[JacobStephens2]

Summary

Each of the three is chosen by the command first, then the User config, then the default:

Choice::of(flags, [harness] section)
  harness = flag ?? harness.default ?? claude
  model   = flag ?? [harness.<chosen>].model   ?? none (Harness's own)
  effort  = flag ?? [harness.<chosen>].effort  ?? none

 main
   command::start            # User config loaded
+  Flags::harness(config)    # one Choice per Command
   run / architect / pickup
     LaunchDirectory::open_for_run      # Origin match
+    Choice::check                      # CLI on PATH, codex "not supported yet",
+                                       # test call `claude -p --model M [--effort E]` if a Model is named
     logs::started
+      "sessions run on claude · opus · high"   (Command log + Activity log start line)
     Delivery
-      Sessions::within(logs, worktree)
+      Sessions::within(logs, worktree, harness)   # --model/--effort on every session, Repair and Resume
+      write "Built with claude · opus · high" in the PR body (Spec PR too), warning on failure
       child_run::start(issue, kind, base_fix
+                       , harness)                # --sessions-harness/-model/-effort hidden args

• harness <name>, model <name>, effort <level> (with or without dashes) on a Run, an Architect run and a Pickup run. They are rejected with --plan-only, when repeated, or when given without a value.
• [harness], [harness.claude] and [harness.codex] are parsed into harness::Settings. Unknown keys and wrong types are User config errors.
• Checks run only for a Run started by hand, or in a Pass once it has decided not to skip. A Spec run's Tickets, a Base fix and a dispatched run never repeat them.
• The Run notification gains the same Built with … line.
• Setup is untouched; that's Setup asks for the Harness, Model and Effort #393.

Evidence

• Before: thirdshift <url> model opus → not a GitHub issue URL: model; every session runs claude -p --permission-mode auto … with no --model.
• After: the new tests/harness_choice.rs and new tests in spec_run, pickup_run, architect_run and base_fix pass. They cover:

  no flag, no config         → sessions have no --model/--effort, no test call
  model X effort Y           → one test call ["-p","--model",X,"--effort",Y] "Reply with OK.", every session gets them
  harness claude, codex default → [harness.claude] used
  model Claude refuses       → exit 1, no Claim, no worktree, no Command log
  claude not on PATH         → "claude is not on PATH … chosen by …"
  harness codex              → "harness codex, chosen by …, is not supported yet"
  Spec run                   → one test call; #21, #22 and the Spec review on the Model
  skipped Pickup/Architect   → no claude call at all
  Base fix                   → its session on the Run's Model and Effort
  Command log, Activity log, PR body, notification → name claude · X · Y
  

• The full suite passes except two failures already on issue-136. The Factory skills were updated in a2afa36 without regenerating them: prompts_page_shows_what_the_binary_sends (stale site/prompts) and loads_every_factory_skill_as_the_thirdshift_plugin (missing resolving-merge-conflicts). Neither involves this change.

Merge Danger

Door: two-way

With no flag and no [harness], sessions run exactly as before. The new visible output is:

• a sessions run on … line in the Command log;
• , on claude · default model · default effort at the end of the Activity log start line;
• one Built with … line, with a hidden marker, appended to each pull request body.

Blast Radius: moderate

Anything that parses the Activity log start line, or a PR body exactly, sees the new text. A Run with a named Model makes one extra small Claude call before work.

Unaddressed findings

Standards

• Choice::check does I/O inside a value type (testability). It is covered end to end by the fake claude on PATH, which is the seam the spec names. Injecting a runner adds a seam no test needs.
• Flags::harness and Started::built_with are middle men. They mirror the existing Flags::notification and Started::took, keeping one way to resolve and record an ask.
• The same model/effort shape is duplicated in the retry command, hidden args and Claude args. Each produces a different syntax (public flags, hidden flags, claude flags), so a shared helper would need its own mapping table.
• Choice flattens model/effort instead of holding ModelAndEffort; Reader keeps them separate. Flat fields read more simply at every use (choice.model). It is cosmetic.
• Naming. The harness: Choice parameters lead to self.harness.harness, and built_with is both a getter and a setter. The parameter name matches the domain term (the Harness choice). Renaming every site is churn without clearer meaning. I renamed the unclear local sessions_on.
• Shotgun surgery from threading &Choice through ~12 signatures. This is explicit by design. The alternative is process-wide state, and the spec requires the choice to reach every session and child Run.
• The --plan-only flag list is repeated in the message, docs and tests. That matches how the existing list was already spelled out.
• The round-trip test pairs kinds with harnesses through zip/cycle. It is a test idiom only.

Spec

• The test call has no timeout. The spec asks for none. A hung claude would equally hang the session that follows.
• A child Run records ChosenBy::Command even when the User config chose. It is never shown, because child Runs make no checks. It is documented at the field.
• Scope creep: the Base fix retry command repeats harness/model/effort. Kept, so the retried Run asks for what the original asked.

Closes #392

🤖 Generated with Claude Code