feat: remove the manual-only Codex dispatch route and resume on the active dispatch route (#62)

.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "model-router",
-  "version": "0.25.1",
+  "version": "0.26.0",
   "description": "Route authorized software work through one durable, capacity-aware subscription policy.",
   "author": {
     "name": "toolboxmd",

.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "model-router",
-  "version": "0.25.1",
+  "version": "0.26.0",
   "description": "Route authorized software work through one durable, capacity-aware subscription policy.",
   "author": {
     "name": "toolboxmd",

.grok-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "model-router",
-  "version": "0.25.1",
+  "version": "0.26.0",
   "description": "Route authorized software work through one durable, capacity-aware subscription policy.",
   "author": {
     "name": "toolboxmd",

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## [0.26.0] - 2026-09-23
+
+### Added
+
+- Remove the manual-only Codex dispatch route and resume Codex on the job's active dispatch route
+
 ## [0.25.1] - 2026-09-23
 
 ### Changed

RUNNER.md

@@ -708,7 +708,7 @@ test keeps the two equal; `policy.main(['validate'])` checks the data.
 Stages in order: planning (Fable 5.1 max in Claude Code, then Astra max on
 Codex; Sonnet medium only as the explicit live-test override), dispatch (Luna
 max on Codex read-only first; `luna-go/max` on Go in OpenCode plan mode as the
-recorded fallback; Terra max is a manual-only option), the implementation
+recorded fallback), the implementation
 lanes default, small, and hard, planner-chosen rungs (Astra medium on Codex,
 Opus 5.5 high on the Claude harness; the planner itself chooses and runs them),
 critical

VERSION

@@ -1 +1 @@
-0.25.1
+0.26.0

runner/controller.py

@@ -287,9 +287,8 @@ def _save_codex_task(state_dir, request_id: str, task_id: str,
                      effort: str = adapters.CODEX_EFFORT,
                      adapter: str = "codex",
                      dispatch_route: str | None = None) -> dict:
-    """Persist the dispatcher task and, for the OpenCode fallback, its route
-    in one transaction, so a crash between two writes cannot leave a session
-    without its harness."""
+    """Persist the dispatcher task and its active route in one transaction,
+    so a crash between two writes cannot leave a session without its harness."""
     con = store.connect(state_dir)
     try:
         con.execute("BEGIN IMMEDIATE")
@@ -743,7 +742,7 @@ def dispatch(state_dir, request_id: str, run_cmd=None, probe=None) -> dict:
     # The thread exists even when the turn failed: save it so recovery
     # resumes it instead of creating a replacement task.
     _save_codex_task(state_dir, request_id, task_id, model=dispatch["model"],
-                     effort=dispatch["variant"])
+                     effort=dispatch["variant"], dispatch_route=dispatch_route)
     if not codex_h.turn_ok("codex_dispatch", rc, out, cmd):
         # A missing or rejected login blocks with its reason: no other
         # route holds the same login, so there is no fallback.
@@ -1008,12 +1007,12 @@ def resume_luna(state_dir, request_id: str, prompt: str, run_cmd=None,
                                      reason="resume", session_id=task_id, phase="resumed")
     last_path = _last_message_path(state_dir, request_id,
                                    "codex-resume-" + _prompt_digest(message))
-    dispatch = policy.ROUTES[policy.stage_routes("dispatch")[0]]
+    dispatch = policy.ROUTES[active]
     cmd = adapters.build_codex_resume_cmd(task_id, message, last_message_path=last_path,
                                           model=dispatch["model"], effort=dispatch["variant"])
     # Resume uses the saved workspace as cwd; never --cd/--reasoning.
     rc, out, err = run_cmd(cmd, job["workspace"], None, kind="codex_resume",
-                           meta={"stage": "dispatch", "route": "luna/max", "reason": "resume"})
+                           meta={"stage": "dispatch", "route": active, "reason": "resume"})
     codex_h = harnesses.harness_named("codex")
     resumed_id, _skind = codex_h.parse_session("codex_resume", out, err, job)
     _record_child(state_dir, request_id, "codex_resume", cmd, rc,

