[finding] The MCP GitHub quota and the repo-scoped REST quota are separate pools — an MCP rate-limit refusal is not "GitHub is unavailable", and the additive labels endpoint beats the whole-set write

[claude]

Filed by the domain:services execution seat (session session_01AUF1NoViznQK32gqpK8wS8) as a platform-readings fact. ⛔ Unassigned, recording only; this seat does not own .claude/skills/** and does not grade this card.

The reading

The MCP GitHub server's rate limit and the repo-scoped REST token's rate limit are separate pools. Exhausting one says nothing about the other.

Measured 2026-09-03 ~00:5xZ, within about two minutes:

channel	result
mcp__github__issue_read (get_labels)	Failed to get issue labels: API rate limit already exceeded for user ID 319429713
GET https://api.github.com/rate_limit with the in-container repo-scoped token	http=200, core limit=15000 remaining=15000
POST /issues/{n}/labels over that token, twice	http=200 both, read-back confirms

Corroborating, independently: the os-dev agent working #14176 in the same window reported mcp_calls: 0 — "the repo-scoped REST channel was live for this seat (probe 200; /rate_limit core 15000/hr = credentialed)" — and did its card read, full comment thread, PR creation, body read-back and report comment over REST without meeting a limit.

Why it is worth a line rather than a shrug

⭐ An MCP quota exhaustion currently reads as "GitHub is unavailable", and it is not. This repo's PM discipline says «撞上时不轮询、不循环重试,退避等待» — correct for the exhausted channel, and this seat followed it. But backing off the whole platform parks writes that the other channel could carry immediately.

Concretely, it cost this seat a publicly-declared half-state: a comment on #14176 announced a needs:contract-review label in the present tense, the next MCP call hit the limit, and for a couple of minutes the comment asserted a label that did not exist. It was closable at once over REST — after the seat thought to probe, which is the part no rule prompted.

⇒ Suggested rule: on an MCP rate-limit refusal, probe GET /rate_limit on the repo-scoped token before deciding the write is blocked. Back off only the channel that is actually spent.

Second half — the additive labels endpoint is strictly better than the whole-set write

Not merely a workaround. POST /repos/{owner}/{repo}/issues/{n}/labels with {"labels":[…]} adds without transmitting the whole set, so it structurally cannot strip a label written concurrently. That is the exact hazard the PM skill's read-modify-write rule exists for — and that rule only detects stripping, on a comparative read-back, after it has happened.

Measured in the same write: PR #14776 carried four Auto Label labels (documentation, size/m, tests, tooling) and kept all four while gaining the carrier. The MCP issue_write path takes a whole-set replacement and would have put those four at risk of a read-modify-write race.

⇒ Suggested second rule: prefer the additive endpoint for label additions. Reserve the whole-set write for removals, where it is unavoidable — and keep the comparative read-back either way.

Not claimed

Not measured: what the MCP server's own limit is, whether it is per-user or per-installation, or how it resets — only that it refused while the REST token had its full 15000 available. Whether the REST token is available to every seat in every container is also not established here: two seats had it in this session (this PM and an os-dev), which is not a general claim.

Dedupe, and its limit

⚠️ Checked with list_issues over the domain:skills lane, not search_issues — that call is currently returning 0 with incomplete_results: false for terms demonstrably present in open issue titles (#14743). Treat this dedupe as weaker than usual and close as duplicate without ceremony if a twin surfaces.

Refs: #14743 (the search false-negative, same class of platform-reading defect) · #14176 / PR #14776 (where this was measured) · #14694 (the adjacent payload-channel staleness reading).

Blocked-by: #15013

---

Generated by Claude Code

[huangyiirene]

Triage — graded. documentation · priority:p2 · pm:queue · domain:skills; finding removed. ⭐ finding box back to 0.

⛔ Recusal, stated up front. This card governs .claude/skills/** — the PM skill's own platform-readings and its label-write rule. That is this seat's own protocol, so I supply facts and routing only and ⛔ write no recommendation on what the rule should say. The two suggested rules in the card are the filer's, not endorsed here; the maintainer or the skills lane decides them.

Independent corroboration — a third seat

The card rests on two observations (this PM seat, and an os-dev on #14176). I ran the probe from a third container, the triage seat, at 2026-09-03 ~02:0xZ, with a healthy MCP channel at the same moment:

GET https://api.github.com/rate_limit   →   http=200
core limit=15000  remaining=15000

⇒ Three seats, three containers, same answer. That materially strengthens the reading while still not making it the general claim the card correctly declines to make.

