Dev Guide Debug Troubleshoot

LingTai Debug & Troubleshoot Reference

Nested lingtai-dev-guide reference. Read this after the top-level router sends you here.

Read the lingtai-kernel-anatomy skill first. It owns the concepts — lifecycle states, memory layers, mail protocol, tool surface — that every section below assumes. This document is only the diagnosis-and-recovery procedure; the per-section concept map is at the end.

Quick Diagnosis Decision Tree

Problem?
├── Process issues?
│   ├── Peer unresponsive → §1.1
│   ├── Peer OOM / crashed → §1.2
│   └── Cannot spawn avatar → §1.3
├── Memory issues?
│   ├── Post-molt amnesia → §2.1
│   ├── Codex entries missing → §2.2
│   ├── Pad not loaded → §2.3
│   └── Molt imminent, critical operations incomplete → §2.4
├── Communication issues?
│   ├── Pigeon not delivered → §3.1
│   ├── Pigeon bounced "No agent at X" → §3.2
│   └── Scheduled pigeon not firing → §3.3
└── Tool issues?
    ├── Tool timeout → §4.1
    ├── Tool not found → §4.2
    └── Tool output truncated → §4.3

1. Process Issues

1.1 Peer Unresponsive

Pigeons go unanswered; the peer is in contacts but silent. Possible states: busy (long LLM turn), stuck (LLM timeout/upstream error), asleep (energy depleted or lulled), suspended (process dead), or wrong address.

Diagnose in order — system(show) (your own health) → email(contacts) (verify the address) → email(send, address=<peer>, message="ping") → the heartbeat:

ls -la <work-dir>/.lingtai/<peer>/.agent.heartbeat
cat <work-dir>/.lingtai/<peer>/.agent.heartbeat

Fresh (< 5 min) = busy, just wait. Stale (> 5 min) = stuck or crashed. No file = probably no agent at that address. For a whole network at once, use the sweep in §5.

Action. With karma: system(interrupt, address=<peer>) for a stuck LLM turn, system(cpr, address=<peer>) to revive a suspended agent. Without karma: report to parent with evidence (heartbeat timestamp, last contact time).

Pitfalls. Repeated probe emails waste resources and cannot wake a suspended process; CPR without nirvana privileges fails silently. Asleep ≠ suspended — asleep agents wake on email, suspended ones need CPR first. Check the heartbeat before choosing to wait, interrupt, or CPR.

1.2 Peer OOM / Crashed

Heartbeat stops abruptly while the working directory survives. Usual causes: host memory exhausted (OS OOM killer), LLM upstream unresponsive past the process timeout, an uncaught Python exception, or a full disk.

# Comprehensive health check for an agent
peer_dir="<work-dir>/.lingtai/<peer>"
echo "=== Process ==="
ls -la "$peer_dir/.agent.heartbeat" 2>/dev/null || echo "No heartbeat"
echo "=== Disk ==="
df -h "$peer_dir" | tail -1
echo "=== Recent logs ==="
tail -30 "$peer_dir/logs/"*.log 2>/dev/null || echo "No logs"
echo "=== OOM scan ==="
grep -il "oom\|killed\|memory" "$peer_dir/logs/"*.log 2>/dev/null || echo "No OOM indicators"

Action. With karma, system(cpr, address=<peer>) revives it — then check context usage immediately, because reviving into a near-full context crashes again. After an OOM, prioritize the context window and attachment file sizes; ignoring disk space leaves the root cause live and the issue recurs.

1.3 Cannot Spawn Avatar

avatar(spawn) errors, or the new process never appears under delegates. Causes: name collision, unwritable working directory, insufficient disk, or a malformed init.json.

# List all current avatars (name collisions, quantity limits)
cat <work-dir>/.lingtai/delegates/ledger.jsonl | python3 -c "
import sys, json
for line in sys.stdin:
    entry = json.loads(line.strip())
    print(f\"{entry.get('name', '?')}: {entry.get('status', '?')}\")
"

# Directory writable?
touch <work-dir>/.lingtai/delegates/.test && rm <work-dir>/.lingtai/delegates/.test

# Disk space
df -h <work-dir>

If those pass, compare the avatar's init.json against the parent's to validate the format.

Pitfalls. Special characters in avatar names (slashes, spaces, leading dots) make spawn fail silently, as do names over 64 characters. Check the ledger first to avoid a collision; use only letters, digits, underscores, and hyphens.

