dotharness/dotcraft/dotcraft-error-diagnosis

DotCraft Error Diagnosis

Use persisted evidence to explain a DotCraft failure. Check operational logs first. For thread failures, treat rollout JSONL as the authority for thread history and use state.db for projections, runtime state, and traces.

Safety

• Keep diagnosis read-only. Do not edit logs, state.db, or rollout files unless the user asks for repair.
• Read SQLite with mode=ro; copy evidence to a temporary directory before using a client that may write.
• Do not dump prompts, model output, tool arguments/results, provider history entries, credentials, or raw trace JSON. Use bounded sanitized previews only when necessary.
• Cite log timestamps/categories, rollout lines/items, or trace event IDs for each conclusion.

Workflow

1. Set the failure window. Record the approximate local time, workspace, component, and any thread, Turn, request, process, or connection ID.
2. Check operational logs. Read the newest files around that window:

• Hub: ~/.craft/logs/dotcraft-hub-*.log
• Workspace host: <workspace>/.craft/logs/dotcraft-*.log
• Read the applicable Logging.Directory setting before assuming logs. Logging.Enabled, retention, or a custom directory can explain missing files.
• For managed AppServer startup, inspect both the Hub log and the target workspace log.
• Correlate timestamp, severity, PID, category, and scopes such as Module, WorkspacePath, RequestMethod, RequestId, ThreadId, and TurnId.

1. Inspect thread evidence when relevant. Run the bundled summarizer against the matching rollout and database:

python path\to\dotcraft-error-diagnosis\scripts\analyze_dotcraft_thread.py `
  --state-db "D:\path\to\workspace\.craft\state.db" `
  --thread "D:\path\to\workspace\.craft\threads\active\thread_x.jsonl"

1. Reconstruct current thread state.

• Apply turn_state_replaced by Turn ID and let it replace earlier incremental state for that Turn.
• Apply thread_rolled_back before reporting Turns; ignore rolled-back tail Turns as current evidence.
• Report model and provider history only as schema, boundary, count, source, reason, and validation metadata.

1. Correlate state and traces.

• Read trace_session_bindings for root_thread_id = thread_id. The main session often has binding_kind = threadMain; subagents and maintenance forks may have different session keys.
• Inspect trace_sessions counters, duration, finish reason, maintenance forks, and prompt drift.
• Query trace_events by session_key and timestamp. Prioritize events with type = 'Error', tool failures, unusual finish reasons, or a failed request immediately before the rollout Error item.

1. Conclude from the strongest evidence. Separate a confirmed root cause from an inference. Recommend the smallest durable fix and relevant regression test.

Output

• Finding: root cause or most likely failure class.
• Evidence: 3-6 bullets with timestamps, rollout line/item IDs, trace event IDs, and table names.
• Fix: concrete remediation steps.
• Residual risk: what remains uncertain and what extra evidence would settle it.