[poc] Demo Persistent sessions/directories across harnesses/sandboxes by junaway · Pull Request #4813 · Agenta-AI/agenta

junaway · 2026-06-24T07:42:39Z

No description provided.

vercel · 2026-06-24T07:42:46Z

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
agenta-documentation	Ready	Preview, Comment	Jun 25, 2026 8:14am

coderabbitai · 2026-06-24T07:42:49Z

Important

Review skipped

Draft detected.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: 9163debd-9e1a-4505-a9e5-3590813eb37e

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests
Commit unit tests in branch poc/persistent-sessions

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands.}

+
+async function sh(sbx, cmd, opts = {}) {
+  const r = await sbx.commands.run(cmd, { timeoutMs: 180000, ...opts });
+  console.log(`$ ${cmd}\n  exit=${r.exitCode}${r.stderr ? "\n  stderr: " + r.stderr.slice(0, 400) : ""}`);


+// GET /files?session_id=&path=  -> proxy sandbox-agent fs listing (live view)
+app.get("/files", async (req, res) => {
+  const { session_id: sid, path = "" } = req.query;
+  const abs = `/work/${sid}/${path}`.replace(/\/+$/, "") || `/work/${sid}`;


Copilot

Pull request overview

Adds a self-contained “persistent sessions” demo that persists a per-session working directory in SeaweedFS (S3 API) via geesefs, and reuses/resumes coding-agent sessions across local and cloud sandboxes via a Node sidecar + FastAPI UI.

Changes:

Introduces a Node sidecar that mounts per-session S3 prefixes into sandboxes and streams sandbox-agent ACP events as NDJSON.
Adds a FastAPI service + static UI to create/resume sessions, persist transcripts to Postgres, and browse durable files from SeaweedFS.
Adds docker-compose wiring plus multiple provider-specific smoke tests (E2B/Modal/Daytona) and an E2B template Dockerfile.

Reviewed changes

Copilot reviewed 33 out of 37 changed files in this pull request and generated 10 comments.

Show a summary per file

File	Description
sessions/demo/sidecar/server.js	Sidecar HTTP API for mounting, running prompts, streaming events, and proxying sandbox filesystem listing.
sessions/demo/sidecar/provider-modal.js	Modal provider integration via a Python bridge subprocess.
sessions/demo/sidecar/provider-e2b.js	E2B provider integration (provision/reconnect, mount geesefs, start sandbox-agent).
sessions/demo/sidecar/package.json	Sidecar Node dependencies and start script.
sessions/demo/sidecar/modal_bridge.py	Modal sandbox provisioning/mounting/agent bootstrap logic (Python SDK).
sessions/demo/sidecar/Dockerfile	Sidecar container image (Node + Python + Modal SDK).
sessions/demo/seaweedfs/s3.json	SeaweedFS S3 identity configuration for the demo bucket.
sessions/demo/sandbox/entrypoint.sh	Sandbox container entrypoint (seed agent auth/trust files; start sandbox-agent).
sessions/demo/sandbox/Dockerfile	Local sandbox image (sandbox-agent + harness installs + geesefs + entrypoint).
sessions/demo/modal-smoketest/smoke.py	Modal FUSE/geesefs write round-trip smoke test.
sessions/demo/matrix_test.py	End-to-end harness×sandbox durability matrix smoke script.
sessions/demo/e2b-template/e2b.Dockerfile	E2B template image build (manual agent layout + geesefs).
sessions/demo/e2b-smoketest/smoke.mjs	E2B FUSE/geesefs mount+write smoke test.
sessions/demo/e2b-smoketest/package.json	E2B smoketest dependency manifest.
sessions/demo/e2b-smoketest/package-lock.json	E2B smoketest dependency lockfile.
sessions/demo/docker-compose.yml	Demo stack orchestration (postgres, seaweedfs, ngrok, sandbox, sidecar, fastapi).
sessions/demo/daytona-smoketest/write.mjs	Daytona geesefs write probe.
sessions/demo/daytona-smoketest/smoke.mjs	Daytona viability gate (egress + geesefs mount/write).
sessions/demo/daytona-smoketest/puttest.mjs	Daytona raw S3 PUT probe via tunnel (no geesefs).
sessions/demo/daytona-smoketest/package.json	Daytona smoketest dependency manifest.
sessions/demo/daytona-smoketest/l4probe.mjs	Daytona TCP-connect vs TLS/egress probe.
sessions/demo/daytona-smoketest/ipv4write.mjs	Daytona detached mount+write probe with IPv4 pinning.
sessions/demo/daytona-smoketest/ipv4.mjs	Daytona IPv4-only tunnel test + geesefs write.
sessions/demo/daytona-smoketest/final.mjs	Daytona consolidated curl+geesefs experiment script.
sessions/demo/daytona-smoketest/egress.mjs	Daytona basic egress probe to common hosts + tunnel.
sessions/demo/daytona-smoketest/curltest.mjs	Daytona GET/PUT curl probes to tunnel (including larger payload).
sessions/demo/daytona-smoketest/allowlist2.mjs	Daytona domainAllowList widening probe using example.com + tunnel.
sessions/demo/daytona-smoketest/allowlist.mjs	Daytona domainAllowList probe focusing on tunnel reachability.
sessions/demo/api/static/index.html	Static UI for session list, invoking runs, transcript view, and file browsing.
sessions/demo/api/requirements.txt	FastAPI service Python dependencies.
sessions/demo/api/main.py	FastAPI endpoints for invoke, session management, transcript persistence, and file reads.
sessions/demo/api/Dockerfile	FastAPI container build.
sessions/demo/api/db.py	Postgres schema + session/transcript persistence helpers.
sessions/demo/.env.example	Example environment variables for running the demo stack.

