GitHub API Limits At openclaw Scale
- Status: active audit reference
- Owner: ClawSweeper maintainers
- Source of truth: the GitHub readers named in each row,
/api/github-egress-observability, - Last verified:
openclaw/clawsweeper@2d6eb4d561on 2026-11-08 (after #2183, #2184 and #2185 merged) - Update when: a reader starts calling a new GitHub route, changes its page size or
and live probes against openclaw/openclaw
pagination, or openclaw crosses another documented limit
Read when adding or changing any GitHub read. On 2026-11-07 openclaw/openclaw outgrew the recursive tree listing and most reviews failed (#2183). This page records, for every route ClawSweeper reads, the hard limit GitHub enforces, how GitHub signals it, what our code does at the limit, and where openclaw stands.
#openclaw scale on 2026-11-08
| Measure | Value |
|---|---|
Tracked tree entries (main) | 67,554 (recursive tree listing now truncated: true) |
| Open pull requests | about 2,900 |
| Check runs on an open PR head | over 100 on 144 of 200 sampled heads (72%); max 299 |
| Changed files on one PR | 5,018 (#194566) |
| Commits on one PR | 2,522 (#129388) |
| Timeline events on one PR | 703 (#129388; 3.8 MB, 13.4 s to read in full) |
| Comments on one PR | 476 (#141004; 1.4 MB, 4.3 s) |
| Repository labels | 405 |
is:pr is:open search results | 3,046 |
#Route table
Ranked by risk: silent means partial data was used as if complete; fails means the reader stops safely at the limit; ok means the reader handles the limit or openclaw is far from it.
| Route | Hard limit and signal | Our reader at the limit | openclaw live probe | Status |
|---|---|---|---|---|
GET git/trees/{sha}?recursive=1 | 100,000 entries or 7 MB; truncated: true | Was: refused (review_blob_metadata_unavailable) | truncated: true since 2026-11-07 ~16:00 UTC | Fixed by #2183: GraphQL object(oid) byte sizes for only the missing blobs |
GET commits/{sha}/check-runs | 100 per page; total_count | Before #2184/#2185: first page only in review hydration, apply guard and dashboard CI. Now: every page, bounded at 10, with completeness reported | 72% of open heads over 100; failures past page 1 on 14 of 30 checked | Was silent; fixed by #2184 (hydration and guard, merged) and #2185 (dashboard, merged) |
GET commits/{sha}/status | statuses paged (default 30, max 100); total_count | Before #2184: guard read the first 30. Now: every page | 3 statuses per head today | Was silent at scale; fixed by #2184 (merged) |
GET pulls/{n}/files | 3,000 files total (page 31+ is empty, no error) | Prompt window marks truncated; capacity guard refuses when count differs from changed_files | #194566: page 30 has 100 files, page 31 returns 0 of 5,018 | ok: no reader treats the list as complete |
GET pulls/{n}/commits | 250 commits (page 4+ is empty, no error) | Prompt window marks truncated, but its "latest commits" tail is empty past 250 | #129388: page 3 returns 50, page 4 returns 0 of 2,522 | ok for decisions (head SHA comes from the PR); prompt context shows only the oldest 250 |
GET pulls/{n} diff media type | 300 files; HTTP 406 too_large | Not used for review source (local git diff) | #194566 returns too_large | ok |
GET compare/{a}...{b} | 250 commits and 300 files in the body | Only status is read | — | ok (single fields; response size bounded) |
GET search/issues | 1,000 results; incomplete_results | Mixed. Count guards (author open-PR budget, bulk-filer counts) use per_page=1 and reject incomplete_results. Placeholder recovery pages within 1,000 and records the flag. The structural relation probe (structuralExternalRelationSensitivity) treats an incomplete search with no match as unknown, so the review is not reused from cache. The related-item and merged-PR prompt-context searches and the recovery-label backfill read items without the flag: they only add context or act on matches they find, and the backfill revisits every page on rotation | is:pr is:open has 3,046 | ok; no decision trusts an incomplete empty result |
GET actions/runs?status=… | 1,000 results per filtered query | Dashboard census bounded at 3 pages and marks census_incomplete. The sweep.yml apply continuation and review backstops queue nothing when a run list cannot be read, because they cannot rule out an active or recent run; scheduled runs continue the lanes | about 60 queued after the cleanup | ok |
GET actions/runs (unfiltered) | history listing; slow at large run counts | Removed from the status refresh by #2154 | 10–11 s, occasional 502 | Fixed by #2154 |
GET issues/{n}/timeline | none documented; paged at 100 | Close-guard head activity reads only force pushes, through GraphQL timelineItems(itemTypes: [HEAD_REF_FORCE_PUSHED_EVENT], last: 100), paging back at most 10 pages and failing closed past that. The low-signal conflict guard still reads the full REST timeline because it also needs author activity | #129388: full timeline 3.8 MB in 11–13 s; force-push read 126 bytes in 0.6 s | ok. GraphQL afterCommit is null for commits GitHub no longer has; the current head always resolves |
GET issues/{n}/comments | none; paged at 100 | Paged in full where completeness matters; prompt windows otherwise; ETag-conditional in the publisher (93% 304s in telemetry) | #141004: 476 comments, 1.4 MB | ok |
GET pulls/{n}/reviews, pulls/{n}/comments | none; paged at 100 | Paged in full or with an explicit limit | — | ok |
GET repos/{repo}/labels | none; paged at 100 | Paged in full | 405 labels, 5 pages | ok |
| GraphQL connections | 100 nodes per page; pageInfo | Review cache, activity cursor and census query pageInfo / totalCount and treat overflow as incomplete | — | ok |
GraphQL closingIssuesReferences(first: 100) | 100 | No pageInfo | no PR close to 100 | ok (follow-up: request pageInfo) |
GET contents/{path} | inline body only up to 1 MB (1–100 MB returns no content; use the raw media type); directories list 1,000 entries | Apply-guard path checks read existence only. The dashboard apply-health reader (readApplyHealthMarker) decodes the inline body of the sweep-status JSON in the state repository | largest status file 7 KB (openclaw-openclaw.json, 2026-11-08) | ok. Over 1 MB the body is absent and the dashboard shows apply health as unavailable, not a wrong value. Raw reads would need media-type support in the dashboard ETag cache, which is not justified at this headroom |
#How the live probes were run
Read-only gh api calls with a maintainer token against openclaw/openclaw on 2026-11-08: the largest open PRs by changed files, commits, comments and review threads came from a GraphQL scan of the 200 most recently updated open PRs, and check-run counts from statusCheckRollup.contexts.checkRunCount. Production call volume comes from /api/github-egress-observability (six-hour window).
#Gaps
/api/github-egress-observabilitydoes not classifygit/trees, so the- The egress feed truncates at 2,000 rows (
rows_truncated: true). Rows are
failing route of 2026-11-07 appeared only as unknown. Routes added to readers must also get a route template, or the telemetry hides them.
ordered by bucket_start ascending, so truncation keeps the earliest buckets of the six-hour window; how much time that covers depends on row density, and the most recent buckets are the ones dropped.