mirrors/hermes-agent
1
0
Fork 0
mirror of https://github.com/NousResearch/hermes-agent.git synced 2026-07-21 16:18:55 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 8

fix(delegation): budget subagent summaries against parent context headroom

Browse source 
Batch delegation returned each subagent's full final_response verbatim
into the parent's context. A fan-out of N children could dump 60k+ tokens
at once, blowing the parent's context window and — on rate-limited
providers — triggering a compression/429 death spiral (429 misread as
context-too-large -> window step-down -> retry loop -> conversation dies).

Cap each summary against the parent's *remaining* context headroom split
across the batch (not a magic char count). When trimming, mirror the
web_extract convention: spill the full text to cache/delegation (mounted
into remote backends via credential_files._CACHE_DIRS) and return a
head+tail window (75/25, line-snapped) plus a footer with the exact
read_file offset to page the omitted middle. Both the subagent's opening
AND its closing (outcomes / files-changed / issues, which live at the end)
survive in-context, and nothing is lost — the parent can read_file the
full version on any backend.

delegation.max_summary_chars (default 24000) is a static ceiling layered
on top as belt-and-suspenders for models that ignore 'be concise'; 0
disables it. Child prompt tightened to lead with outcomes / bullets.

Co-authored-by: rc-int <rcint@klaith.com>
This commit is contained in:
teknium1 2026-06-30 01:30:47 -07:00 • committed by Teknium
parent 3b2bb30c5d
commit 35a0803a3b
5 changed files with 341 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

16
hermes_cli/config.py
View file

@ -2087,6 +2087,22 @@ DEFAULT_CONFIG = {
"inherit_mcp_toolsets": True,
"max_iterations": 50, # per-subagent iteration cap (each subagent gets its own budget,
# independent of the parent's max_iterations)
# Subagent summaries return to the parent's context verbatim. A batch
# fan-out (N children) returns N summaries at once, which can exceed
# the parent's context window and trigger a compression/429 death
# spiral. delegate_task sizes each summary against the parent's
# remaining context headroom (split across the batch); when it must
# trim, the full text is spilled to ~/.hermes/cache/delegation/
# (mounted into remote backends) and the in-context summary becomes a
# head+tail window plus a footer with the exact read_file offset to
# page the omitted middle — the same convention web_extract uses for
# large pages. Nothing is lost. max_summary_chars is a hard per-summary
# character ceiling layered on top of that dynamic budget
# (belt-and-suspenders for models that ignore the "be concise"
# instruction). 0 disables the hard ceiling; the dynamic headroom
# budget still applies.
"max_summary_chars": 24000,
"child_timeout_seconds": 0, # optional wall-clock cap per child agent. 0 (default)
# = no timeout: children fail only from real errors
# (API, tools, iteration budget), never a delegation