runner/policy.py

@@ -18,12 +18,13 @@
 from pathlib import Path
 
 POLICY_ID = "durable-runner-policy-v2"
-POLICY_VERSION = "2.5.0"
+POLICY_VERSION = "2.6.0"
 # Provenance: who decided this policy and where the evidence lives.
 POLICY_SOURCE = ("human decision, toolboxmd/model-router#10 (amended 2026-09-19), "
                  "#12, #26 (Go plan, 2026-09-20), #28/#29 (role kits, 2026-09-20), "
                  "#34 (usage probes, 2026-09-20), #38 (planner handoff, 2026-09-20), "
-                 "and #17 (Opus 5.5, human request, 2026-09-23)")
+                 "#17 (Opus 5.5, human request, 2026-09-23), "
+                 "and #62 (manual-only dispatch route removed, human direction, 2026-09-23)")
 POLICY_EVIDENCE = "https://github.com/toolboxmd/model-router/issues/29"
 
 # Subscription pools only. No Zen balance overflow, no pay-per-token APIs.
@@ -189,10 +190,6 @@
                       "variant": "medium", "role": "planning", "family": "claude",
                       "override_only": True, "context_window": 200_000,
                       "note": "explicit live-test override only, never a default"},
-    "terra/max": {"harness": "codex", "pool": "codex", "model": "gpt-5.6-terra",
-                  "variant": "max", "role": "dispatch", "family": "gpt",
-                  "manual": True, "context_window": 400_000,
-                  "note": "manual-only option; never selected automatically"},
     "luna/max": {"harness": "codex", "pool": "codex", "model": "gpt-5.6-luna",
                  "variant": "max", "role": "dispatch", "family": "gpt",
                  "sandbox": "read-only", "context_window": 400_000},
@@ -309,10 +306,9 @@
                  "note": "Fable 5.1 in Claude Code, then Astra max on Codex; "
                          "Sonnet medium only as the explicit live-test override"},
     "dispatch": {"executor": "runner", "routes": ["luna/max", "luna-go/max"],
-                 "manual": ["terra/max"],
                  "capabilities": ["read_only", "session_resume", "structured_output"],
                  "note": "read-only Codex sandbox first; the same model on Go in OpenCode plan mode "
-                         "when Codex cannot start the task; Terra max is a manual-only option"},
+                         "when Codex cannot start the task"},
     "implementation_default": {"executor": "runner",
                                "routes": ["muse-spark-xhigh-free", "muse-spark-xhigh-go",
                                           "glm-5.3-flash-go", "qwen3.8-flash-go",

skills/model-routing/references/codex.md

@@ -1,12 +1,12 @@
 # Routing policy
 
-Generated from `runner/policy.py` (`durable-runner-policy-v2` 2.5.0); do not edit by hand.
-Source: human decision, toolboxmd/model-router#10 (amended 2026-09-19), #12, #26 (Go plan, 2026-09-20), #28/#29 (role kits, 2026-09-20), #34 (usage probes, 2026-09-20), #38 (planner handoff, 2026-09-20), and #17 (Opus 5.5, human request, 2026-09-23). Evidence: https://github.com/toolboxmd/model-router/issues/29.
+Generated from `runner/policy.py` (`durable-runner-policy-v2` 2.6.0); do not edit by hand.
+Source: human decision, toolboxmd/model-router#10 (amended 2026-09-19), #12, #26 (Go plan, 2026-09-20), #28/#29 (role kits, 2026-09-20), #34 (usage probes, 2026-09-20), #38 (planner handoff, 2026-09-20), #17 (Opus 5.5, human request, 2026-09-23), and #62 (manual-only dispatch route removed, human direction, 2026-09-23). Evidence: https://github.com/toolboxmd/model-router/issues/29.
 
