[mvanhorn]

Summary

An idle watch-mode source stops re-sending the same envelope to every sink every 30 seconds, and slow links stop failing at a hardcoded 60-second push budget. On the reporter's install, cookie-only pushes still cost about 1.8 MB per 30 s (about 5 GB a day) with nothing changing. On uplinks slower than about 12 Mbit/s, each push hit context deadline exceeded (6,356 times in one log), well under the 5-minute SyncClient timeout v1.1.0 set.

• The source remembers, per sink (peer and URL), a hash of the last envelope that sink accepted. A sink that already holds the same payload is skipped, with a heartbeat at least every 10 minutes. Sequence and cookie access times are left out of the hash, so Chrome bumping last_access_utc on open tabs does not count as a change. The sent payload still carries the access times.
• Only a successful POST counts as delivered. A failed POST forgets that sink's last delivery, so it is always retried and never skipped on an older acknowledgement.
• Watch-mode pushes get the same budget as --once (SyncClient timeout plus 30 s). The watcher skips a trigger while a push is still running, so slow pushes cannot pile up; the next 30 s tick picks up anything skipped.
• A Linux sink with live CDP on now re-injects the cookies from its last /sync into the running Chrome every 30 s, locally, applying its current cookie policy. The source's 30 s re-push used to do this as a side effect, covering browser contexts an agent opened since the last sync and a Chrome that restarted. Without it, dedup would leave those logged out until the heartbeat. Re-injects are serialized with /sync, so a newer set always lands last.

Reported by @Dannyxlm in #149 (items 2 and 3).

Related: #149

Verified

• go build ./...
• go vet ./...
• go test -race ./...
• golangci-lint v2.11.4 (the CI pin): 0 issues

Dedup tests drive the real fan-out on a stub clock:

• two identical pushes inside 10 minutes make one POST, and one past 10 minutes POSTs again
• a changed cookie POSTs at once, while a change only to last_access_utc does not
• a 500 is re-sent, and a payload that follows a failed POST is not skipped on an earlier acknowledgement
• with two sinks where one failed, only the failed sink is re-sent
• Sequence strictly increases across real POSTs
• a slow POST does not bring the heartbeat forward

Watcher tests check that the push deadline is the configured budget and that only one push runs at a time. Re-injector tests cover re-inject after /sync and nothing before it, no ticker when live CDP is off or on --dry-run, failure and recovery each logged once, the current policy applied each tick, and a newer /sync landing after an in-flight tick.

Code review: ce-code-review at full depth, with correctness, reliability, testing and maintainability reviewers plus an independent Codex adversarial read (gpt-6-luna, independence verified).

• Verdict: Ready with fixes.
• All six actionable findings are fixed on this branch: the tick//sync race, forgetting a delivery after a failed POST, timestamping deliveries per sink, the shared endpoint helper, splitting the new code into source_dedup.go and sink_reinject.go, and a dry-run guard test.

Notes

• A restarted source sends everything once at startup, since the delivery log is in memory only.
• A sink restart loses the re-inject set until the source's next change or heartbeat (at most 10 minutes).
• doctor only warns after 24 h without a push, so status stays fresh.
• No config keys and no wire-format change.
• Item 1 of Source ships ~90 MB of Local Storage per push to Linux sinks (plus three related sync-path findings) #149 is the --skip-local-storage change in fix(source): cookie-only source with --skip-local-storage and --skip-secrets #153, and item 4 is fix(config): honor live_cdp.enabled: false on Linux sinks #152, so Source ships ~90 MB of Local Storage per push to Linux sinks (plus three related sync-path findings) #149 can close once all three merge.

---

🤖 Generated with Claude Code

https://claude.ai/code/session_01LgVWtkXgDGWMh6L95d8piN

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new blocking issue was found.

Summary

The source skips unchanged deliveries per sink until a 10-minute heartbeat and gives watch pushes the same budget as one-shot pushes.

• Sources skip unchanged payloads and send a heartbeat every 10 minutes.
• Watch pushes get the source --once time budget without piling up.
• Live-CDP sinks re-inject their last synced cookies into Chrome every 30 seconds.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Source trigger] --> B{Push already running?}
  B -->|Yes| C[Skip trigger]
  B -->|No| D[Read current payload]
  D --> E{Sink has same payload within heartbeat?}
  E -->|Yes| F[Skip POST]
  E -->|No| G[POST to sink]
  G -->|Success| H[Remember delivery and time]
  G -->|Failure| I[Forget delivery]
  H --> J[Hold accepted cookies at sink]
  J --> K[Every 30 seconds: check current policy]
  K --> L[Inject into live Chrome]

Loading

_{Reviews (4) · Last reviewed commit: "Merge remote-tracking branch 'origin/mai..." · Reviewed by Greptile}