feat(agent-manager): add incremental conversation tailing

[codeaholicguy]

Summary

• add one asynchronous conversation-tail API with bounded caching and deterministic read diagnostics
• incrementally tail Codex JSONL by file identity and byte offset, including partial records, truncation, rotation, parse errors, LRU eviction, and mirrored-message deduplication
• use SQL LIMIT for OpenCode, worker-thread JSON parsing for Gemini, and stale-result rejection in useAgentConversation
• retain the 20-message preview, unchanged-source caching, and polling fallback

Benchmark

82.7 MiB Codex fixture, 86,682,730 bytes and 11,747 records:

• legacy full parse median: 300.5 ms across five runs; prior measured hotspot: 336.6 ms
• incremental initial load: 241.4 ms
• one appended 130-byte record: 0.246 ms
• unchanged refresh: 0.029 ms

Run with:

npm run benchmark:conversation-tail --workspace packages/agent-manager -- <session.jsonl>

Validation

• agent-manager: 27 files, 526 tests passed sequentially
• CLI after rebase onto current main: 81 files, 970 tests passed
• focused conversation-tail tests: 5 files, 104 tests passed
• agent-manager and CLI lint/typecheck passed; CLI retains five pre-existing warnings outside touched files
• six-project monorepo build passed
• feature-doc lint and diff check passed
• regression proof confirmed the same-size rewrite test fails without reset detection and passes with it restored

Adapter support

Optimized now:

• Codex: reusable incremental JSONL cache
• OpenCode: displayable-row SQL query with LIMIT
• Gemini: off-thread monolithic JSON read/parse

Compatibility fallback:

• Claude, Copilot, Grok, and Pi preserve current conversation semantics and unchanged-file caching through the shared async manager API; changed files still receive a deferred full parse

Follow-up: migrate those four JSONL adapters to the reusable reducer/cache with parity tests. No second UI or adapter architecture is required.

Risks

• Codex mirror-key state grows for the lifetime of each cached session; session count is bounded by the 50-entry LRU
• the four compatibility adapters still do full changed-file parses until their reducer follow-ups land
• the unrelated print-agent integration requires a longer allowance on heavily loaded hosts because process inspection can exceed its default 5-second timeout