE_CONTEXT_WINDOW_TOKENS` (useful for
/// future models or when the operator knows the exact limit). The env /// future models or when the operator knows the exact limit). The env
/// var takes precedence over the model-name heuristic. /// var takes precedence over the model-name heuristic.
///
/// Resolution order (first match wins):
/// 1. `HIVE_CONTEXT_WINDOW_TOKENS_<KEY>` — per-model override where KEY
/// (case-insensitive) is a substring of the active model name.
/// Set by `hyperhive.contextWindowTokens.<key>` in `agent.nix`.
/// 2. `HIVE_CONTEXT_WINDOW_TOKENS` — global override (any model).
/// 3. Auto-derive: haiku → 200 000, sonnet / opus → 1 000 000.
#[must_use] #[must_use]
pub fn context_window_tokens(model: &str) -> u64 { pub fn context_window_tokens(model: &str) -> u64 {
let m = model.to_ascii_lowercase();
// Per-model overrides: HIVE_CONTEXT_WINDOW_TOKENS_<KEY_UPPER> where
// KEY (lowercased) must be a non-empty substring of the model name.
for (key, val) in std::env::vars() {
if let Some(suffix) = key.strip_prefix("HIVE_CONTEXT_WINDOW_TOKENS_") {
if !suffix.is_empty() && m.contains(&suffix.to_ascii_lowercase()) {
if let Ok(v) = val.trim().parse::<u64>() {
if v > 0 {
return v;
}
}
}
}
}
// Global override.
if let Ok(s) = std::env::var("HIVE_CONTEXT_WINDOW_TOKENS") { if let Ok(s) = std::env::var("HIVE_CONTEXT_WINDOW_TOKENS") {
if let Ok(v) = s.trim().parse::<u64>() { if let Ok(v) = s.trim().parse::<u64>() {
if v > 0 { if v > 0 {
@ -312,7 +334,7 @@ pub fn context_window_tokens(model: &str) -> u64 {
} }
} }
} }
let m = model.to_ascii_lowercase(); // Auto-derive from model family.
if m.contains("sonnet") || m.contains("opus") { if m.contains("sonnet") || m.contains("opus") {
1_000_000 1_000_000
} else { } else {

View file

@ -357,6 +357,12 @@ struct StateSnapshot {
/// the operator can see what they just switched to (and what's /// the operator can see what they just switched to (and what's
/// in flight). Mutable at runtime via `POST /api/model`. /// in flight). Mutable at runtime via `POST /api/model`.
model: String, model: String,
/// Effective context-window token budget for the current model.
/// Derived from `events::context_window_tokens(&model)` — respects
/// per-model and global `HIVE_CONTEXT_WINDOW_TOKENS_*` overrides then
/// falls back to model-family heuristic. Consumers (e.g. dashboard
/// badge) use this to render the ctx-usage percentage.
context_window_tokens: u64,
/// Last-inference token usage from the most recent completed /// Last-inference token usage from the most recent completed
/// turn — represents the current context-window size at turn-end. /// turn — represents the current context-window size at turn-end.
/// `null` until the first turn finishes. /// `null` until the first turn finishes.
@ -451,6 +457,7 @@ async fn api_state(State(state): State<AppState>) -> axum::Json<StateSnapshot> {
let inbox = recent_inbox(&state.socket, state.flavor()).await; let inbox = recent_inbox(&state.socket, state.flavor()).await;
let (turn_state, turn_state_since) = state.bus.state_snapshot(); let (turn_state, turn_state_since) = state.bus.state_snapshot();
let model = state.bus.model(); let model = state.bus.model();
let context_window_tokens = crate::events::context_window_tokens(&model);
let ctx_usage = state.bus.last_ctx_usage(); let ctx_usage = state.bus.last_ctx_usage();
let cost_usage = state.bus.last_cost_usage(); let cost_usage = state.bus.last_cost_usage();
axum::Json(StateSnapshot { axum::Json(StateSnapshot {
@ -463,6 +470,7 @@ async fn api_state(State(state): State<AppState>) -> axum::Json<StateSnapshot> {
turn_state, turn_state,
turn_state_since, turn_state_since,
model, model,
context_window_tokens,
ctx_usage, ctx_usage,
cost_usage, cost_usage,
gui_enabled: state.gui_vnc_port.is_some(), gui_enabled: state.gui_vnc_port.is_some(),

