Mills run record

Factory Evidence

Mills is the agent delivery pipeline that maintains my repositories: a council plans backlog, sandboxed agents implement, gates decide what may merge, and a human resolves what the gates refuse. This page is its run record, sanitized, for the last 30 days. Escalations are counted, not hidden.

Live from the operator· 1 min agogenerated 2026-10-15 20:27 UTCJSON →How the factory found its own gaps →Mills docs →

Pipeline runs: 83030 days
Merged: 40048% of runs
Escalated: 40650% escalation rate
Billed per merge: $0.03$22.43 metered; $5.07 at list price
Gate pass rate: 93%all gate evaluations
Retry burn: $1,204attempts after the first, at list price

Who paid for the compute

30-day window, list-price equivalent

At list price: $3,697every token, priced as if metered
On plans: $3,67499%, flat-rate, paid either way
Metered: $22.431%, the only slice that is a bill

Claude Code plan$1,997 · 54%
Codex plan$1,675 · 45%
subscription, harness not recorded$2.22 · 0%
metered API$22.43 · 1%

Mills runs its agents under Claude Code and Codex subscriptions. Those plans cost the same each month whether the factory uses them or not, so their share is capacity already paid for, not spend. Every figure on this page is the vendor list-price equivalent of the tokens used, which keeps runs comparable regardless of who paid. Plan shares are the operator's per-harness tallies scaled to the window's subscription total.

Listed runs per day

UTC days, 30-day window ending 2026-10-15

Sep 16Oct 15

mergedescalatedretired or other104 of 215 listed runs fall on these days; 111 carry no timestamp and are listed from their summary card.

Runs per day as a table

Listed runs per UTC day with at least one run
Day (UTC)	Merged	Other	Total
2026-09-22	1	0	1
2026-09-29	3	0	3
2026-10-06	13	0	13
2026-10-07	15	0	15
2026-10-08	29	0	29
2026-10-09	15	0	15
2026-10-10	8	0	8
2026-10-11	1	0	1
2026-10-13	2

The six figures and the cost split come from the operator's 30-day counters and include runs against repositories this page does not list. The chart of runs per day and everything below it use the 215 runs listed here.

Run tracer

Trace a run stage by stage

Pick a run to see every stage attempt in the order it ran, what each cost and which plan paid for it, and which gates passed or refused the change. The open run is in the URL, so a trace can be linked.

Listed runs

Live from the operator: stage detail is published for the most recent 24 runs; older runs are listed from their summary card.

Repository

Outcome

Escalation class

Sort by

Only runs with a stage trace (24)

Showing 215 of 215 listed runs

Mergedloom-corerun 529321593fa9

Mills canary: update heartbeat fixture

started 2026-10-15 09:00 UTC · merged 2026-10-15 10:17 UTC · 1h 16m wall clock

plan_slice

$0.401m 55s

research

<$0.0151s

implement

$0.051m 49s

tests

$0.004m 51s

pr_self_review

$0.211m 49s

$0.004s

ci_watch×3

$0.001h 04m

merge

$0.0040s

cleanup

$0.001s

run total

1h 16m in stages of a 1h 16m run

$0.671h 16m

9 stages, 11 attempts (2 retries) · 1h 16m in stages · $0.67 on subscription plans. Hover, focus, or tap a bar for its detail.

Bars run left to right in the order the operator recorded them. Width is proportional to the square root of duration so a 1-second cleanup stays visible next to a 17-minute implement; exact times are in the table. Time between stages, mostly model judges grading the change and back-off before a retry, has no bar. Costs are the vendor list-price equivalent of the tokens used; the gate judges' cost belongs to no stage, so it gets its own row and the column adds up to the run's total. Striped bars reused an earlier try's branch or merge request rather than redoing the work.

