Summary

CLI session IDs are stored under cliSessionIds["unknown"] but looked up via cliSessionIds["claude"] (or whichever runtime). This mismatch means --resume is never called, causing all CLI sessions to collapse into a single Claude project session.

Root Cause

resolveDefaultModel() in directive-handling.persist.ts:126 returns defaultProvider: "unknown" (model system was gutted). This flows through:

1. get-reply.ts:44 → provider = "unknown"
2. agent-runner.ts:414 → providerUsed = "unknown"
3. session-usage.ts:22 → setCliSessionId(entry, "unknown", cliSessionId)

Read side uses resolveAgentRuntimeOrThrow() → "claude", so the lookup always misses.

Fix

Replace resolveDefaultModel() stub with runtime-based resolution. Use resolveAgentRuntimeOrThrow() (or resolveAgentRuntime()) as the provider identifier throughout the session usage pipeline. This is the first step of gutting the vestigial model/provider plumbing.

Acceptance Criteria

• Given a thread session runs with Claude runtime
  When the CLI session ID is persisted
  Then cliSessionIds is keyed by "claude" (not "unknown")

• Given a thread session has a stored CLI session ID
  When a second message arrives in the same thread
  Then --resume <sessionId> is passed to the Claude CLI

• Given two different Slack threads
  When each receives a message
  Then each gets a separate Claude CLI session (different session IDs)

Files

• src/auto-reply/reply/directive-handling.persist.ts:119-127 — resolveDefaultModel() stub
• src/auto-reply/reply/agent-runner.ts:441-442 — providerUsed derivation
• src/auto-reply/reply/session-usage.ts:22 — setCliSessionId call