2026-10-07 16:19:09 +00:00
14 changed files with 12196 additions and 245 deletions
--- a/agent/dashboard/options.py
+++ b/agent/dashboard/options.py
@ -49,27 +49,6 @@ SUPPORTED_MODELS: list[ModelOption] = [
        "default_effort": "high",
        "supports_images": False,
    },
-    {
-        "id": "fireworks:accounts/fireworks/models/minimax-m3",
-        "label": "MiniMax M3",
-        "efforts": ["medium", "high"],
-        "default_effort": "high",
-        "supports_images": True,
-    },
-    {
-        "id": "fireworks:accounts/fireworks/models/gpt-oss-120b",
-        "label": "gpt-oss-120b",
-        "efforts": ["low", "medium", "high"],
-        "default_effort": "medium",
-        "supports_images": False,
-    },
-    {
-        "id": "fireworks:accounts/fireworks/models/deepseek-v4-flash",
-        "label": "DeepSeek V4 Flash",
-        "efforts": ["none", "medium", "high"],
-        "default_effort": "high",
-        "supports_images": False,
-    },
 ]

 SUPPORTED_MODEL_IDS: frozenset[str] = frozenset(m["id"] for m in SUPPORTED_MODELS)
--- a/deploy/MIGRATION.md
+++ b/deploy/MIGRATION.md
@ -1,9 +1,7 @@
 # Open SWE — Migration Plan: Self-Hosted AWS → Managed LangGraph Cloud + Vercel

 **Repo:** `Sea-Haven-Industries/open-swe` (private) · **AWS:** 328440206208 / us-east-1
-**Author:** Adam Moussa · **Date:** 2026-06-29 (final topology added 2026-06-30) · **Status:** EXECUTED — managed cutover live; §§3–11 below are the original (now-historical) phased plan, **superseded by §1a for all current-state facts (URLs, project layout, env)**.
-
-> **READ §1a FIRST.** The phased plan (§§2–11) and the Phase A/C spike notes capture how we got here and still hold for rationale, cost, and rollback. But the spike-era specifics they cite — the single `open-swe-dashboard` Vercel project, the `open-swe-dev-hosted-…`/`open-swe-v3-…` deployment URLs, the `ui/vercel.json` same-origin rewrite, the single GitHub App — are **stale**. §1a is the authoritative final topology and wins on every conflict.
+**Author:** Adam Moussa · **Date:** 2026-06-29 · **Status:** DRAFT — owes a `/sh-plan-review` before prod cutover (Phase C gate)

 ---

@ -32,75 +30,6 @@ A self-hosted `langgraph up` + RDS plan (already `/sh-plan-review`'d to APPROVE-

 ---

-## 1a. Final, verified topology (AUTHORITATIVE — supersedes spike-era values)
-
-This is the live managed deployment as of 2026-06-30. Where any later section disagrees (old URLs, a single Vercel project, a single GitHub App, the `ui/vercel.json` rewrite), **this section wins**.
-
-### Backend — managed LangGraph Cloud (two deployments, one LangSmith workspace)
-Both deployments live in the **same LangSmith workspace**; the **same workspace API key authenticates both** (including the Store API — so per-deployment store writes use that one key with the per-deployment URL).
-
-| Deployment | URL | Git connection |
-|---|---|---|
-| **dev** | `https://open-swe-dev-fb737aa219605c8bbdb30ecbb33f30c0.us.langgraph.app` | branch `dev` |
-| **prod** | `https://open-swe-prod-d6c7bb63aaa651b6a1d92f9492b1d983.us.langgraph.app` | branch `main` (auto-deploys on push to `main`) |
-
-> The dev deployment was **renamed `open-swe-dev`, deleted, and recreated** — which minted the **new URL hash** above. The spike-era `open-swe-v3-…` / `open-swe-dev-hosted-…` URLs are **dead/superseded**. Deleting + recreating a deployment is the one operation that changes the URL hash (otherwise stable across revisions) — when it happens, update **every** reference (Vercel env, GitHub App webhooks, OAuth callbacks, docs).
-
-### UI — Vercel (ONE project, two environments)
-**One** Vercel project `open-swe-prod` (team `sea-haven`, id `prj_OOh6yjXMp4ah3Ws3Y7XRQxjmMmQU`). The old separate `open-swe-dashboard` project was **DELETED**.
-
-| Vercel environment | Branch | Backend | Custom domain |
-|---|---|---|---|
-| production | `main` | prod deployment URL | `openswe.seahaven.com` |
-| custom **`dev`** (id `env_SMI23PULAJXk0GhwE0HLVhp5J3ZS`) | `dev` | dev deployment URL | `openswe-dev.seahaven.com` |
-
- A **per-environment** env var `LANGGRAPH_BACKEND_URL` (prod env = prod URL, dev env = dev URL) drives the `/dashboard/api/*` proxy.
- Project settings: `framework=null`, `outputDirectory` cleared, root directory `ui`.
- **Proxy mechanism (current, after PR #76):** **Nitro `routeRules`** in `ui/vite.config.ts` read `process.env.LANGGRAPH_BACKEND_URL` and Nitro's Vercel preset compiles them into `.vercel/output/config.json` (Build Output API) at build time — a CDN-level proxy (not redirect, so the `osw_session` cookie stays first-party). PR #75's hand-rolled `ui/scripts/build-vercel-output.mjs` was the broken first attempt and is **gone** (`ui/scripts/` no longer exists). **Never** add a manual script that `rm`s `.vercel/output` — Nitro's Vercel preset auto-emits it.