Trace as a table

Every stage attempt of this run in execution order
Step	Stage	Attempt	Outcome	Duration	Cost	Billed to
1	plan_slice	1	success	1m 55s	$0.40	Claude Code plan
2	research	1	success	51s	<$0.01	local GPU
3	implement	1	success	1m 49s	$0.05	Codex plan
4	tests	1	success	4m 51s	$0.00	—
5	pr_self_review	1	success	1m 49s	$0.21	Claude Code plan
6	mr	1	success	4s	$0.00	—
7	ci_watch	1	error	30m 00s	$0.00	—
8	ci_watch	2	error	30m 00s	$0.00	—
9	ci_watch	3	success	4m 45s	$0.00	—
10	merge	1	success	40s	$0.00	—
11	cleanup	1	success	1s	$0.00	—
Run total				1h 16m	$0.67	—

Gate verdicts

after post_implement_gate9 pass · 0 fail

nonempty_diff: pass
branch_pushed: pass
diff_size: pass
scope: pass
fabricated_slice: pass
path_policy: pass
secret_scan: pass
commit_format: pass
docs_guardrail: pass

after post_tests_gate1 pass · 0 fail

tests_verdict: pass

after post_review_gate3 pass · 0 fail

tested_head: pass
spec_conformance: pass
pr_self_review: pass

Run facts

Diff: 1 file, +1 / -1
Merge request: !2783
Grade: not graded
Escalation: none
Tested commit: d62e13eca818
Run cost: $0.67$0.67 on subscription plans
Backlog item: first tryno earlier try

Reading a run

Read escalations alongside merges. A stopped run records where a gate rejected the change and what it cost. The listed merges passed 13 gates and a CI pipeline. The later keep-or-regret assessment records what happened after merge. A further 51 merge-queue entries in this window were merge requests I opened by hand; they are counted above but not listed.

The most useful fact in a trace is the tested commit: the commit the tests stage actually ran against. Of the runs with stage detail here, 22 record a pinned commit and 0 record "unresolved". Until 2026-09-02 an unresolved pin meant the gate had quietly run against the sandbox copy of main, so the local gate and CI could check different code. Since the fix that afternoon, the stage fails when it cannot resolve the intended commit. Inspect the recorded test commit when comparing a run with its published change.

Reliability

Where attempts fail and where runs stop

Aggregates from the operator's stage and gate telemetry over the same 30 days, across every run, listed or not.

Stage reliability

Error rate per stage over all recorded attempts. plan_slice fails most often: 48% of 1423 attempts.

plan_slice48% of 1423
p50 2m 20s · p90 10m 23s · 594 retries
merge46% of 1235
p50 29m 04s · p90 30m 02s · 774 retries
ci_watch19% of 712
p50 4s · p90 27m 46s · 160 retries
tests17% of 1426
p50 8m 54s · p90 21m 24s · 785 retries
implement12% of 1361
p50 10m 53s · p90 35m 53s · 628 retries
pr_self_review7% of 786
p50 3m 48s · p90 17m 34s · 196 retries
research3% of 753
p50 55s · p90 2m 03s · 19 retries
mr1% of 639
p50 2s · p90 5s · 85 retries
cleanup1% of 399
p50 1s · p90 2s · 1 retries
post_implement_gate0% of 60
p50 0s · p90 0s · 11 retries
operator_pause0% of 1
p50 0s · p90 0s

A failed attempt is usually retried; the retry count is how many attempts came after the first. p50 and p90 are attempt durations.

Stage reliability as a table

Stage	Attempts	Error rate	Retries	p50	p90
tests	1426	17%	785	8m 54s	21m 24s
plan_slice	1423	48%	594	2m 20s	10m 23s
implement	1361	12%	628	10m 53s	35m 53s
merge	1235	46%	774	29m 04s	30m 02s
pr_self_review	786	7%	196	3m 48s	17m 34s
research	753	3%	19	55s	2m 03s
ci_watch	712	19%	160	4s	27m 46s
mr	639	1%	85	2s	5s
cleanup	399	1%	1	1s	2s
post_implement_gate	60	0%	11	0s	0s
operator_pause	1	0%	0	0s	0s

