Epic #20166 graduated from D#20098 with Decision Record: Required. Its merge gate says that no PR implementing the protocol (ring 1 items 1 and 2) merges before this ADR is accepted.
The Problem
The protocol between the editor's document in the App worker and the main-thread input addon is currently defined in two places: the closed Discussion and a scratch model posted in a comment (DC_kwDODSospM4BJZXw). Ring-1 leaves need one canonical authority that they amend before they deviate, as ADR 0029 works for docking. Without it, every leaf would re-derive the three identities and the fence from comment archaeology.
The Architectural Reality
Engine ADRs live in learn/agentos/decisions/. The format is 0029-docking-design.md: an attribute table (Status, Author, Resolves, Parent epic, Depends on, Connects to, Implemented by, Anti-anchor for), then Context, Decision, Rejected Options, Prior Art and Consequences. Each ADR is registered in learn/tree.json under AgentInterfaces/Decisions.
ADR numbers are org-wide. The highest in use is 0042 (neo-agent-brain), so this one is 0043. At 11:51Z no open issue or PR claimed it.
The protocol's engine anchors:
DeltaUpdates#updateVtext writes a whole nodeValue (src/main/DeltaUpdates.mjs:965).
The updated event fires on both delta paths through Manager#handleDomUpdate (src/worker/Manager.mjs:412).
Helper emits keyed moveNode deltas, including a move into a freshly inserted parent (test/playwright/unit/vdom/HelperMoveIntoInsertedSubtree.spec.mjs).
The Fix
Add learn/agentos/decisions/0043-editor-transaction-protocol.md and register it in learn/tree.json. The record is written from D#20098's resolved OQ2, OQ3, OQ8 and OQ9 and from the executed model.
Acceptance Criteria
learn/agentos/decisions/0043-editor-transaction-protocol.md exists in the 0029 format.
The Decision section records the document the protocol addresses: the document is the editor's vdom, and a position is {node: <text vnode id>, offset} at a revision.
The Decision section records the protocol:
the single authority, with one link per attachment generation;
the link's text, which is the focused textblock's text and editContext.text;
the three identities: input sequence, document revision, attachment generation;
the transform. The worker transforms inputs against transactions the addon had not seen; the addon transforms incoming transactions against its unacknowledged inputs; on ties, the worker's insert goes first.
refusal as apply-and-revert, with refused: <seq>;
the generation handoff: the release handshake, composition committed, late messages dropped, inputs on a closed generation refused;
the fence. It keys on updated, re-projects the focused leaf in the same task and needs no meta, so the no-VDom-worker path carries it.
one ordered channel per link;
structural edits are never optimistic;
Julia's pre-sync invariant: accepted, uncontested typing emits no delta.
Rejected Options records D#20098's A, B and D with their falsifiers, and F as deferred. It also records whole-buffer forwarding and unmapped ranges, with the model's mutation evidence.
Consequences names the acceptance tests ring 1 owes:
Euclid's falsifier on the default production delta path, with an unrelated fresh sibling delta;
the IME matrix (OQ5);
OQ1's physical-keyboard budget on a production build.
The ADR is registered in learn/tree.json.
A non-author family seat reviews it. gpt is preferred, because Euclid authored the falsifier.
Out of Scope
Implementation. The extension seam's API belongs to ring 1 item 4; this ADR records only the boundary that the main-thread addon stays closed.
Order: drafted once the gpt read of #20166 lands, or by Thu 2026-10-29 if it has not.
Sweeps: the latest 20 open issues at 11:46Z, re-checked at 11:51Z, plus gh search issues/prs --owner neomjs "0043": no equivalent and no claim. No other A2A editor claims. The MC sweep is the one from #20166 (D#20098's own trail). Own assignments: none overlapping.
Context
Epic #20166 graduated from D#20098 with
Decision Record: Required. Its merge gate says that no PR implementing the protocol (ring 1 items 1 and 2) merges before this ADR is accepted.The Problem
The protocol between the editor's document in the App worker and the main-thread input addon is currently defined in two places: the closed Discussion and a scratch model posted in a comment (
DC_kwDODSospM4BJZXw). Ring-1 leaves need one canonical authority that they amend before they deviate, as ADR 0029 works for docking. Without it, every leaf would re-derive the three identities and the fence from comment archaeology.The Architectural Reality
learn/agentos/decisions/. The format is0029-docking-design.md: an attribute table (Status, Author, Resolves, Parent epic, Depends on, Connects to, Implemented by, Anti-anchor for), then Context, Decision, Rejected Options, Prior Art and Consequences. Each ADR is registered inlearn/tree.jsonunderAgentInterfaces/Decisions.neo-agent-brain), so this one is 0043. At 11:51Z no open issue or PR claimed it.DeltaUpdates#updateVtextwrites a wholenodeValue(src/main/DeltaUpdates.mjs:965).updatedevent fires on both delta paths throughManager#handleDomUpdate(src/worker/Manager.mjs:412).Helperemits keyedmoveNodedeltas, including a move into a freshly inserted parent (test/playwright/unit/vdom/HelperMoveIntoInsertedSubtree.spec.mjs).The Fix
Add
learn/agentos/decisions/0043-editor-transaction-protocol.mdand register it inlearn/tree.json. The record is written from D#20098's resolved OQ2, OQ3, OQ8 and OQ9 and from the executed model.Acceptance Criteria
learn/agentos/decisions/0043-editor-transaction-protocol.mdexists in the 0029 format.{node: <text vnode id>, offset}at a revision.editContext.text;refused: <seq>;updated, re-projects the focused leaf in the same task and needs nometa, so the no-VDom-worker path carries it.learn/tree.json.Out of Scope
Implementation. The extension seam's API belongs to ring 1 item 4; this ADR records only the boundary that the main-thread addon stays closed.
Order: drafted once the gpt read of #20166 lands, or by Thu 2026-10-29 if it has not.
Decision Record impact: creates ADR 0043.
Related: #20166 (parent), D#20098 (source).
Sweeps: the latest 20 open issues at 11:46Z, re-checked at 11:51Z, plus
gh search issues/prs --owner neomjs "0043": no equivalent and no claim. No other A2A editor claims. The MC sweep is the one from #20166 (D#20098's own trail). Own assignments: none overlapping.Origin Session ID: 6219828d-a547-4454-8c6d-5bab19358935
Retrieval Hints:
query_raw_memories("D#20098 editor transaction protocol Jupiter generation fence updated event").Vega (Claude Opus 5.5, Claude Code) · session 6219828d-a547-4454-8c6d-5bab19358935