hermes-agent/tools/delegation_live_log.py
Frowtek 183712ab82 fix(delegation): redact credentials in live subagent transcripts
The live transcripts added with delegate_task write each child's events to
<hermes_home>/cache/delegation/live/<delegation_id>/. Nothing on that path is
redacted, and the rendered events are precisely the secret-bearing surfaces:
tool args, tool results and streamed assistant text.

That location is not incidental — delegate_tool.py:1620 documents cache/
delegation as "mounted read-only into remote backends", so every line lands
in a file readable from inside the sandbox.

Observed on the writer today:

    tool   | -> terminal(curl -H "Authorization: Bearer sk-ant-api03-...")
    result | terminal ok 0.4s: OPENAI_API_KEY=sk-proj-... AWS_SECRET_ACCESS_KEY=wJalr...

The same three values through the canonical redactor:

    curl -H "Authorization: Bearer ***" https://api.internal
    OPENAI_API_KEY=*** AWS_SECRET_ACCESS_KEY=***

Every other sink for this data already routes through that redactor — search
results via redact_sensitive_text(file_read=True), terminal output via
redact_terminal_output — so the transcript was the one place an operator's
keys reached disk in the clear.

Redact at three points, covering every artefact the dispatch writes into that
directory:

  * event() — every typed helper (assistant_text, thinking, tool_start,
    tool_result, marker, finalize, the stream flush) funnels through it, so
    one call covers them all and a helper added later cannot bypass it.
  * the .log header — written directly rather than through event(), and a
    caller can paste a key into the task text.
  * manifest.json — _write_manifest serialises the same goal, and the manifest
    sits in the same mounted directory, so redacting only the header would
    have left the credential exposed one file over.

force=True because this is a safety boundary and must redact even with the
global toggle off. On the (impossible-in-practice) import failure the line is
withheld instead of written raw: losing a debug line costs less than writing
a live credential into a sandbox-readable file.

Key names, tool names, statuses, durations and ordinary prose are untouched,
so tail -f stays as useful as before — pinned by its own test, alongside a
whole-directory sweep asserting no file under live/<id>/ carries the raw key.
2026-07-20 06:50:31 -07:00

