fix(agents): prevent OOM from cumulative tool history (#4838)

* fix(agents): prevent OOM from cumulative tool history

Repeated progress updates retained complete historical copies of growing tool output. Retain projected items on ingestion and serve them for every history fetch, preserving source sequence coverage for catch-up. Keep intermediate payloads only in streaming; confine compatibility degradation to child transcripts and overwritten fork checkpoints.

* test(client): expect projected subagent capability in hello

* fix(app): report unexpected child history failures

Keep connection failures recoverable while forwarding unexpected fetch and reconciliation errors to diagnostics. Name observer arguments to keep server and agent identities explicit.

* test(agents): assert projected history after rewind

* docs(protocol): tag projected child compatibility gates
This commit is contained in:
Mohamed Boudra
2026-09-14 16:51:10 +02:00
committed by GitHub
parent d1a53255fc
commit cdc669fde6
26 changed files with 1190 additions and 713 deletions
+22 -3
View File
@@ -5,8 +5,10 @@ Agent chat delivery has two paths:
1. **Live stream** — `agent_stream` WebSocket messages for immediacy. These may be delta-shaped lifecycle updates.
2. **Authoritative history** — `fetch_agent_timeline_request` for correctness. This always returns full projected timeline items, never lifecycle deltas.
The daemon keeps canonical rows only for its runtime. Provider history is the durable transcript
authority and repopulates those rows when an agent resumes.
The daemon retains projected items in memory. Each source event advances the stream sequence,
then replaces the previous tool state or merges into the current text item. Intermediate payloads
are never retained for history or catch-up. Provider history is the durable transcript authority
and rebuilds the projection when an agent resumes.
The invariants are:
@@ -67,7 +69,10 @@ Provider message IDs are not guaranteed for every displayed item. Paseo-generate
Actions that address a point in chat history, such as Fork, use the daemon timeline `epoch` plus the projected item's `seqEnd`. The app carries that position on the rendered assistant item for both live and fetched history. When adjacent projected chunks merge, the merged item retains the newer chunk's position.
The daemon validates that the epoch is current and the exact source sequence still exists before slicing rows. It slices before projection so later lifecycle updates cannot leak into the selected context.
The daemon validates the epoch and locates the projected item at the selected position. A fork
includes projected items through that checkpoint. If an item spans the checkpoint and changed
afterward, the daemon refuses that fork with an actionable error: discarded historical payloads
cannot be reconstructed from sequence metadata. Forking the current context remains available.
## Resume behavior
@@ -93,6 +98,20 @@ history replay. The pending approval UI can hide that tool from presentation, bu
must retain its position. Creating a new history card on rejection places it after the prompt that
rejected it; changing steer-event ordering would also put new assistant output before that prompt.
## Provider child history
Child transcripts use the same projected-page reconciliation as the main conversation. The client
retains rendered items and sequence cursors, never a second cache of source events. Pagination
uses the projected display anchor; live updates advance the source cursor without moving tools.
Clients advertising `projected_subagent_timeline` receive child streams and projected fetches.
Updated clients require `features.projectedSubagentTimeline` for child history; older hosts show
an update-host notice for that pane.
Older clients retain child names/status and receive an upgrade message when opening a child
conversation. This gate affects only the child transcript; it does not gate the connection,
main conversation, or current-context forks. A legacy `canonical` root request still receives
projected items.
## Client replica lifetime
The session projection remains host-scoped for as long as the host is registered. The viewed-timeline