Files not reviewed (1)

sessions/demo/e2b-smoketest/package-lock.json: Generated file

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

+        ) as resp:
+            async for line in resp.aiter_lines():
+                line = line.strip()
+                if not line:
+                    continue
+                evt = json.loads(line)


+        resp = await s3.list_objects_v2(Bucket=S3_BUCKET, Prefix=f"{sid}/")
+        keys = [{"Key": o["Key"]} for o in resp.get("Contents", [])]
+        if keys:
+            await s3.delete_objects(Bucket=S3_BUCKET, Delete={"Objects": keys})


+async def set_sandbox_id(sid: str, sandbox_id: str):
+    async with (await pool()).acquire() as con:
+        await con.execute(
+            "UPDATE sessions SET sandbox_id = $2 WHERE id = $1", sid, sandbox_id
+        )


+async def append_event(sid: str, event_index, sender, session_update, payload: dict):
+    # seq = current count of this session's events (dense, monotonic per session).
+    async with (await pool()).acquire() as con:
+        seq = await con.fetchval(
+            "SELECT count(*) FROM session_transcripts WHERE session_id = $1", sid
+        )
+        await con.execute(
+            """INSERT INTO session_transcripts
+                 (id, session_id, seq, event_index, sender, session_update, payload)
+               VALUES ($1, $2, $3, $4, $5, $6, $7)
+               ON CONFLICT (session_id, seq) DO NOTHING""",
+            uuid7(),
+            sid,
+            seq,
+            event_index,
+            sender,
+            session_update,
+            json.dumps(payload),
+        )
+        await con.execute("UPDATE sessions SET updated_at = now() WHERE id = $1", sid)


+app.get("/files", async (req, res) => {
+  const { session_id: sid, path = "" } = req.query;
+  const abs = `/work/${sid}/${path}`.replace(/\/+$/, "") || `/work/${sid}`;
+  try {
+    const r = await fetch(`${AGENT_URL}/v1/fs/entries?path=${encodeURIComponent(abs)}`);
+    res.status(r.status).type("application/json").send(await r.text());
+  } catch (e) {
+    res.status(502).json({ error: String(e) });
+  }
+});


+// Auto-approve permission backstop (modes above usually skip prompts).
+function autoApprove(session) {
+  session.onPermissionRequest?.((reqEvt) => {
+    const opts = reqEvt?.options || [];
+    const pick =
+      opts.find((o) => /allow.*always|always/i.test(o.kind || o.name || "")) ||
+      opts.find((o) => /allow/i.test(o.kind || o.name || "")) ||
+      opts.find((o) => !/reject|deny/i.test(o.kind || o.name || "")) ||
+      opts[0];
+    session.respondPermission?.(reqEvt.id, { optionId: pick?.optionId ?? pick?.id, allow: true });
+  });
+}