-
-### DNS — Route 53 zone `seahaven.com` (`Z06652411XKH89KTZD3XA`)
- `openswe.seahaven.com` → CNAME `cname.vercel-dns.com` (prod env)
- `openswe-dev.seahaven.com` → CNAME to Vercel (dev env)
-
-### GitHub Apps — TWO (dev/prod isolated; each its own webhook URL)
-| App | app_id | install | client_id | org | members scope | repos |
-|---|---|---|---|---|---|---|
-| **prod** `seahaven-openswe` | `4146115` | `142615168` | `Iv23lil96pKQNNDUn5yp` | `Sea-Haven-Industries` | members:**write** | all |
-| **dev** `seahaven-openswe-dev` | `4162963` | `143023302` | `Iv23licQwJvGAPJj1HJe` | `seahaven-open-swe-dev` | members:**read** | all |
-
- **Promotion App** `seahaven-promotion` (actor `4170147`) is the sole non-admin fast-forward-push bypass on the `main` ruleset `18238334` — its FF-push of `dev → main` is what triggers the managed prod build.
-
-### Env per deployment (set in LangGraph Cloud config + Vercel env — NOT Secrets Manager)
-| Var | dev | prod |
-|---|---|---|
-| `LANGGRAPH_URL` | own (dev) deployment URL | own (prod) deployment URL |
-| `DASHBOARD_BASE_URL` / `DASHBOARD_API_BASE_URL` | `https://openswe-dev.seahaven.com` | `https://openswe.seahaven.com` |
-| `VITE_DASHBOARD_API_BASE_URL` | empty (same-origin via Vercel proxy) | empty |
-| `ALLOWED_GITHUB_ORGS` | dev org (`seahaven-open-swe-dev`) | `Sea-Haven-Industries` |
-| `CONFIGURED_ADMINS` | `amoussa1229,adam@seahavenind.com` | `amoussa1229,adam@seahavenind.com` |
-
-`DASHBOARD_BASE_URL` / `DASHBOARD_API_BASE_URL` **must include `https://`** (see gotcha 3). Secret **values** are still sourced from `open-swe-{dev,prod}/*` Secrets Manager + SSM (the remaining source of truth) and set into the LangGraph Cloud + Vercel env stores — the accepted `secrets-and-config.md` deviation.
-
-### Bedrock IAM (PR #74, still OPEN)
-Two IAM users `open-swe-dev-bedrock` + `open-swe-prod-bedrock`, each attached to customer-managed policy `open-swe-bedrock-invoke` (least-privilege `bedrock:InvokeModel[WithResponseStream]` on the `us.anthropic.claude-opus-4-8` inference-profile ARN + its 3 routed foundation-model ARNs in us-east-1/us-east-2/us-west-2). Default model `bedrock_converse:us.anthropic.claude-opus-4-8` + 3 Fireworks models. Static access keys live only in the deployment env (dev key → dev, prod key → prod).
-
-### User store — per-deployment
-**Each** managed deployment has its **own** Store. The GitHub→email mapping `amoussa1229 → adam@seahavenind.com` (namespace `["user_mappings"]`, key = lowercased login, record `{github_login, work_email, status:"active", source, created_at, updated_at}`) was written to **both** the dev and prod stores directly. New users need a mapping **per-deployment** (write each store directly, or use the dashboard admin User-mappings UI — the `work_email` field was added by PR #65 fix #4).
-
-### AWS decommission (PR #64)
-Self-host CDK stacks destroyed. Residual: `CDKToolkit` (shared, **preserved**); ~50 `RETAIN`'d Secrets Manager shells + 3 S3 asset buckets (**pending cleanup**); AWS **Bedrock** (live dependency, kept).
-
-### Operational gotchas (hard-won — carry these into any runbook)
-1. **Per-deployment store → seed user mappings per-deployment.** A missing mapping makes `process_github_issue` silently early-return ("No email mapping … skipping"): the webhook returns 200/accepted but produces **no reaction and no run**. Seed dev **and** prod.
-2. **Org-login gate uses the App *installation* token**, so the App must be org-installed with **Members:read**. OAuth working ≠ membership check working — they use **separate creds** (CLIENT_ID/SECRET for OAuth vs APP_ID/INSTALLATION_ID/PRIVATE_KEY for the install token). A mangled multi-line `GITHUB_APP_PRIVATE_KEY` breaks the install token (and thus the gate) while OAuth still works.
-3. **`DASHBOARD_API_BASE_URL` must be `https://`** — an `http://` value makes GitHub reject the OAuth callback with "redirect_uri not associated."
-4. **`osw_oauth_state` cookie is host-only** — start login on the **same host** as `DASHBOARD_API_BASE_URL`, or you get "oauth state mismatch."
-5. **Webhooks go DIRECT to the langgraph URL** (`/webhooks/*`). Vercel only proxies `/dashboard/api/*`. The app is **same-origin only** (no CORS).
-6. **On Vercel CI, Nitro's Vercel preset auto-emits `.vercel/output`** — drive the proxy via Nitro `routeRules` from `LANGGRAPH_BACKEND_URL`; never a manual script that `rm`s `.vercel/output`.
-7. **Deleting + recreating a LangGraph deployment mints a NEW URL hash** (otherwise stable across revisions) — update every reference (Vercel env, webhooks, OAuth, docs).
-
---
-
 ## 2. Architecture: Before → After

 ### Before (self-hosted AWS — LIVE as of 2026-06-29)
@ -120,29 +49,23 @@ Browser (dashboard) ─────────────▶ openswe.seahaven.
