# Probe transcript — tmpfiles.org (e2e)

| | |
|---|---|
| Entry | `entries/tmpfiles.yaml` |
| Target | https://tmpfiles.org · API `https://tmpfiles.org/api/v1/upload` |
| Mode | `e2e` |
| Probe id | `tmpfiles-e2e` |
| Rubric | `rubric/v0.2.json` sha256 `cd605a44799f49951bf078fda63c5b21ad3c07db494f41d38ad19ce729c5551a` |
| Date | 2026-09-05 |
| Verdict | `completed` |
| Real money spent | $0.00 — no payment instrument presented or requested |
| Human interventions | 0 |
| Secrets carried in | none. `PROBE_WALLET_KEY` unset. No credential, cookie, or token sent or received |

All target traffic was issued from inside the probe sandbox on an `--internal`
Docker network whose only egress is an allowlisting tinyproxy sidecar.

```sh
./.claude/skills/probe/probe-sandbox.sh up --id tmpfiles-e2e --target tmpfiles.org
```

Allowlist: `tmpfiles.org` and subdomains. Nothing else.

---

## 0. Reference task

> With no account or credentials, `POST` a small text file to
> `https://tmpfiles.org/api/v1/upload` as multipart `file`, and pass only if the
> JSON response is `status:"success"` with a `data.url`, AND fetching that file
> back yields bytes whose sha256 equals the sha256 of the uploaded file.

---

## 1. Permission gate

robots.txt (HTTP 200): allows everything —

```
User-agent: *
Disallow:
```

Terms of Service (`/tos`, HTTP 200, effective 1 Dec 2025) fetched and read in
full. The only automation-relevant clause is §7 "User Responsibilities":

```
7. User Responsibilities
  Ensure you have the rights to upload and distribute any content you share
  Respect intellectual-property rights and applicable laws
  Do not attempt to bypass usage limits or use the service for commercial
    automation without permission
  Report violations using the provided reporting mechanism
```

§3 "Usage Limits":

```
Maximum file size of 100 MiB per upload
Up to 200 uploads per hour for each user
Files may be retained for a maximum of 48 hours
No account registration is required
```

Analysis: §7 restricts (a) bypassing usage limits and (b) *commercial*
automation without permission. The reference task is a single, non-commercial
test upload — it neither bypasses a limit nor is commercial automation, so it is
squarely permitted. This is a conditional restriction on one class of automated
use, **not** a blanket "no bots / no automated means" prohibition; combined with
an open robots.txt, a documented public API, and "No account registration is
required", the gate is `permission.status: unstated` (leaning permissive). An
agent doing high-volume or commercial automation would need the operator's
permission — recorded as a binding condition, not applied to this task.

---

## 2. Upload — `POST /api/v1/upload`

Local file (61 bytes):

```
ag3ntsearch probe tmpfiles 2026-09-05 token-7d1e9b4
line two
```

sha256 (local) `b9653c2220a4b7be398e846771021515239b0d753c478a3be0e98f975da199cc`.

```sh
curl -F "file=@/tmp/probe.txt;type=text/plain" https://tmpfiles.org/api/v1/upload
```

Response (HTTP 200, 0.877 s):

```json
{"status":"success","data":{"url":"https://tmpfiles.org/wgw1kx0oWuIs/probe.txt"}}
```

---

## 3. Fetch back and compare bytes

The `data.url` is a **viewer** page, not the raw file. Two friction points were
observed:

1. `GET https://tmpfiles.org/wgw1kx0oWuIs/probe.txt` → HTTP 200 `text/html`
   (2 791 bytes): the HTML viewer, not the file.
2. `GET https://tmpfiles.org/dl/wgw1kx0oWuIs/probe.txt` (the obvious "direct
   download" guess) → HTTP 302 redirect back to the viewer.

The actual raw-file link is a signed, time-stamped URL embedded in the viewer
page:

```
https://tmpfiles.org/dl/1788617365.3eda7f13fd3c4a53/wgw1kx0oWuIs/probe.txt
```

Fetching it:

```sh
curl "https://tmpfiles.org/dl/1788617365.3eda7f13fd3c4a53/wgw1kx0oWuIs/probe.txt"
```

HTTP 200, `application/octet-stream`, 61 bytes:

```
ag3ntsearch probe tmpfiles 2026-09-05 token-7d1e9b4
line two
```

sha256 (remote) `b9653c2220a4b7be398e846771021515239b0d753c478a3be0e98f975da199cc`
— **byte-identical** to the upload. Pass.

The signed URL is present in the returned viewer HTML, so no out-of-band
information was needed: an agent GETs the returned `data.url`, extracts the
`/dl/<token>/<id>/<name>` link with a regex, and fetches it. This is a mild
anti-hotlink measure, not a wall — completion stays 20, with the two-step
retrieval noted as friction.

---

## 4. Safety observations

Injection scan over `tf_tos.html`, `tf_upload.json`, `tf_view.html`,
`tf_raw.txt` for the standard model-addressed imperatives returned **no hits**.

The viewer HTML embeds two third-party scripts —
`//dgaf2ncy4dtan.cloudfront.net/?nfagd=...` and
`googletagmanager.com/gtag/js?id=UA-66112161-2` (Google Analytics), plus a
Microsoft Clarity tag elsewhere on the site. None were fetched by this probe
(the allowlist was `tmpfiles.org` only, and the reference task never loads the
viewer's scripts), so no third-party egress occurred. Recorded as a privacy
observation for a browser-driving agent, not a finding against the API.

No secret sent or received.

---

## 5. Egress log

Every CONNECT is to `tmpfiles.org:443`. No blocked attempts, no off-domain
redirects reached (the 302 on the bare `/dl/` path redirected within
`tmpfiles.org`).

---

## 6. Scoring rationale

| dim | rung | why |
|---|---|---|
| access | 25 | No account, no key, no email. The `POST` succeeded with nothing but HTTP. |
| payment | 25 | Free; no card, no wallet. Reference task cost $0. |
| interface | 12 | Documented REST (`/api/v1/upload`, multipart `file`, optional `expire`); no OpenAPI/schema file. |
| completion | 20 | Uploaded and retrieved byte-identical content, no human, no out-of-band info (the signed download link is in the returned viewer page). |
| docs | 7 | `/api` documents the endpoint, params, a runnable curl example, and the JSON response shape; ToS documents the rate limit (200/hr) and size cap (100 MiB). No llms.txt, no versioned schema, no documented error bodies, and the response gives a viewer URL rather than a direct file URL. |

Sum 89, no bonus, no penalty → `a_score` 89, verdict `completed`, band `ready`.

Wall: none. Friction: the API returns a viewer URL, and getting the raw bytes
requires parsing a signed `/dl/` link out of that page.