[session-manager] session_info and session_read return inconsistent results due to transient SDK API failures

[B1INGO]

Bug Description

session_info and session_read tools return inconsistent results for the same session ID when the OpenCode SDK HTTP API experiences transient failures. Specifically:

1. session_info succeeds and returns session metadata
2. session_read immediately fails with "Session not found" for the same session ID

Root Cause Analysis

After deep investigation, the issue stems from dual storage architecture + transient SDK API failures:

Storage Architecture

• SDK Storage (SQLite DB): ~/.local/share/opencode/opencode.db - contains full session data
• File Storage: ~/.local/share/opencode-data/opencode/storage/message/ - fallback storage

For sessions created purely in SQLite backend (isSqliteBackend() === true), no file storage exists as fallback.

Failure Chain

When SDK API transiently fails:

1. session_info �� getSessionInfo() �� SDK fails �� fallback to getFileSessionInfo() �� file doesn't exist �� returns null �� "Session not found"
2. session_read �� sessionExists() �� SDK's session.list() fails �� fallback to fileSessionExists() �� file doesn't exist �� returns false �� "Session not found"

Both paths fail because:

• SDK API is temporarily unavailable (timeout, network hiccup, race condition)
• File storage has no data for pure-SQLite sessions

Key Code Locations

In dist/index.js:

• Line 107196-107208: sessionExists2() - existence check with SDK + file fallback
• Line 107210-107223: readSessionMessages2() - message read with fallback
• Line 107102-107132: getSdkSessionMessages() - SDK message fetch (no retry)
• Line 107496-107521: session_read tool definition

Evidence

# Session exists in SQLite with full data
sqlite3 ~/.local/share/opencode/opencode.db "SELECT COUNT(*) FROM message WHERE session_id='ses_187699fafffe5TjZ0wrO97d21b'"
# Result: 157 messages, 485 parts

# But file storage is empty for this session
ls ~/.local/share/opencode-data/opencode/storage/message/ses_187699fafffe5TjZ0wrO97d21b
# Result: directory does not exist

Proposed Solution

Add retry logic for SDK API calls in session tools, since transient failures are expected in networked environments:

// In session_read tool (line ~107504)
execute: async (args, _context) => {
  const maxRetries = 2;
  const retryDelay = 500; // ms
  
  for (let attempt = 0; attempt < maxRetries; attempt++) {
    if (await resolvedDeps.sessionExists(args.session_id)) {
      break;
    }
    if (attempt < maxRetries - 1) {
      await new Promise(r => setTimeout(r, retryDelay));
    }
  }
  
  if (!await resolvedDeps.sessionExists(args.session_id)) {
    return `Session not found: ${args.session_id}`;
  }
  // ... rest unchanged
}

Alternatively, add retry logic directly in getSdkSessionMessages() and getSdkSessionInfo() functions.

Environment

• OpenCode version: 1.15.12
• oh-my-openagent: latest (installed via npm)
• OS: Windows 11
• Storage backend: SQLite (isSqliteBackend() === true)

Workaround

Wait a few seconds and retry the session_read call - the SDK API typically recovers quickly.

Related

This issue particularly affects sessions that were created in SQLite backend without corresponding file storage. The fallback mechanism works correctly, but for pure-SQLite sessions there's nothing to fall back TO when SDK fails.