unitbob 0.4.2 → 0.4.4

package/README.md

@@ -30,7 +30,7 @@ claude plugin install unitbob@unitbob
 ```
 codex plugin marketplace add sergeygershun/unitbob-connector
 codex plugin add unitbob@unitbob
-npx -y unitbob@0.4.2 codex-install
+npx -y unitbob@0.4.4 codex-install
 ```
 
 Start a new Claude Code or Codex thread so the installed skill and named agents

package/dist/runner/worldProbe.js

@@ -28,13 +28,14 @@ export async function probeBehavioralWorld(projectRoot, deps = {
             env: {
                 ...process.env,
                 RAILS_ENV: 'test',
+                CUCUMBER_PUBLISH_QUIET: 'true',
                 UNITBOB_REPO_ROOT: projectRoot,
                 BUNDLE_GEMFILE: join(projectRoot, '.unitbob', 'behavioral', 'Gemfile'),
             },
         });
         if (result.code === 0)
             return { status: 'ok' };
-        const detail = [result.stderr, result.stdout].find((text) => text.trim())?.trim() ?? `exit ${result.code}`;
+        const detail = [result.stdout, result.stderr].map((text) => text.trim()).filter(Boolean).join('\n') || `exit ${result.code}`;
         return { status: 'fixable', message: `The connector-owned Ruby/Cucumber World probe failed: ${detail}` };
     }
     catch (error) {
@@ -58,6 +59,14 @@ PROBE_OTHER_TIME_ZONE = PROBE_TIME_ZONE&.name == 'UTC' ? 'Hawaii' : 'UTC'
 PROBE_LOCALE = I18n.locale
 PROBE_RECEIVER = Object.new
 
+# The probe runs in a dedicated process, so replacing its in-memory route set
+# cannot affect the host application after the process exits. Keeping these
+# endpoints here makes the profile check independent of routes the host owns.
+Rails.application.routes.draw do
+  match '/__unitbob_world_probe__', to: proc { [200, { 'Content-Type' => 'text/plain' }, ['ok']] }, via: :all
+  match '/__unitbob_world_probe_redirect__', to: redirect('/__unitbob_world_probe_target__'), via: :all
+end
+
 Given('the first World probe scenario mutates supported state') do
   @unitbob_connection.execute("INSERT INTO schema_migrations (version) VALUES ('#{PROBE_VERSION}')")
   Time.zone = PROBE_OTHER_TIME_ZONE

package/dist/verbs/runLocal.js

@@ -33,24 +33,11 @@ export async function runLocal(config, args = [], deps) {
         // examined the suite.
         if (!ran)
             continue;
-        const spent = spend(config.projectRoot, `run-local:${suiteKind}`);
-        if (request.budget && spent > request.budget.repair_rounds) {
-            d.stdout.write(polishedEnoughNotice(suiteKind, spent, request.budget.repair_rounds));
-        }
+        // Kept as compatibility diagnostics only. The bounded repair role owns the
+        // mechanical ceiling; this counter never stops execution or classifies reds.
+        spend(config.projectRoot, `run-local:${suiteKind}`);
     }
 }
-// Not a refusal, and deliberately not about the budget either. After eight
-// rounds of repair the interesting fact is not that a number ran out — it is
-// what the remaining reds most likely are. Both real logs show 3-5 runs of a
-// branch as ordinary work, so a branch on its ninth has already been repaired
-// past the point where the harness is the usual explanation.
-function polishedEnoughNotice(suiteKind, spent, allowed) {
-    return (`\nThat was run ${spent} of the ${suiteKind} branch; this build budgeted ${allowed}. ` +
-        'Reds that survive this many rounds of repair are far more likely to be defects of your product ' +
-        'than of the harness around it.\n' +
-        'Publish the suite as it stands rather than keep polishing. A first suite that comes out red is ' +
-        'a finding, not a failure — finding those reds is what it was written to do.\n');
-}
 // Which branches to run. No argument runs every branch the request asked for —
 // the same "one suite, one run" shape both recipes insist on, so the default
 // never teaches the habit the recipes forbid. A named branch is for the repair

package/dist/verbs/suitePrepare.js

@@ -173,7 +173,7 @@ export async function suitePrepare(config, args = [], deps) {
     // A fixable runner blocker is not a failure: the structural suite still builds this run. Tell the
     // vibecoder the one command that unblocks the behavioral peer, then re-run suite-prepare.
     if (fixableNotices.length > 0) {
-        actual.stdout.write('\nBehavioral suite skipped this run — its BDD runner is not installed yet. ' +
+        actual.stdout.write('\nBehavioral suite skipped this run — its runner or connector-owned World profile is not ready. ' +
             'This is a fixable setup step, not a build failure, and it does not affect the structural suite:\n' +
             fixableNotices.join('\n') +
             '\nFix the above, then re-run `unitbob suite-prepare` to build the behavioral peer.\n');