2. Memory Issues

2.1 Post-Molt Amnesia

After a molt you don't know what you were doing, and pad/lingtai are empty or partial. (A vanished conversation history is normal — that is what molt does.) Causes: durable layers not updated before molting, a system-forced molt that left no summary, or appended files exceeding the 100K token limit and failing to load.

Recover in this order:

psyche(pad, load)                                    # 1. reload working notes
codex(filter)                                        # 2. browse archived knowledge
psyche(lingtai, load)                                # 3. reload identity
email(check)                                         # 4. mail that arrived during the molt
codex(export, ids=[...]) → psyche(pad, edit, files=[<paths>])   # 5. rebuild an empty pad

If the molt was system-forced (no summary), the activity log is the only trail:

tail -200 <work-dir>/.lingtai/<name>/logs/events.jsonl
grep "molt" <work-dir>/.lingtai/<name>/logs/events.jsonl | tail -5

Prevention. Follow the fixed pre-molt checklist — codex → pad edit → lingtai update → molt summary. Start preparing past 70% context and act immediately on a level-1 warning. Never treat conversation history as storage; it is all lost. Self-email survives molts, so mail yourself anything critical and unfinished.

2.2 Codex Entries Missing

An entry you remember creating no longer appears in codex(filter). Causes: submission failed silently, consolidate merged it into another entry (originals are deleted), manual deletion, or a lost export file.

find <work-dir> -name "*.codex.*" -mtime -1                                  # export files
grep "codex" <work-dir>/.lingtai/<name>/logs/events.jsonl | tail -20         # operation records

Also re-run codex(filter) and look for the content under a different title.

Pitfalls. Originals do not survive consolidate — export critical entries first. Confirm every submit succeeded; a network error can fail silently.

2.3 Pad Not Loaded

The system prompt has no pad content and working notes are gone. Causes: pad.md is empty, total appended file size exceeds the 100K token limit, or a loading error.

psyche(pad, load)

cat <work-dir>/.lingtai/<name>/system/pad.md      # does it have content?
du -sh <work-dir>/.lingtai/<name>/system/          # appended size, if it does

If the pad is genuinely empty, rebuild it: codex(export, ids=[...]) → psyche(pad, edit, files=[<paths>]).

Pitfalls. Appending too many large files breaks loading at the 100K limit. Review the appended list periodically — psyche(pad, append) with no files parameter prints it — and check pad files before molting, not after.

2.4 Molt Imminent, Critical Operations Incomplete

Context warnings, difficulty recalling earlier conversation, sluggish tool calls. Spend what remains in priority order:

Self-email survives molt, so use it for critical unfinished items. If you can only do one thing, write the most detailed molt summary possible.

Pitfalls. Do not start long operations (file analysis, web search) past 80% context — overflow is guaranteed. Ignoring system warnings ends in a forced molt with no summary.

3. Communication Issues

3.1 Pigeon Not Delivered

Sent successfully, but the recipient never got it. Causes: address format error (an internal address containing @), send used where reply was needed, a misspelled recipient directory, or a suspended recipient (mail is delivered but never processed).

email(check, folder=sent)     # confirm it actually went out

ls -la <work-dir>/.lingtai/<recipient>/mailbox/inbox/    # does their inbox exist?

Address format: human, researcher, some-peer (bare path) are correct; human@example.com is not an internal address — the @ routes it through the IMAP channel instead of the LingTai pigeon.

Pitfalls. Use reply for incoming messages and send only for new conversations; send on a reply can route into the wrong address space. Repeatedly mailing a suspended recipient just piles up unprocessed mail.

3.2 Pigeon Bounced "No agent at X"

email(send) returns "No agent at X". Either X contains @ (wrong channel — switch to the IMAP tool), or X is a bare path with no agent behind it: renamed/migrated, nirvana'd (permanently deleted), or mid-molt and briefly unavailable.

cat <work-dir>/.lingtai/delegates/ledger.jsonl    # renamed, migrated, or gone?

Ask the parent or peers whether the agent was nirvana'd. If it was just molting, wait a few seconds and retry.

Pitfalls. "No agent" is not proof of deletion — it is often temporary. Determine the address type first, then pick the channel.

3.3 Scheduled Pigeon Not Firing

A schedule stops sending at the expected interval. Causes: paused/cancelled, count exhausted, or interval/count set wrong.

email(schedule={action: "list"})                                  # status: paused / active / exhausted
email(schedule={action: "reactivate", schedule_id: "<id>"})       # if paused
email(schedule={action: "cancel", schedule_id: "<id>"})           # if parameters are wrong:
email(schedule={action: "create", interval: N, count: M}, address=..., message=...)

