Session hangs indefinitely: filesystem tools block forever on unresponsive UNC path; parent delegate never returns

Session hangs indefinitely: filesystem tools block forever on unresponsive UNC path; parent delegate never returns

Summary

An interactive session hung silently for ~14.5 hours (until manually killed) because a delegated sub-agent issued filesystem tool calls (grep, read_file) against an RDP-redirected UNC share (\\tsclient\Z\...) that became unresponsive. There is no tool-level timeout, watchdog, or heartbeat anywhere in the chain:

blocked filesystem I/O → sub-session never produces its next LLM request → parent's delegate tool call never receives a result → entire session hangs with no error, no log output, no indication anything is wrong.

A secondary but independently serious finding: even on a healthy share, an unscoped repo-root grep over RDP drive redirection took 55 minutes 9 seconds and returned 0 matches (a near-hang by itself, with a misleading result — see "Contributing factors").

Environment

• OS: Windows 11 (10.0.26200)
• Working directory: \\tsclient\Z\home\dan\code\freshell — an RDP drive-redirection UNC share pointing at a Linux client filesystem
• Repo characteristics: repo root contains node_modules, .git, dist, release, playwright-report, and a .worktrees directory holding 44 worktree copies of the repo
• WSL: not installed (relevant to a secondary finding below)
• Provider/model: anthropic / claude-fable-5
• Bundle: anchors (delegate to anchors:explorer)

Timeline (UTC; local = UTC-7)

Time	Event
2026-07-08 23:40:23	session:start — prompt asks to read up on a tauri port and propose a plan
23:40:39	Turn-1 tools (read_file repo root, todo) complete in ~1.4 s — share healthy
23:40:58	LLM iter-2 response → delegate to anchors:explorer
23:40:59.664	delegate:agent_spawned — last parent event before the hang
23:41:09	Sub-agent LLM responds → 3 parallel tool calls
23:41:10.181	Sub-session tool:pre × 3: repo-root grep -i tauri, read_file repo root, read_file .worktrees
—	Log ends. No tool:post, no error, no timeout. Nothing for ~14 h 35 m.
2026-07-09 14:16:31	session:resume — session manually killed and restarted by user
14:17:16	Resumed explorer sub-session re-runs the identical repo-root grep
15:12:26	tool:post for that grep: 3,309 s (55 m 09 s), 0 matches

Evidence (verbatim event-log excerpts; large payloads trimmed as [...], no secrets present)

1. Parent session — last events before the hang

# parent events.jsonl, line 21 (LLM iter-2 response; raw content trimmed)
{"ts": "2026-07-08T23:40:58.664193200+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "llm:response", "duration_ms": 17880, "redaction": {"applied": true, "rules": ["secrets", "pii-basic"]}, "session_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "status": "ok", "data": {"model": "claude-fable-5", "parent_id": null, "provider": "anthropic", "rate_limits": {"input_tokens_limit": 4000000, "input_tokens_remaining": 3992000, [...]}, "raw": {"container": null, "content": [{"signature": "[...]", "thinking": "", "type": "thinking"}, [...]}}}

# parent events.jsonl, line 29 (LAST parent event before 14.5h gap)
{"ts": "2026-07-08T23:40:59.664168100+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "delegate:agent_spawned", "session_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "data": {"agent": "anchors:explorer", "context_depth": "none", "context_scope": "conversation", "model_role": null, "parallel_group_id": "3406152b-bbdb-4789-a042-4ee5a182fca6", "parent_id": null, "parent_session_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "provider_preferences": null, "sub_session_id": "0000000000000000-d7e764a0c0b74e86_anchors-explorer", "tool_call_id": "toolu_01XNLUqpMX6xtzzVymC7Eeg8"}}

2. Hung explorer sub-session — three tool:pre events, then the log simply ends

These are the final lines of the file; no tool:post ever follows for any of the three:

