ForgePlan · explosivebit · Jun 3, 2026 · Jun 3, 2026 · Jun 3, 2026
diff --git a/Makefile b/Makefile
@@ -13,7 +13,7 @@ endif
         smoke-run smoke-dry resume postmortem \
         eval-core-test litellm-up litellm-down stack-up stack-down stack-status \
         openrouter-smoke env-check \
-        harness-image-aider sandbox-net-up
+        harness-image-aider harness-image-goose sandbox-net-up
 
 demo-run:
 	python -m pollmevals_eval_core.demo_run --tasks evals/tasks --output artifacts
@@ -118,6 +118,11 @@ stack-status:
 harness-image-aider:
 	docker build -t pollmevals-harness-aider:0.1.0 infra/docker/harness-aider/
 
+# Build the goose harness image (RFC-006 Phase 5). No spend — pulls base +
+# the Block goose release binary. goose is model-agnostic (OpenAI-compatible).
+harness-image-goose:
+	docker build -t pollmevals-harness-goose:0.1.0 infra/docker/harness-goose/
+
 # Build the Python eval sandbox image (Phase 5 — PythonCorrectnessEvaluator,
 # runs BigCodeBench unittest suites). No spend — pulls base + pip scientific stack.
 eval-image-py:

diff --git a/apps/eval-core-py/scripts/build_real_board.py b/apps/eval-core-py/scripts/build_real_board.py
@@ -83,6 +83,12 @@
     "grok-4",
 ]
 _AIDER_MODELS = ["qwen-3-14b", "qwen3-coder-30b", "codestral", "devstral"]
+# goose runs the SAME coder models as aider, so each (model) gives a directly
+# comparable pair — isolating the harness variable (aider L4 vs goose L2 on
+# identical models). Diverge this list later if goose handles models aider can't.
+_GOOSE_MODELS = ["qwen-3-14b", "qwen3-coder-30b", "codestral", "devstral"]
+# Per-stack candidate model lists for --add-stack (merge ONE harness column in).
+_STACK_MODELS = {"aider": _AIDER_MODELS, "goose": _GOOSE_MODELS}
 _SEEDS = [1, 2]
 _TASK = "be_01_jwt_auth"
 _JUDGES = ["claude-sonnet-4-6-judge", "gpt-5-mini-judge", "gemini-3-flash"]
@@ -171,6 +177,14 @@ async def _main() -> int:
         "existing board.json (fills previously-failed cells without re-running "
         "the rest). e.g. --fill grok-4",
     )
+    ap.add_argument(
+        "--add-stack",
+        default="",
+        help="run ONLY this stack's grid (on _STACK_MODELS[stack]) and MERGE its "
+        "new harness column (cells + harness metadata) into the existing "
+        "board.json, without re-spending on the other harnesses. e.g. "
+        "--add-stack goose",
+    )
     args = ap.parse_args()
 
     os.chdir(REPO)
@@ -275,6 +289,53 @@ def caller_for(stack_id: str) -> object:
             print(f"    {c.stack_id} x {c.model_id}: score={c.mean_score} cost=${c.mean_cost_usd}")
         return 0
 
+    # --add-stack: run ONLY this harness's grid and merge its NEW column (cells +
+    # harness metadata) into the existing board.json, without re-spending on the
+    # other harnesses. Unlike --fill (replace cells in place), this also unions
+    # the new harness into board.harnesses so the matrix renders the column.
+    if args.add_stack:
+        stack_id = args.add_stack
+        models = _STACK_MODELS.get(stack_id, _AIDER_MODELS)
+        print(f"ADD-STACK: {stack_id} x {models} → merge column into {out.name} ...")
+        result = await runner.run(
+            GridSpec(
+                run_hash=_RUN_HASH,
+                models=models,
+                tasks=[_TASK],
+                stacks=[stack_id],
+                seeds=_SEEDS,
+                task_timeout_s=task_timeout,
+            )
+        )
+        rows = _rows_from_result(result, _PRICING)
+        partial = build_board(rows, stacks_root=stacks_root, run_hash=_RUN_HASH, run_type="smoke")
+        existing = Board.model_validate_json(out.read_text(encoding="utf-8"))
+        # Replace-or-append cells by (model, stack); union harnesses + models.
+        by_key = {(c.model_id, c.stack_id): c for c in existing.cells}
+        for c in partial.cells:
+            by_key[(c.model_id, c.stack_id)] = c
+        h_by_id = {h.stack_id: h for h in existing.harnesses}
+        for h in partial.harnesses:
+            h_by_id[h.stack_id] = h
+        m_by_id = {m.model_id: m for m in existing.models}
+        for m in partial.models:
+            m_by_id.setdefault(m.model_id, m)
+        cells = list(by_key.values())
+        scored_now = sum(1 for c in cells if c.mean_score is not None)
+        merged_board = existing.model_copy(
+            update={
+                "cells": cells,
+                "harnesses": list(h_by_id.values()),
+                "models": list(m_by_id.values()),
+                "scored": scored_now > 0,
+            }
+        )
+        out.write_text(merged_board.model_dump_json(indent=2) + "\n", encoding="utf-8")
+        print(f"  board now {scored_now}/{len(cells)} cells scored. new {stack_id} cells:")
+        for c in partial.cells:
+            print(f"    {c.stack_id} x {c.model_id}: score={c.mean_score} cost=${c.mean_cost_usd}")
+        return 0
+
     # Two specs so the grid is non-cartesian: raw-llm on every candidate, aider
     # only on the models that follow its edit format. Same run_hash + runner so
     # rows merge into one board.