Pitfalls. Omitting count can make a schedule fire once and stop; cancelling without recreating loses the task. List immediately after creating to confirm the parameters.

4. Tool Issues

4.1 Tool Timeout

A tool call hangs or times out. Causes: I/O-intensive work (bash, web_search) exceeding the default timeout, an unavailable external API, an oversized file, or host resource shortage.

bash(command="...", timeout=120)                       # raise the bash timeout
read(file_path="...", offset=1, limit=100)             # chunk large reads

# Best pattern for long output: redirect to a file, then read in chunks
bash(command="long-running-command > /tmp/output.txt 2>&1", timeout=300)
read(file_path="/tmp/output.txt", offset=1, limit=100)

For web operations, test connectivity with a simple query first; for systemic timeouts, check host load. vision is compute-intensive while bash/web_search are I/O-intensive — they fail differently.

Pitfalls. The default 30-second bash timeout guarantees failure on long tasks; single-call reads of large files guarantee truncation.

See also web_search(action="manual") / the web-search-manual skill.

4.2 Tool Not Found

A tool returns "not available", or a newly installed MCP tool is invisible. Causes include a changed MCP/init input not yet applied to the current runtime, a missing capability in init.json, malformed servers.json, an MCP boot failure, or a module-source mismatch.

Start from the exact error and inspect only the named owner input:

cat <work-dir>/.lingtai/<name>/mcp/servers.json

mcp-manual owns MCP registry/config validation and repair. For a known changed MCP input that only needs a same-runtime reload, kernel system-manual → reference/refresh-precheck/SKILL.md owns the transaction and changed-MCP assertion. If a refresh already failed or the tool remains absent, route the exact error to that reference's failure recovery and then to the owning MCP fault. Do not run a generic show/refresh/show loop or retry blindly.

Pitfalls. Installing an MCP server or editing init.json does not by itself prove the changed surface is live. Preserve the error, validate the changed owner input, and follow the kernel transaction or failure route instead of inventorying unrelated tools.

4.3 Tool Output Truncated

Output is incomplete or ends in a truncation marker — the file exceeded the single-return limit, grep hit max_matches, or an email preview was clipped.

Pitfalls. Never assume output is complete — silent truncation drops exactly the detail you needed. Re-reading the same large file wastes context; write it out once, then read what you need.

5. Health Checks

One-Click Network Diagnosis

# Check heartbeats for all agents
for dir in <network-dir>/.lingtai/*/; do
  name=$(basename "$dir")
  hb="$dir/.agent.heartbeat"
  if [ -f "$hb" ]; then
    age=$(( $(date +%s) - $(stat -f %m "$hb" 2>/dev/null || stat -c %Y "$hb") ))
    if [ "$age" -lt 300 ]; then
      echo "✅ $name: alive (${age}s ago)"
    else
      echo "⚠️  $name: stale heartbeat (${age}s ago)"
    fi
  else
    echo "❌ $name: no heartbeat"
  fi
done

# Check disk space
df -h <network-dir>

# Check inbox sizes
for dir in <network-dir>/.lingtai/*/mailbox/inbox/; do
  name=$(echo "$dir" | sed 's|.*\.lingtai/\(.*\)/mailbox/.*|\1|')
  count=$(ls "$dir" 2>/dev/null | wc -l)
  if [ "$count" -gt 50 ]; then
    echo "⚠️  $name: inbox has $count messages (possible overflow)"
  fi
done

✅ = healthy · ⚠️ = warning, needs attention · ❌ = error, requires immediate action.

6. Escalation Protocol

When you cannot resolve a problem on your own:

1. Gather evidence: heartbeat timestamps, log excerpts, error messages.
2. Report to parent: email(send, address=<parent>) with a subject prefixed [Issue].
3. Include: what happened, what was attempted, what was expected, and the relevant file paths.
4. If the parent is also unresponsive: check whether other peers are alive — the issue may be network-wide.
5. Never send repeated probe emails to a seemingly unresponsive peer; escalate upward instead.

Concept references

Each section's underlying concepts live in lingtai-kernel-anatomy:

Appendix: Five Lifecycle States Quick Reference

Key distinction: ASLEEP agents still have a running body and can be woken by email; SUSPENDED agents have a dead process and require CPR before they can process mail.