fix(server): stop PR status lookups amplifying GitHub rate limits

[t3dotgg]

Problem

I keep getting throttled on the GitHub API. Digging through logs, most of it is
my own agents running gh in bursts, but T3 Code makes it worse in two ways.

The PR status badge caches a successful lookup for 2 minutes and a failed
one for 20 seconds. GitHub rejects a throttled request instantly, so once you're
rate limited the poller starts asking six times faster than it does when
everything is healthy. On top of that, lookupStatusPr catches the failure and
returns the last known PR, so the poll reports success and
VcsStatusBroadcaster's exponential backoff never engages. Nothing ever slows
down, and a transient 429 turns into a sustained one.

Separately, the badge asks the provider about every branch, including branches
that were never pushed anywhere. Those queries cannot match a PR. On this machine
66% of sampled t3code worktrees are in that state, so two thirds of the baseline
API spend was guaranteed-empty.

Solution

Failed lookups now back off per branch (20s, 40s, 80s, ... capped at 15 min)
instead of retrying on a flat 20s timer, so a failing branch drops below the
healthy refresh rate after a couple of attempts rather than above it.

Branches git has no record of on any remote skip the provider call entirely. The
check looks for a remote-tracking ref rather than upstream config, because
git push writes that ref even without -u, which is how most terminal and
agent pushes land. Repositories that track no remotes at all can't answer the
question, so they keep the old behaviour, as does any failed probe: hiding a PR
badge is worse than spending the lookup.

Tests cover the backoff curve, the skip, and the case that makes the naive
version wrong (a branch pushed without --set-upstream still gets looked up).

Not in this PR

Two bigger wins from the same investigation, kept separate because they change
lookup semantics and deserve their own review:

• Batch the sidebar into one gh pr list per repo matched locally by branch,
  instead of one call per branch. Caps the worst case (expanded thread list,
  262 distinct cwds here) at a few calls instead of thousands.
• Stop worktree dev servers from polling GitHub at all. Four servers were live
  during the investigation, each with independent pollers and caches against the
  same token.

---

Made by Claude Opus 5 (1M context) in Claude Code, on T3 Code.

---

Note

Medium Risk
Changes PR badge lookup timing and when GitHub is contacted; incorrect unpublished detection could hide a PR badge, though failures fall back to the old lookup path.

Overview
Reduces GitHub API load from VCS PR status polling by changing how failed lookups are cached and when provider calls run.

Failed PR lookups no longer use a flat 20s cache TTL. They use per-branch exponential backoff (20s → 40s → … capped at 15 minutes), with streak tracking cleared on success, so rate-limited branches retry less often than the healthy 2-minute refresh instead of more often.

For branches without upstream config, status can skip gh pr list when git shows remotes exist but no refs/remotes/*/<branch> for the current head—treating never-pushed local branches as having no PR. Branches that were pushed without -u still get looked up because remote-tracking refs exist. Ambiguous cases (no remotes, or probe failure) still call the provider.

Tests cover the backoff curve, the never-pushed skip, and push-without-upstream still querying GitHub.

^{Reviewed by Cursor Bugbot for commit a76cbc6. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Fix PR status lookups amplifying GitHub rate limits via exponential backoff and branch skip logic

• Repeated PR lookup failures now back off exponentially per branch (base 20s, max 15min) using a new prLookupFailureTtl(consecutiveFailures) function, resetting on success.
• Branches that have never been pushed to any remote skip the provider PR lookup entirely, using a new isUnpublishedBranch check via git for-each-ref.
• Branches pushed without --set-upstream still perform the provider lookup as before.
• Behavioral Change: branches with no remote-tracking ref and no push history return no PR without making any GitHub API calls.

^{Macroscope summarized a76cbc6.}

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 74956aaf-5cb4-41d9-8824-56b90fc108fc

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	11.3 KiB	11.3 KiB	+23 B (+0.2%)	15.1 KiB	✅
Codex	Thread snapshot wire	5.5 KiB	5.5 KiB	+3 B (+0.1%)	7.3 KiB	✅
Codex	Live turn WebSocket wire	5.9 KiB	5.9 KiB	+20 B (+0.3%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	49.7 KiB	49.7 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	16	16	0 (0.0%)	21	✅
Claude	Total thread wire	11.3 KiB	11.3 KiB	−19 B (−0.2%)	15.1 KiB	✅
Claude	Thread snapshot wire	5.5 KiB	5.5 KiB	0 B (0.0%)	7.3 KiB	✅
Claude	Live turn WebSocket wire	5.9 KiB	5.8 KiB	−19 B (−0.3%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	50.6 KiB	50.6 KiB	0 B (0.0%)	66.4 KiB	✅
Claude	Live turn messages	16	16	0 (0.0%)	21	✅

Baseline: 31891a1 · PR result: a76cbc6 · Source CI: success

Scenario and decoded snapshot size

10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.

• Codex decoded thread snapshot: 94.6 KiB
• Claude decoded thread snapshot: 95.4 KiB

Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.

[macroscopeapp]

Approvability

Verdict: Approved a76cbc6

Self-contained performance optimization that adds exponential backoff for failed PR lookups and skips unnecessary API calls for branches never pushed to a remote. Well-tested with clear, conservative logic.

^{You can customize Macroscope's approvability policy. Learn more.}