You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: .claude-plugin/marketplace.json
+1-1Lines changed: 1 addition & 1 deletion
Original file line number
Diff line number
Diff line change
@@ -9,7 +9,7 @@
9
9
"name": "cuey",
10
10
"source": "./plugins/cuey",
11
11
"description": "Cuey is an agentic financial intelligence system that analyzes prompts and Excel workbooks, improves decision accuracy through cross-checking and fact-checking, and delivers evidence-backed recommendations.",
Copy file name to clipboardExpand all lines: plugins/cuey/.claude-plugin/plugin.json
+1-1Lines changed: 1 addition & 1 deletion
Original file line number
Diff line number
Diff line change
@@ -1,6 +1,6 @@
1
1
{
2
2
"name": "cuey",
3
-
"version": "0.4.10",
3
+
"version": "0.4.12",
4
4
"description": "Cuey is an agentic financial intelligence system that analyzes prompts and Excel workbooks, improves decision accuracy through cross-checking and fact-checking, and delivers evidence-backed recommendations.",
Copy file name to clipboardExpand all lines: plugins/cuey/skills/cuey/SKILL.md
+64-3Lines changed: 64 additions & 3 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,6 +1,6 @@
1
1
---
2
2
name: cuey
3
-
description: Run Cuey only when the user explicitly invokes /cuey. Cuey analyzes prompts and Excel workbooks, cross-checks key claims, fact-checks assumptions, and returns evidence-backed recommendations. Use /cuey probe only for attachment transport diagnostics.
3
+
description: Run Cuey only when the user explicitly invokes /cuey. Cuey analyzes prompts and Excel workbooks, cross-checks key claims, fact-checks assumptions, and returns evidence-backed recommendations. Use /cuey probe or /cuey probe https only for attachment transport diagnostics.
4
4
argument-hint: <question>
5
5
---
6
6
@@ -10,13 +10,67 @@ Run this skill only after the user explicitly invokes `/cuey`. Do not invoke Cue
10
10
11
11
Before using any tool, classify the invocation using this routing table:
12
12
13
-
1. If `$ARGUMENTS` is exactly `probe`, starts with `probe `, or is an explicit
13
+
1. If `$ARGUMENTS` is exactly `probe https` or starts with `probe https `,
14
+
this is an HTTPS attachment probe. You must follow **HTTPS Attachment Probe**
15
+
and must not call `cuey:ask_cuey`.
16
+
2. If `$ARGUMENTS` is exactly `probe`, starts with `probe `, or is an explicit
14
17
attachment transport diagnostic request such as `attachment probe`, this is
15
18
a probe request. You must follow **Attachment Probe** and must not call
16
19
`cuey:ask_cuey`.
17
-
2. Otherwise, this is a normal Cuey request. Follow **Ask Cuey** and call
20
+
3. Otherwise, this is a normal Cuey request. Follow **Ask Cuey** and call
18
21
`cuey:ask_cuey`.
19
22
23
+
## HTTPS Attachment Probe
24
+
25
+
This section applies only to `/cuey probe https`. It validates the direct,
26
+
large-file transport path. Do not search for or call `cuey:ask_cuey`, do not
27
+
build an Ask Cuey payload, and do not analyze the attachment.
28
+
29
+
Use exactly one raw attachment from the current user message. Never reuse an
30
+
attachment from an earlier message. Claude must have a readable raw file path
31
+
for that current attachment, normally under `/root/.claude/uploads/...`. If no
32
+
such path is available, say that direct upload cannot be tested for this
33
+
attachment and stop.
34
+
35
+
1. Read the file size with `stat -c%s` and SHA-256 with `sha256sum` in the
36
+
Claude sandbox. Do not base64 encode or chunk the file.
37
+
2. Call only `cuey:probe_claude_attachment_https` with:
38
+
39
+
```json
40
+
{
41
+
"action": "begin",
42
+
"filename": "original filename",
43
+
"mimeType": "original MIME type",
44
+
"sizeBytes": 123,
45
+
"sha256": "lowercase sha256"
46
+
}
47
+
```
48
+
49
+
3. From the returned `upload_url`, `headers`, and `probe_token`, use `curl` in
50
+
the Claude sandbox to stream the exact original file:
51
+
52
+
```bash
53
+
curl --fail --silent --show-error --request PUT \
54
+
--upload-file "$ATTACHMENT_PATH" \
55
+
-H "Content-Type: $MIME_TYPE" \
56
+
-H "x-amz-meta-cuey-sha256: $SHA256" \
57
+
"$UPLOAD_URL"
58
+
```
59
+
60
+
4. Call only `cuey:probe_claude_attachment_https` again:
61
+
62
+
```json
63
+
{
64
+
"action": "complete",
65
+
"probeToken": "token returned by begin"
66
+
}
67
+
```
68
+
69
+
The upload URL and probe token are short-lived secrets. Use them only in tool
70
+
calls and commands; never include either in the user-visible answer. Return the
71
+
completion result exactly and stop. This probe does not run fanout, persist a
72
+
Cuey document, or produce a normal Ask Cuey answer.
73
+
20
74
## Attachment Probe
21
75
22
76
This section applies only to probe requests.
@@ -51,6 +105,13 @@ by the local Cuey MCP runtime. First read that current-message file in Claude's
51
105
sandbox and pass its bytes inline as base64. Include the path only as metadata
52
106
alongside the base64.
53
107
108
+
Do not run a size preflight and then decide to send a sandbox path instead of
109
+
bytes. For this probe, attempt the MCP call with inline base64 when Claude can
110
+
read the current-message attachment bytes. If the attachment is too large for a
111
+
single inline MCP tool call, stop and report that this file requires the direct
112
+
HTTPS transport path. In that failure case, do not
113
+
retry with a `/root/.claude/uploads/...` path-only payload.
114
+
54
115
Only if Claude exposes a non-sandbox file path or handle that the MCP runtime
55
116
can read may you pass a path or handle without base64. If Claude exposes only
56
117
extracted text and no raw file reference or bytes, leave `attachments` and
0 commit comments