adeelahmad · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026 · Oct 18, 2026
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -23,6 +23,7 @@ tag in lockstep. Tag releases as `vMAJOR.MINOR.PATCH`.
 ## [Unreleased]
 
 ### Added
+- `gate-dispatch` now refuses a dispatch of any aloop role whose `model` is not the project's confirmed one (`aloop models resolve <role>`, matched by model family so `haiku` and `claude-haiku-5-5` agree), or that passes no model (which would fall back to the agent file's default, opus for the planning roles). Reason: live e2e run 20 dispatched planners on opus and workers on sonnet against a project whose defaults.md said haiku for both, and cost $2.79 instead of $0.99 for the same 12/12 result. Test: `scripts/test/dispatch.test.py`. Live run 21 with the gate: 12/12, 8m17s, $0.66, every dispatch on haiku (ledger rows for runs 20 and 21 in `docs/evidence/README.md`).
 - Evidence ledger (ROADMAP §4.11): `docs/evidence/README.md` lists every live run behind a public
   number (intent, plugin version, checks, wall, cost, wakes, plugin reads, gate blocks, reproduce
   command); `bin/run-summary <run-dir>` writes a run's `summary.json`, `summary.md` and ledger row from

diff --git a/docs/ROADMAP.md b/docs/ROADMAP.md
diff --git a/docs/TRACKING.md b/docs/TRACKING.md
@@ -12,6 +12,7 @@ Status of every requirement in [ROADMAP.md](ROADMAP.md), in the four states §8
 | 3.4 | Hard breaches classified with evidence before acting | IMPLEMENTED_TESTED | `scripts/test/budget-classify.test.py` |
 | 3.5 | No post-kill narration; harness writes the re-plan block | IMPLEMENTED_TESTED | `scripts/test/budget-hard-stop.test.py` |
 | 3.6 | Abstract model routes | IMPLEMENTED_TESTED | `scripts/test/model-routes.test.py` |
+| 3.6 | The dispatch model is the project's resolved route, enforced by `gate-dispatch` | IMPLEMENTED_TESTED | `scripts/test/dispatch.test.py` |
 | 4.1 | Intent Contract with harness-issued ids | IMPLEMENTED_TESTED | `scripts/test/intent-contract.test.py` |
 | 4.1 | Clarification policy (assume / default / ask / spike) | IMPLEMENTED_TESTED | `scripts/test/intent-contract.test.py` |
 | 4.1 | Explicit autonomous mode | IMPLEMENTED_TESTED | `scripts/test/intent-contract.test.py` |

diff --git a/docs/evidence/README.md b/docs/evidence/README.md
@@ -24,11 +24,15 @@ stories). What each run found and fixed is in ROADMAP M0.8.
 | run 17 | wordfreq | v0.11.x (master, 2026-10-18) | 13/13 | 12m58s | $0.95 | 2 | 0 | 6 | `make e2e-live YES=1 INTENT=wordfreq` |
 | run 18 | wordfreq | v0.12.0 (fd1fd31) | 13/13 | 21m21s | $1.03 | 6 | 2 | 6 | `make e2e-live YES=1 INTENT=wordfreq` |
 | run 19 | wordfreq | v0.12.0 + #29 (b93e6fd) | 13/13 | 15m24s | $1.01 | 1 | 0 | 3 | `make e2e-live YES=1 INTENT=wordfreq` |
+| run 20 | semver | v0.12.0 + #31 (915810c) | 12/12 | 10m47s | $2.79 | 1 | 2 | 1 | `make e2e-live YES=1 INTENT=semver` |
+| run 21 | semver | v0.12.0 + #33 (1ad94e2) | 12/12 | 8m17s | $0.66 | 1 | 0 | 3 | `make e2e-live YES=1 INTENT=semver` |
 