424 lines
18 KiB
Python

"""Live, tail-able transcripts for delegated subagents.
Every ``delegate_task`` dispatch creates one append-only, human-readable log
per child under::
<hermes_home>/cache/delegation/live/<delegation_id>/task-<n>.log
The files are pre-created with a header at dispatch time (so ``tail -f``
attaches immediately) and then stream one line per child event: assistant
text, thinking, tool calls, tool results, and lifecycle markers. The paths
are returned from ``delegate_task`` so the parent agent (or the user) can
watch a child work instead of waiting blind for the consolidated summary.
Placement under ``cache/delegation`` is deliberate: that directory is
mounted read-only into remote terminal backends (Docker/Modal/SSH) via
``credential_files._CACHE_DIRS``, so the logs are readable from any backend.
Design constraints:
* **Never raise into the agent loop.** Every write is wrapped; the first
failure disables the writer and degrades to a debug log.
* **Survive child crashes.** Files are opened in append mode per write —
no long-lived handle to lose, every event is flushed when written.
* **Side-channel only.** Nothing here touches message content, so prompt
caching is unaffected.
* **No config knobs.** Retention is a module constant (7 days), pruned
opportunistically on each new dispatch.
"""
from __future__ import annotations
import json
import logging
import shutil
import threading
import time
import uuid
from pathlib import Path
from typing import Any, Dict, List, Optional
logger = logging.getLogger(__name__)
# Live transcript directories older than this are pruned on new dispatches.
LIVE_RETENTION_DAYS = 7
# Per-line truncation budgets (chars). The .log is a compact operational
# view, not the full-fidelity record — the child's SessionDB transcript and
# the summary spill files carry complete text.
_ASSISTANT_MAX = 600
_THINKING_MAX = 300
_ARGS_MAX = 220
_RESULT_MAX = 400
_KICKOFF_MAX = 500
# Stream deltas are buffered and flushed as one assistant line when another
# event type arrives (or on completion). Cap the buffer so a huge streamed
# reply can't hold memory hostage.
_STREAM_BUFFER_FLUSH_CHARS = 4000
def live_transcript_root() -> Path:
"""Root directory for live transcripts (profile-safe, never ~/.hermes)."""
from hermes_constants import get_hermes_dir
return get_hermes_dir("cache/delegation", "delegation_cache") / "live"
def new_live_delegation_id() -> str:
"""Same shape as async_delegation's ids so the dir name matches the handle."""
return f"deleg_{uuid.uuid4().hex[:8]}"
def _one_line(text: Any, limit: int) -> str:
"""Collapse to a single line and truncate with an elided-chars note."""
s = str(text or "")
s = " ".join(s.split()) # collapse newlines/runs of whitespace
if len(s) > limit:
omitted = len(s) - limit
s = s[:limit] + f" …(+{omitted} chars)"
return s
def _redact(text: str) -> str:
"""Mask credentials before anything reaches the transcript file.
These logs live under ``cache/delegation``, which ``delegate_tool`` mounts
READ-ONLY into remote terminal backends — so every line written here is
readable from inside the sandbox. The events rendered here carry exactly
the data that tends to hold secrets: tool args (a bearer header on a
curl), tool results (a ``.env`` dump, a provider error echoing the key
back) and streamed assistant text. Every other sink for that data already
routes through this same redactor — search results via
``redact_sensitive_text``, terminal output via ``redact_terminal_output``
— so a transcript that skipped it is the one place the operator's keys
land in plaintext.
``force=True``: this is a safety boundary, so it must redact even when the
global toggle is off. Withholds the line rather than emitting raw text if
the redactor is somehow unavailable — losing a debug line costs less than
writing a live credential into a sandbox-readable file.
"""
if not text:
return text
try:
from agent.redact import redact_sensitive_text
return redact_sensitive_text(text, force=True) or ""
except Exception: # pragma: no cover - core module; never leak on failure
return "[line withheld: redaction unavailable]"
class LiveTranscriptWriter:
"""Append-only human-readable event log for ONE subagent task.
All methods are best-effort: the first write failure flips ``_ok`` off
and subsequent calls become no-ops (debug-logged). Never raises.
"""
def __init__(self, delegation_id: str, task_index: int, goal: str,
context: Optional[str] = None, root: Optional[Path] = None):
self.delegation_id = delegation_id
self.task_index = task_index
self._ok = True
self._lock = threading.Lock()
self._stream_buf: List[str] = []
self._stream_len = 0
try:
base = (root if root is not None else live_transcript_root())
d = base / delegation_id
d.mkdir(parents=True, exist_ok=True)
self.path: Optional[Path] = d / f"task-{task_index}.log"
header = [
