MythEclipse/pr-agent-server
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

feat(worker): per-fork session header + 3600s sync timeout + merge salvage

Browse Source 
- _hermes_api_post sends X-Hermes-Session-Id so each fork sync uses a
  fresh gateway transcript (was: derived session reused 200k-token history
  across upstream tips, making later runs pay growing context)
- SYNC_CLAUDE_TIMEOUT 2700→3600: a fresh session needs ~25min/120 calls
  to resolve 17 files + verify + commit; 2700s cut the HTTP call while the
  agent was still committing
- salvage-on-timeout: when the agent call ends early (timeout/transport)
  but the merge is already committed (MERGE_HEAD gone), finish+push it
  instead of aborting — verified live: 17 conflicts resolved by agent,
  push ok, fork 27 ahead / 0 behind
- _sync_finish_merge blocks committing files that still carry conflict
  markers (guard against git add of unresolved files)
- tests: 68/68 (session header, salvage completed merge, abort unfinished,
  marker guard; fake-leak hardening in _install_git_fake)
This commit is contained in:
asepharyana
2026-09-21 21:02:33 +07:00
parent 99d47c8ff8
commit 66b567a6bc
3 changed files with 291 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
scripts/pr-queue-worker.py
+63 -21
View File
@@ -73,7 +73,7 @@ AI_FIX_ENABLED = True
# skips when the binary exists in PATH but not at the hardcoded location.
CLAUDE_BIN = shutil.which("claude") or "/usr/local/bin/claude"
AI_FIX_MAX_TURNS = 100
AI_FIX_TIMEOUT = 600 # seconds per PR
AI_FIX_TIMEOUT = 1800 # seconds per PR (agent turns + tool calls, not one inference)
# Hermes gateway API server (OpenAI-compatible). The worker drives the running
# gateway instead of spawning a CLI: the gateway already holds the provider
# (9router) config and a full toolset (terminal/file/web).
@@ -985,7 +985,10 @@ def run_ai_fix(repo_full, pr_num, title, head_sha, head_ref, base_ref, token):
SYNC_STATE_FILE = Path("/tmp/pr-queue-sync-state.json")
SYNC_TMP_BASE = Path("/tmp/pr-queue-sync-work")
SYNC_PR_PREFIX = "upstream-sync-"
SYNC_CLAUDE_TIMEOUT = 900 # seconds: conflict resolution + verification run
SYNC_CLAUDE_TIMEOUT = 3600 # seconds: conflict resolution + verification run.
# Observed 2026-09-21: a fresh session needs ~25 min (56+ API calls) to resolve
# 17 files + verify + commit. 2700s cut the HTTP call while the agent was still
# committing. 3600 leaves room; salvage still catches a genuinely hung call.
UPSTREAM_SYNC = {
"enabled": os.environ.get("PR_AGENT_UPSTREAM_SYNC", "1") != "0",
"interval_h": 1.0, # check upstream hourly (user decision 2026-09-21)
@@ -1188,7 +1191,7 @@ def _push_ref(workdir, fork, source, dest, app_token, force=False):
return False, last, False
def _hermes_api_post(prompt, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label="hermes"):
def _hermes_api_post(prompt, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label="hermes", session_id=None):
"""Run a Hermes agent task through the local gateway API server
(OpenAI-compatible POST /v1/chat/completions on 127.0.0.1:8642).
@@ -1198,6 +1201,12 @@ def _hermes_api_post(prompt, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label="hermes
transport (Anthropic Messages mismatch / JSON-not-a-Message / exit 1)
and spawned hanging MCP servers.
`session_id` (optional) is sent as X-Hermes-Session-Id. When the gateway
sees that header it continues that session's transcript (per SHA == one
session: a retry of the same upstream tip resumes where the previous run
stopped, and a NEW upstream tip gets a fresh context instead of paying a
growing 200k-token history from older runs).
Returns (ok, snippet). snippet is the agent's final answer text.