 Columns: *checks* are the e2e script's own assertions (12 for `semver`, 13 for `wordfreq` since
 the commit-trace check); *wakes* are supervisor turns that ended and resumed (each re-reads
 context; the main cost driver); *plugin reads* are sub-agent tool calls into the plugin's own
 source; *gate blocks* are `exit 2` verdicts in the run's `gates.jsonl` (host hooks and
 self-checks). A dash means the run predates that counter.
+Run 20 cost $2.79 because the supervisor dispatched planners on opus and workers on sonnet against a
+project set to haiku; `gate-dispatch` now refuses that (#33), and run 21 is the same intent on haiku.
 Run 19's reads were first reported as 5: all five were the planner's `md-db describe --schema` on
 the plugin's schemas through a shell variable, which the counter now recognises as schema use.
diff --git a/plugin/bin/README.md b/plugin/bin/README.md
@@ -25,7 +25,7 @@ ctx-symbols) → WARN + grep fallback; never a silent pass.
 | script | event · matcher | checks | exit 2 when |
 |--------|-----------------|--------|-------------|
 | gate-supervisor-scope     | PreToolUse · Write\|Edit\|MultiEdit\|Bash | MAIN agent (no `agent_type`) may not write production source mid-sprint, nor hand-author planning artifacts (stories/tasks/validate/plan/intake/standards.md — dispatch the activity), nor `git commit` on main (commits on main arrive through `task-worktree merge`); sub-agents (`agent_type` present) exempt | supervisor writes source / hand-writes a planner-owned artifact / commits on main |
-| gate-dispatch             | PreToolUse · Agent\|Task | the MAIN agent's dispatch of an execution role opens with `TASK_ID: <id>` and that task's contract `<run-data>/state/tasks/<id>.env` exists (repos with a run-data dir only) | no TASK_ID line / contract not written yet |
+| gate-dispatch             | PreToolUse · Agent\|Task | the MAIN agent's dispatch of any aloop role passes the project's confirmed model (`aloop models resolve <role>`, matched by family; no model or another model is refused), and a dispatch of an execution role opens with `TASK_ID: <id>` and that task's contract `<run-data>/state/tasks/<id>.env` exists (repos with a run-data dir only) | no TASK_ID line / contract not written yet |
 | gate-tooling              | SubagentStart · exec roles + supervisor self-check | md-db + ctx-symbols installed (execution may not run grep-degraded); runs `ensure-tools --sync` first | a backend is missing and can't be built |
 | gate-validate-artifact    | PostToolUse · Write         | md-db-validate the written artifact | artifact malformed |
 | gate-red-verify           | SubagentStop · red-worker   | worktree-isolated; appended output.md block (validate_comms); every new test FAILS BY ASSERTION; no regression; diff = tests+shims | no/stale output.md block / shared-tree run / a new test passes / prod code / regression |

diff --git a/plugin/bin/gate-dispatch b/plugin/bin/gate-dispatch
@@ -15,21 +15,46 @@ HERE="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 payload="$(cat 2>/dev/null || true)"
 
 command -v python3 >/dev/null 2>&1 || exit 0
-read -r role tid < <(printf '%s' "$payload" | python3 -c '
+read -r role tid model < <(printf '%s' "$payload" | python3 -c '
 import json, re, sys
 p = json.load(sys.stdin)
 if p.get("agent_type"):
-    print("- -"); sys.exit()
+    print("- - -"); sys.exit()
 ti = p.get("tool_input") or {}
 role = (ti.get("subagent_type") or "").split(":")[-1]
 m = re.match(r"\s*TASK_ID:\s*([A-Za-z0-9._-]+)", ti.get("prompt") or "")
-print(role or "-", m.group(1) if m else "-")' 2>/dev/null) || exit 0
+model = re.sub(r"\s+", "", str(ti.get("model") or "")) or "-"
+print(role or "-", m.group(1) if m else "-", model)' 2>/dev/null) || exit 0
 
-case "$role" in red-worker|scaffolder|green-worker|structural-reviewer|final-gate) ;; *) exit 0 ;; esac
+case "$role" in red-worker|scaffolder|green-worker|structural-reviewer|final-gate|planner|intake|standards|archivist) ;; *) exit 0 ;; esac
 # shellcheck source=/dev/null
 . "$HERE/_paths.sh"
 [ -d "$AGENTIC_HOME" ] || exit 0
 