# d7e764a0c0b74e86_anchors-explorer events.jsonl, lines 17-19 (end of file)
{"ts": "2026-07-08T23:41:10.181105500+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "tool:pre", "session_id": "0000000000000000-d7e764a0c0b74e86_anchors-explorer", "data": {"parallel_group_id": "dfb5232d-96a5-4a3f-9ee4-68c81f5ccad2", "parent_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "tool_call_id": "toolu_015uAasGVYJzUn7TW2vM6Mka", "tool_input": {"-i": true, "output_mode": "files_with_matches", "path": "\\\\tsclient\\Z\\home\\dan\\code\\freshell", "pattern": "tauri"}, "tool_name": "grep"}}
{"ts": "2026-07-08T23:41:10.181162700+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "tool:pre", "session_id": "0000000000000000-d7e764a0c0b74e86_anchors-explorer", "data": {"parallel_group_id": "dfb5232d-96a5-4a3f-9ee4-68c81f5ccad2", "parent_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "tool_call_id": "toolu_01M93zX6KfpJPusEyoezgPqe", "tool_input": {"file_path": "\\\\tsclient\\Z\\home\\dan\\code\\freshell\\.worktrees"}, "tool_name": "read_file"}}
{"ts": "2026-07-08T23:41:10.181128500+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "tool:pre", "session_id": "0000000000000000-d7e764a0c0b74e86_anchors-explorer", "data": {"parallel_group_id": "dfb5232d-96a5-4a3f-9ee4-68c81f5ccad2", "parent_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "tool_call_id": "toolu_019qjZFX4dmoJadxHRY6HNsT", "tool_input": {"file_path": "\\\\tsclient\\Z\\home\\dan\\code\\freshell"}, "tool_name": "read_file"}}

Note that two of the three blocked calls are trivial directory listings that completed in 0.9–1.8 s in the resumed session — this is not "slow tool", it is I/O blocked indefinitely on a dead redirected share.

3. session:resume next day (~14 h 35 m gap after line 29)

# parent events.jsonl, line 30
{"ts": "2026-07-09T14:16:31.850838200+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "session:resume", "session_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "data": {"parent_id": null}}

4. Post-resume: the identical grep on a healthy share — 55 m 09 s, 0 matches

# 21a73eff3f104c60_anchors-explorer events.jsonl, line 17 (tool:pre)
{"ts": "2026-07-09T14:17:16.790788300+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "tool:pre", "session_id": "0000000000000000-21a73eff3f104c60_anchors-explorer", "data": {"parallel_group_id": "75e00007-45b6-4a57-8d7a-d8da39a445ea", "parent_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "tool_call_id": "toolu_01RN1vexS9GH4xbHgyHASnVu", "tool_input": {"-i": true, "output_mode": "files_with_matches", "path": "\\\\tsclient\\Z\\home\\dan\\code\\freshell", "pattern": "tauri"}, "tool_name": "grep"}}

# 21a73eff3f104c60_anchors-explorer events.jsonl, line 20 (matching tool:post, +55m09s, 0 matches)
{"ts": "2026-07-09T15:12:26.212156200+00:00", "lvl": "INFO", "schema": {"name": "amplifier.log", "ver": "1.0.0"}, "event": "tool:post", "session_id": "0000000000000000-21a73eff3f104c60_anchors-explorer", "data": {"parallel_group_id": "75e00007-45b6-4a57-8d7a-d8da39a445ea", "parent_id": "6094c395-790b-4d8f-b97d-70901b099e3e", "result": {"error": null, "output": {"files": [], "matches_count": 0, "output_mode": "files_with_matches", "pattern": "tauri", "total_matches": 0}, "success": true}, "tool_call_id": "toolu_01RN1vexS9GH4xbHgyHASnVu", "tool_input": {"-i": true, "output_mode": "files_with_matches", "path": "\\\\tsclient\\Z\\home\\dan\\code\\freshell", "pattern": "tauri"}, "tool_name": "grep"}}