+  // start the sandbox-agent server as root with agent creds in its env.
+  // background:true returns immediately without an exit code — do NOT run()-check it.
+  const envExports = Object.entries(agentEnv).map(([k, v]) => `${k}='${v}'`).join(" ");
+  sbx.commands
+    .run(`sudo sh -c "${envExports} exec sandbox-agent server --no-token --host 0.0.0.0 --port ${AGENT_PORT} >/tmp/sa.log 2>&1"`,
+      { background: true, timeoutMs: 0 })


+    const d = document.createElement('div');
+    d.className = 'sess' + (s.id === current ? ' active' : '');


+      <div class="meta">
+        <span class="badge b-provider" title="provider">${s.provider}</span>${modelShort(s) ? `<span class="badge b-model" title="model">${modelShort(s)}</span>` : ''}${s.reasoning && s.reasoning !== 'none' ? `<span class="badge b-reasoning" title="reasoning">${s.reasoning}</span>` : ''}
+      </div>`;
+    d.onclick = () => select(s.id);


+RUN apt-get update && apt-get install -y --no-install-recommends python3 python3-pip \
+    && pip3 install --break-system-packages --no-cache-dir modal \
+    && rm -rf /var/lib/apt/lists/*


…rollbars Daytona's egress tier was lifted, so the geesefs-over-tunnel durability gate (real S3 PUT that used to hang) now passes — host-verified durable in SeaweedFS. Built it out as a first-class provider, mirroring the E2B path: - sidecar/provider-daytona.js: provision from snapshot, mount geesefs via ngrok, start the agent server, connect through the preview link (url + token header). Long-lived server uses a process session (runAsync) — executeCommand always blocks until exit, so a backgrounded & times out. - sidecar/daytona_snapshot.js: builds the snapshot (template-equivalent) with sandbox-agent + claude/codex/opencode/pi + geesefs baked in. Idempotent. - Wire daytona into server.js /run + /kill, compose env, UI dropdown (first of the non-local sandboxes), and matrix_test.py defaults. - Matrix: daytona x {claude,codex,opencode,pi} all PASS (durable). - Cleanup: drop 11 dead egress-investigation probe scripts, keep durability.mjs. - UI: hide scrollbar chrome on the session list / main pane while keeping scroll. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…terface Collapse the hand-rolled per-provider glue onto rivet sandbox-agent's SandboxProvider abstraction. SandboxAgent.start({ sandbox }) now drives the full create/reconnect/getUrl/destroy lifecycle; a single withGeesefs() wrapper adds the reusable geesefs-over-ngrok durable cwd (mount demo:<sid> + seed auth + start the credentialed agent server). makeProvider(name) returns the wrapped provider. - sidecar/sandbox-provider.js: withGeesefs wrapper + custom bases for daytona, e2b, docker. Custom (not SDK built-in) because: the built-in e2b() calls a removed Sandbox.betaCreate; built-in docker() clobbers HostConfig (can't add FUSE caps without losing the port mapping); and crucially all three built-ins start a CREDENTIAL-LESS agent server, which our restart can't reliably replace (the exec shell is a child of that server). Our bases start no server — withGeesefs starts the only one, with creds. Verified the server env carries ANTHROPIC_API_KEY. - Add `docker` provider: fresh container per session via the host daemon (docker.sock mounted into the sidecar), /dev/fuse + SYS_ADMIN for geesefs, host.docker.internal for the agent URL. New docker-image/Dockerfile (agenta-sandbox-agent:local, built --platform linux/amd64 to match the x64 agent/geesefs binaries). UI dropdown after local; matrix default. - modal stays on the Python bridge (provider-modal.js) — the Node modal SDK needs a separately-baked Modal image. local stays the persistent compose container. - server.js /run + /kill route daytona/e2b/docker through makeProvider; modal + local unchanged. - Delete the now-dead provider-daytona.js / provider-e2b.js; fix Dockerfile COPY; drop the unused Node modal dep; bump @e2b/code-interpreter to 2.x. Matrix (durable agent write, host-verified): local/docker/daytona/modal/e2b all PASS. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

