perf(agents): bound autonomous sessions and reply overhead

[malkovitc]

Summary

• reduce default ACP conversation context from 12 messages to 6 and proactively rotate sessions after 8 turns
• disable Codex app/plugin injection for autonomous managed agents and start bypass-permission Codex sessions directly in full-access mode
• reduce prompt-driven ceremony and duplicate validation; let buzz messages send --reply-to infer Forum comment kind from its parent

Production canary

Linza managed-agent [QUESTION] canary after refreshing the 1.4.0 Codex ACP adapter:

• exact response, no duplicate publication
• 26 seconds end-to-end
• 2 model calls
• 1 tool call (final Buzz publication)
• 44,006 cumulative tokens, 32,256 cached input tokens
• historical unbounded baseline: 112 seconds, 10 model calls, 325.7k processed tokens

Validation

• cargo fmt --all -- --check
• cargo test -p buzz-acp -p buzz-cli — 802 ACP tests passed before one load-sensitive keepalive timing test failed while the host was compiling; isolated retry passed after load settled
• cargo test -p buzz-acp acp::tests::keepalive_resets_idle_past_deadline -- --exact
• cargo clippy -p buzz-acp -p buzz-cli -- -D warnings
• pnpm --dir desktop typecheck
• desktop test suite: 5,397 passed
• git diff --check origin/main...HEAD

The keepalive test was also run on unmodified origin/main; it is timing-sensitive to host load and passed there without source differences in acp.rs.

[malkovitc]

Final packaged-runtime evidence: combined current-main build e4ab80873 was installed as Buzz Linza Patched 0.5.18. All 13 production Linza agents started with context=6, max turns=8, full-access initial mode, apps/plugins/remote_plugin disabled, and no stale adapter nudge. A fresh Caliper [QUESTION] canary completed in 18s with exactly one published reply, 2 model calls, 1 Buzz publication tool call, and 43,999 cumulative tokens (32,256 cached). The historical baseline was 112s / 10 model calls / 325.7k processed tokens.

[malkovitc]

Follow-up production finding and fix (1b9e0bf90): a Conductor Kanban status request previously consumed 361,430 cumulative tokens across 11 model calls / 10 tool calls in 157s and published an empty reply. Root causes were repeated full-board MCP downloads (including a 31k-token tool payload), waits/retries, and --content - with no stdin. The CLI now rejects blank message bodies (while preserving attachment-only messages), with regression tests. Operationally, Kanban reads now use one compact cached filtered command and safe piped publication. Final same-question canary: correct non-empty single reply in 29s, 3 model calls, 2 tool calls, 68,056 cumulative tokens (53,504 cached; 13,808 fresh input): ~81% fewer processed tokens and ~58% less fresh input.

[Chessing234]

this silently flips max_turns_per_session from 0 (disabled) to 8 and cuts context 12→6. that's a product default change, not just a bound — please call it out or leave the defaults and make the cap opt-in.

[malkovitc]

Closing: these agent-efficiency changes are part of our internal downstream Buzz customization and will be maintained in our fork rather than submitted upstream.