[JacobStephens2]

Summary

0.26.0 adds process Experiments: compare Built-in or Custom Ticket processes on merged work, grade their final commits, and report paired quality, cost and time Verdicts (#1915).

• thirdshift triage <Issue URL or number> [base <branch>] investigates and updates an existing issue in place (Support targeted in-place issue triage #1911). With terminal input or empty stdin at EOF, thirdshift triage [base <branch>] selects one eligible needs-triage issue, including from cron (Run unattended Triage Passes from terminal or EOF input #1918).
• thirdshift evals propose-process <repository> drafts tasks from merged Tickets. evals approve process <task> and evals prepare process <task> <destination> approve and screen them: tests must fail at base and pass at merge; preparation refuses changed inline Rust tests and produces a base-only Trial repository (Build and screen process Eval tasks from merged Tickets #1926).
• Built-in Ticket Trials run implementation and Security review locally, retain Steps, Sessions and traces, and stop before publication (Isolate the Process engine’s Ticket source and publication #1907, Run Built-in Ticket processes as Experiment Trials #1919). Reference tests grade the Trial’s final committed work offline under one shared timeout; process manifests now require process_grading (Grade process Trials with the reference fix's tests #1927, Spec: process Trials and a code-graded implement Eval suite #1915).
• Custom process Arms freeze Process files, referenced home skills and each Step’s Harness, Model and Effort; frozen files are verified before every Trial (Freeze Custom process Arms and home skills for Experiment Trials #1928). Changed Custom input content selects a new Experiment, and Experiment JSON advances to format 8 (Spec: process Trials and a code-graded implement Eval suite #1915).
• --prices supplies frozen, dated token prices for Session, Trial, role and Arm accounting, with paired median cost/time ratios. Missing usage or unpriced Models make a Trial Void (Price Trial sessions by role and report paired cost and time #1932). Opt-in --cost-ceiling and --time-ceiling constrain Candidate/Baseline ratios; Verdicts add cheaper/faster annotations, and unknown measurements cannot pass a configured ceiling (Add cost and time ceilings and cheaper/faster Experiment Verdicts #1938).
• Process Trials scan retained Session evidence for the real fix, including stopped Sessions after resume. Pending hits hold the Verdict; human dismissal retains an auditable false positive, while confirmation voids the Trial and recalculates the paired Verdict (Check process Trial Session evidence for the real fix #1934).
• Triage preserves category labels, holds readiness at ready-for-human unless permitted, and cleans up Claims and reservations across success, failure, interruption and Stop (Support targeted in-place issue triage #1911, Unattended triage of existing needs-triage issues #1917). Passes honor Command/Failure limits and configured issue order, retaining failed Triage in failure accounting (Run unattended Triage Passes from terminal or EOF input #1918). Sessions receive bounded cached images from trusted GitHub attachments (Pre-download trusted issue images for triage #1922); Report forms honor the selected base branch, and private Security-finding evidence stays out of email (Unattended triage of existing needs-triage issues #1917).
• Prepared grader snapshot directories are now owner-only, preventing other accounts from reading the prepared repository there (Create the prepared grader snapshot directory as owner-only #1941).

Also: site and docs updates (#1917, #1915), plus the 0.25.0 release PR (#1937).

Version diff

diff --git a/Cargo.lock b/Cargo.lock
index fd5c8cb2..f9895513 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -2432,3 +2432,3 @@ dependencies = [
 name = "thirdshift"
-version = "0.25.0"
+version = "0.26.0"
 dependencies = [
diff --git a/Cargo.toml b/Cargo.toml
index be3ead3a..e8f80093 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -2,3 +2,3 @@
 name = "thirdshift"
-version = "0.25.0"
+version = "0.26.0"
 edition = "2024"