"""
import httpx
@@ -1209,6 +1218,8 @@ def _hermes_api_post(prompt, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label="hermes
"Authorization": "Bearer " + key,
"Content-Type": "application/json",
}
if session_id:
headers["X-Hermes-Session-Id"] = session_id
body = {
"model": "hermes-agent",
"messages": [
@@ -1289,7 +1300,8 @@ def _run_hermes_sync(workdir, prompt, label, fork, dry=False):
" cd " + str(workdir) + "\n"
"Then complete the task below.\n\n" + prompt
)
ok, snippet = _hermes_api_post(prompt_with_dir, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label=label)
ok, snippet = _hermes_api_post(prompt_with_dir, workdir, timeout=SYNC_CLAUDE_TIMEOUT, label=label,
session_id="sync_" + str(fork).replace("/", "_"))
if ok:
try:
out_path.write_text(label + " ok: " + snippet + "\n")
@@ -1307,11 +1319,28 @@ def _sync_unmerged_files(workdir):
return [f for f in (r.stdout or "").split("\n") if f.strip()]
def _sync_has_conflict_markers(workdir):
"""Tracked files that still contain conflict markers.
A file the agent `git add`ed mid-resolution would pass the unmerged check
while still carrying `<<<<<<<` — never commit that.
"""
r = _sync_git(["grep", "-l", "-E", r"^(<<<<<<<|>>>>>>>|=======)$"],
workdir, 60)
# git grep exits 1 when there are no matches (that is the healthy case).
if r.returncode not in (0, 1):
return []
return [f for f in (r.stdout or "").split("\n") if f.strip()]
def _sync_finish_merge(workdir):
"""Complete an in-progress merge once every conflict is resolved."""
unmerged = _sync_unmerged_files(workdir)
if unmerged:
return False, f"{len(unmerged)} file(s) still unmerged: {', '.join(unmerged[:5])}"
marked = _sync_has_conflict_markers(workdir)
if marked:
return False, f"conflict markers left in: {', '.join(marked[:5])}"
_sync_git(["add", "-A"], workdir, 120)
r = _sync_git(["commit", "--no-edit"], workdir, 120)
if r.returncode != 0:
@@ -1543,23 +1572,36 @@ def sync_fork_repo(token, fork, parent, local_branch, upstream_branch, upstream_
workdir, _conflict_prompt(fork, parent, upstream_branch, local_branch, conflicted),
"hermes_sync_conflicts", fork, dry)
if not ok:
_sync_git(["merge", "--abort"], workdir, 60)
note = f"conflict resolution failed — {snippet[:200]}"
if not dry:
entry.update({"last_sync_ts": time.time(), "last_attempt_sha": upstream_sha,
"skip_reason": note, "notified": False})
save_sync_state(state)
return "conflict-failed", note
done, why = _sync_finish_merge(workdir)
if not done:
_sync_git(["merge", "--abort"], workdir, 60)
note = f"conflict resolution incomplete — {why}"
if not dry:
entry.update({"last_sync_ts": time.time(), "last_attempt_sha": upstream_sha,
"skip_reason": note, "notified": False})
save_sync_state(state)
return "conflict-failed", note
resolution = f"Hermes resolved {len(conflicted)} conflict(s)"
# Salvage: a timeout/transport error does NOT mean the agent
# failed. Resolving N conflicts takes many minutes (observed:
# 17 files ≈ 16 min, 128 API calls), so the HTTP call can time
# out AFTER the agent finished resolving and committed. Only
# abort when the workdir still shows unresolved state.
salvaged, why_salvage = _sync_finish_merge(workdir)
if salvaged:
BUFFER.append(" ♻️ agent call ended early but the merge is complete — salvaged")
resolution = f"Hermes resolved {len(conflicted)} conflict(s); " \
f"agent call ended early ({snippet[:80]}) — merge salvaged"
else:
_sync_git(["merge", "--abort"], workdir, 60)
note = f"conflict resolution failed — {snippet[:200]}"
if not dry:
entry.update({"last_sync_ts": time.time(), "last_attempt_sha": upstream_sha,
"skip_reason": note, "notified": False})
save_sync_state(state)
return "conflict-failed", note
done, why = True, "salvaged"
else:
done, why = _sync_finish_merge(workdir)
if not done:
_sync_git(["merge", "--abort"], workdir, 60)
note = f"conflict resolution incomplete — {why}"
if not dry:
entry.update({"last_sync_ts": time.time(), "last_attempt_sha": upstream_sha,
"skip_reason": note, "notified": False})
save_sync_state(state)
return "conflict-failed", note
resolution = f"Hermes resolved {len(conflicted)} conflict(s)"
else:
resolution = "clean merge"
if cfg.get("ai_fix_after_merge", True):