[incognitojam]

Note

Pull requests now run only the Fork CI jobs their changed files can affect. Replaying the last 40 merged PRs through the classifier gives an estimated 30% fewer runner-minutes. Main and intake pushes still run everything. Lint now runs on every PR, docs-only ones included.

Fork CI had one switch: any change outside *.md and docs/ started every lane. A typical full run uses about 45 runner-minutes, and on a narrow PR most of that is spent on suites the change can't reach. Upstream gates its CI by changed packages in pingdotgg#19587. The fork already carries that classifier, .github/scripts/ci-changes.cjs, but only the disabled inherited ci.yml used it.

What changes

• Fork Changes runs the classifier. It maps changed files to workspace packages and their dependents, and each job is gated on the result:
  • Typecheck and build: any affected package, or a desktop, server or web change.
  • Linux browser calibration: desktop or web changes.
  • Server shards, the previous-nightly upgrade test and Windows device tests: server (@styal/cli) changes.
  • Workspace suites:
    • Web 1/2 and 2/2 for web or client-runtime changes. Web 2/2 also runs for Rust crate changes.
    • Clients for desktop or mobile changes.
    • E2E for e2e, server or web changes.
    • Other tests only the affected remaining packages.
  • Release smoke: changes to scripts/.
• Full runs where they matter. Pushes to main and intake branches still run every job, so main stays the backstop for anything the gate misses. Package manifests, paths the classifier doesn't recognise, and an unreadable diff also run everything.
• Fork Lint runs for every change, so docs-only and ledger-only PRs now get format and lint checks too, as upstream does. When the ledger or the provider canary workflow changes, it also runs their tests. Before, those ran in a separate path inside Fork Changes.
• The required checks keep their names. The aggregate jobs (Fork Check, Fork Test (Server), Fork Test (Workspace) and Fork Release Smoke) now run even after a cancellation. They accept a skipped job only when the classifier said it wasn't needed. Fork Release Smoke becomes an aggregate over a new Fork Release Smoke (Run) job, so a skipped run can't pass as a success.

Validation

• Replay of the last 40 merged PRs. The estimate uses each job's median duration. Before: 1,831 runner-minutes. After: 1,292. Saved: 540 (29.5%).
  • 15 PRs touched workflows, manifests or shared paths and would still run everything.
  • Mobile-only PRs drop to Clients, Other and typecheck, saving about 34 minutes each.
  • Web-only PRs skip the server shards and Windows device tests, saving about 20 minutes.
  • Server-only PRs skip both web shards and Clients, saving about 15 minutes.
• Failures a skip would have hidden. The replay checked each PR's real CI failures against the jobs it would skip. One case came up: fix(mobile): composer attachment uploads retry after a connection failure #1099, a mobile-only change, hit an unrelated server storage-cleanup assertion that the gate would now skip. Main's full run still catches failures like that.
• Classifier tests. scripts/ci-changes.test.ts checks the classifier's decisions for representative paths. It also runs each aggregate job's jq rule against simulated job results, and checks that failures, cancellations, and skips the classifier didn't allow all fail.
• Not exercised by this PR's CI. This PR edits fork-ci.yml and the classifier, so its own Fork CI run takes the run-everything path. The gated path first runs on the next narrow PR after merge.

---

Written by an agent (Claude Code, claude-opus-5-5), with implementation by Codex (gpt-6.1-sol).

[github-actions]

Upstream divergence

Both sides are compared with upstream at the in-order tip b912bd2da0. Excludes .repos/ and pnpm-lock.yaml.

Measure	Base	This PR	Change
Upstream files the fork changes	1,742	1,742	0
Lines changed in them	97,776	97,826	+50
Upstream files the fork removes	49	49	0
Fork-only files	694	693	−1
Fork-only lines	124,878	125,058	+180

Upstream files whose difference changes (1)

File	Change	Now differs by
.github/scripts/ci-changes.cjs	+50	68 lines

[github-actions]

Thread transfer impact

✅ Thread transfer remains within every enforced ceiling.

ℹ️ No successful main baseline artifact is available yet. This run establishes the initial measurement.

Provider	Metric	Main baseline	This PR	Impact	PR ceiling
Codex	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Codex	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Codex	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Codex	Live turn WebSocket decoded	—	20.9 KiB	—	29.3 KiB	✅
Codex	Live turn messages	—	2	—	8	✅
Claude	Total thread wire	—	5.0 KiB	—	6.8 KiB	✅
Claude	Thread snapshot wire	—	3.8 KiB	—	4.9 KiB	✅
Claude	Live turn WebSocket wire	—	1.2 KiB	—	2.0 KiB	✅
Claude	Live turn WebSocket decoded	—	21.2 KiB	—	29.3 KiB	✅
Claude	Live turn messages	—	2	—	8	✅

Baseline: unavailable · PR result: 64af9d2 · 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: 108.5 KiB
• Claude decoded thread snapshot: 108.8 KiB

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