+# The model is the project's confirmed one (docs/agents/defaults.md → `aloop models resolve <role>`,
+# ROADMAP §3.6), never the supervisor's pick: live e2e run 20 dispatched planners on opus and
+# workers on sonnet with defaults.md saying haiku for both, and cost $2.79 instead of $0.99. A
+# pinned id or a route is matched by model family (haiku / sonnet / opus), so `haiku` and
+# `claude-haiku-5-5` are the same answer; an absent model would fall back to the agent file's
+# default (opus for the planning roles), so it is refused too. An empty resolution (Codex: the
+# session's own model) checks nothing.
+want="$("$HERE/models" resolve "$role" 2>/dev/null)" || want=""
+if [ -n "$want" ]; then
+  fam() { case "$(printf '%s' "$1" | tr '[:upper:]' '[:lower:]')" in *haiku*) echo haiku ;; *sonnet*) echo sonnet ;; *opus*) echo opus ;; *) echo "" ;; esac; }
+  bad=0
+  if [ "$model" = "-" ]; then bad=1
+  elif [ "$model" != "$want" ]; then
+    wf="$(fam "$want")"; mf="$(fam "$model")"
+    { [ -z "$wf" ] || [ -z "$mf" ] || [ "$wf" != "$mf" ]; } && bad=1
+  fi
+  if [ "$bad" = 1 ]; then
+    echo "BLOCK[$GATE_NAME]: dispatch $role with model: $want (the project's confirmed model for this role: \`aloop models resolve $role\`, from docs/agents/defaults.md)${model:+; this call passes $( [ "$model" = "-" ] && echo "no model, which would fall back to the agent file's default" || echo "model: $model")}." >&2
+    echo "  The model per role is a project setting (confirmed at init), not a per-dispatch choice. To change it: re-run the init skill or edit .aladdin/config.yaml and \`aloop config sync\`." >&2
+    exit 2
+  fi
+fi
+case "$role" in planner|intake|standards|archivist) exit 0 ;; esac
+
 if [ "$tid" = "-" ]; then
   echo "BLOCK[$GATE_NAME]: start the $role dispatch prompt with the line \`TASK_ID: <id>\` — its gates find the task contract through it." >&2
   exit 2

diff --git a/plugin/skills/aloop/SKILL.md b/plugin/skills/aloop/SKILL.md
@@ -38,8 +38,10 @@ Roles map to a ROUTE (`strongest` · `strong` · `economical`), and each host re
 its own id (`plugin/hosts/routes.json`). The project's confirmed choice lives in
 `docs/agents/defaults.md` (`AGENTIC_MODEL_PLANNING`, `AGENTIC_MODEL_WORKER`: a route or a
 pinned id). On every dispatch pass `aloop models resolve planning|worker` as the Agent
-call's `model`; never hard-code an id. The agent frontmatter carries the Claude Code
-resolution of the defaults.
+call's `model`; never hard-code an id. `gate-dispatch` refuses a dispatch on any other model,
+or with none (live run 20: planners on opus and workers on sonnet against a haiku project cost
+$2.79 instead of $0.99). The agent frontmatter carries the Claude Code resolution of the
+defaults.
 
 ## Project setup comes first (init) — and where everything lives
 

diff --git a/scripts/test/dispatch.test.py b/scripts/test/dispatch.test.py
@@ -25,16 +25,21 @@ def ok(msg):
     print("ok - " + msg)
 
 
-def run(subagent_type, prompt, agent_type=None):
+def run(subagent_type, prompt, agent_type=None, model="haiku"):
     p = {"session_id": "s", "cwd": T, "tool_name": "Agent",
          "tool_input": {"subagent_type": subagent_type, "prompt": prompt, "isolation": "worktree"}}
+    if model:
+        p["tool_input"]["model"] = model
     if agent_type:
         p["agent_type"] = agent_type
     return subprocess.run([GATE], input=json.dumps(p), cwd=T, env=ENV, capture_output=True, text=True)
 
 
 subprocess.run(["git", "init", "-q"], cwd=T, check=True)
 os.makedirs(os.path.join(T, ".aladdin", "state", "tasks"))
+os.makedirs(os.path.join(T, "docs", "agents"))
+open(os.path.join(T, "docs", "agents", "defaults.md"), "w").write(
+    "---\ntype: agentic-defaults\n---\n```conf\nAGENTIC_MODEL_PLANNING=haiku\nAGENTIC_MODEL_WORKER=haiku\n```\n")
 
