mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-27 17:58:07 +00:00
374 lines
15 KiB
Python
374 lines
15 KiB
Python
"""Prompt-size diagnostic: ``hermes prompt-size``.
|
|
|
|
Reports a byte/char breakdown of the system prompt the agent would build for
|
|
a fresh session — system prompt total, the ``<available_skills>`` index,
|
|
memory + user profile, and tool-schema JSON. Lets users see where their fixed
|
|
prompt budget goes (issue #34667) without parsing a saved session JSON by hand.
|
|
|
|
The diagnostic builds a real inspection agent (so the numbers match what
|
|
actually ships on the wire) but never makes a network call: it passes dummy
|
|
credentials so ``AIAgent.__init__`` takes the direct-construction path, then
|
|
calls ``build_system_prompt_parts`` / inspects ``agent.tools`` offline.
|
|
"""
|
|
|
|
from __future__ import annotations
|
|
|
|
import json
|
|
import re
|
|
from pathlib import Path
|
|
from typing import Any, Dict, List, Optional, Tuple
|
|
|
|
# The skills index is wrapped in this tag pair inside the stable tier.
|
|
_SKILLS_BLOCK_RE = re.compile(r"<available_skills>.*?</available_skills>", re.DOTALL)
|
|
|
|
# A rendered skill entry inside <available_skills> is `` - name: desc`` (or
|
|
# `` - name`` when the skill has no description). Category headers use two
|
|
# leading spaces, so the four-space + ``- `` prefix isolates skill lines.
|
|
_SKILL_LINE_PREFIX = " - "
|
|
|
|
# Posture-demoted categories render all visible skill names on one shared line.
|
|
_NAMES_ONLY_LINE_RE = re.compile(r"^ .+ \[names only\]: (?P<names>.+)$")
|
|
|
|
# Cap the human-readable "Skills by size" table; ``--json`` always has them all.
|
|
_SKILLS_TABLE_LIMIT = 20
|
|
|
|
|
|
def _bytes(s: str) -> int:
|
|
return len(s.encode("utf-8"))
|
|
|
|
|
|
def _tool_name(tool: Any) -> str:
|
|
"""Return the callable name of a tool schema (OpenAI ``function`` shape)."""
|
|
if not isinstance(tool, dict):
|
|
return ""
|
|
fn = tool.get("function")
|
|
if isinstance(fn, dict) and fn.get("name"):
|
|
return str(fn["name"])
|
|
return str(tool.get("name", ""))
|
|
|
|
|
|
def _build_inspection_agent(platform: str) -> Any:
|
|
"""Construct an offline AIAgent for prompt inspection.
|
|
|
|
Dummy ``api_key`` + ``base_url`` force the direct-construction path in
|
|
``run_agent.py`` (no provider auto-detection, no network). Toolsets and
|
|
platform come from the caller so the breakdown matches a real session.
|
|
"""
|
|
from run_agent import AIAgent
|
|
from hermes_cli.config import load_config
|
|
from hermes_cli.tools_config import _get_platform_tools
|
|
|
|
cfg = load_config()
|
|
model_cfg = cfg.get("model", {}) if isinstance(cfg.get("model"), dict) else {}
|
|
model = model_cfg.get("default") or model_cfg.get("model") or ""
|
|
|
|
# Resolve platform-specific toolsets the same way the gateway does.
|
|
enabled_toolsets = sorted(_get_platform_tools(cfg, platform))
|
|
agent_cfg = cfg.get("agent") or {}
|
|
disabled_toolsets = agent_cfg.get("disabled_toolsets") or None
|
|
|
|
return AIAgent(
|
|
model=model,
|
|
api_key="inspect-only",
|
|
base_url="https://openrouter.ai/api/v1",
|
|
quiet_mode=True,
|
|
save_trajectories=False,
|
|
platform=platform,
|
|
enabled_toolsets=enabled_toolsets,
|
|
disabled_toolsets=disabled_toolsets,
|
|
)
|
|
|
|
|
|
def _skill_md_paths_by_name() -> Dict[str, Path]:
|
|
"""Map each installed skill's name to its ``SKILL.md`` path on disk.
|
|
|
|
Keyed by both the frontmatter ``name`` (what the index renders) and the
|
|
skill directory name, so either resolves. Local skills win over external
|
|
dirs (``get_all_skills_dirs`` yields local first), matching the index's own
|
|
precedence. Used to attribute the real on-disk read cost per skill.