                  Sandbox: LangSmith cloud (DEFAULT_SANDBOX_SNAPSHOT_ID + GitHub proxy)
 ```

-### After (managed — FINAL, see §1a for exact values)
+### After (managed)
 ```
-                              ┌──────────────── DEV lane ────────────────┐   ┌──────────────── PROD lane ───────────────┐
-GitHub(dev org)/Slack/Linear  │ webhook → open-swe-dev-….us.langgraph.app │   │ webhook → open-swe-prod-….us.langgraph.app│  GitHub(SHI org)/Slack/Linear
-  App seahaven-openswe-dev ───┘    (DIRECT to langgraph URL, /webhooks/*)  │   │   (DIRECT to langgraph URL, /webhooks/*)  └─── App seahaven-openswe
-                                                                          ▼   ▼
-Browser ▶ openswe-dev.seahaven.com ─┐                          ┌─▶ openswe.seahaven.com ◀ Browser
-                                    │   ONE Vercel project `open-swe-prod` (team sea-haven)
-                                    │   ├─ env `dev`  (branch dev)  → proxies /dashboard/api/* → dev langgraph URL
-                                    │   └─ env production (branch main) → proxies /dashboard/api/* → prod langgraph URL
-                                    └─ proxy compiled by Nitro routeRules from per-env LANGGRAPH_BACKEND_URL (PR #76)
+GitHub/Slack/Linear ──webhook──▶ *.langgraph.app  (or hooks.seahaven.com CNAME → TODO §10)
+Browser (dashboard) ─────────────▶ open-swe-dashboard.vercel.app (stable alias / custom domain)
+                                          │ same-origin rewrite /dashboard/api/* (ui/vercel.json)
                                          ▼
-            TWO LangGraph Cloud deployments (same LangSmith workspace; one workspace key auths both incl. Store)
-              ├─ dev  ← branch `dev`     ·  prod ← branch `main` (push-to-main auto-deploys prod)
-              ├─ each serves the graphs + the custom http.app (agent.webapp:app = webhooks + dashboard API + OAuth)
-              ├─ each has its OWN durable Postgres store + checkpointer (issue #9 SOLVED) — user mappings per-deployment
-              └─ env/secrets in the Deployment + Vercel config (NOT Secrets Manager) — Adam accepted deviation
+            LangGraph Cloud "Deployment" (managed, git-connected to `main` for prod / `dev` for dev)
+              ├─ serves the 6 graphs (agent, reviewer, analyzer, chat, scheduler, ci_monitor)
+              ├─ serves the custom http.app (agent.webapp:app = webhooks + dashboard API + OAuth)
+              ├─ durable Postgres store + checkpointer (issue #9 SOLVED) — autoscaled 1→10 replicas
+              └─ env/secrets in the Deployment config (NOT Secrets Manager) — Adam accepted deviation
                                          ▼
            Sandbox: LangSmith cloud (UNCHANGED — DEFAULT_SANDBOX_SNAPSHOT_ID + GitHub-App proxy)
-            Bedrock: IAM users open-swe-{dev,prod}-bedrock + policy open-swe-bedrock-invoke (static keys in deploy env)
+            Auth: GitHub App seahaven-openswe (UNCHANGED — App 4146115 / Install 142615168)
 ```

-**What changes shape:** runtime host (EC2 → managed PaaS), durability (in-memory → managed Postgres, one store **per deployment**), CD (bespoke S3/SSM/packer → git-connected auto-build), config home (Secrets Manager/SSM → Deployment+Vercel env), ingress topology (shared ALB + one App → two dev/prod-isolated GitHub Apps each hitting its own `*.langgraph.app` directly), UI proxy (`ui/vercel.json` rewrite → Nitro `routeRules`). **What stays:** the LangSmith sandbox plane, CI (lint/format/unit/Playwright), the app code itself, and Secrets Manager/SSM as the secret-**value** source of truth.
+**What changes shape:** runtime host (EC2 → managed PaaS), durability (in-memory → managed Postgres), CD (bespoke S3/SSM/packer → git-connected auto-build), config home (Secrets Manager/SSM → Deployment+Vercel env). **What stays:** the GitHub App, the LangSmith sandbox plane, CI (lint/format/unit/Playwright), the app code itself.

 ---

@ -151,11 +74,11 @@ Browser ▶ openswe-dev.seahaven.com ─┐                          ┌─▶ o
 ### Phase A — Dev spike (MOSTLY DONE)
 Goal: prove managed serves our custom app + durability, at $0, before committing prod $.

-**Proven this session** (spike-era specifics — superseded by §1a; URLs/project below are DEAD):
+**Proven this session:**
 - ✅ Dev backend deployed to LangGraph Cloud, connected to branch `dev`:
-  ~~`https://open-swe-dev-hosted-e76c2b0e8a7955fe8ad3110a7a54e5d0.us.langgraph.app`~~ → final dev URL in §1a (`open-swe-dev-fb737aa…`; the deployment was later deleted + recreated)
+  `https://open-swe-dev-hosted-e76c2b0e8a7955fe8ad3110a7a54e5d0.us.langgraph.app`
  (Bedrock + a Fireworks key set; AWS creds for Bedrock deferred — see §10 open decision).
- ✅ UI deployed to Vercel — team `sea-haven`, ~~project `open-swe-dashboard`, `https://open-swe-dashboard.vercel.app`~~ (DELETED); now the **single** project `open-swe-prod` with dev/prod environments (§1a); proxy is Nitro `routeRules`, not the `ui/vercel.json` rewrite.
+- ✅ UI deployed to Vercel — team `sea-haven`, project `open-swe-dashboard`, `https://open-swe-dashboard.vercel.app`; `ui/vercel.json` rewrite repointed at the dev deployment.
 - ✅ Managed serves the custom `http.app` (dashboard API + webhooks) — **no platform auth gate** in front of our routes (webhook returns 401 sig-enforced, so signatures still govern).
