fix(server): find a branch's Forgejo PR by head instead of paging every PR

[pseudoseed]

Why

Fixes #41.

The Forgejo source-control provider found a branch's PR by listing /pulls 50 at a time and comparing head.ref. A branch with no PR read every page. On git.pseudoseed.com (Forgejo 16.0.5) that was 153,134 list requests in one day from one desktop client, about 2 s each on the server.

What changed

apps/server/src/sourceControl/ForgejoSourceControlProvider.ts, listChangeRequests:

• Every page request now sends head=<branch>. Forgejo 16 narrows /pulls to that head branch, so a lookup takes 1 request, or 2 when the caller asks for more than the PRs that exist.
• The loop stops after 5 pages (250 PRs). Older Forgejo and Gitea ignore head and return every PR, so this bounds the scan there.
• The existing head.ref, owner and repository checks are unchanged, so owner:branch selectors still filter forks the same way.

Tests in ForgejoSourceControlProvider.test.ts use a fake server that either honours head or ignores it. They cover: PR found in 1 filtered request (branch name with / and #, to check encoding), no PR in 1 request, owner selector sends the bare branch and keeps the owner check, the 5-page cap when head is ignored, and a recent PR still found when head is ignored.

Why not GET /pulls/{base}/{head}

The issue suggested it. I did not use it, for 2 reasons:

• Callers (GitManager.findOpenPr, findLatestPrForHeadContext) pass a head selector and a state, never a base branch.
• It returns one arbitrary PR, ignoring state. Forgejo 16 source (GetPullRequestByBaseHeadInfo) has no state filter and no ORDER BY. Live, pulls/main/renovate/ghcr.io-shelf-nu-shelf.nu-latest returned refactor(contracts): remove unused OrchestrationPersistedEvent schema pingdotgg/t3code#601, a merged PR last updated 2026-08-31, while 17 PRs share that head and the newest is feat: Add provider health and per-provider controls pingdotgg/t3code#1289. On a reused branch name it would hide a newer open PR.

Suggestion 1 (fetch by known PR number) already holds: getChangeRequest calls pulls/{index}. listChangeRequests callers have no number.

How it was verified

Live API, git.pseudoseed.com, repo pseudoseed/infra (1,093 PRs), read-only GETs with curl:

• /api/v1/version returns 16.0.5+gitea-1.22.0. The served swagger lists head and base query parameters on GET /repos/{owner}/{repo}/pulls.
• pulls?state=all&head=no-such-branch-zzz returns an empty array, x-total-count: 0, 86 ms. The unfiltered first page took 1.36 s.
• pulls?state=open&head=fix%2Fgideon-vm-discard returns an array with 1 object: fix(web): apply sticky model when reusing existing draft thread pingdotgg/t3code#1291, state=open, head.ref=fix/gideon-vm-discard.
• pulls?state=all&head=renovate%2Fghcr.io-shelf-nu-shelf.nu-latest returns 17 objects, all merged. Their head.ref is refs/pull/N/head, because Forgejo reports that once the branch is deleted, so the provider's head.ref check drops them. The old code dropped them too.
• head=pseudoseed%3Afix%2Fgideon-vm-discard (owner form) returns an empty array, so the provider sends the bare branch.
• pulls/main/no-such-branch-zzz returns 404 with {"message":"The target couldn't be found.",...}. A missing repository also returns 404.

Patched provider run against the live server through the real ForgejoCli layer (fj credentials), via a throwaway script that is not committed:

Lookup	Before	After
no-such-branch-zzz, state all, limit 20	23 requests, 31.3 s	1 request, 93 ms
no-such-branch-zzz, state open, limit 1	not run	1 request, 81 ms
fix/gideon-vm-discard, state open, limit 1	not run	1 request, 214 ms, found pingdotgg#1291
fix/gideon-vm-discard, state all, limit 20	not run	2 requests, 163 + 85 ms, found pingdotgg#1291

Times are client round trips from this machine, not server-side.

Commands, run in apps/server with Node 24.20.0 and pnpm 11.10.0:

• vp test run src/sourceControl/ForgejoSourceControlProvider.test.ts: 11 passed. Against the unpatched provider, 4 of the 5 new tests fail. For example, the no-PR case made 21 requests instead of 1.
• vp test run src/sourceControl/ src/git/GitManager.test.ts: 20 files, 327 tests passed.
• vp run typecheck (tsc --noEmit): exit 0, 0 errors.
• vp lint --report-unused-disable-directives and vp fmt --check on the 2 changed files: exit 0, no findings.

Not verified

• No desktop or dev-server run. The real user path was checked only through the provider script above.
• The tea api transport was not run live. Only the fj HTTP path was. Both receive the same path string.
• No server-side request count after the change. That needs a build in daily use.
• Older servers were not run live. That Forgejo 11.0 to 15.0 and Gitea (main, release/v1.24, release/v1.22) ignore head on the list endpoint comes from reading their source. On those servers, a PR older than the 250 most recently updated is no longer found.
• Suggestion 4 (poll idle threads less often) is not in this PR. GitManager already caches lookups for 60 s with an open PR and 5 min without one.

🤖 Generated with Claude Code

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ The exact PR base did not have a successful artifact. Baseline uses the latest successful main measurement shown below.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	13.9 KiB	14.0 KiB	+22 B (+0.2%)	15.1 KiB	✅
Codex	Thread snapshot wire	7.3 KiB	7.3 KiB	+5 B (+0.1%)	7.4 KiB	✅
Codex	Live turn WebSocket wire	6.6 KiB	6.7 KiB	+17 B (+0.2%)	7.8 KiB	✅
Codex	Live turn WebSocket decoded	57.6 KiB	57.6 KiB	0 B (0.0%)	66.4 KiB	✅
Codex	Live turn messages	9	9	0 (0.0%)	21	✅
Claude	Total thread wire	14.0 KiB	13.9 KiB	−36 B (−0.3%)	15.1 KiB	✅
Claude	Thread snapshot wire	7.3 KiB	7.3 KiB	−8 B (−0.1%)	7.4 KiB	✅
Claude	Live turn WebSocket wire	6.7 KiB	6.6 KiB	−28 B (−0.4%)	7.8 KiB	✅
Claude	Live turn WebSocket decoded	58.4 KiB	58.3 KiB	−44 B (−0.1%)	66.4 KiB	✅
Claude	Live turn messages	9	8	−1 (−11.1%)	21	✅

Baseline: e399c14 · PR result: dc6b9e2 · 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: 116.3 KiB
• Claude decoded thread snapshot: 117.0 KiB

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