|
|
"""
|
|
from agent.skill_utils import (
|
|
get_all_skills_dirs,
|
|
iter_skill_index_files,
|
|
parse_frontmatter,
|
|
)
|
|
|
|
mapping: Dict[str, Path] = {}
|
|
for skills_dir in get_all_skills_dirs():
|
|
if not skills_dir.exists():
|
|
continue
|
|
for skill_file in iter_skill_index_files(skills_dir, "SKILL.md"):
|
|
frontmatter_name = skill_file.parent.name
|
|
try:
|
|
frontmatter, _ = parse_frontmatter(
|
|
skill_file.read_text(encoding="utf-8")
|
|
)
|
|
frontmatter_name = str(frontmatter.get("name") or frontmatter_name)
|
|
except Exception:
|
|
pass
|
|
# setdefault keeps the first (local) occurrence on name collisions.
|
|
mapping.setdefault(frontmatter_name, skill_file)
|
|
mapping.setdefault(skill_file.parent.name, skill_file)
|
|
return mapping
|
|
|
|
|
|
def _compute_skills_breakdown(skills_block: str) -> List[Dict[str, Any]]:
|
|
"""Per-skill byte breakdown parsed from the rendered ``<available_skills>``.
|
|
|
|
Two honest, distinct numbers per skill:
|
|
|
|
* ``index_line_bytes`` — the skill's attributed bytes in the always-on
|
|
index (the fixed per-call cost of *listing* the skill). For a compact
|
|
``[names only]`` line, each name keeps its own bytes and receives an
|
|
even share of the category prefix and separators. The attributed bytes
|
|
therefore sum exactly to the shared rendered line.
|
|
* ``skill_md_bytes`` — the on-disk size of the skill's ``SKILL.md`` (the
|
|
real token cost paid only when the model loads it via ``skill_view``).
|
|
``None`` when the name can't be mapped to a file (e.g. a plugin skill
|
|
whose source lives outside the scanned skill dirs).
|
|
|
|
Sorted largest-first by ``skill_md_bytes`` (the read cost that dominates
|
|
pruning decisions), tie-broken by name.
|
|
"""
|
|
name_to_path = _skill_md_paths_by_name()
|
|
entries: List[Dict[str, Any]] = []
|
|
|
|
def append_entry(
|
|
name: str,
|
|
*,
|
|
attributed_bytes: int,
|
|
total_bytes: int,
|
|
shared_bytes: int,
|
|
skill_count: int,
|
|
) -> None:
|
|
path = name_to_path.get(name)
|
|
md_bytes: Optional[int] = None
|
|
if path is not None:
|
|
try:
|
|
md_bytes = path.stat().st_size
|
|
except OSError:
|
|
md_bytes = None
|
|
entries.append({
|
|
"name": name,
|
|
"index_line_bytes": attributed_bytes,
|
|
"index_line_total_bytes": total_bytes,
|
|
"index_line_shared_bytes": shared_bytes,
|
|
"index_line_skill_count": skill_count,
|
|
"skill_md_bytes": md_bytes,
|
|
"path": str(path) if path is not None else "",
|
|
})
|
|
|
|
for line in skills_block.splitlines():
|
|
compact_match = _NAMES_ONLY_LINE_RE.match(line)
|
|
if compact_match is not None:
|
|
names = [
|
|
name.strip()
|
|
for name in compact_match.group("names").split(",")
|
|
if name.strip()
|
|
]
|
|
if not names:
|
|
continue
|
|
total_bytes = _bytes(line)
|
|
name_bytes = [_bytes(name) for name in names]
|
|
shared_total = total_bytes - sum(name_bytes)
|
|
shared_base, shared_remainder = divmod(shared_total, len(names))
|
|
for index, name in enumerate(names):
|
|
shared_bytes = shared_base + (1 if index < shared_remainder else 0)
|
|
append_entry(
|
|
name,
|
|
attributed_bytes=name_bytes[index] + shared_bytes,
|
|
total_bytes=total_bytes,
|
|
shared_bytes=shared_bytes,
|
|
skill_count=len(names),
|
|
)
|
|
continue
|
|
|
|
if not line.startswith(_SKILL_LINE_PREFIX):
|
|
continue
|
|
rest = line[len(_SKILL_LINE_PREFIX):]
|
|
# ``name: desc`` — the first ``": "`` separates name from description.
|
|
# Namespaced names (``codex:rescue``) have no space after their colon,
|
|
# so partitioning on ``": "`` keeps the full name intact.
|
|
name = rest.partition(": ")[0].strip()
|
|
if not name:
|
|
continue
|
|
line_bytes = _bytes(line)
|
|
append_entry(
|
|
name,
|
|
attributed_bytes=line_bytes,
|
|
total_bytes=line_bytes,
|
|
shared_bytes=0,
|
|
skill_count=1,
|
|
)
|
|
entries.sort(key=lambda e: (-(e["skill_md_bytes"] or 0), e["name"]))
|
|
return entries
|
|
|
|
|
|
def _compute_toolsets_breakdown(tools: List[Any]) -> List[Dict[str, Any]]:
|
|
"""Per-toolset schema-byte breakdown of the resolved tool list.