 | Stage | Routes in order | Notes |
 | --- | --- | --- |
 | planning | `fable-5.1/max`, `astra/max` (override: `sonnet/medium`) | host; Fable 5.1 in Claude Code, then Astra max on Codex; Sonnet medium only as the explicit live-test override |
-| dispatch | `luna/max`, `luna-go/max` (manual-only: `terra/max`) | runner; read-only Codex sandbox first; the same model on Go in OpenCode plan mode when Codex cannot start the task; Terra max is a manual-only option |
+| dispatch | `luna/max`, `luna-go/max` | runner; read-only Codex sandbox first; the same model on Go in OpenCode plan mode when Codex cannot start the task |
 | implementation_default | `muse-spark-xhigh-free`, `muse-spark-xhigh-go`, `glm-5.3-flash-go`, `qwen3.8-flash-go`, `deepseek-v4.1-flash-go`, `hy3-go`, `minimax-m3-go`, `mimo-v2.5-go`, `minimax-m2.7-go`, `longcat-2.0-go`, `glm-5.2-go`, `kimi-k2.6-go`, `glm-5.1-go` | runner; the full Go implementer chain in intelligence order, every new job starting on Muse free (no concurrency cap, parallel sessions by design); Qwen 3.7 Plus and 3.6 Plus are absent: no known tier. Policy declares exhaustion moves the same model to the next pool where declared (free Muse to Go Muse, Go Grok to xAI Grok), otherwise to the next family; overload moves to the next model family within one minute |
 | implementation_small | `muse-spark-xhigh-free`, `muse-spark-xhigh-go`, `glm-5.3-flash-go`, `qwen3.8-flash-go`, `deepseek-v4.1-flash-go`, `hy3-go`, `minimax-m3-go`, `mimo-v2.5-go`, `minimax-m2.7-go`, `longcat-2.0-go`, `glm-5.2-go`, `kimi-k2.6-go`, `glm-5.1-go` | runner; small bounded edits; every new job starts on Muse free (no concurrency cap, parallel sessions by design), then Muse Go, then from GLM 5.3 Flash onward |
 | implementation_hard | `muse-spark-xhigh-free`, `muse-spark-xhigh-go`, `glm-5.3-go`, `deepseek-v4-pro-go`, `grok-4.6-go`, `grok-4.6-build`, `grok-4.6-xai` | runner; $15 models get one turn per job; every new job starts on Muse free (no cap); Grok 4.6 sits here as the hard lane's last Go rung, then native Grok Build on the xAI subscription, then OpenCode's xAI provider |
@@ -21,7 +21,7 @@ Source: human decision, toolboxmd/model-router#10 (amended 2026-09-19), #12, #26
 
 The planner itself chooses one rung and runs it in its own session; the runner never selects these automatically: `astra/medium` = `gpt-6-astra` medium; `opus-5.5/high` = `claude-opus-5-5` high.
 