 - ✅ Vercel same-origin rewrite → backend works.
 - ✅ GitHub OAuth dashboard login end-to-end.
@ -179,7 +102,7 @@ Land the 6 code fixes (§5), codify env/config, and resolve the Bedrock-auth dec

 ### Phase C — Prod deployment
 - [ ] C1. Create a **prod LangGraph Cloud deployment** tracking branch `main` (the durable autoscaled 1→10 tier, not the free Dev tier). Record its `*.langgraph.app` URL.
- [x] C2. ~~Create the **prod Vercel project/target** (or promote the existing `open-swe-dashboard` to production)~~ — **DONE differently:** one project `open-swe-prod` with a production env (`main`) + a custom `dev` env (`dev`), each with its own `LANGGRAPH_BACKEND_URL`. See §1a.
+- [ ] C2. Create the **prod Vercel project/target** (or promote the existing `open-swe-dashboard` to production); set its env (same-origin mode: `VITE_DASHBOARD_API_BASE_URL` empty).
 - [ ] C3. Set the **prod env triad** (§4) on the prod deployment + Vercel:
  - `LANGGRAPH_URL` = the prod `*.langgraph.app` URL
  - `DASHBOARD_BASE_URL` + `DASHBOARD_API_BASE_URL` = the prod Vercel origin (with `https://` scheme — fix #2)
@ -216,19 +139,15 @@ Only after managed prod is proven + soaked. See §7 for the precise retire-vs-ke

 ## 4. Env / Config Reference

-### The prod triad (per INSTALLATION.md §10, lines 630–658) — see §1a for the exact dev/prod values
-| Var | Value (per deployment/env) | Notes |
+### The prod triad (per INSTALLATION.md §10, lines 630–658)
+| Var | Prod value | Notes |
 |---|---|---|
-| `LANGGRAPH_URL` | the **own** deployment URL (`https://...langgraph.app`) | **NOT** localhost. Drives `thread_ops.langgraph_url()` (fix #1). dev→dev URL, prod→prod URL. |
-| `DASHBOARD_BASE_URL` | the own dashboard origin, **with `https://`** (`https://openswe-dev.seahaven.com` / `https://openswe.seahaven.com`) | |
-| `DASHBOARD_API_BASE_URL` | same as above, **with `https://` scheme** | scheme required or OAuth `redirect_uri` is schemeless and GitHub rejects (fix #2) |
-| `VITE_DASHBOARD_API_BASE_URL` | **empty** | same-origin mode; UI calls relative `/dashboard/api/*`, the Vercel Nitro proxy rewrites to backend |
-| `LANGGRAPH_BACKEND_URL` | **Vercel env var**, per Vercel environment (dev env = dev URL, prod env = prod URL) | drives the Nitro `routeRules` `/dashboard/api/*` proxy at build time (PR #76) — required on Vercel builds |
-| `ALLOWED_GITHUB_ORGS` | `Sea-Haven-Industries` (prod) / `seahaven-open-swe-dev` (dev) | org-login gate; checked via the App **installation** token (gotcha 2) |
+| `LANGGRAPH_URL` | the deployment URL (`https://...langgraph.app`) | **NOT** localhost. Drives `thread_ops.langgraph_url()` (fix #1). |
+| `DASHBOARD_BASE_URL` | the Vercel origin | same-origin rewrite mode |
+| `DASHBOARD_API_BASE_URL` | the Vercel origin, **with `https://` scheme** | scheme required or OAuth `redirect_uri` is schemeless and GitHub rejects (fix #2) |
+| `VITE_DASHBOARD_API_BASE_URL` | **empty** | same-origin mode; UI calls relative `/dashboard/api/*`, Vercel rewrites to backend |

-GitHub App dashboard OAuth callback = `<DASHBOARD_API_BASE_URL>/dashboard/api/auth/callback` (the own dashboard origin — prod App on `openswe.seahaven.com`, dev App on `openswe-dev.seahaven.com`).
-
-**UI proxy mechanism (final, PR #76):** the `/dashboard/api/*` proxy is **Nitro `routeRules`** in `ui/vite.config.ts` reading `process.env.LANGGRAPH_BACKEND_URL`, compiled by Nitro's Vercel preset into `.vercel/output/config.json`. This **supersedes** the spike-era `ui/vercel.json` same-origin rewrite and PR #75's hand-rolled `ui/scripts/build-vercel-output.mjs` (deleted). `ui/vercel.json` now only carries `framework:null` + `buildCommand: bun run build`.
+GitHub App dashboard OAuth callback = `<DASHBOARD_API_BASE_URL>/dashboard/api/auth/callback` (the Vercel prod origin).

 ### Where secrets/env live now
 **LangGraph Cloud Deployment config + Vercel env** — NOT AWS Secrets Manager. This **deviates from the Sea Haven `secrets-and-config.md` "Secrets Manager for all sensitive" handbook rule** — **Adam ACCEPTED this deviation** (managed has no instance role / no fetch-config boot hook; the platform's own secret store is the mechanism).
--- a/tests/e2e/fake_llm.py
+++ b/tests/e2e/fake_llm.py
@ -226,7 +226,7 @@ def _save_plan_step(_messages: list[BaseMessage]) -> AIMessage:
        tool_calls=[
            {
                "name": "save_plan",
-                "args": {"plan_markdown": PLAN_MARKDOWN},
+                "args": {"plan_file_path": PLAN_FILE_PATH},
                "id": "call-save-plan",
            }
        ],
@ -297,6 +297,22 @@ SCRIPT_LIBRARY: dict[str, tuple[StepSpec, ...]] = {
        ),
        _dynamic_step(_reply_step),
    ),
+    "breakout": (
+        _tool_step(
+            "Starting a separate Slack thread for the breakout task.",
+            {
+                "title": "Add greet() helper",
+                "instructions": "Please add a greet() helper and open a draft PR in the default repository. Use the current Slack request as context, and report progress in this new thread.",
+            },
+            "call-breakout",
+        ),
+        _tool_step(
+            "Confirming the breakout thread was started.",
+            "slack_thread_reply",
+            {"message": "I started a separate Open SWE thread for that aspect."},
+            "call-breakout-reply",
+        ),
+    ),
    "plan": (
        _tool_step(
            "This is worth planning first — entering plan mode.",
@ -319,6 +335,11 @@ def _is_plan_request(text: str) -> bool:
    return "plan" in text.lower()


+def _is_breakout_request(text: str) -> bool:
+    t = text.lower()
+    return "break out" in t or "separate thread" in t or "split out" in t
+
+
 def _is_approval(text: str) -> bool:
    t = text.lower()
    return "approved" in t and "implement" in t
@ -333,6 +354,9 @@ SCRIPT_RULES: tuple[ScriptRule, ...] = (
    ScriptRule("implement", lambda ctx: _is_approval(ctx.last_text)),
    ScriptRule("plan", lambda ctx: _is_revision(ctx.last_text)),
    ScriptRule("plan", lambda ctx: ctx.human_count <= 1 and _is_plan_request(ctx.first_text)),
+    ScriptRule(
+        "breakout", lambda ctx: ctx.human_count <= 1 and _is_breakout_request(ctx.first_text)
+    ),
    ScriptRule("implement", lambda ctx: ctx.human_count <= 1),
    ScriptRule("followup", lambda _ctx: True),
 )
--- a/tests/e2e/global-setup.ts
+++ b/tests/e2e/global-setup.ts
@ -14,9 +14,9 @@ export default function globalSetup() {
  if (existsSync(shell) && !process.env.E2E_FORCE_UI_BUILD) return;

  if (!existsSync(resolve(ui, "node_modules"))) {
-    execSync("bun install", { cwd: ui, stdio: "inherit" });
+    execSync("corepack pnpm install --frozen-lockfile", { cwd: ui, stdio: "inherit" });
  }
-  execSync("bun run build", {
+  execSync("corepack pnpm run build", {
    cwd: ui,
    stdio: "inherit",
    env: { ...process.env, VITE_DASHBOARD_API_BASE_URL: `http://127.0.0.1:${port}` },
--- a/tests/e2e/tests/full_flow.spec.ts
+++ b/tests/e2e/tests/full_flow.spec.ts
@ -37,6 +37,26 @@ test.describe("Open SWE full flow", () => {
    await expect(page.locator('.pr[data-pr="1"]')).toContainText("greet.py");
  });

+  test("Slack breakout request starts a new top-level Open SWE thread", async ({ page }) => {
+    await page.locator("#text").fill("<@U0BOT> please break out adding a greet() helper into a separate thread");
+    await page.locator("#send").click();
+
+    const breakout = page
+      .locator(".msg.bot")
+      .filter({ hasText: /Open SWE breakout thread:\* Add greet\(\) helper/ });
+    await expect(breakout).toBeVisible({ timeout: 60_000 });
+    const breakoutThreadTs = await breakout.getAttribute("data-thread-ts");
+    expect(breakoutThreadTs).toBeTruthy();
+
+    const breakoutThreadMessages = page.locator(`.msg.bot[data-thread-ts="${breakoutThreadTs}"]`);
+    await expect(breakoutThreadMessages.locator('a[href*="/agents/"]')).toBeVisible({
+      timeout: 60_000,
+    });
+    await expect(
+      page.locator(".msg.bot").filter({ hasText: "I started a separate Open SWE thread" }),
+    ).toBeVisible({ timeout: 60_000 });
+  });
+
  test("a message that does not mention the bot produces no run and no PR", async ({ page }) => {
    await page.locator("#mention").uncheck();
    await page.locator("#text").fill("just chatting with the team, nothing for the bot");
--- a/tests/e2e/tests/plan_review.spec.ts
+++ b/tests/e2e/tests/plan_review.spec.ts
@ -126,10 +126,11 @@ test.describe("Plan review (HTTP comments)", () => {
    await addComment(collab, "Reviewer: please also add a docstring.");
    await expect(collab.getByTestId("plan-comment")).toHaveCount(2);

-    // 5. The owner sees the collaborator's comment (polled), then approves.
+    // 5. The owner sees the collaborator's comment (polled), then approves and
+    //    returns to the main conversation while implementation starts.