|
|
|
|
Each tool is attributed to its single canonical toolset from the registry,
|
|
so ``json_bytes`` sums are fully attributable: the grand total equals the
|
|
sum of the individual tool serializations (which is the array total from
|
|
``tools['json_bytes']`` minus JSON framing of ``2 * count`` bytes). Sorted
|
|
largest-first by ``json_bytes``, tie-broken by toolset name.
|
|
"""
|
|
from tools.registry import registry
|
|
|
|
tool_to_toolset = registry.get_tool_to_toolset_map()
|
|
groups: Dict[str, Dict[str, Any]] = {}
|
|
for tool in tools:
|
|
name = _tool_name(tool)
|
|
toolset = tool_to_toolset.get(name) or "(unknown)"
|
|
group = groups.setdefault(
|
|
toolset, {"toolset": toolset, "tool_count": 0, "json_bytes": 0}
|
|
)
|
|
group["tool_count"] += 1
|
|
group["json_bytes"] += _bytes(json.dumps(tool, ensure_ascii=False))
|
|
out = list(groups.values())
|
|
out.sort(key=lambda g: (-g["json_bytes"], g["toolset"]))
|
|
return out
|
|
|
|
|
|
def compute_prompt_breakdown(platform: str = "cli") -> Dict[str, Any]:
|
|
"""Return a dict of prompt-size measurements for a fresh session.
|
|
|
|
Keys: ``system_prompt`` (chars/bytes), ``skills_index``, ``memory``,
|
|
``user_profile``, ``tools`` (count + json bytes), ``sections`` (a list of
|
|
(label, chars, bytes) for the three prompt tiers), ``skills_breakdown``
|
|
(per-skill index-line + on-disk SKILL.md bytes, largest-first), and
|
|
``toolsets_breakdown`` (per-toolset tool count + schema json bytes,
|
|
largest-first). The last two answer "what should I disable to cut tokens?".
|
|
"""
|
|
from agent.system_prompt import build_system_prompt, build_system_prompt_parts
|
|
|
|
agent = _build_inspection_agent(platform)
|
|
|
|
parts = build_system_prompt_parts(agent)
|
|
full = build_system_prompt(agent)
|
|
|
|
stable = parts.get("stable", "")
|
|
context = parts.get("context", "")
|
|
volatile = parts.get("volatile", "")
|
|
|
|
# Skills index — the <available_skills> block (the largest single block
|
|
# when many skills are installed). Measured inside the stable tier.
|
|
skills_match = _SKILLS_BLOCK_RE.search(stable)
|
|
skills_index = skills_match.group(0) if skills_match else ""
|
|
|
|
# Memory + user profile live in the volatile tier. We re-derive their
|
|
# blocks directly from the memory store so the numbers are attributable
|
|
# even though they're joined into ``volatile``.
|
|
memory_block = ""
|
|
user_block = ""
|
|
store = getattr(agent, "_memory_store", None)
|
|
if store is not None:
|
|
try:
|
|
if getattr(agent, "_memory_enabled", True):
|
|
memory_block = store.format_for_system_prompt("memory") or ""
|
|
if getattr(agent, "_user_profile_enabled", True):
|
|
user_block = store.format_for_system_prompt("user") or ""
|
|
except Exception:
|
|
pass
|
|
|
|
# Tool-schema JSON — the other half of the fixed per-call payload.
|
|
tools = getattr(agent, "tools", None) or []
|
|
tools_json = json.dumps(tools, ensure_ascii=False)
|
|
|
|
sections: List[Tuple[str, int, int]] = [
|
|
("stable (identity/guidance/skills)", len(stable), _bytes(stable)),
|
|
("context (AGENTS.md/cwd files)", len(context), _bytes(context)),
|
|
("volatile (memory/profile/timestamp)", len(volatile), _bytes(volatile)),
|
|
]
|
|
|
|
return {
|
|
"platform": platform,
|
|
"model": getattr(agent, "model", "") or "",
|
|
"system_prompt": {"chars": len(full), "bytes": _bytes(full)},
|
|
"skills_index": {"chars": len(skills_index), "bytes": _bytes(skills_index)},
|
|
"memory": {"chars": len(memory_block), "bytes": _bytes(memory_block)},
|
|
"user_profile": {"chars": len(user_block), "bytes": _bytes(user_block)},
|
|
"tools": {"count": len(tools), "json_bytes": _bytes(tools_json)},
|
|
"sections": sections,
|
|
"skills_breakdown": _compute_skills_breakdown(skills_index),
|
|
"toolsets_breakdown": _compute_toolsets_breakdown(tools),
|
|
}
|
|
|
|
|
|
def _fmt_kb(n: int) -> str:
|
|
return f"{n / 1024:.1f} KB"
|
|
|
|
|
|
def render_breakdown(data: Dict[str, Any]) -> str:
|
|
"""Render the breakdown as plain text suitable for a terminal."""
