fix(goal_frontier): carry the agent-owned frontier identity into writeback ACK checkpoints - #4645
huangruiteng wants to merge 1 commit into
Conversation
…k ACKs The long-chain ACK fence already matches on the identity over the rows this agent owns (`frontier_owned_identity`, added in #4610), but the typed-delta writeback path never carried that identity: - `LongTodoChainObservation.to_trigger()` omitted it, so the recorded obligation trigger only exposed `frontier_revision`; - `replan_obligation_trigger_checkpoints()` rebuilds the ACK checkpoints from those triggers, so writeback ACKs recorded a revision-only checkpoint. The revision covers the whole selectable set, which also contains rows nobody has claimed yet. Any peer lane claiming or editing such a row moves the revision, so a writeback ACK re-armed the pending `long_todo_chain` replan obligation on the next wake even though this agent's own work basis had not changed. Observed live: the 15:24 ACK for `replan-132c272361cab7f1` recorded `trigger_checkpoints: [{kind, frontier_revision}]` only, and the next guard reported `rearmed_after_obligation_id: replan-132c272361cab7f1`. The prompt-path ACK (`replan_successor_transition_ack`) was unaffected because it composes `long_todo_chain_source_checkpoint`, which already carries the identity; this change makes both ACK paths record the same fence. Validation: `tests/control_plane/test_goal_frontier_replan_rules.py::test_writeback_ack_owned_identity_absorbs_a_peer_claim` reproduces the peer-claim re-arm and fails on either dropped hop (verified by reverting each half independently); 151 focused control-plane tests pass. Signed-off-by: huangruiteng <14976749+huangruiteng@users.noreply.github.com>
There was a problem hiding this comment.
This exact-head review is superseded by the current owner-requested implementation assessment. Head: 32082e9.
The patch repairs two dropped-field hops, but checkpoint construction still has separate Python paths and the successor path repeats frontier reads. A replacement is being developed with a single TypeScript checkpoint owner and real writeback/readback validation. The previous review also overstated timestamp semantics: maintenance timestamps do not change the material revision; owned material fields do.
Historical validation reported 151 focused tests for this patch. That evidence does not qualify the forthcoming replacement head, live deployment, or the complete collaboration journey. No merge readiness is asserted here.
English verdict: Superseded assessment; hold this PR for the typed checkpoint replacement and its exact-head validation.
|
Superseded by #4655. The replacement retains the owned-frontier ACK fix, unifies checkpoint construction in the existing TypeScript owner, and covers successor causality through real CLI and canonical File-provider readback. Wrong origins, stale/ambiguous successors and incomplete sources remain rejected. |
1 participant
A semantic writeback ACK drops the agent-owned frontier identity between long-chain observation and persisted checkpoint. A peer claiming shared unclaimed work can therefore rearm an already acknowledged chain even though the agent-owned material basis is unchanged.
The two-field propagation fix is being superseded by a cohesive typed checkpoint refactor: one TypeScript owner for observation, source and writeback checkpoint construction; Python retains legacy source encoding and IO adaptation. The replacement will qualify both ACK paths through synthetic real-entrypoint tests before this PR is retired.
Compatibility must preserve revision-only historical ACKs and rearm on changes to owned material work. Maintenance timestamps alone do not alter material identity. The initial patch does not qualify end-to-end recovery or complete the wider collaboration roadmap. Maintainer review and merge remain separate.