    await expect(owner.getByTestId("plan-comment")).toHaveCount(2, { timeout: 30_000 });
    await owner.getByTestId("approve-plan").click();
-    await expect(owner.getByTestId("plan-decision")).toContainText(/implementing/i);
+    await expect(owner).toHaveURL(new RegExp(`/agents/${threadId}$`));

    // 6. The agent implements, opens a PR, and links it back in the Slack thread,
    //    echoing the reviewers' feedback — which proves the comments were stored
--- a/tests/test_fireworks_model.py
+++ b/tests/test_fireworks_model.py
@ -1,27 +1,12 @@
 import pytest

-from agent.dashboard.options import (
-    SUPPORTED_MODEL_IDS,
-    SUPPORTED_MODELS,
-    model_supports_effort,
-    model_supports_images,
-)
+from agent.dashboard.options import SUPPORTED_MODELS
 from agent.utils.model import (
    fallback_model_id_for,
    fireworks_reasoning_effort_for,
    provider_model_kwargs,
 )

-_FIREWORKS_PREFIX = "fireworks:accounts/fireworks/models/"
-
-NEW_FIREWORKS_MODELS = {
-    "minimax-m3": (["medium", "high"], "high", True),
-    "gpt-oss-120b": (["low", "medium", "high"], "medium", False),
-    "deepseek-v4-flash": (["none", "medium", "high"], "high", False),
-}
-
-_ALL_EFFORTS = ("none", "low", "medium", "high", "xhigh", "max")
-

 def test_fireworks_reasoning_effort_maps_effort() -> None:
    for effort in ("none", "low", "medium", "high", "xhigh", "max"):
@ -71,50 +56,6 @@ def test_provider_model_kwargs_for_fireworks_unknown_effort_omits_reasoning() ->
    assert "model_kwargs" not in kwargs


-@pytest.mark.parametrize("slug", sorted(NEW_FIREWORKS_MODELS))
-def test_new_fireworks_model_is_supported(slug: str) -> None:
-    model_id = _FIREWORKS_PREFIX + slug
-    assert model_id in SUPPORTED_MODEL_IDS
-    model = next(m for m in SUPPORTED_MODELS if m["id"] == model_id)
-    efforts, default_effort, supports_images = NEW_FIREWORKS_MODELS[slug]
-    assert model["efforts"] == efforts
-    assert model["default_effort"] == default_effort
-    assert model["supports_images"] is supports_images
-
-
-@pytest.mark.parametrize("slug", sorted(NEW_FIREWORKS_MODELS))
-def test_new_fireworks_model_supports_only_listed_efforts(slug: str) -> None:
-    model_id = _FIREWORKS_PREFIX + slug
-    efforts = NEW_FIREWORKS_MODELS[slug][0]
-    for effort in efforts:
-        assert model_supports_effort(model_id, effort) is True
-    for effort in _ALL_EFFORTS:
-        if effort not in efforts:
-            assert model_supports_effort(model_id, effort) is False
-
-
-@pytest.mark.parametrize(
-    "slug",
-    [
-        "qwen3-coder-480b-a35b-instruct",
-        "kimi-k2-thinking",
-        "kimi-k2-instruct-0905",
-        "glm-4p6",
-        "mistral-large-3-fp8",
-        "deepseek-v3p2",
-        "qwen3-30b-a3b-instruct-2507",
-    ],
-)
-def test_unavailable_fireworks_models_are_gated_out(slug: str) -> None:
-    assert _FIREWORKS_PREFIX + slug not in SUPPORTED_MODEL_IDS
-
-
-@pytest.mark.parametrize("slug", sorted(NEW_FIREWORKS_MODELS))
-def test_only_minimax_m3_supports_images(slug: str) -> None:
-    model_id = _FIREWORKS_PREFIX + slug
-    assert model_supports_images(model_id) is (slug == "minimax-m3")
-
-
 def test_fireworks_falls_back_to_bedrock() -> None:
    assert (
        fallback_model_id_for("fireworks:accounts/fireworks/models/deepseek-v4-pro")
--- a/ui/AGENTS.md
+++ b/ui/AGENTS.md
@ -0,0 +1,12 @@
+# AGENTS.md
+
+This file applies to all work under `ui/`.
+
+## Package manager
+
+- Use **pnpm** for dashboard dependency management and script execution.
+- Run UI scripts with `pnpm run <script>` (for example, `pnpm run typecheck`, `pnpm run lint`, `pnpm run test`, `pnpm run build`).
+- Install or update UI dependencies with `pnpm install` / `pnpm add` only.
+- Do **not** use npm in this directory: no `npm install`, `npm ci`, `npm run`, `npx`, or npm lockfile changes.
+- Do **not** use Bun in this directory: no `bun install`, `bun add`, `bun run`, `bunx`, or Bun lockfile changes.
+- If a command must use npm or Bun, it belongs outside `ui/` in a subtree that explicitly owns that package-manager configuration and lockfiles.
--- a/ui/README.md
+++ b/ui/README.md
@ -7,7 +7,7 @@ This is a template for a new TanStack Start project with React, TypeScript, and
 To add components to your app, run the following command:

 ```bash
-npx shadcn@latest add button
+pnpm dlx shadcn@latest add button
 ```

 This will place the ui components in the `components` directory.
--- a/ui/package.json
+++ b/ui/package.json
@ -1,10 +1,11 @@
 {
  "name": "open-swe-dashboard",
  "private": true,
+  "packageManager": "pnpm@11.9.0",
  "type": "module",
  "scripts": {
    "dev": "vite dev --port 3000",
-    "build": "vite build",
+    "build": "node --max-old-space-size=4096 ./node_modules/vite/bin/vite.js build",
    "preview": "vite preview",
    "test": "vitest run",
    "lint": "eslint",
@ -55,6 +56,7 @@
    "@types/react": "^19.2.14",
    "@types/react-dom": "^19.2.3",
    "@vitejs/plugin-react": "^5.2.0",
+    "httpxy": "0.5.3",
    "jsdom": "^27.4.0",
    "prettier": "^3.8.1",
    "prettier-plugin-tailwindcss": "^0.8.0",
--- a/ui/pnpm-lock.yaml
+++ b/ui/pnpm-lock.yaml
--- a/ui/src/components/agents/PlanReview.tsx
+++ b/ui/src/components/agents/PlanReview.tsx
@ -1,4 +1,5 @@
 import { useCallback, useEffect, useState } from "react"
+import { useNavigate } from "@tanstack/react-router"

 import type { PlanComment, PlanData } from "@/lib/plan"
 import {
@ -7,6 +8,7 @@ import {
  deletePlanComment,
  getPlanComments,
  rejectPlan,
+  updatePlan,
 } from "@/lib/plan"
 import { Button } from "@/components/ui/button"
 import { Markdown } from "@/components/agents/ported"
@ -47,6 +49,7 @@ async function copyToClipboard(text: string): Promise<boolean> {
 }

 export function PlanReview({ plan }: { plan: PlanData }) {
+  const navigate = useNavigate()
  const resolvedTheme = useResolvedTheme()
  const [comments, setComments] = useState<Array<PlanComment>>([])
  const [draft, setDraft] = useState("")
@ -55,6 +58,50 @@ export function PlanReview({ plan }: { plan: PlanData }) {
  const [busy, setBusy] = useState<"approve" | "reject" | null>(null)
  const [error, setError] = useState<string | null>(null)
  const [copied, setCopied] = useState(false)
+  // Locally track the displayed markdown so a manual edit shows immediately; the
+  // route's query stops polling once a plan exists, so the prop won't refetch.