Gate verdicts

Pass, fail, and skip share per gate. tested_head refuses most often: 262 of 1026 evaluations failed (26%).

tested_head74% pass · 262 fail
scope86% pass · 158 fail
spec_conformance90% pass · 75 fail · 2 skip
commit_format91% pass · 93 fail
nonempty_diff91% pass · 100 fail
tests_verdict93% pass · 87 fail
diff_size96% pass · 44 fail
pr_self_review97% pass · 23 fail
path_policy99% pass · 15 fail
branch_pushed99% pass · 7 fail
docs_guardrail100% pass · 5 fail
secret_scan100% pass · 2 fail
fabricated_slice76% pass · 0 fail · 256 skip

passfailskip

Gate verdicts as a table

Gate	Evaluations	Pass	Fail	Skip
tests_verdict	1161	93%	87	0
nonempty_diff	1151	91%	100	0
scope	1128	86%	158	0
branch_pushed	1057	99%	7	0
commit_format	1057	91%	93	0
diff_size	1057	96%	44	0
docs_guardrail	1057	100%	5	0
fabricated_slice	1057	76%	0	256
path_policy

Where escalated runs stopped

Where escalated runs stopped, by last stage and its outcome. The most common stop is plan_slice (error): 87 of the 374 stops shown.

A run can stop after a stage that succeeded when a gate or budget refuses the next step.

Escalation stops as a table

Last stage	Outcome	Runs	Share
plan_slice	error	87	23%
implement	success	49	13%
tests	error	48	13%
ci_watch	error	41	11%
merge	error	41	11%
ci_watch	success	29	8%
post_implement_gate	gate_fail	28	7%
implement	error	19	5%
mr	success	17	5%
tests	success	15	4%

Merges by repository

400 merges; services/loom-core accounts for 395 (99%).

Repositories outside the published set are folded into 'other'.

Merges by repository as a table

Repository	Count	Share
services/loom-core	395	99%
libs/fi-fhir	4	1%
services/flexinfer	1	0%

Escalations by class

406 escalations; code accounts for 139 (34%).

The classified reason the operator recorded when a run stopped. Free-text reasons and log tails stay internal.

Escalations by class as a table

Class	Count	Share
code	139	34%
infra	87	21%
config	61	15%
transient	41	10%
transient_quota	40	10%
substrate	17	4%
external_dependency_incident	15	4%
unclassified	6	1%

Entry points

Recent merges and escalations

Each row opens that run in the tracer. Repositories: services/loom-core, libs/mcp-go, libs/fi-fhir, libs/edilint, services/flexinfer, services/flexdeck, services/flexinfer-site, services/procmodel.

Recent merges

109 listed, newest first

Pipeline merges against published repositories.

Recent escalations

83 listed, newest first

Each one is a run the gates refused to merge.

Sanitized from the Mills operator: identifiers are hashed, titles are scrubbed of paths and URLs, judge prose and log tails are dropped, and only runs against published repositories are listed. Escalations are counted, not hidden. The merge requests themselves live on a private GitLab; the commits that reach public repositories are on GitHub, and the tested commit shown in a trace is how you match a run to one.

Factory Evidence

Thirty-day ledger

Who paid for the compute

Listed runs per day

Trace a run stage by stage

Listed runs

Mills canary: update heartbeat fixture

Gate verdicts

Run facts

Reading a run

Where attempts fail and where runs stop

Stage reliability

Gate verdicts

Where escalated runs stopped

Merges by repository

Escalations by class

Recent merges and escalations

Recent merges

Recent escalations

Thirty-day ledger

Who paid for the compute

Listed runs per day

Trace a run stage by stage

Listed runs

Mills canary: update heartbeat fixture

Gate verdicts

Run facts

Reading a run

Where attempts fail and where runs stop

Stage reliability

Gate verdicts

Where escalated runs stopped

Merges by repository

Escalations by class

Recent merges and escalations

Recent merges

Recent escalations