-Route models, in policy order: `fable-5.1/max` = `claude-fable-5-1` max; `astra/max` = `gpt-6-astra` max; `sonnet/medium` = `claude-sonnet-5` medium; `terra/max` = `gpt-5.6-terra` max; `luna/max` = `gpt-5.6-luna` max; `luna-go/max` = `opencode-go/gpt-5.6-luna`; `luna-max-review` = `gpt-5.6-luna` max; `luna-go-review` = `opencode-go/gpt-5.6-luna`; `astra/high-review` = `gpt-6-astra` high; `opus-5.5/high-review` = `claude-opus-5-5` high; `muse-spark-xhigh-free` = `opencode/muse-spark-1.3-contributor-free` xhigh; `muse-spark-xhigh-go` = `opencode-go/muse-spark-1.3-contributor` xhigh; `glm-5.3-go` = `opencode-go/glm-5.3`; `deepseek-v4-pro-go` = `opencode-go/deepseek-v4-pro`; `kimi-k2.7-code-go` = `opencode-go/kimi-k2.7-code`; `glm-5.3-flash-go` = `opencode-go/glm-5.3-flash`; `qwen3.8-flash-go` = `opencode-go/qwen3.8-flash`; `deepseek-v4.1-flash-go` = `opencode-go/deepseek-v4.1-flash`; `hy3-go` = `opencode-go/hy3`; `minimax-m3-go` = `opencode-go/minimax-m3`; `mimo-v2.5-go` = `opencode-go/mimo-v2.5`; `minimax-m2.7-go` = `opencode-go/minimax-m2.7`; `longcat-2.0-go` = `opencode-go/longcat-2.0`; `glm-5.2-go` = `opencode-go/glm-5.2`; `kimi-k2.6-go` = `opencode-go/kimi-k2.6`; `glm-5.1-go` = `opencode-go/glm-5.1`; `astra/medium` = `gpt-6-astra` medium; `opus-5.5/high` = `claude-opus-5-5` high; `grok-4.6-go` = `opencode-go/grok-4.6` medium; `grok-4.6-build` = `grok-4.6` medium; `grok-4.6-xai` = `xai/grok-4.6` medium.
+Route models, in policy order: `fable-5.1/max` = `claude-fable-5-1` max; `astra/max` = `gpt-6-astra` max; `sonnet/medium` = `claude-sonnet-5` medium; `luna/max` = `gpt-5.6-luna` max; `luna-go/max` = `opencode-go/gpt-5.6-luna`; `luna-max-review` = `gpt-5.6-luna` max; `luna-go-review` = `opencode-go/gpt-5.6-luna`; `astra/high-review` = `gpt-6-astra` high; `opus-5.5/high-review` = `claude-opus-5-5` high; `muse-spark-xhigh-free` = `opencode/muse-spark-1.3-contributor-free` xhigh; `muse-spark-xhigh-go` = `opencode-go/muse-spark-1.3-contributor` xhigh; `glm-5.3-go` = `opencode-go/glm-5.3`; `deepseek-v4-pro-go` = `opencode-go/deepseek-v4-pro`; `kimi-k2.7-code-go` = `opencode-go/kimi-k2.7-code`; `glm-5.3-flash-go` = `opencode-go/glm-5.3-flash`; `qwen3.8-flash-go` = `opencode-go/qwen3.8-flash`; `deepseek-v4.1-flash-go` = `opencode-go/deepseek-v4.1-flash`; `hy3-go` = `opencode-go/hy3`; `minimax-m3-go` = `opencode-go/minimax-m3`; `mimo-v2.5-go` = `opencode-go/mimo-v2.5`; `minimax-m2.7-go` = `opencode-go/minimax-m2.7`; `longcat-2.0-go` = `opencode-go/longcat-2.0`; `glm-5.2-go` = `opencode-go/glm-5.2`; `kimi-k2.6-go` = `opencode-go/kimi-k2.6`; `glm-5.1-go` = `opencode-go/glm-5.1`; `astra/medium` = `gpt-6-astra` medium; `opus-5.5/high` = `claude-opus-5-5` high; `grok-4.6-go` = `opencode-go/grok-4.6` medium; `grok-4.6-build` = `grok-4.6` medium; `grok-4.6-xai` = `xai/grok-4.6` medium.
 
 Rules:
 

tests/test_bridge.py

@@ -17,6 +17,7 @@
 import threading
 import time
 import unittest
+from unittest import mock
 from pathlib import Path
 
 ROOT = Path(__file__).resolve().parents[1]
@@ -259,6 +260,43 @@ def fake_resume(cmd, cwd=None, timeout=120, **kw):
         # Never contains a different task ID.
         self.assertNotIn("codex-other", " ".join(cmd))
 