|
|
lines: List[str] = []
|
|
sp = data["system_prompt"]
|
|
lines.append(f"Prompt-size breakdown (platform={data['platform']}, model={data['model'] or 'unset'})")
|
|
lines.append("")
|
|
lines.append(f" System prompt total : {sp['bytes']:>8,} B ({_fmt_kb(sp['bytes'])}, {sp['chars']:,} chars)")
|
|
lines.append("")
|
|
lines.append(" Major blocks:")
|
|
si = data["skills_index"]
|
|
mem = data["memory"]
|
|
up = data["user_profile"]
|
|
lines.append(f" skills index : {si['bytes']:>8,} B ({_fmt_kb(si['bytes'])})")
|
|
lines.append(f" memory : {mem['bytes']:>8,} B ({_fmt_kb(mem['bytes'])})")
|
|
lines.append(f" user profile : {up['bytes']:>8,} B ({_fmt_kb(up['bytes'])})")
|
|
lines.append("")
|
|
lines.append(" Prompt tiers:")
|
|
for label, chars, byts in data["sections"]:
|
|
lines.append(f" {label:<36}: {byts:>8,} B ({_fmt_kb(byts)})")
|
|
lines.append("")
|
|
tools = data["tools"]
|
|
lines.append(f" Tool schemas : {tools['json_bytes']:>8,} B ({_fmt_kb(tools['json_bytes'])}, {tools['count']} tools)")
|
|
|
|
# Per-toolset schema cost — which toolset's tools cost the most to ship.
|
|
toolsets = data.get("toolsets_breakdown") or []
|
|
if toolsets:
|
|
lines.append("")
|
|
lines.append(" Toolsets by size (tool-schema JSON, largest first):")
|
|
lines.append(f" {'toolset':<22} {'tools':>5} {'schema':>10}")
|
|
for ts in toolsets:
|
|
lines.append(
|
|
f" {ts['toolset']:<22} {ts['tool_count']:>5} "
|
|
f"{ts['json_bytes']:>8,} B ({_fmt_kb(ts['json_bytes'])})"
|
|
)
|
|
|
|
# Per-skill cost — index line (always shipped) vs SKILL.md (read on load).
|
|
skills = data.get("skills_breakdown") or []
|
|
if skills:
|
|
lines.append("")
|
|
lines.append(
|
|
" Skills by size (SKILL.md on-disk = read cost; index cost = "
|
|
"attributed always-on bytes, largest first):"
|
|
)
|
|
lines.append(f" {'skill':<28} {'SKILL.md':>10} {'index cost':>10}")
|
|
shown = skills[:_SKILLS_TABLE_LIMIT]
|
|
for sk in shown:
|
|
md = sk["skill_md_bytes"]
|
|
md_str = f"{md:>8,} B" if md is not None else f"{'n/a':>10}"
|
|
name = sk["name"]
|
|
if len(name) > 28:
|
|
name = name[:27] + "…"
|
|
lines.append(
|
|
f" {name:<28} {md_str} {sk['index_line_bytes']:>8,} B"
|
|
)
|
|
remaining = len(skills) - len(shown)
|
|
if remaining > 0:
|
|
lines.append(f" … and {remaining} more (use --json for the full list)")
|
|
return "\n".join(lines)
|
|
|
|
|
|
def cmd_prompt_size(args: Any) -> None:
|
|
"""Entry point for ``hermes prompt-size``."""
|
|
platform = getattr(args, "platform", "cli") or "cli"
|
|
as_json = getattr(args, "json", False)
|
|
try:
|
|
data = compute_prompt_breakdown(platform)
|
|
except Exception as e:
|
|
print(f"Could not compute prompt-size breakdown: {e}")
|
|
return
|
|
if as_json:
|
|
print(json.dumps(data, ensure_ascii=False, indent=2))
|
|
else:
|
|
print(render_breakdown(data))
|