runfoundry Write to Marcel on LinkedInWrite on LinkedIn

Run record R-20261113-bau-1

Foundry's first build run on Navigo Bau to reach a merge, 13 November 2026: one requirement, from the plan to a merged pull request in 101 minutes, with 13 agent rounds, 16 checks and $5.19 at API prices. It was a small piece of work: 4 builder rounds of 17 to 64 seconds, 2 of them on fix tasks from the review, and 2 of its 3 acceptance tests were already met before the build, by a fix shipped earlier that day.

Taken from the run's own files, with the numbers kept and all text, names and paths removed. G1.F1.T2 means goal 1, feature 1, task 2. Times are UTC.

R-20261113-bau-1 The run on a time line, 0 to 101 minutes From events.jsonl
  1. a3.3 min: test writer, the longest round
  2. b40.6 min: acceptance red, green on a re-run at 46.1
  3. c47.5 min: review red: 2 fix tasks
  4. d66.2 min: background red 3×, no task blamed
  5. e101.4 min: gate two, I approved the merge

Gate one, the plan: 1.0 to 1.8 min, approved for me by a Claude session. Agent rounds: 13; the longest, the test writer's, ran 3.3 to 9.8 min. Checks: 16, 2 red (the acceptance check at 40.6 min, the feature review at 47.5 min); on the run's path they took 73 minutes. Background suite: red at 66.2, 87.8, 93.8 min. Gate two, the merge: 89.3 to 101.4 min, approved by me at a terminal. Every mark is a row in the table below.

R-20261113-bau-1 The run, in numbers From run.json and events.jsonl, redacted
product
Navigo Bau, one requirement
started
2026-11-13 14:36:50
ended
2026-11-13 16:18:17
stop reason
done, merged once I approved it
wall time
101 minutes, plan to merge
rounds
13: planner 1, test writer 1, builder 4, blame 4, reviewer 3
time
checks 73 minutes, agent rounds 15, waiting at the two gates for a decision 13, the foreman's own steps (room, pull request, merge) under 1
cost
$5.19 at API prices
cost by role
planner $0.34, test writer $2.31, builder $0.57, blame $0.13, reviewer $1.84

How the time is split: a span of the run counts as agent time while a round is open, as waiting while a gate is open, and as check time from the step before a check to its result (install, task tests, acceptance tests, the goal's suite and CI). The background suite runs beside the run and is not counted. Costs are the Agent SDK's figures at API prices.

Every event, in order

Rounds stand at their start, checks at their result. Cost updates and the building and checking marks of each node are left out. A small number points to a note under the table.