+    def test_dispatch_persists_its_codex_route_and_resume_uses_it(self):
+        # A second Codex dispatch route listed first: dispatch must persist
+        # the route it actually ran on, and resume must use that route even
+        # after the stage order changes back.
+        alt = dict(policy.ROUTES["luna/max"], model="gpt-5.6-luna-alt")
+        orig_routes = list(policy.STAGES["dispatch"]["routes"])
+        captured = {}
+
+        def fake_dispatch(cmd, cwd=None, timeout=120, **kw):
+            captured["dispatch_cmd"] = list(cmd)
+            return 0, codex_out("codex-active-1", {"action": "planner_question",
+                                                   "qid": "q1", "prompt": "Confirm?"}), ""
+
+        def fake_resume(cmd, cwd=None, timeout=120, **kw):
+            captured["cmd"] = list(cmd)
+            captured["meta"] = kw["meta"]
+            return 0, codex_out("codex-active-1", {"action": "completion", "output": "done"}), ""
+
+        with mock.patch.dict(policy.ROUTES, {"luna-alt/max": alt}):
+            policy.STAGES["dispatch"]["routes"] = ["luna-alt/max"] + orig_routes
+            try:
+                core.submit(self.sd, "r1", {"goal": "t"}, self.ws(), "claude-1")
+                controller.dispatch(self.sd, "r1", run_cmd=fake_dispatch,
+                                    probe=lambda *a: None)
+            finally:
+                policy.STAGES["dispatch"]["routes"] = orig_routes
+            job = core.get_job(self.sd, "r1")
+            self.assertEqual(controller._load_controller_state(dict(job))["dispatch_route"],
+                             "luna-alt/max")
+            self.assertEqual(policy.stage_routes("dispatch")[0], "luna/max")
+            controller.resume_luna(self.sd, "r1", "follow up", run_cmd=fake_resume)
+        dispatch_cmd = captured["dispatch_cmd"]
+        self.assertEqual(dispatch_cmd[dispatch_cmd.index("--model") + 1], "gpt-5.6-luna-alt")
+        self.assertEqual(captured["cmd"][captured["cmd"].index("-m") + 1], "gpt-5.6-luna-alt")
+        self.assertIn('model_reasoning_effort="max"', captured["cmd"])
+        self.assertEqual(captured["meta"]["route"], "luna-alt/max")
+
     def test_planner_resume_never_forks_new_session(self):
         core.submit(self.sd, "r1", {"goal": "t"}, self.ws(), "claude-exact-77")
         captured = {}

tests/test_policy_v2.py

@@ -120,8 +120,7 @@ def test_stage_table_matches_parent_decision(self):
         self.assertEqual(s["correction"]["routes"], ["kimi-k2.7-code-go"])
         self.assertEqual(s["recovery"]["routes"], ["grok-4.6-go", "grok-4.6-build", "grok-4.6-xai"])
         self.assertEqual(s["dispatch"]["routes"], ["luna/max", "luna-go/max"])
-        self.assertEqual(s["dispatch"]["manual"], ["terra/max"])
-        self.assertTrue(policy.ROUTES["terra/max"]["manual"])
+        self.assertNotIn("manual", s["dispatch"])
         self.assertEqual(policy.ROUTES["luna-go/max"]["agent"], "plan")
         # Planner and review fallback routes on other subscriptions.
         self.assertEqual(s["planning"]["routes"], ["fable-5.1/max", "astra/max"])
@@ -148,7 +147,7 @@ def test_planner_chosen_rungs_are_not_implementation_lanes(self):
                 self.assertFalse(policy.ROUTES[route].get("planner_chosen"), (stage, route))
                 self.assertFalse(policy.ROUTES[route].get("manual"), (stage, route))
         for route in policy.implementation_routes() + policy.stage_routes("correction"):
-            self.assertNotEqual(policy.route_spec(route).get("variant"), "terra")
+            self.assertFalse(policy.route_spec(route).get("manual"))
 
     def test_route_params_and_pool_moves(self):
         self.assertEqual(policy.opencode_route_params("glm-5.3-flash-go"),
@@ -542,15 +541,15 @@ def test_max_concurrent_must_be_exactly_one(self):
     def test_guarded_route_in_lane_recovery_or_dispatch_fails(self):
         orig_routes = list(policy.STAGES["implementation_default"]["routes"])
         try:
-            policy.STAGES["implementation_default"]["routes"] = orig_routes + ["terra/max"]
+            policy.STAGES["implementation_default"]["routes"] = orig_routes + ["astra/medium"]
             problems = policy.validate_policy()
             self.assertTrue(any("implementation_default" in p and "never" in p for p in problems),
                             problems)
         finally:
             policy.STAGES["implementation_default"]["routes"] = orig_routes
         orig_rec = list(policy.STAGES["recovery"]["routes"])
         try:
-            policy.STAGES["recovery"]["routes"] = orig_rec + ["terra/max"]
+            policy.STAGES["recovery"]["routes"] = orig_rec + ["astra/medium"]
             problems = policy.validate_policy()
             self.assertTrue(any("recovery" in p and "never" in p for p in problems), problems)
         finally: