ci: charge newer runs one root runner each when gui runners are on

[teamleaderleo]

Problem

The PR runner picker charged each newer run's whole marker peak on an owned pool against that pool's root runners. With gui runners configured, a newer run's shards, lag and cli-product take the gui label and its side lanes never hold a root runner, so the run holds exactly one root runner (its admission).

In run 36371179217 three newer runs with a combined 32-machine peak were charged against 15 root runners (12 busy, 2 replayed): the queue bound came out at 15 x 3 - 12 - 32 - 2 = -1, so the run got a root budget of 0. Only claude-wrapper stayed owned; admission, the 7 app-host shards, lag and cli-product went to Blacksmith while owned capacity sat idle.

Fix

• choose() reads, per owned pool, whether its gui label has a count in the slots (the same test gui_runner() uses). On such a pool it passes the newer runs' run count (Routed.runs()) to decide() as their root charge, for the bound and for "now". With the runners read live those are the live window's runs; older runs' admissions show busy on the runners or queued through older.
• decide() gains root_since / root_now, and pick() gains root_taken / root_taken_now. The root room, the "starts now" filter and the "N of M root runners free" text use them. None keeps the old charge (the machines), so the E2E and iOS pickers are unchanged.
• On a pool without gui runners the whole peak stays the charge: a run's root share there is its peak less the side lanes on the pool, which a marker does not split out.
• Docstrings updated; the "whole marker peak" claim in decide() now describes the new parameters.

Tests

python3 -m unittest tests/test_ci_pr_runner_pool.py: 216 tests OK.

• test_newer_runs_hold_one_root_runner_each_with_gui_runners reproduces the run's numbers (15 root online, 12 busy, 2 replayed, 3 newer runs with a 32-machine peak). With gui runners the run takes std with a root budget of 28 and its admission is placed owned. With a gui count only on another pool, std keeps the whole-peak charge and the budget stays 0, as before.
• test_decide_charges_the_root_runners_newer_runs_hold checks the root text reads the passed root charge.

🤖 Generated with Claude Code

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Fixes the PR runner picker charging each newer run's whole marker peak against root runners when gui runners are on; it now charges one root runner per newer run (their admission).

With gui runners, a newer run's shards, lag and cli-product take the gui label and side lanes never hold a root runner, so it holds exactly one root runner. Charging the whole peak left a run with a root budget of 0 out of 15 root runners while three newer runs held only 3, sending its admission to Blacksmith (run 36371179217).

• choose() passes the newer runs' run count as their root charge per pool where that pool's gui label has a slot count; on other pools the whole peak stays the charge since a marker does not split side lanes from root jobs.
• decide() and pick() now accept the root charge separately; None keeps the old behavior, so the E2E and iOS pickers are unchanged.
• Tests reproduce run 36371179217 and verify the root text.

^{Written for commit 69a4efe. Summary will update on new commits.}

Summary by CodeRabbit

• Bug Fixes
  • Runner-pool capacity calculations now account separately for root-runner usage by newer runs when GUI runners are enabled. This improves estimates of available root runners and can affect which runs are admitted to the pool.
  • When GUI runners are not enabled, root-runner accounting continues to follow existing machine-usage behavior. Calls that omit the new accounting information retain the previous peak-based fallback.
• Tests
  • Added coverage for root-runner availability and capacity decisions with newer runs.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: Repository: manaflow-ai/cmux/.coderabbit.yaml

Review profile: ASSERTIVE

Plan: Advanced

Run ID: be8be423-73a5-4854-baad-205eac36675e

📥 Commits

Reviewing files that changed from the base of the PR and between 3606617 and 254430a.

📒 Files selected for processing (2)

• scripts/ci/pr_runner_pool.py
• tests/test_ci_pr_runner_pool.py

 __________________________
< Grammarly for your code. >
 --------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

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]

Merge receipt for 69a4efe764: every check was green at merge (13 verified; 18 skipped by policy). Full suite runs on main after merge.