openclaw - 💡(How to fix) Fix Runtime model selection sticks on fallback after `reason=abort`; `/new` does not clear active model

Summary

After a Claude live session closes with reason=abort, the Telegram gateway flips to the configured fallback model (openai-codex/gpt-5.5) and remains there across subsequent turns. A user-issued /new resets the conversation transcript but does not restore the policy primary — the new session continues on the fallback. Active model only returns to primary after manual intervention (e.g. /model claude-cli/claude-opus-4-7).

This violates the implicit contract that primary is the steady-state model and fallback is a temporary recovery state.

Version

[email protected]

Repro

1. Agent main configured with policy agent.chat.claude-cli → primary claude-cli/claude-opus-4-7, fallback openai-codex/gpt-5.5.
2. Abort an active Claude turn (or send SIGINT mid-turn).
3. Send a new prompt several minutes later.

Observed

• 19:07:04 UTC — gateway log: [agent/cli-backend] claude live session close: provider=claude-cli model=claude-opus-4-7 reason=abort
• 19:30:25 UTC — gateway log: [agent/embedded] [timeout-compaction] compaction succeeded for openai-codex/gpt-5.5; retrying prompt (already on fallback by this point — switch happened on the next-turn fallback selection between 19:07 and 19:30)
• 19:31:13 UTC — turn rendered as [Reghar-openai-codex/gpt-5.5] (Telegram banner)
• 19:42:44 UTC — user runs /new. New session opens on gpt-5.5/openai-codex (per session reset banner).
• 19:44 UTC and subsequent turns — continued on fallback until manual /model command.

Expected

• After the abort cooldown elapses or after /new, the next turn re-attempts the primary model.
• Either:
  • /new clears the runtime active_model so the next turn re-reads the policy primary; or
  • The fallback marker carries a TTL / success-count and reverts to primary after N successful fallback turns or T minutes.
• Banner surfaces drift (e.g. [Reghar-... ⚠️ on fallback]) when active != primary so users notice without checking config.

Code surfaces (from dist/ inspection of v2026.5.12)

• dist/pi-embedded-CPEBK2iK.js:~2579 — [timeout-compaction] falls through to failover rotation once MAX_TIMEOUT_COMPACTION_ATTEMPTS is reached; rotation appears to select fallbacks[0] without first re-attempting primary.
• dist/model-fallback-CFP2h22a.js:~504 — suspendSession; suspected to be marking the primary in a cooldown state that survives session reset.
• dist/model-fallback-CFP2h22a.js:~893 — shouldProbePrimaryDuringCooldown exists, suggesting the auto-revert path is implemented but not firing in this scenario. This is the likely root cause: the probe path isn't being entered after an abort-induced fallback.
• Banner template: openclaw.json responsePrefix: \"[<Agent>-{provider}/{model}]\" — no drift flag in template variables.

Proposed fixes (any subset)

1. Compaction retry: on falling through to failover rotation, retry primary before any fallback unless primary has a recent hard failure (not just an abort).
2. /new clears active model marker: session reset hook nulls the cached active model so the next turn re-evaluates policy.
3. Fallback TTL / revert: stamp the fallback marker with set_at and revert_after (default 15 min OR 2 successful fallback turns). On new turn after revert condition, attempt primary.
4. Drift banner: extend template to expose {is_fallback} or render ⚠️ on fallback (primary={...}) when active != policy.primary.

Workarounds for users today

• /model claude-cli/claude-opus-4-7 to force-restore primary.
• Restart the gateway (clears session state entirely).

Repro-environment context

• macOS Darwin 24.6.0
• [email protected]
• Telegram channel, agent main, agentRuntime.id = claude-cli.