"=== Hermes subagent live transcript ===",
f"delegation: {delegation_id} task: {task_index}",
# Header bypasses event(), so redact here too — a goal string
# can carry a key the caller pasted into the task.
f"goal: {_redact(_one_line(goal, _KICKOFF_MAX))}",
f"started: {time.strftime('%Y-%m-%d %H:%M:%S')}",
"(append-only; streams while the subagent runs — tail -f me)",
"=" * 40,
]
self.path.write_text("\n".join(header) + "\n", encoding="utf-8")
self.event("user", "kickoff: " + _one_line(goal, _KICKOFF_MAX)
+ (f" | context: {_one_line(context, _KICKOFF_MAX)}" if context else ""))
except Exception as exc:
logger.debug("Live transcript init failed (%s task %s): %s",
delegation_id, task_index, exc)
self._ok = False
self.path = None
# ── low-level ────────────────────────────────────────────────────────
def event(self, role: str, text: str) -> None:
"""Append one ``HH:MM:SS role ⟩ text`` line. Flushed per event."""
if not self._ok or self.path is None:
return
# Single choke point: every typed helper funnels through here, so
# redacting once covers args, results, thinking and streamed text —
# and a helper added later can't bypass it.
line = f"{time.strftime('%H:%M:%S')} {role:<9}| {_redact(text)}\n"
try:
with self._lock:
# Append mode per write: no held handle, survives child crash,
# and the close() acts as the flush.
with open(self.path, "a", encoding="utf-8") as fh:
fh.write(line)
except Exception as exc:
self._ok = False
logger.debug("Live transcript write failed (%s): %s", self.path, exc)
# ── typed helpers ────────────────────────────────────────────────────
def assistant_text(self, text: str) -> None:
t = _one_line(text, _ASSISTANT_MAX)
if t:
self.event("assistant", t)
def thinking(self, text: str) -> None:
t = _one_line(text, _THINKING_MAX)
if t:
self.event("think", t)
def tool_start(self, name: str, args_preview: Any = None) -> None:
self.flush_stream()
args = _one_line(args_preview, _ARGS_MAX)
self.event("tool", f"-> {name or '?'}({args})")
def tool_result(self, name: str, result: Any = None,
duration: Any = None, is_error: bool = False) -> None:
status = "ERROR" if is_error else "ok"
dur = ""
try:
if duration is not None:
dur = f" {float(duration):.1f}s"
except (TypeError, ValueError):
pass
self.event("result", f"{name or '?'} {status}{dur}: "
f"{_one_line(result, _RESULT_MAX)}")
def marker(self, text: str) -> None:
"""Lifecycle marker: start / final / error / interrupt / budget."""
self.flush_stream()
self.event("final", _one_line(text, _ASSISTANT_MAX))
# ── streamed reply buffering ─────────────────────────────────────────
def add_stream_delta(self, delta: str) -> None:
"""Buffer streamed assistant reply text; flushed as one line."""
if not delta or not self._ok:
return
self._stream_buf.append(delta)
self._stream_len += len(delta)
if self._stream_len >= _STREAM_BUFFER_FLUSH_CHARS:
self.flush_stream()
def flush_stream(self) -> None:
if not self._stream_buf:
return
text = "".join(self._stream_buf)
self._stream_buf = []
self._stream_len = 0
self.assistant_text(text)
# ── event demux (the tool_progress_callback surface) ─────────────────
def observe(self, event_type: Any, tool_name: Any = None,
preview: Any = None, args: Any = None, **kwargs: Any) -> None:
"""Map a child tool_progress_callback event onto transcript lines.
Mirrors the shapes emitted by agent/tool_executor.py,
agent/conversation_loop.py, and tools/delegate_tool._run_single_child.
Unknown events are ignored. Never raises (event() swallows I/O).
"""
et = str(event_type or "")
if et == "tool.started":
self.tool_start(str(tool_name or ""), preview if preview else args)
elif et == "tool.completed":
self.tool_result(
str(tool_name or ""),
result=kwargs.get("result"),
duration=kwargs.get("duration"),
is_error=bool(kwargs.get("is_error")),
)
elif et == "_thinking":
# Fired as cb("_thinking", <text>) — the text rides in the
# tool_name positional slot (see conversation_loop.py).
self.thinking(str(tool_name or preview or ""))
elif et == "reasoning.available":
# cb("reasoning.available", "_thinking", <text>, None)
self.thinking(str(preview or ""))
elif et == "subagent.text":
self.add_stream_delta(str(preview or ""))
elif et == "subagent.start":
self.event("start", _one_line(preview, _KICKOFF_MAX))
elif et == "subagent.complete":
self.flush_stream()
status = kwargs.get("status", "?")
dur = kwargs.get("duration_seconds")
parts = [f"status={status}"]
if dur is not None:
parts.append(f"duration={dur}s")
summary = kwargs.get("summary") or preview
if summary:
parts.append(f"summary: {_one_line(summary, _RESULT_MAX)}")
self.marker(" ".join(parts))
def finalize(self, entry: Dict[str, Any]) -> None:
"""Terminal marker from the aggregated result entry.
Adds exit-reason detail the subagent.complete event doesn't carry
(budget exhaustion via exit_reason=max_iterations, errors, etc.).
"""
parts = [f"end status={entry.get('status', '?')}"]