diff --git a/apps/eval-core-py/src/orchestrator/stack_executor.py b/apps/eval-core-py/src/orchestrator/stack_executor.py
@@ -231,10 +231,45 @@ def _aider_invocation(
     )
 
 
+def _goose_invocation(
+    proxy_base_url: str, api_key: str, model_alias: str, prompt: str
+) -> ProxyInvocation:
+    """Goose recipe — PROVEN (2026-06-03 isolation smoke; memory).
+
+    Block's goose is model-agnostic via an OpenAI-compatible provider. Unlike
+    aider's single ``OPENAI_API_BASE``, goose splits the endpoint into
+    ``OPENAI_HOST`` (scheme+host, NO path) + ``OPENAI_BASE_PATH`` (the chat
+    route), and selects provider/model via ``GOOSE_PROVIDER`` / ``GOOSE_MODEL``.
+    ``GOOSE_DISABLE_KEYRING`` (no system keyring in a sandbox container) and
+    ``GOOSE_MODE=auto`` (never block waiting on a tool-call confirmation) are
+    baked into the image; set here too so the recipe is self-contained. The
+    model is chosen by env, so ``extra_args`` is empty; the prompt rides ``-t``
+    and the ``goose run --no-session --with-builtin developer`` scaffolding
+    lives in stack.yaml ``execution.args``.
+    """
+    base = proxy_base_url.rstrip("/")
+    return ProxyInvocation(
+        env={
+            "GOOSE_PROVIDER": "openai",
+            "GOOSE_MODEL": model_alias,
+            "GOOSE_MODE": "auto",
+            "GOOSE_DISABLE_KEYRING": "1",
+            "OPENAI_API_KEY": api_key,
+            "OPENAI_HOST": base,
+            "OPENAI_BASE_PATH": "v1/chat/completions",
+        },
+        config_files={},
+        extra_args=[],
+        prompt_args=["-t", prompt],
+    )
+
+
 # Proven recipes (validated end-to-end via the proxy). aider is the RFC-006
-# first slice (aider x qwen x be_01).
+# first slice (aider x qwen x be_01); goose is the second harness (2026-06-03),
+# a model-agnostic peer that runs the same coder models for a clean comparison.
 _PROVEN_RECIPES: dict[str, _RecipeBuilder] = {
     "aider": _aider_invocation,
+    "goose": _goose_invocation,
 }
 
 # Known harnesses whose recipe is proven in spikes but lands at its per-stack
@@ -244,7 +279,6 @@ def _aider_invocation(
         "claude-code",
         "codex",
         "opencode",
-        "goose",
         "openhands",
         "hermes",
         "cline",

diff --git a/apps/eval-core-py/tests/test_stack_executor.py b/apps/eval-core-py/tests/test_stack_executor.py
@@ -188,7 +188,36 @@ def test_aider_strips_trailing_slash(self) -> None:
         )
         assert inv.env["OPENAI_API_BASE"] == "http://localhost:4000/v1"
 