+  const [markdown, setMarkdown] = useState(plan.markdown)
+  const [editing, setEditing] = useState(false)
+  const [editDraft, setEditDraft] = useState(plan.markdown)
+  const [saving, setSaving] = useState(false)
+
+  // Reflect external plan updates (e.g. an agent revision) while not editing.
+  useEffect(() => {
+    if (!editing) setMarkdown(plan.markdown)
+  }, [plan.markdown, editing])
+
+  const canEdit =
+    plan.isOwner && plan.status !== "approved" && plan.status !== "cancelled"
+
+  const startEditing = useCallback(() => {
+    setEditDraft(markdown)
+    setEditing(true)
+    setError(null)
+  }, [markdown])
+
+  const cancelEditing = useCallback(() => {
+    setEditing(false)
+    setError(null)
+  }, [])
+
+  const saveEdit = useCallback(async () => {
+    const next = editDraft.trim()
+    if (!next) {
+      setError("The plan cannot be empty.")
+      return
+    }
+    setSaving(true)
+    setError(null)
+    try {
+      const result = await updatePlan(plan.threadId, next)
+      setMarkdown(result.markdown)
+      setEditing(false)
+    } catch (e) {
+      setError((e as Error).message)
+    } finally {
+      setSaving(false)
+    }
+  }, [editDraft, plan.threadId])

  // Poll so reviewers see each other's comments without a realtime transport.
  useEffect(() => {
@ -108,39 +155,42 @@ export function PlanReview({ plan }: { plan: PlanData }) {
      setBusy(kind)
      setError(null)
      try {
-        if (kind === "approve") await approvePlan(plan.threadId)
-        else await rejectPlan(plan.threadId)
-        setDecision(
-          kind === "approve"
-            ? "Plan approved — the agent is implementing it."
-            : "Changes requested — the agent is revising the plan."
-        )
+        if (kind === "approve") {
+          await approvePlan(plan.threadId)
+          await navigate({
+            to: "/agents/$threadId",
+            params: { threadId: plan.threadId },
+          })
+          return
+        }
+        await rejectPlan(plan.threadId)
+        setDecision("Changes requested — the agent is revising the plan.")
      } catch (e) {
        setError((e as Error).message)
      } finally {
        setBusy(null)
      }
    },
-    [plan.threadId]
+    [navigate, plan.threadId]
  )

  const copyPlan = useCallback(async () => {
    setError(null)
-    if (await copyToClipboard(plan.markdown)) {
+    if (await copyToClipboard(markdown)) {
      setCopied(true)
      window.setTimeout(() => setCopied(false), 1500)
    } else {
      setError("Couldn't copy the plan to the clipboard.")
    }
-  }, [plan.markdown])
+  }, [markdown])

  return (
    <div
      data-testid="plan-review"
      className="flex min-h-0 flex-1 flex-col bg-[var(--ui-bg)] text-[var(--ui-text)]"
    >
-      <div className="flex items-center justify-between gap-4 border-b border-[var(--ui-border)] px-6 py-3">
-        <div>
+      <div className="flex flex-col gap-3 border-b border-[var(--ui-border)] px-4 py-3 md:flex-row md:items-center md:justify-between md:gap-4 md:px-6">
+        <div className="min-w-0">
          <h1 className="text-base font-semibold text-[var(--ui-text)]">
            Implementation plan
          </h1>
@ -150,58 +200,105 @@ export function PlanReview({ plan }: { plan: PlanData }) {
            <span data-testid="plan-status">{plan.status}</span>
          </p>
        </div>
-        <div className="flex shrink-0 items-center gap-2">
+        <div className="flex min-w-0 flex-wrap items-center gap-2 md:shrink-0 md:justify-end">
          {decision && (
            <span
              data-testid="plan-decision"
-              className="text-xs text-[var(--ui-text-dim)]"
+              className="w-full text-xs text-[var(--ui-text-dim)] md:w-auto"
            >
              {decision}
            </span>
          )}
-          <Button
-            data-testid="copy-plan"
-            variant="secondary"
-            disabled={!plan.markdown.trim()}
-            onClick={() => void copyPlan()}
-          >
-            {copied ? "Copied!" : "Copy markdown"}
-          </Button>
-          {plan.isOwner && (
-            <Button
-              data-testid="approve-plan"
-              disabled={busy !== null || decision !== null}
-              onClick={() => void decide("approve")}
-            >
-              Approve
-            </Button>
+          {editing ? (
+            <>
+              <Button
+                data-testid="cancel-edit-plan"
+                variant="secondary"
+                disabled={saving}
+                onClick={cancelEditing}
+              >
+                Cancel
+              </Button>
+              <Button
+                data-testid="save-plan"
+                disabled={saving || !editDraft.trim()}
+                onClick={() => void saveEdit()}
+              >
+                {saving ? "Saving…" : "Save"}
+              </Button>
+            </>
+          ) : (
+            <>
+              {canEdit && (
+                <Button
+                  data-testid="edit-plan"
+                  variant="secondary"
+                  disabled={busy !== null || decision !== null}
+                  onClick={startEditing}
+                >
+                  Edit
+                </Button>
+              )}
+              <Button
+                data-testid="copy-plan"
+                variant="secondary"
+                disabled={!markdown.trim()}
+                onClick={() => void copyPlan()}
+              >
+                {copied ? "Copied!" : "Copy markdown"}
+              </Button>
+              {plan.isOwner && (
+                <Button
+                  data-testid="approve-plan"
+                  disabled={busy !== null || decision !== null}
+                  onClick={() => void decide("approve")}
+                >
+                  Approve
+                </Button>
+              )}
+              <Button
+                data-testid="reject-plan"
+                variant="secondary"
+                // Requesting changes feeds the comments to the agent, so it's
+                // meaningless with none — disable until at least one is left.