exit_reason = entry.get("exit_reason")
if exit_reason:
parts.append(f"exit_reason={exit_reason}")
if exit_reason == "max_iterations":
parts.append("(iteration budget exhausted)")
if entry.get("error"):
parts.append(f"error: {_one_line(entry['error'], _RESULT_MAX)}")
self.marker(" ".join(parts))
def wrap_progress_callback(inner_cb, writer: LiveTranscriptWriter):
"""Wrap a child's tool_progress_callback so events also land in the log.
``inner_cb`` may be None (no parent display) — the wrapper still records.
Writer failures never propagate; inner callback behavior is unchanged
(its own exceptions are handled by callers exactly as before).
Preserves the ``_flush`` attribute contract used by _run_single_child.
"""
def _cb(event_type, tool_name=None, preview=None, args=None, **kwargs):
try:
writer.observe(event_type, tool_name, preview, args, **kwargs)
except Exception as exc: # noqa: BLE001 — must never hit the agent loop
logger.debug("Live transcript observe failed: %s", exc)
if inner_cb is not None:
inner_cb(event_type, tool_name, preview, args, **kwargs)
def _flush():
try:
writer.flush_stream()
except Exception:
pass
inner_flush = getattr(inner_cb, "_flush", None)
if callable(inner_flush):
inner_flush()
_cb._flush = _flush
return _cb
# ── dispatch-time helpers ────────────────────────────────────────────────
def create_live_transcripts(
task_list: List[Dict[str, Any]],
context: Optional[str] = None,
delegation_id: Optional[str] = None,
) -> tuple[Optional[str], List[Optional[LiveTranscriptWriter]], List[str]]:
"""Create one pre-headered writer per task + a manifest.json.
Returns ``(delegation_id, writers, paths)``. On any top-level failure
returns ``(None, [None]*n, [])`` so delegation proceeds untouched.
Also opportunistically prunes stale live dirs (retention).
"""
n = len(task_list)
try:
prune_stale_live_dirs()
except Exception:
pass
try:
deleg_id = delegation_id or new_live_delegation_id()
writers: List[Optional[LiveTranscriptWriter]] = []
paths: List[str] = []
for i, t in enumerate(task_list):
w = LiveTranscriptWriter(
deleg_id, i, str(t.get("goal", "")),
context=t.get("context") or context,
)
writers.append(w if w.path is not None else None)
if w.path is not None:
paths.append(str(w.path))
if not paths:
return None, [None] * n, []
_write_manifest(deleg_id, task_list, paths)
return deleg_id, writers, paths
except Exception as exc:
logger.debug("Live transcript creation failed: %s", exc)
return None, [None] * n, []
def _manifest_path(delegation_id: str) -> Path:
return live_transcript_root() / delegation_id / "manifest.json"
def _write_manifest(delegation_id: str, task_list: List[Dict[str, Any]],
paths: List[str]) -> None:
try:
manifest = {
"delegation_id": delegation_id,
"started": time.strftime("%Y-%m-%d %H:%M:%S"),
"task_count": len(task_list),
"tasks": [
{
"index": i,
# manifest.json sits in the same mounted
# cache/delegation/live/<id>/ directory as the .log files,
# so it needs the same treatment — redacting the header
# while serialising the goal verbatim here would leave the
# credential exposed one file over.
"goal": _redact(str(t.get("goal", ""))[:500]),
"log": paths[i] if i < len(paths) else None,
"status": "running",
}
for i, t in enumerate(task_list)
],
}
_manifest_path(delegation_id).write_text(
json.dumps(manifest, indent=2, ensure_ascii=False), encoding="utf-8"
)
except Exception as exc:
logger.debug("Live transcript manifest write failed: %s", exc)
def update_manifest_statuses(delegation_id: Optional[str],
results: List[Dict[str, Any]]) -> None:
"""Best-effort per-task status update once the batch has aggregated."""
if not delegation_id:
return
try:
mp = _manifest_path(delegation_id)
manifest = json.loads(mp.read_text(encoding="utf-8"))
by_index = {r.get("task_index"): r for r in results if isinstance(r, dict)}
for task in manifest.get("tasks", []):
r = by_index.get(task.get("index"))
if r is not None:
task["status"] = r.get("status", task.get("status"))
if r.get("exit_reason"):
task["exit_reason"] = r["exit_reason"]
manifest["completed"] = time.strftime("%Y-%m-%d %H:%M:%S")
mp.write_text(json.dumps(manifest, indent=2, ensure_ascii=False),
encoding="utf-8")
except Exception as exc:
logger.debug("Live transcript manifest update failed: %s", exc)
def prune_stale_live_dirs(max_age_days: int = LIVE_RETENTION_DAYS) -> int:
"""Remove live/<delegation_id> dirs older than the retention window.
Returns how many were removed. Fully best-effort.
"""
removed = 0
try:
root = live_transcript_root()
if not root.is_dir():
return 0
cutoff = time.time() - max_age_days * 86400
for child in root.iterdir():
try:
if child.is_dir() and child.stat().st_mtime < cutoff:
shutil.rmtree(child, ignore_errors=True)
removed += 1
except OSError:
continue
except Exception as exc:
logger.debug("Live transcript pruning failed: %s", exc)
return removed