-    @pytest.mark.parametrize("cli", ["claude-code", "codex", "goose", "openhands", "opencode"])
+    def test_goose_recipe_is_proven(self) -> None:
+        inv = build_proxy_invocation(
+            "goose",
+            proxy_base_url="http://pollmevals-litellm-proxy:4000",
+            api_key="sk-local-xyz",
+            model_alias="qwen-3-14b",
+            prompt="do the thing",
+        )
+        assert inv.env["GOOSE_PROVIDER"] == "openai"
+        assert inv.env["GOOSE_MODEL"] == "qwen-3-14b"
+        assert inv.env["OPENAI_API_KEY"] == "sk-local-xyz"
+        # goose splits the endpoint: host (NO path) + the chat route separately.
+        assert inv.env["OPENAI_HOST"] == "http://pollmevals-litellm-proxy:4000"
+        assert inv.env["OPENAI_BASE_PATH"] == "v1/chat/completions"
+        assert inv.env["GOOSE_DISABLE_KEYRING"] == "1"
+        assert inv.extra_args == []  # model selected by env, not a CLI flag
+        assert inv.prompt_args == ["-t", "do the thing"]
+        assert inv.config_files == {}
+
+    def test_goose_strips_trailing_slash(self) -> None:
+        inv = build_proxy_invocation(
+            "goose",
+            proxy_base_url="http://h:4000/",
+            api_key="k",
+            model_alias="m",
+            prompt="p",
+        )
+        assert inv.env["OPENAI_HOST"] == "http://h:4000"
+
+    @pytest.mark.parametrize("cli", ["claude-code", "codex", "openhands", "opencode"])
     def test_known_but_pending_harness_raises_pending(self, cli: str) -> None:
         with pytest.raises(HarnessRecipePending, match="Phase 5"):
             build_proxy_invocation(
@@ -207,8 +236,8 @@ def test_none_cli_raises_unsupported(self) -> None:
                 None, proxy_base_url="x", api_key="k", model_alias="m", prompt="p"
             )
 
-    def test_supported_harnesses_is_aider_only_in_phase_1(self) -> None:
-        assert supported_harnesses() == frozenset({"aider"})
+    def test_supported_harnesses_are_aider_and_goose(self) -> None:
+        assert supported_harnesses() == frozenset({"aider", "goose"})
 
 
 # ---------------------------------------------------------------------------

diff --git a/apps/site/public/board.json b/apps/site/public/board.json
@@ -26,6 +26,16 @@
         "L0_bare_llm"
       ],
       "family": "baseline"
+    },
+    {
+      "stack_id": "goose",
+      "name": "Goose",
+      "level": 2,
+      "layers": [
+        "L1_system_prompt",
+        "L2_tools"
+      ],
+      "family": "agnostic"
     }
   ],
   "models": [
@@ -629,6 +639,99 @@
         "type_safety": 5.25
       },
       "on_frontier": false
+    },
+    {
+      "model_id": "codestral",
+      "stack_id": "goose",
+      "mean_score": null,
+      "mean_cost_usd": 0.0,
+      "mean_latency_ms": 9353,
+      "pass_hat_k": null,
+      "quality_per_dollar": null,
+      "per_task": {
+        "be_01_jwt_auth": {
+          "score": null,
+          "cost_usd": 0.0,
+          "pass_hat_k": null
+        }
+      },
+      "per_criterion": {},
+      "on_frontier": false
+    },
+    {
+      "model_id": "devstral",
+      "stack_id": "goose",
+      "mean_score": 7.12,
+      "mean_cost_usd": 0.0,
+      "mean_latency_ms": 200855,
+      "pass_hat_k": null,
+      "quality_per_dollar": null,
+      "per_task": {
+        "be_01_jwt_auth": {
+          "score": 7.12,
+          "cost_usd": 0.0,
+          "pass_hat_k": null
+        }
+      },
+      "per_criterion": {
+        "code_clarity": 7.5,
+        "correctness": 7.75,
+        "error_handling": 7.0,
+        "security_posture": 8.0,
+        "test_alignment": 7.0,
+        "type_safety": 5.5
+      },
+      "on_frontier": false
+    },
+    {
+      "model_id": "qwen-3-14b",
+      "stack_id": "goose",
+      "mean_score": 6.0,
+      "mean_cost_usd": 0.0,
+      "mean_latency_ms": 59121,
+      "pass_hat_k": null,
+      "quality_per_dollar": null,
+      "per_task": {
+        "be_01_jwt_auth": {
+          "score": 6.0,
+          "cost_usd": 0.0,
+          "pass_hat_k": null
+        }
+      },
+      "per_criterion": {
+        "code_clarity": 7.0,
+        "correctness": 4.5,
+        "error_handling": 6.0,
+        "security_posture": 6.5,
+        "test_alignment": 7.0,
+        "type_safety": 5.0
+      },
+      "on_frontier": false
+    },
+    {
+      "model_id": "qwen3-coder-30b",
+      "stack_id": "goose",
+      "mean_score": 7.17,
+      "mean_cost_usd": 0.0,
+      "mean_latency_ms": 48202,
+      "pass_hat_k": null,
+      "quality_per_dollar": null,
+      "per_task": {
+        "be_01_jwt_auth": {
+          "score": 7.17,
+          "cost_usd": 0.0,
+          "pass_hat_k": null
+        }
+      },
+      "per_criterion": {
+        "code_clarity": 7.5,
+        "correctness": 8.0,
+        "error_handling": 7.5,
+        "security_posture": 8.25,
+        "test_alignment": 6.5,
+        "type_safety": 5.25
+      },
+      "on_frontier": false
     }
   ]
 }
diff --git a/apps/site/src/components/HarnessModelMatrix.tsx b/apps/site/src/components/HarnessModelMatrix.tsx
@@ -10,7 +10,7 @@ import {
   frontierKeys,
 } from "@/lib/board";
 import { heat, heatText, norm } from "@/lib/color";
-import { formatUsd, formatScore } from "@/lib/format";
+import { formatUsd, formatCost, formatScore } from "@/lib/format";
 
