swarm: a soft context limit per agent, as a percent of the window, applied at turn end
This commit is contained in:
damocles
mara
parent
c56faf9dff
commit
1e03979836
3 changed files with 214 additions and 35 deletions
|
|
@ -83,7 +83,7 @@ Manage agents across the swarm
|
|||
* `mint-forge-token` — Check one agent's forge token, and mint it if it's missing or stale
|
||||
* `set-state` — Declare the state a hive converges an agent to: `up`, `offline`, `paused` or `destroyed`
|
||||
* `watch` — Follow an agent's terminal rows live: the newest retained rows, then each new one, one line per row, until Ctrl-C or the stream ends
|
||||
* `inference` — Show the model, effort, subagent models and subagent cap the swarm sets for an agent, or change the model, effort and subagent models
|
||||
* `inference` — Show the model, effort, subagent models, subagent cap and soft context limit the swarm sets for an agent, or change the model, effort and subagent models
|
||||
* `permissions` — Show or change the permissions the swarm grants an agent
|
||||
* `agents-md` — Show an agent's AGENTS.md, or edit it with a diff
|
||||
|
||||
|
|
@ -192,9 +192,9 @@ The swarm-level counterpart of `hivectl agent watch`, read from the swarm-contro
|
|||
|
||||
## `swarmctl agent inference`
|
||||
|
||||
Show the model, effort, subagent models and subagent cap the swarm sets for an agent, or change the model, effort and subagent models.
|
||||
Show the model, effort, subagent models, subagent cap and soft context limit the swarm sets for an agent, or change the model, effort and subagent models.
|
||||
|
||||
`get` prints the subagent cap (`max_subagents`); `set` has no flag for it and keeps a stored one. Only swarm-ui sets it.
|
||||
`get` prints the subagent cap (`max_subagents`) and the soft context limit (`soft_context_percent`); `set` has no flag for them and keeps stored ones. Only swarm-ui sets them.
|
||||
|
||||
A field the swarm sets wins over the agent's own value (its agent.nix or a `hyperhive-model`/`hyperhive-effort` file) from its next turn on; the agent reads its settings every ten seconds. A field the swarm doesn't set, or clears, leaves the agent on its own value.
|
||||
|
||||
|
|
|
|||
|
|
@ -1,8 +1,8 @@
|
|||
# Per-agent inference settings
|
||||
|
||||
The swarm can set the model and reasoning effort each agent's turns run on,
|
||||
plus the models it may start subagents on and how many subagents it may run
|
||||
at once. A field you set wins over the
|
||||
plus the models it may start subagents on, how many subagents it may run
|
||||
at once, and a soft limit on how full its context may get before it compacts. A field you set wins over the
|
||||
agent's own value; a field you leave unset keeps the agent on its own value,
|
||||
so an agent you set nothing for runs as it does today.
|
||||
|
||||
|
|
@ -25,6 +25,14 @@ swarmctl agent inference clear iris
|
|||
dialog; `set` has no flag for it and keeps the stored one. It's a whole
|
||||
number from 0 to 32, where 0 allows no subagents; unset means no cap
|
||||
([`subagent.md`](../agent-tools/subagent.md#how-many-subagents-may-run-at-once)).
|
||||
- You set the soft context limit, `soft_context_percent`, in the same
|
||||
dialog; `set` has no flag for it either. It is the percent of the model's
|
||||
context window at which the agent compacts at the end of a turn, in place of
|
||||
the harness's own 75%: 20 on a 1M window compacts once the context passes
|
||||
200k. A turn can still run past it, and it changes nothing else (the idle
|
||||
auto-reset, the context shown in the UI). It's a whole number from 5 to 100;
|
||||
unset keeps the harness's own watermark
|
||||
([`claude-invocation.md`](../turn-loop/claude-invocation.md)).
|
||||
- `get` shows each field, naming the unset ones (or says the swarm sets
|
||||
nothing for the agent), and then the model and effort the agent reports
|
||||
running on, each with where it comes from.
|
||||
|
|
@ -45,6 +53,7 @@ model: opus
|
|||
effort: (not set; the agent's own)
|
||||
subagent_models: (not set; the agent's own)
|
||||
max_subagents: (not set; no cap)
|
||||
soft_context_percent: (not set; the harness's own watermark)
|
||||
running model: opus (swarm)
|
||||
running effort: medium (local_file)
|
||||
```
|
||||
|
|
@ -93,7 +102,8 @@ which swarm-controller alone writes:
|
|||
"model": "opus",
|
||||
"effort": "high",
|
||||
"subagent_models": ["haiku", "sonnet"],
|
||||
"max_subagents": 3
|
||||
"max_subagents": 3,
|
||||
"soft_context_percent": 20
|
||||
}
|
||||
```
|
||||
|
||||
|
|
@ -106,7 +116,7 @@ warning in its log, until the next `PUT` replaces it.
|
|||
| method | path | body |
|
||||
| -------- | ------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ |
|
||||
| `GET` | `/api/agents/<agent>/inference` | —; answers `{"configured": <the stored value, or null>, "effective": <what the agent reported, or null>}` |
|
||||
| `PUT` | `/api/agents/<agent>/inference` | the whole value, replacing the stored one; the controller refuses unknown fields, an unknown effort, or a model name with whitespace or a leading `-`, or a `max_subagents` above 32 |
|
||||
| `PUT` | `/api/agents/<agent>/inference` | the whole value, replacing the stored one; the controller refuses unknown fields, an unknown effort, or a model name with whitespace or a leading `-`, a `max_subagents` above 32, or a `soft_context_percent` outside 5 to 100 |
|
||||
| `DELETE` | `/api/agents/<agent>/inference` | —; answers 204 |
|
||||
|
||||
An agent's own queue credential grants it
|
||||
|
|
@ -118,6 +128,10 @@ The harness (`hive-agent`'s `swarm_inference`) applies a set field to the
|
|||
in-memory model or effort without writing `hyperhive-model` or
|
||||
`hyperhive-effort`, which keep the agent's own value.
|
||||
|
||||
The same loop hands `soft_context_percent` to the Bus
|
||||
(`set_soft_context_percent`), which the compaction policy reads when it decides
|
||||
at turn end; a key with no limit, or no key, clears it on the next read.
|
||||
|
||||
The subagent daemon (`hive-subagent-mcp`'s `swarm_models`) reads
|
||||
`subagent_models` and `max_subagents` with the same direct get, under the
|
||||
same credential, on each `start` or `continue`.
|
||||
|
|
|
|||
|
|
@ -164,7 +164,10 @@ The **when** for hyperhive's own policy is a `hive_claude::CompactionPolicy` inj
|
|||
`turn::make_session` builds a `PercentPolicy` that compacts once the
|
||||
model-reported context fill reaches `HIVE_COMPACT_WATERMARK_PERCENT`
|
||||
(default **75%**), falling back to `harness_state::context_window_tokens(model)`
|
||||
for the window on turns the model didn't report one. `0` disables proactive
|
||||
for the window on turns the model didn't report one. The swarm can replace
|
||||
that percent per agent with its `soft_context_percent`
|
||||
([`inference.md`](../swarm/inference.md)); `turn::SoftLimitPolicy` reads it
|
||||
for each decision. `0` disables proactive
|
||||
compaction (the reactive path always applies). The proactive path is
|
||||
best-effort — a failed checkpoint or `/compact` never fails the turn that
|
||||
already succeeded. An ACP agent gets the same policy; how it compacts is in
|
||||
|
|
|
|||
Loading…
Reference in a new issue