 r = run("aloop:green-worker", "TASK_ID: S1-01-T1\nDo GREEN.")
 assert r.returncode == 2 and "S1-01-T1.env" in r.stderr, r
@@ -75,7 +80,35 @@ def run(subagent_type, prompt, agent_type=None):
 for st in ("aloop:planner", "aloop:intake", "general-purpose", "Explore"):
     r = run(st, "Plan it.")
     assert r.returncode == 0, (st, r)
-ok("planning roles and non-aloop agents are not affected")
+for st in ("general-purpose", "Explore"):
+    r = run(st, "Look around.", model="opus")
+    assert r.returncode == 0, (st, r)
+ok("planning roles need no contract; non-aloop agents are not affected at all")
+
+# The dispatch model is the project's confirmed one (ROADMAP §3.6), not the supervisor's pick.
+# Live e2e run 20: defaults.md said haiku for every role and `aloop models resolve` printed it,
+# yet the supervisor dispatched planners on opus and workers on sonnet: $2.79 instead of $0.99.
+for st, m in (("aloop:green-worker", "sonnet"), ("aloop:green-worker", "claude-opus-5-5"), ("aloop:planner", "opus"),
+              ("aloop:intake", "claude-sonnet-5-5"), ("aloop:archivist", "sonnet")):
+    r = run(st, "TASK_ID: S1-01-T1\nDo it.", model=m)
+    assert r.returncode == 2 and "model" in r.stderr and "haiku" in r.stderr and "aloop models resolve" in r.stderr, (st, m, r)
+r = run("aloop:green-worker", "TASK_ID: S1-01-T1\nDo GREEN.", model=None)
+assert r.returncode == 2 and "model: haiku" in r.stderr, r
+r = run("aloop:planner", "Plan it.", model=None)
+assert r.returncode == 2 and "model: haiku" in r.stderr, r
+ok("a dispatch on a model other than the project's resolved one, or with no model, is denied for every aloop role")
+for st, m in (("aloop:green-worker", "haiku"), ("aloop:green-worker", "claude-haiku-5-5"), ("aloop:planner", "haiku")):
+    r = run(st, "TASK_ID: S1-01-T1\nDo it.", model=m)
+    assert r.returncode == 0, (st, m, r)
+open(os.path.join(T, "docs", "agents", "defaults.md"), "w").write(
+    "---\ntype: agentic-defaults\n---\n```conf\nAGENTIC_MODEL_PLANNING=strongest\nAGENTIC_MODEL_WORKER=strong\n```\n")
+r = run("aloop:planner", "Plan it.", model="claude-opus-5-5")
+assert r.returncode == 0, r
+r = run("aloop:green-worker", "TASK_ID: S1-01-T1\nDo GREEN.", model="sonnet")
+assert r.returncode == 0, r
+r = run("aloop:green-worker", "TASK_ID: S1-01-T1\nDo GREEN.", model="haiku")
+assert r.returncode == 2 and "sonnet" in r.stderr, r
+ok("the resolved model (pinned id or route, by family) is accepted; a cheaper-than-confirmed pick is denied too")
 
 r = run("aloop:green-worker", "TASK_ID: S1-09-T9\nDo GREEN.", agent_type="aloop:planner")
 assert r.returncode == 0, r

diff --git a/scripts/test/layout.test.py b/scripts/test/layout.test.py
@@ -113,13 +113,13 @@ def sh_home(t, **env):
 assert out.splitlines()[0] == os.path.join(C, ".agentic", "state", "tasks", "T1.env") and out.splitlines()[-1] == "rc=1", out
 assert run(C, "python3", "-c", 'import sys; sys.path.insert(0, "%s"); import _paths; print(_paths.contract_path("%s", "T1"))' % (BIN, C)).stdout.strip() == os.path.join(C, ".agentic", "state", "tasks", "T1.env")
 payload = json.dumps({"hook_event_name": "PreToolUse", "tool_name": "Agent", "cwd": C,
-                      "tool_input": {"subagent_type": "aloop:green-worker", "prompt": "TASK_ID: T1\ndo it"}})
+                      "tool_input": {"subagent_type": "aloop:green-worker", "prompt": "TASK_ID: T1\ndo it", "model": "sonnet"}})
 r = run(C, os.path.join(BIN, "gate-dispatch"), input=payload) if False else subprocess.run(
     [os.path.join(BIN, "gate-dispatch")], cwd=C, input=payload, capture_output=True, text=True,
     env=dict(os.environ, HOME=os.path.join(C, "home"), AGENTIC_NO_AUTOINSTALL="1"))
 assert r.returncode == 0, (r.returncode, r.stderr)
 payload2 = json.dumps({"hook_event_name": "PreToolUse", "tool_name": "Agent", "cwd": C,
-                       "tool_input": {"subagent_type": "aloop:green-worker", "prompt": "TASK_ID: T9\ndo it"}})
+                       "tool_input": {"subagent_type": "aloop:green-worker", "prompt": "TASK_ID: T9\ndo it", "model": "sonnet"}})
 r = subprocess.run([os.path.join(BIN, "gate-dispatch")], cwd=C, input=payload2, capture_output=True, text=True,
                    env=dict(os.environ, HOME=os.path.join(C, "home"), AGENTIC_NO_AUTOINSTALL="1"))
 assert r.returncode == 2 and ".aladdin/state/tasks/T9.env" in r.stderr, (r.returncode, r.stderr)