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "unitbob",
-  "version": "0.4.2",
+  "version": "0.4.4",
   "description": "Unitbob connector — thin local hands for the Unitbob Rails brain. Owns no domain logic: it runs tools, relays bytes over the wire, and prints what the server returns.",
   "type": "module",
   "bin": {

package/plugin/codex/agents/suite-repair-worker.toml

@@ -1,16 +1,37 @@
 name = "suite-repair-worker"
-description = "Completes one bounded Unitbob failure packet using the prior checkpoint and owned files, without widening scope or running the suite."
+description = "Completes and executes one bounded Unitbob owned-slice repair packet without widening scope."
 model = "gpt-5.6-terra"
 model_reasoning_effort = "medium"
 developer_instructions = '''
-You receive one failure packet: one validated plan item, its checkpoint, owned
-paths, and only related failures or stack traces. This is your complete scope.
+You receive one failure packet: one validated plan item, its checkpoint, branch,
+owned paths and case markers, and only related failures or stack traces. This is
+your complete write scope.
 
 Complete `unresolved_promises` first while preserving every completed file and
-decision. Then repair only harness problems whose stack points into this slice's
-owned files. Do not expand capabilities, promises, planned cases, or paths. Do
-not edit host-owned shared files, the connector-owned harness, application
-production code, or another slice.
+decision. You may read the plan item's source paths and, only as needed for the
+owned diagnosis, stack-referenced project source, the runner setup and harness
+actually in use, helpers, and factories. Repair only this slice's owned generated
+files and its checkpoint. Do not expand capabilities, promises, planned cases,
+markers, or paths. Do not edit production code, host-owned shared files, the
+connector-owned harness, or another slice.
+
+After every owned edit, run
+`npx -y --loglevel=error unitbob@0.4.4 run-local <branch>` and inspect the machine
+report. Look only at examples or scenarios matching your owned paths or case
+markers. Do not require a green exit code from the whole branch: foreign failures
+and an already-confirmed product red do not widen your scope. Repeat the bounded
+`edit → run-local → inspect` loop until every owned case passes or is diagnosed
+as a product defect. Do not run the project's suite directly, boot a dev server,
+or invoke an arbitrary runner command.
+
+A product-defect diagnosis must briefly name the violated business contract, the
+reason, and production source references. Never delete a planned case, marker,
+capability binding, or assertion; never add `skip`, `pending`, `todo`, or weaken a
+business promise for green. You may correct a generated expectation only when
+the business promise remains intact and source confirms the correction. If the
+harness is still wrong, continue the loop. If the outcome is ambiguous, the
+runner is unusable, or the turn ceiling stops unfinished work, leave an honest
+branch `build_error`, never a product red. No strict JSON handoff is required.
 
 Update the same checkpoint as promises complete. Keep facts compact and
 source-referenced. The normative JSON shape of one facts entry is:
@@ -19,15 +40,12 @@ source-referenced. The normative JSON shape of one facts entry is:
 ```
 Every facts entry is an object in that shape, never a string. Before handoff,
 make one final read of the checkpoint and confirm every `facts` entry is an
-object in the normative shape above. Never run the suite or boot the
-application; the coordinator
-owns the single final run. Do not delegate a second repair, continue another
-agent, or request another repair round. If work remains at the turn ceiling,
-record it in `unresolved_promises` so the coordinator can produce an honest
-branch `build_error`.
+object in the normative shape above. Do not delegate repair or auto-resume after
+the ceiling. Preserve files and checkpoint for the coordinator's existing
+`Continue once / Stop` choice; record unfinished work in `unresolved_promises`.
 '''
 
 [features.rollout_budget]
 enabled = true
-limit_tokens = 15000
+limit_tokens = 40000
 reminder_at_remaining_tokens = [3000]