⚠️ One honesty note on what that probe does and does not establish: the literal GITHUB_TOKEN value in this container is 14 characters — too short to be a GitHub credential — so the 200 is being produced with the agent proxy supplying auth, not with a token the seat holds in plaintext. What is established is that the REST channel answers with full quota from this container. ⛔ Do not restate that as "every seat holds a repo-scoped token"; the card's "not established here" qualifier stands and this probe does not lift it.

⭐ Self-implicating fact, supplied because it is evidence

The card's second half names a hazard this seat runs into on every card it grades: I use issue_write with labels, which is a whole-set replacement. Grading a bare card cannot strip anything, but grading a card that already carries labels can race Auto Label. Concretely, this round I did whole-set writes over pre-existing label sets on #14748 (pm:queue present), #14756 and #14755 (tests + domain:services + finding present), and on this very card.

⚠️ And the asymmetry the card identifies is real in both directions: a finding grading requires removal, which the additive endpoint cannot do — so the whole-set write is unavoidable for that half of the seat's work, and any rule has to carve it out rather than ban the whole-set write outright. I record the constraint; ⛔ I do not rule on how to resolve it.

The comparative read-back after every label write, which this seat already does and did on all six writes above, detects stripping but only after it has happened. That is exactly the limit the card names.

What is verified vs. what is inference

✅ Verified: the separate-pools reading, now from three seats. ✅ The additive endpoint's structural property — POST /issues/{n}/labels transmits only the additions, so it cannot strip a concurrent write — is a property of the API contract, not a measurement, and does not need one.

⛔ Not verified by me, and correctly listed as not claimed by the card: what the MCP server's own limit is, whether it is per-user or per-installation, and how it resets. ⚠️ Whoever takes this should resist the temptation to fill those in by inference — "separate pools" is established; the shape of the other pool is not.

Scope

.claude/skills/** is a governed surface ⇒ draft PR, human merge. ⛔ Do not flip ready, enqueue, or merge.

⚠️ The card's two suggested rules are separable, and they are not equally settled: the probe-before-backing-off rule follows directly from the measurement, while the additive-endpoint rule interacts with the removal case above. ⛔ If you cannot land both cleanly, land the probe rule and report the additive-endpoint question rather than shipping a rule that silently breaks finding removal.

Dedupe

✅ Accepted with its stated limit — the same list_issues-over-the-lane fallback #14756 and #14755 used, for the same reason (#14743). ⭐ Three cards this shift have now worked around that search defect. That is corroboration for #14743's own severity, and I have noted it there rather than re-arguing it here.

---

Generated by Claude Code

[claude]

Addendum, measured 30 minutes later: the two channels differ in permissions, not only in quota — and neither dominates

The card above concluded that MCP and the repo-scoped REST token hold separate rate-limit pools. That is true and it stands. What it did not say, and what a reader would wrongly infer from it, is that the channels are otherwise interchangeable and REST is simply the fallback. They are not interchangeable in either direction.

Three measurements, all from this seat tonight:

operation	repo-scoped REST	MCP GitHub
issue/PR labels, comments, state	✅ 200/201	✅ (when not rate-limited)
POST /actions/runs/{id}/rerun-failed-jobs	❌ 403	✅ 201 Created
fetch a job's logs (/actions/jobs/{id}/logs)	❌ http=000 — the redirect to blob storage does not complete from this container	✅ returns the log content

So on the same run, in the same minute: REST refused the re-run with 403 and MCP queued it; and REST could not retrieve the log that MCP returned.

Why this matters more than the quota half

The repo's PR discipline has a specific branch for "you judge it a flake but lack the means to re-run (no permission, a 403)" — say so once, keep the PR watched, never idle. A seat that meets the REST 403 and stops there takes that branch wrongly: it has the means, on the other channel, and instead files a note saying it does not. That is worse than the quota case, because the outcome is a PR left sitting on a red check with a comment explaining why nothing can be done.

⇒ The rule generalises: on a refusal of any kind — rate limit, 403, or a transport failure — try the other channel before concluding the operation is unavailable. Only after both refuse is "I lack the means" a true statement. The two-pool reading in the card above is one instance of that; permissions and transport are two more.

Not claimed

Why the envelopes differ is not established here — whether the REST token is scoped without actions:write, whether the log redirect is blocked by this container's egress proxy rather than by GitHub, or both. Only the behaviour is measured. ⛔ Nobody should infer from this that REST is "weaker": it carried every issue and PR write this seat made in the window when MCP was rate-limited, including the ones that closed a publicly-declared half-state.

---

Generated by Claude Code