UTCmineventnodedetail
14:36:500.0run: started, the plan is being made
14:37:060.3room: being made
14:37:070.3room: up
14:37:070.3round: plannerG1large model, 43 s, ok, $0.34
14:37:501.0check: plan formatG1green (3 passed, 0 failed)
14:37:501.0plan proposed: goal, features, tasks (titles withheld)G1
14:37:501.0gate one: the plan waits for a decision
14:38:381.8approved for me by a Claude session (delegated)
14:38:381.8run: queued
14:38:411.9room: up
14:38:411.9run: running
14:40:063.3check: installG1green (1 passed, 0 failed), 1 min 25 s
14:40:063.3round: test writerG1.F1large model, 6 min 34 s, ok, $2.31
14:46:409.8note from the test writer, for me (text withheld)G1.F1
14:46:409.8note from the test writer, for me (text withheld)G1.F1
14:46:409.8check: acceptance tests written, REQ (id and text withheld)1G1.F1green (7 passed, 0 failed)
14:46:409.8round: builderG1.F1.T1mid model, 1 min 04 s, ok, $0.20
14:53:0516.3check: task tests2G1.F1.T1green (24,010 passed, 0 failed), 5 min 21 s
14:53:0516.3task frozen: its tests are greenG1.F1.T1
14:53:0516.3round: builderG1.F1.T2mid model, 45 s, ok, $0.19
14:58:5122.0check: task testsG1.F1.T2green (24,011 passed, 0 failed), 5 min 01 s
14:58:5122.0task frozen: its tests are greenG1.F1.T2
15:17:2540.6check: acceptance tests, REQ (id withheld)3G1.F1red (24,011 passed, 5 failed), 18 min 34 s
15:22:5646.1check: acceptance tests, REQ (id withheld)G1.F1green (24,016 passed, 0 failed), 5 min 31 s
15:22:5646.1round: reviewerG1.F1large model, 1 min 22 s, ok, $0.74
15:24:1847.5check: feature reviewG1.F1red (2 findings)
15:24:1847.5question from the review, for me (text withheld)G1.F1
15:24:1847.5fix task added from the review (text withheld)G1.F1.T3
15:24:1847.5fix task added from the review (text withheld)G1.F1.T4
15:24:1847.5round: builderG1.F1.T3mid model, 17 s, ok, $0.06
15:29:4052.8check: task testsG1.F1.T3green (24,011 passed, 0 failed), 5 min 05 s
15:29:4052.8task frozen: its tests are greenG1.F1.T3
15:29:4152.9round: builderG1.F1.T4mid model, 41 s, ok, $0.12
15:36:2259.5check: task testsG1.F1.T4green (24,011 passed, 0 failed), 6 min 00 s
15:36:2259.5task frozen: its tests are greenG1.F1.T4
15:41:5565.1check: acceptance tests, REQ (id withheld)G1.F1green (24,016 passed, 0 failed), 5 min 33 s
15:41:5565.1round: reviewerG1.F1large model, 1 min 06 s, ok, $0.57
15:43:0166.2check: feature reviewG1.F1green (no open finding)
15:43:0166.2question from the review, for me (text withheld)G1.F1
15:43:0166.2check: full suite, background room4red (0 passed, 1 failed)
15:43:0166.2round: blameG1small model, 25 s, ok, $0.01
15:43:2666.6round: blameG1mid model, 14 s, ok, $0.07
15:44:3667.8check: acceptance tests, REQ (id withheld)2G1.F1green (7 passed, 0 failed), 56 s
15:44:3667.8feature passedG1.F1
15:49:5273.0check: goal: full suite2G1green (14 passed, 0 failed), 5 min 16 s
16:04:3687.8check: goal: the product's CIG1green (1 passed, 0 failed), 14 min 44 s
16:04:3687.8check: full suite, background room4red (0 passed, 1 failed)
16:04:3687.8round: blameG1small model, 32 s, ok, $0.01
16:05:0988.3round: blameG1mid model, 14 s, ok, $0.05
16:05:2488.6note: background suite red, blame names no task; the goal check decides4G1
16:05:2488.6round: reviewerG1large model, 42 s, ok, $0.53
16:06:0689.3check: run review5G1green (no open finding)
16:06:0689.3question from the review, for me (text withheld)G1
16:06:0689.3goal passedG1
16:06:1189.3pull request opened (draft)
16:06:1189.3gate two: the merge waits for a decision6
16:06:1189.3room: being given back
16:06:1189.3run: waiting at a gate
16:06:1789.5room: gone
16:10:3993.8check: full suite, background room4red (0 passed, 1 failed)
16:18:12101.4I approved, at a terminal
16:18:17101.5pull request merged
16:18:17101.5room: being given back
16:18:17101.5run: donestop reason done
16:18:17101.5room: gone

Notes on the odd lines

  1. This check is green when each new acceptance test does what it should before any code is written: it fails, unless the test writer declares it already met by today's product and says why. Here 1 of 3 failed, as it must; the other 2 were declared already met (the reasons came to me as the notes just above) and were green. The 7 counts those 3 tests and 4 steps that start and stop the test setup.
  2. The counts add up a check's commands, not only tests. A step that starts or stops the test setup, a type check or a lint counts as one, each acceptance test as one, and a test command adds the tests it ran. So the first task check's 24,010 is 2 steps and 24,008 product tests; the acceptance check's 24,016 is 4 setup steps, 3 acceptance tests and 24,009 product tests. At 67.8 min the feature had been merged into the run's branch and its acceptance tests ran there once more, alone: 7, that is 4 steps and 3 tests. The goal's full suite shows 14, and that number is wrong. The suite is one command that runs several test runners. The counter only reads Node's built-in test summaries when there are any, and ignores the other runners. Only the design system's lint-rule tests print that format, so it counted their 14 and nothing else. Its log shows 28,108 tests passed in all and none failed. The verdict comes from the command's exit code, so the green was right; only the number is wrong.
  3. The test setup did not come up for this check, so its 3 acceptance tests could not run: the 5 failed are 2 setup steps and those 3 tests. A red outside the tests says nothing about the code, so the foreman runs the same check once more, without an agent round. The second go, at 46.1 min, was green.
  4. The background suite runs the product's whole suite in a room of its own, on a plain copy of the code without its git history. 19 product tests and 3 test setups need that history, so the job was red every time, for that reason alone (a job counts as one: 0 passed, 1 failed). Blame found no task to reopen, so the run did not stop: the goal check decides. That check, the same suite on the run's own checkout, was green at 73.0 min, and the product's CI at 87.8 min.
  5. On 13 November 2026 the run review came after the suite and CI. It comes before them now, so a review fix doesn't pay for the whole suite again.
  6. On 13 November 2026 gate two was the merge: the pull request waited until I approved it, 12 minutes later. Runs that merge themselves came later.

Back to the results