 const METRICS: { id: Metric; label: string; caption: string }[] = [
   {
@@ -175,7 +175,7 @@ function MatrixCell({
   const fg = v === null ? "#6e6e7a" : heatText(t);
   const title =
     `${cell.model_id} × ${cell.stack_id}\n` +
-    `score ${cell.mean_score ?? "—"} · ${formatUsd(
+    `score ${cell.mean_score ?? "—"} · ${formatCost(
       cell.mean_cost_usd
     )}/task · ` +
     `${

diff --git a/apps/site/src/components/PerTaskWinners.tsx b/apps/site/src/components/PerTaskWinners.tsx
@@ -1,5 +1,5 @@
 import type { Board, Cell } from "@/lib/board";
-import { formatUsd, formatScore } from "@/lib/format";
+import { formatCost, formatScore } from "@/lib/format";
 
 /**
  * Per-task winners — because the best stack is NOT the same for every task.
@@ -61,7 +61,7 @@ export function PerTaskWinners({ board }: { board: Board }) {
                     </span>
                   </span>
                   <span className="psc tnum">{formatScore(r.score)}</span>
-                  <span className="pcost tnum muted">{formatUsd(r.cost)}</span>
+                  <span className="pcost tnum muted">{formatCost(r.cost)}</span>
                 </li>
               ))}
             </ol>

diff --git a/apps/site/src/components/StackDrawer.tsx b/apps/site/src/components/StackDrawer.tsx
@@ -2,7 +2,7 @@
 
 import { useEffect } from "react";
 import type { Board, Cell } from "@/lib/board";
-import { formatUsd, formatScore } from "@/lib/format";
+import { formatCost, formatScore } from "@/lib/format";
 
 /**
  * Right-side drawer: everything about one stack (model × harness) in one place —
@@ -60,7 +60,7 @@ export function StackDrawer({
       value:
         cell.quality_per_dollar === null ? "—" : grp(cell.quality_per_dollar),
     },
-    { label: "Cost / task", value: formatUsd(cell.mean_cost_usd) },
+    { label: "Cost / task", value: formatCost(cell.mean_cost_usd) },
     {
       label: "Speed",
       value: cell.mean_latency_ms

diff --git a/apps/site/src/components/StackMasterTable.tsx b/apps/site/src/components/StackMasterTable.tsx
@@ -2,7 +2,7 @@
 
 import { useMemo, useState } from "react";
 import type { Board, Cell } from "@/lib/board";
-import { formatUsd, formatScore } from "@/lib/format";
+import { formatCost, formatScore } from "@/lib/format";
 
 // Locale-independent thousands separator — `toLocaleString()` differs between
 // the Node server render and the browser, which breaks hydration.
@@ -174,7 +174,7 @@ export function StackMasterTable({ board }: { board: Board }) {
                 <td className="num strong">
                   {r.score === null ? "—" : formatScore(r.score)}
                 </td>
-                <td className="num">{formatUsd(r.cost)}</td>
+                <td className="num">{formatCost(r.cost)}</td>
                 <td className="num">
                   {r.latency ? (r.latency / 1000).toFixed(1) + "s" : "—"}
                 </td>