Context
docs/goals/post-release/13-low-latency-turn-delta-sync.md (committed in 314a9f4, status urgent, not implemented, owners Tendwire and Herdres) specifies the cache-only, host-scoped change feed that replaces full turn-list traversal on every herdres sync. We hit exactly the production evidence it describes (11–25s idle passes, minutes of user-visible delay) and want to take on the implementation, treating the goal doc as the authoritative spec.
What we'd build (per the doc)
- Change feed over the current turn projection reporting insertions, updates (Working text/status/revision/ownership), and removals as explicit tombstones — durable opaque mutation watermark, bounded pages under the 1 MiB frame cap, atomic page/watermark reads, explicit invalid/expired/cross-host/incompatible outcomes, no private identifiers, no canonical content copies (Goal 05 stays authoritative), no change to Goal 10 final-delivery authority.
- Herdres side: persisted watermark, checkpoint-after-durable-apply, single bounded bootstrap on invalid/expired, fast delivery lane decoupled from slow maintenance.
- Acceptance per the doc's 11 criteria, including the 10k-turn seed proofs and the p95 targets (≤350 ms API work, ≤5 s ingested-change → provider operation).
Process
Same flow as #8: multi-model design review first, implementation on the jerryfane fork with adversarial review + mutation-tested regressions, then a PR here. Flagging before writing code in case you have design preferences beyond the doc (watermark encoding, RPC naming, journal vs trigger capture) — happy to align on anything upfront.
🤖 Generated with Claude Code
Context
docs/goals/post-release/13-low-latency-turn-delta-sync.md(committed in 314a9f4, status urgent, not implemented, owners Tendwire and Herdres) specifies the cache-only, host-scoped change feed that replaces full turn-list traversal on every herdres sync. We hit exactly the production evidence it describes (11–25s idle passes, minutes of user-visible delay) and want to take on the implementation, treating the goal doc as the authoritative spec.What we'd build (per the doc)
Process
Same flow as #8: multi-model design review first, implementation on the jerryfane fork with adversarial review + mutation-tested regressions, then a PR here. Flagging before writing code in case you have design preferences beyond the doc (watermark encoding, RPC naming, journal vs trigger capture) — happy to align on anything upfront.
🤖 Generated with Claude Code