File size: 5,480 Bytes
22decf5 | 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 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 | """Readable narratives: what a run actually did, in order, on one page.
The numeric features say how much; this says what. Each narrative keeps only
the load-bearing moments -- the plan it wrote down, every training and
evaluation job, every error and the move that followed it, and the summary it
signed off with -- so a run that took ten hours and 180 tool calls can be read
in a minute and compared against another.
"""
from __future__ import annotations
import json
import os
import re
import textwrap
from .parse import Run, ToolCall
from .taxonomy import classify_action
MILESTONE_ACTIONS = {"train", "generate", "eval_official", "eval_custom", "data_prep"}
SCORE_RES = [
re.compile(r"(?<![\w_])accuracy\s{2,}([01]?\.\d+)"),
re.compile(r"(?<![\w_])[\"']accuracy[\"']\s*:\s*([01]?\.\d+)"),
]
def _hhmm(t: float) -> str:
return f"{int(t // 3600):d}:{int(t % 3600) // 60:02d}"
def _one_line(s: str, n: int = 150) -> str:
s = re.sub(r"\s+", " ", (s or "").strip())
return s[:n] + ("…" if len(s) > n else "")
def _scores_in(text: str) -> list[str]:
return [m for pat in SCORE_RES for m in pat.findall(text or "")]
def _error_kind(tc: ToolCall) -> str | None:
"""Name the failure in a tool result, if there is one.
`is_error` alone is too narrow: a training script that dies with a
traceback still exits through a normal tool result, and that is exactly
the kind of failure we want to watch the agent recover from.
"""
t = tc.result_text or ""
if "CUDA out of memory" in t or "OutOfMemoryError" in t:
return "CUDA OOM"
if re.search(r"^Traceback \(most recent call last\)", t, re.M):
last = [ln for ln in t.strip().splitlines() if ln.strip()][-1:]
return f"exception: {_one_line(last[0], 90)}" if last else "exception"
if "ModuleNotFoundError" in t or "ImportError" in t:
return "missing package"
if "EAI_AGAIN" in t or "Temporary failure in name resolution" in t:
return "network blocked (offline sandbox)"
if "No space left on device" in t:
return "disk full"
if tc.is_error:
return f"tool error: {_one_line(t, 90)}"
return None
def narrative(run: Run) -> str:
out: list[str] = []
out.append(f"# {run.run_id}")
out.append("")
out.append(f"- task: **{run.task}**, base model: **{run.base_model}**")
out.append(f"- harness accuracy: **{run.accuracy}**")
out.append(f"- wall clock: {_hhmm(run.wall_s)}, tool calls: {len(run.tool_calls)}")
if run.result:
out.append(
f"- terminal: {run.result.get('terminal_reason')} "
f"(api_error={run.result.get('api_error_status')}), "
f"cost ${run.result.get('total_cost_usd', 0):.2f}"
)
out.append("")
# The plan, as the agent wrote it down.
todos = [tc for tc in run.tool_calls if tc.name in ("TaskCreate", "TaskUpdate")]
if todos:
out.append("## Plan as first written")
seen = 0
for tc in todos:
desc = tc.inp.get("description") or tc.inp.get("prompt") or json.dumps(tc.inp)[:200]
out.append(f"- {_one_line(desc, 120)}")
seen += 1
if seen >= 12:
break
out.append("")
first_text = next(
(
c.get("text", "")
for ev in run.events
if ev.kind == "assistant"
for c in (ev.raw.get("message", {}).get("content") or [])
if isinstance(c, dict) and c.get("type") == "text" and len(c.get("text", "")) > 80
),
"",
)
if first_text:
out.append("## First stated intent")
out.append(textwrap.indent(_one_line(first_text, 700), "> "))
out.append("")
out.append("## Milestones")
out.append("")
out.append("| t | action | what | scores seen |")
out.append("|---|---|---|---|")
for tc in run.tool_calls:
act = classify_action(tc)
if act not in MILESTONE_ACTIONS:
continue
payload = tc.command or tc.inp.get("file_path", "")
scores = _scores_in(tc.result_text)
out.append(
f"| {_hhmm(tc.t_rel)} | {act} | `{_one_line(payload, 110)}` | "
f"{', '.join(scores[:4]) if scores else ''} |"
)
out.append("")
out.append("## Failures and what happened next")
out.append("")
n = 0
for i, tc in enumerate(run.tool_calls):
kind = _error_kind(tc)
if not kind:
continue
n += 1
nxt = run.tool_calls[i + 1] if i + 1 < len(run.tool_calls) else None
follow = _one_line(nxt.text_payload, 140) if nxt else "(end of run)"
out.append(f"- **{_hhmm(tc.t_rel)} {kind}**")
out.append(f" - was running: `{_one_line(tc.text_payload, 120)}`")
out.append(f" - next move: `{follow}`")
if n >= 25:
out.append(f"- … and more (showing first 25)")
break
if n == 0:
out.append("_none detected_")
out.append("")
if run.final_message:
out.append("## Final summary the agent wrote")
out.append("")
out.append(run.final_message)
out.append("")
return "\n".join(out)
def write_all(runs: list[Run], out_dir: str) -> None:
os.makedirs(out_dir, exist_ok=True)
for r in runs:
with open(os.path.join(out_dir, f"{r.run_id}.md"), "w") as fh:
fh.write(narrative(r))
print(f" wrote {len(runs)} narratives to {out_dir}")
|