View file

@ -37,20 +37,28 @@
}; };
options.hyperhive.contextWindowTokens = lib.mkOption { options.hyperhive.contextWindowTokens = lib.mkOption {
type = lib.types.int; type = lib.types.attrsOf lib.types.int;
default = 0; default = { };
example = 1000000; example = {
haiku = 150000;
sonnet = 900000;
};
description = '' description = ''
Context-window size in tokens for this agent's model. `0` (the Per-model context-window overrides. Each attribute name is a
default) means "auto-derive from the model name": haiku 200 000, model-family short name (e.g. `"haiku"`, `"sonnet"`) matched as a
sonnet / opus 1 000 000. Set an explicit value here when you are case-insensitive substring of the active model name at runtime, so
using a model the harness does not recognise, or when Anthropic `"sonnet"` matches `"claude-sonnet-4-5"` and any future variant.
changes the window for an existing model family. Empty map (the default) means auto-derive: haiku 200 000,
sonnet / opus 1 000 000.
Sets the `HIVE_CONTEXT_WINDOW_TOKENS` environment variable; the Each entry is rendered as
harness reads it at runtime and uses it to compute the default `HIVE_CONTEXT_WINDOW_TOKENS_<KEY_UPPER>` (e.g.
compaction and auto-reset watermarks (75% and 50% of the window `HIVE_CONTEXT_WINDOW_TOKENS_SONNET = "900000"`). The harness checks
respectively). these per-model vars first, then the global
`HIVE_CONTEXT_WINDOW_TOKENS`, then the model-family heuristic.
At runtime, the effective window drives compaction (75%) and
auto-reset (50%) watermarks, and is exposed via `/api/state` as
`context_window_tokens`.
''; '';
}; };
@ -249,13 +257,23 @@
# Model + context-window env vars consumed by the harness at boot. # Model + context-window env vars consumed by the harness at boot.
# HIVE_DEFAULT_MODEL seeds the initial model selection when no persisted # HIVE_DEFAULT_MODEL seeds the initial model selection when no persisted
# model choice exists in the state dir. HIVE_CONTEXT_WINDOW_TOKENS # model choice exists in the state dir.
# overrides the auto-derived window size (only set when the NixOS option # HIVE_CONTEXT_WINDOW_TOKENS_<KEY_UPPER> provides per-model overrides
# is non-zero so an unset env var lets the harness use its own heuristic). # (e.g. HIVE_CONTEXT_WINDOW_TOKENS_SONNET) from contextWindowTokens attrset.
environment.variables.HIVE_DEFAULT_MODEL = config.hyperhive.model; # SHELL must be set so claude's Bash tool finds a POSIX shell.
environment.variables = lib.mkIf (config.hyperhive.contextWindowTokens != 0) { environment.variables = lib.mkMerge (
HIVE_CONTEXT_WINDOW_TOKENS = toString config.hyperhive.contextWindowTokens; [
}; {
HIVE_DEFAULT_MODEL = config.hyperhive.model;
SHELL = "${pkgs.bashInteractive}/bin/bash";
}
]
++ lib.mapAttrsToList
(model: tokens: {
"HIVE_CONTEXT_WINDOW_TOKENS_${lib.toUpper model}" = toString tokens;
})
config.hyperhive.contextWindowTokens
);
boot.isNspawnContainer = true; boot.isNspawnContainer = true;
@ -356,9 +374,6 @@
}; };
}; };
# claude's Bash tool refuses to run without a POSIX shell + $SHELL set.
environment.variables.SHELL = "${pkgs.bashInteractive}/bin/bash";
system.stateVersion = "25.11"; system.stateVersion = "25.11";
}; };
} }