+                disabled={
+                  busy !== null || decision !== null || comments.length === 0
+                }
+                title={
+                  comments.length === 0
+                    ? "Leave a comment first to request changes"
+                    : undefined
+                }
+                onClick={() => void decide("reject")}
+              >
+                Request changes
+              </Button>
+            </>
          )}
-          <Button
-            data-testid="reject-plan"
-            variant="secondary"
-            // Requesting changes feeds the comments to the agent, so it's
-            // meaningless with none — disable until at least one is left.
-            disabled={busy !== null || decision !== null || comments.length === 0}
-            title={
-              comments.length === 0
-                ? "Leave a comment first to request changes"
-                : undefined
-            }
-            onClick={() => void decide("reject")}
-          >
-            Request changes
-          </Button>
        </div>
      </div>

-      <div className="flex min-h-0 flex-1 overflow-hidden">
+      <div className="flex min-h-0 flex-1 flex-col overflow-y-auto md:flex-row md:overflow-hidden">
        <div
-          className="min-h-0 flex-1 overflow-auto px-6 py-4"
+          className="min-w-0 px-4 py-4 md:min-h-0 md:flex-1 md:overflow-auto md:px-6"
          data-testid="plan-document"
          data-color-scheme={resolvedTheme}
        >
-          {plan.markdown.trim() ? (
-            <Markdown content={plan.markdown} />
+          {editing ? (
+            <div className="flex h-full flex-col gap-2">
+              {error && (
+                <p className="text-xs text-[color:var(--ui-danger)]">{error}</p>
+              )}
+              <textarea
+                data-testid="plan-editor"
+                value={editDraft}
+                onChange={(e) => setEditDraft(e.target.value)}
+                spellCheck={false}
+                className="min-h-[20rem] w-full flex-1 resize-none rounded-md border border-[var(--ui-border)] bg-[var(--ui-bg)] px-3 py-2 font-mono text-sm text-[var(--ui-text)] outline-none focus:border-[var(--ui-accent)]"
+              />
+            </div>
+          ) : markdown.trim() ? (
+            <Markdown content={markdown} />
          ) : (
            <p className="text-sm text-[var(--ui-text-dim)]">
              The plan hasn't been written yet.
@ -209,14 +306,14 @@ export function PlanReview({ plan }: { plan: PlanData }) {
          )}
        </div>

-        <aside className="flex w-80 shrink-0 flex-col border-l border-[var(--ui-border)]">
+        <aside className="flex shrink-0 flex-col border-t border-[var(--ui-border)] md:w-80 md:border-t-0 md:border-l">
          <div className="border-b border-[var(--ui-border)] px-4 py-3">
            <h2 className="text-sm font-semibold text-[var(--ui-text)]">
              Comments
            </h2>
          </div>
          <div
-            className="min-h-0 flex-1 space-y-3 overflow-auto px-4 py-3"
+            className="max-h-80 space-y-3 overflow-auto px-4 py-3 md:max-h-none md:min-h-0 md:flex-1"
            data-testid="plan-comments"
          >
            {comments.length === 0 ? (
--- a/ui/src/lib/plan.ts
+++ b/ui/src/lib/plan.ts
@ -113,6 +113,16 @@ export function deletePlanComment(
  )
 }

+export function updatePlan(
+  threadId: string,
+  markdown: string
+): Promise<{ status: PlanStatus; markdown: string }> {
+  return req(`/plan/${encodeURIComponent(threadId)}`, {
+    method: "PUT",
+    body: JSON.stringify({ markdown }),
+  })
+}
+
 export function approvePlan(threadId: string): Promise<{ status: string }> {
  return req(`/plan/${encodeURIComponent(threadId)}/approve`, {
    method: "POST",
--- a/ui/src/routes/agents/$threadId_.plan.tsx
+++ b/ui/src/routes/agents/$threadId_.plan.tsx
@ -13,7 +13,7 @@ export const Route = createFileRoute("/agents/$threadId_/plan")({

 function Centered({ children }: { children: React.ReactNode }) {
  return (
-    <div className="flex min-w-0 flex-1 items-center justify-center p-6">
+    <div className="flex min-w-0 flex-1 items-center justify-center px-4 py-6 max-md:pt-14 md:p-6">
      {children}
    </div>
  )
@ -100,7 +100,7 @@ function PlanPage() {

  return (
    <div className="flex min-w-0 flex-1 flex-col">
-      <div className="border-b border-[var(--ui-border)] px-6 pt-3">
+      <div className="border-b border-[var(--ui-border)] px-4 pt-14 md:px-6 md:pt-3">
        <BackLink threadId={threadId} />
      </div>
      <PlanReview plan={plan} />