[JacobStephens2]

Summary

Freeze a Custom process and its referenced home skills for either Arm, retaining each file's SHA-256 and every Step's Harness, Model and Effort. Baseline defaults to the Built-in Ticket run.

prepare Arms → freeze Process + home skills + Step choices
Trial       → verify both Arms → copy into fresh home
            → execute Steps → stop before publishing
mismatch    → stop Experiment before another Trial; name changed file

Closes #1877

Evidence

Execution evidence

• Before: cargo test --test eval_arms custom_process_and_home_skill_are_frozen_with_step_choices failed with the prepare-arms usage error.
  After: Same command passes (1 test); frozen Process, instructions, helper and Step choices recorded.
• Before: cargo test --test experiments custom_process_trials_use_frozen_home_skills_beside_the_builtin_baseline failed when a later Trial reread the changed original Process (TOML parse error).
  After: Same command passes: 6 delivered Trials, Custom Candidate Steps and Built-in Baseline Steps.
• Before: The Spec review's exact commands python3 .thirdshift-review-FzmsRH/repro_spec.py issue and python3 .thirdshift-review-FzmsRH/repro_spec.py sandbox both failed with codex exited 55; the latter reported gitfile does not point to a valid repository.
  After: Both exact commands exit 0 with outcome: delivered. Retained regression cargo test --test trial custom_throwaway_steps_receive_the_issue_and_keep_git_access_on_another_harness -- --nocapture passes both pass-through and bwrap modes.
• cargo clippy --bin thirdshift -- -D warnings passes. Frozen-file tampering, mixed pinned Harnesses and older Built-in-only Baseline protocol regressions pass.
• CI is green for the exact head 524206a9: 3,947 Rust tests passed, 91 skipped, plus Site, Console, formatting, all-target clippy and Python tooling. cargo nextest run took 19m45s including 1m03s compilation (1,121.179s test execution). Both Linux release builds also passed.
• Local scripts/affected-tests origin/issue-1484 selected site checks and the full Rust suite: 68 site checks and 1,663 Rust unit tests passed. The first full run stopped because the host selected Python 3.9 without tomllib; cargo test --test bug_origins settlement::historical_review_and_reclosed_origin_regressions -- --nocapture passed with Python 3.11. The Python 3.11 retry was stopped after successful CI covered the exact head, reusing CI as docs/testing.md instructs.

Test seams: the evals prepare-arms, experiment and hidden --trial CLI interfaces, plus the rendered site documentation's format contract.

Visual evidence

Documentation captures use the same expanded sections and viewport states. Before is the pinned issue-1484 base; After is the implementation. Public documentation only.

Home, Experiment section

1440×900 desktop

390×844 mobile

How it works, Process Trials

1440×900 desktop

390×844 mobile

Manual, Process Trials

1440×900 desktop

390×844 mobile

Architecture, Trial execution

1440×900 desktop

390×844 mobile

Review coverage

Reviewed against issue-1484 at 8ad2733c with thirdshift-code-review. Standards: 1 finding, addressed. Spec: 2 findings, addressed. Every behavioral reproduction was run as written before deciding. Both axes read all 26 changed files; no unread changed files and no generated-output exclusions. Reports and finding counts are retained in .thirdshift-review-FzmsRH/{standards.md,spec.md,summary.json} beside the Session evidence.

Unaddressed findings

• Standards: None.
• Spec: None.
• Security: None.

Security review outcome

No unaddressed introduced findings.

Merge Danger

Door: two-way

Blast Radius: Evals

Experiment JSON advances to format 5 to describe frozen Process inputs; older saved Arm records remain readable. Process files needing GitHub before their first publishing Step remain refused.

Built with codex codex-cli 0.166.0 · gpt-6.1-sol · xhigh