Repository navigation
Sonny: fleet observation costs grow with what is watched, not the square of the fleet - #538
Conversation
…are of the fleet From the fleet performance review of 1aeb7e5: - Lookout: one subscriber registry every lookout shares, instead of a copy of all the other subscribers in each (S x S names); new subscribers drop news by looking up only what a lookout holds; bots are seated at start through the watcher index, and a new lookout joins the index alone. - /status: lists every page of agents (the 257th was missing), and an idle agent whose record (head, status, turn) has not moved is not read again. - Sky: the replay tape stays within TAPE while the past loads and while paused (unplayed events go too, and the replay is seated again from the new start); a frame drains at most half a frame's work; the hour's pair counts are kept as messages come, not counted again each frame. - A lead's layout follows every agent from now, for the event kinds it acts on, instead of replaying the whole log from cursor 0 to skip it. - Draws for the daemon's events are capped at 30 a second (Ink's cap); keys still draw at once. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 971562bc2d
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- sonny.tool_processes (0 = twice the CPUs): tool calls the tool host answers in `sonny tool` processes and goal checks take a place before they start and hold it until done, no agent more than half, waiting within the call's or check's own time. - /status reads every agent's turns each refresh again (a queued turn moves no field of the bot's record) and keeps only the latest history items, by history head. - A replay keeps who its let-go tape start made apart, so the listing read once the past is in does not drop an agent deleted later on the tape. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA
# Conflicts: # sonny/FEATURES.md
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f7a462a8ba
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…o share /status kept a failed history read as an empty one, so an idle agent's latest step stayed hidden after a reconnect. The read now says when it failed, and only a good read is kept. The process budget's per-agent share is now counted by its holders and waiters, so an admission that times out before its place comes removes the share too. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 5d7c3d6901
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The last listing dropped a folded agent whenever it named someone with the same name. Folded agents are now matched by name and id, as seek() tells incarnations apart, so the original stays seated beside its replacement. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 1b3d208d5f
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The tool host and the lookout are separate processes, each with its own budget, so together they run at most twice sonny.tool_processes. The module, the /config line and the feature log now say so. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA
Requested by George · project thread
This PR covers the Sonny-side findings from the external fleet performance review of 1aeb7e5. The daemon-side findings are being built separately, in the order agreed with George.
What changes
others()used to clone every other subscriber's name into each lookout's ownHashSet, which is S × (S − 1) owned names: about 1.9 GiB of names alone at 10k subscribers. Now there is one sharedArc<HashSet<String>>, and the "not this lead" check is a predicate. When new subscribers appear,Updates::subscribedonly looks up the names that lookout already holds news of. At startup, bots are seated through the watcher index instead of every lookout asking about every bot, and a newly placed lookout is added to the index by itself./statusreads every page. It read one 256-agent page without followingnext_after, so the 257th agent never showed. It now pages through. An agent's latest history items (2 of its 3 calls per refresh) are re-read only once its history head moves, and a failed read is retried on the next refresh. Its turns are still read every time, because a queued turn leaves the bot record unchanged.Replay::keeponly evicted once 10% of the tape had played, so neither the initial load nor a paused replay was bounded (200k loaded kept 200k; paused plus 200k more kept 280k). It now evicts the oldest tenth pastTAPEregardless. If unplayed events go, acutflag reseats the replay from the new start. The creations it folds are kept separately, matched by name and id, so the final listing doesn't drop them.drain()does at most half aFRAMEof work per call. The hour's pair counts are kept assaidchanges, sohear(),land(), the curves and the stats look a pair up instead of scanning the hour.Fleetused to follow*from cursor 0 with no kind filter, replay the whole log only to skip it, and decode every token event of every project. It now follows from now (since_ms) with only the 11 kinds it acts on. Where agents currently stand still comes from the listing (reconcile).sonny.tool_processes, in/config; 0 means twice the CPUs). Each companion has its own: the tool host's tool calls insonny toolprocesses, and the lookout's goal checks. A call or check takes a place before it starts and holds it until done. No agent holds more than half, the wait counts against the call's or check's own time, and a wait that is given up leaves nothing behind (sonny/src/budget.rs).sonny/FEATURES.md: history entries on the lookout,/status, full screen, Rex layout, sky, goals and/configrows.Tests
status_lists_every_page_and_reads_again_only_the_histories_that_moved(it also covers a failed read being retried),the_hour_s_pair_counts_follow_what_was_said,who_the_tape_let_go_of_stays_in_the_cast_past_the_last_listing,who_the_tape_let_go_of_stays_though_its_name_is_used_again, the replay tape test extended with the reviewer's two cases, andbudgettests (one agent takes at most half; no share is left behind, even when a wait is given up).cargo test --bin sonnypasses.cargo clippy --workspace --all-targets -D warningsis clean.Still to come: a limit on auto-opened task windows with LRU close (after #533), browser sessions under a budget, and the perf gate covering
sonny/**.🤖 Generated with Claude Code
https://claude.ai/code/session_013jkdGF9fn2EXiBvzYrQtaA