docker containers are AutoRemove, so a session's container is gone once it stops; SandboxAgent.start() doesn't auto-recreate — it 404s on the dead container id when resuming (also possible for a GC'd cloud sandbox). On resume failure, fall back to starting a fresh sandbox, which remounts the same demo:<sid> prefix so the durable cwd in SeaweedFS (and all prior files) is preserved. Verified: write marker in run 1 (container auto-removed), resume same session in run 2 (fresh container) reads the marker back. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docker containers run `sleep infinity` + the agent server, so they never stop on their own and AutoRemove never fires — every session leaked a live container. docker is fresh-per-turn, so tear the container down after the turn: dispose() the agent connection cleanly FIRST (killing mid-stream caused "other side closed" -> HTTP 500 on the turn), then destroy the container, then clear the persisted sandbox_id. The cwd is durable in SeaweedFS (--fsync-on-close), so resume recreates + remounts. Verified: a docker turn leaves 0 containers; write marker -> resume same session into a fresh container -> reads the marker back. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

+    except httpx.HTTPError as e:
+        note = f"sidecar kill error (cleared anyway): {e}"
+    await db.set_sandbox_id(sid, None)
+    return {"killed": row["sandbox_id"], "note": note}


+// real prompt rides in data.inputs.messages[0], dims in data.parameters.
+function envelope({ sid, prompt, force }) {
+  const body = { force: !!force };
+  if (sid) body.session_id = sid;


+    await fetch(`${API_URL}/sessions/${sid}/sandbox-id`, {
+      method: "PUT",
+      headers: { "content-type": "application/json" },
+      body: JSON.stringify({ sandbox_id: sandboxId }),
+    });


+    }
+    res.end();
+  } catch (err) {
+    console.error("[/run] error", err);


Copilot

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[poc] persistent sessions/directories across harnesses/sandboxes

44675c7

Copilot AI review requested due to automatic review settings June 24, 2026 07:42

Copilot started reviewing on behalf of junaway June 24, 2026 07:43 View session

vercel Bot deployed to Preview June 24, 2026 07:43 View deployment

github-advanced-security AI found potential problems Jun 24, 2026

View reviewed changes

Copilot AI reviewed Jun 24, 2026

View reviewed changes

vercel Bot deployed to Preview June 24, 2026 09:08 View deployment

Copilot AI review requested due to automatic review settings June 24, 2026 10:25

Copilot started reviewing on behalf of junaway June 24, 2026 10:25 View session

Copilot AI reviewed Jun 24, 2026

vercel Bot deployed to Preview June 24, 2026 10:26 View deployment

vercel Bot deployed to Preview June 24, 2026 10:30 View deployment

Copilot AI review requested due to automatic review settings June 24, 2026 10:37

Copilot started reviewing on behalf of junaway June 24, 2026 10:37 View session

Copilot AI reviewed Jun 24, 2026

vercel Bot deployed to Preview June 24, 2026 10:38 View deployment

working send/steer/cancel/attach/detach

35a5fc3

vercel Bot deployed to Preview June 25, 2026 07:25 View deployment

github-advanced-security AI found potential problems Jun 25, 2026

View reviewed changes

Comment thread sessions/demo/api/main.py

except httpx.HTTPError as e:

note = f"sidecar kill error (cleared anyway): {e}"

await db.set_sandbox_id(sid, None)

return {"killed": row["sandbox_id"], "note": note}

github-advanced-security AI found potential problems Jun 25, 2026

View reviewed changes

swap force for cancel/attach

eccd1da

Copilot AI review requested due to automatic review settings June 25, 2026 08:13

Copilot started reviewing on behalf of junaway June 25, 2026 08:14 View session

Copilot AI reviewed Jun 25, 2026

vercel Bot deployed to Preview June 25, 2026 08:14 View deployment

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[poc] Demo Persistent sessions/directories across harnesses/sandboxes#4813

[poc] Demo Persistent sessions/directories across harnesses/sandboxes#4813
junaway wants to merge 7 commits into
mainfrom
poc/persistent-sessions

junaway commented Jun 24, 2026

Uh oh!

vercel Bot commented Jun 24, 2026 •

edited

Loading

Uh oh!

coderabbitai Bot commented Jun 24, 2026 •

edited

Loading

Review skipped

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

		const d = document.createElement('div');
		d.className = 'sess' + (s.id === current ? ' active' : '');

Uh oh!

Conversation

junaway commented Jun 24, 2026

Uh oh!

vercel Bot commented Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

coderabbitai Bot commented Jun 24, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

vercel Bot commented Jun 24, 2026 •

edited

Loading

coderabbitai Bot commented Jun 24, 2026 •

edited

Loading