5. Transcript repair on resume (artifact, not cause — included for completeness)

# parent transcript.jsonl.backup, lines 3-5
{"role": "tool", "tool_call_id": "toolu_015Y4wB4sE9KX8eD5L3DNDSW", "name": "read_file", "content": "{\"error\": \"unknown_error\", \"message\": \"Tool execution was interrupted and no result was captured.\"}"}
{"role": "tool", "tool_call_id": "toolu_01At8CZrTvTtUdvdHmpukbY8", "name": "todo", "content": "{\"error\": \"unknown_error\", \"message\": \"Tool execution was interrupted and no result was captured.\"}"}
{"role": "assistant", "content": [{"type": "text", "text": "The previous tool calls were interrupted. This response was automatically repaired."}]}

Root cause

Filesystem tool execution (grep, read_file) has no timeout. When the underlying I/O blocks indefinitely — which is standard behavior for RDP drive-redirection (\\tsclient) when the RDP session disconnects or the client sleeps — the tool call never completes and never errors. Because the sub-session is awaited synchronously by the parent's delegate call, the block propagates all the way up and the session hangs silently with zero observability.

Contributing factors

1. No delegate heartbeat/timeout. The parent had no way to notice its sub-session had produced no events for 14+ hours.
2. Unscoped grep is pathological on network shares. A repo-root grep over \\tsclient scanned node_modules, .git, dist, and 44 worktree copies file-by-file, taking 55 minutes even when healthy. Ripgrep may also be unavailable/slower in this path (fallback engine), compounding it.
3. Misleading 0-match result. After 55 minutes, the grep returned 0 matches for "tauri" even though a rust-tauri-port worktree exists under .worktrees/ — hidden dot-directories are skipped by default ignore rules, so the agent burned an hour and got a wrong answer. The default-exclusion behavior interacting with dot-directories deserves a look of its own.
4. Secondary: bash tool routes to WSL and fails on machines without a distribution. Every bash call fails with Windows Subsystem for Linux has no installed distributions (returncode 1, UTF-16LE stdout). This is a fast failure rather than a hang, but it (a) breaks agents whose instructions assume a working shell (e.g. git branch -a), (b) makes gh-based workflows impossible, and (c) the UTF-16LE output is garbled when interpreted as UTF-8. Consider native PowerShell/cmd fallback or capability detection at startup.

Suggested fixes

1. Hard timeout on filesystem tool execution (grep, read_file, glob), with a shorter default for UNC/network paths. A blocked read should fail loudly in seconds-to-minutes, not never.
2. Delegate watchdog: parent-side timeout and/or sub-session heartbeat (e.g. "no event emitted in N minutes → fail the delegate call with diagnostics").
3. Grep guardrails: warn or require confirmation for repo-root scans on network paths; surface elapsed time; consider a per-call time budget with partial results.
4. Ignore-rule transparency: when a search returns 0 matches and default exclusions skipped directories (dot-dirs, node_modules, ...), say so in the result payload.
5. Bash/WSL capability detection with a clear error (or native-shell fallback) on Windows machines without WSL distributions.

Session references (local to reporter's machine)

• Parent session: 6094c395-790b-4d8f-b97d-70901b099e3e
C:\Users\Admin\.amplifier\projects\--tsclient-Z-home-dan-code-freshell\sessions\6094c395-790b-4d8f-b97d-70901b099e3e\events.jsonl
• Hung explorer sub-session: 0000000000000000-d7e764a0c0b74e86_anchors-explorer
• Post-resume explorer sub-session: 0000000000000000-21a73eff3f104c60_anchors-explorer

Full event logs available on request (they contain workspace paths but no secrets).

Start by tracing the grep, read_file, glob, and delegate entry points against the event-log excerpts in the issue. Reproduce a blocked UNC call if possible, then verify that filesystem calls and delegated sub-sessions fail within a bounded time with diagnostics; the suggested grep guardrails and WSL capability behavior are additional scope to assess.