Remove Claude attachment probes (#18)

Author: wilsonccccc

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
       "name": "cuey",
       "source": "./plugins/cuey",
       "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.",
-      "version": "0.4.13",
+      "version": "0.4.14",
       "author": {
         "name": "TensorBlock"
       },

plugins/cuey/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "cuey",
-  "version": "0.4.13",
+  "version": "0.4.14",
   "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.",
   "author": {
     "name": "TensorBlock",

plugins/cuey/skills/cuey/SKILL.md

@@ -1,193 +1,20 @@
 ---
 name: cuey
-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.
+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.
 argument-hint: <question>
 ---
 
 # Cuey
 
 Run this skill only after the user explicitly invokes `/cuey`. Do not invoke Cuey automatically for financial decisions, model analysis, business assumptions, or any other request that does not include `/cuey`.
 
-Before using any tool, classify the invocation using this routing table:
+When invoked, call the local MCP tool `cuey:ask_cuey`. Do not use bash, recall memory, search, or answer directly before calling the tool.
 
-1. If `$ARGUMENTS` is exactly `probe https` or starts with `probe https `,
-   this is an HTTPS attachment probe. You must follow **HTTPS Attachment Probe**
-   and must not call `cuey:ask_cuey`.
-2. If `$ARGUMENTS` is exactly `probe`, starts with `probe `, or is an explicit
-   attachment transport diagnostic request such as `attachment probe`, this is
-   a probe request. You must follow **Attachment Probe** and must not call
-   `cuey:ask_cuey`.
-3. Otherwise, this is a normal Cuey request. Follow **Ask Cuey** and call
-   `cuey:ask_cuey`.
+Preserve Claude's normal attachment workflow. Users attach files in Claude's composer; Cuey consumes only files and attachment content that Claude already exposes in the current request. Do not search local paths, upload or send files separately, invent file handles, or replace available attachment content with Claude-generated summaries. Do not decide the merge route. Cuey backend chooses the final synthesis or artifact merge after fanout returns.
 
-## HTTPS Attachment Probe
+If the request includes Claude-visible attachments of any type, read them with Claude's available file, spreadsheet, document, or vision capability and add compact, relevant attachment context to `context`. This applies to Excel, PDF, image, document, text, CSV, and other Claude-readable files. For each relevant attachment, include the filename, file type, extracted content or observations needed for the user's request, and any omitted sections when the file is too large to include fully. Do not include only a filename when actual attachment content is available.
 
-This section applies only to `/cuey probe https`. It validates the direct,
-large-file transport path. Do not search for or call `cuey:ask_cuey`, do not
-build an Ask Cuey payload, and do not analyze the attachment.
-
-Use exactly one raw attachment from the current user message. Never reuse an
-attachment from an earlier message. Claude must have a readable raw file path
-for that current attachment, normally under `/root/.claude/uploads/...`. If no
-such path is available, say that direct upload cannot be tested for this
-attachment and stop.
-
-1. Read the file size with `stat -c%s` and SHA-256 with `sha256sum` in the
-   Claude sandbox. Do not base64 encode or chunk the file.
-2. Call only `cuey:probe_claude_attachment_https` with:
-
-```json
-{
-  "action": "begin",
-  "filename": "original filename",
-  "mimeType": "original MIME type",
-  "sizeBytes": 123,
-  "sha256": "lowercase sha256"
-}
-```
-
-3. From the returned `upload_url`, `headers`, and `probe_token`, use the
-   following exit-safe `curl` command in the Claude sandbox to stream the exact
-   original file. It preserves S3's error details without exposing the upload
-   URL or probe token:
-
-```bash
-response_headers="$(mktemp)"
-response_body="$(mktemp)"
-
-set +e
-http_status="$(curl --http1.1 --fail-with-body --silent --show-error \
-  --request PUT \
-  --upload-file "$ATTACHMENT_PATH" \
-  -H "Content-Type: $MIME_TYPE" \
-  -H "x-amz-meta-cuey-sha256: $SHA256" \
-  --dump-header "$response_headers" \
-  --output "$response_body" \
-  --write-out '%{http_code}' \
-  "$UPLOAD_URL")"
-curl_exit=$?
-set -e
-
-if [ "$curl_exit" -ne 0 ]; then
-  s3_request_id="$(awk 'tolower($1) == "x-amz-request-id:" {gsub(/\r/, "", $2); print $2}' "$response_headers")"
-  s3_host_id="$(awk 'tolower($1) == "x-amz-id-2:" {gsub(/\r/, "", $2); print $2}' "$response_headers")"
-  s3_error_code="$(sed -n 's:.*<Code>\([^<]*\)</Code>.*:\1:p' "$response_body")"
-  s3_error_message="$(sed -n 's:.*<Message>\([^<]*\)</Message>.*:\1:p' "$response_body")"
-  printf 'cuey_s3_upload_failed\nhttp_status=%s\ncurl_exit=%s\ns3_error_code=%s\ns3_error_message=%s\ns3_request_id=%s\ns3_host_id=%s\n' \
-    "$http_status" "$curl_exit" "$s3_error_code" "$s3_error_message" "$s3_request_id" "$s3_host_id"
-  exit 0
-fi
-```
-
-4. If and only if the command succeeded, call only
-   `cuey:probe_claude_attachment_https` again:
-
-```json
-{
-  "action": "complete",
-  "probeToken": "token returned by begin"
-}
-```
-
-The upload URL and probe token are short-lived secrets. Use them only in tool
-calls and commands; never include either in the user-visible answer. On an
-upload failure, return only the printed `cuey_s3_upload_failed` diagnostic and
-stop; do not call `complete`. On success, return the completion result exactly
-and stop. This probe does not run fanout, persist a Cuey document, or produce a
-normal Ask Cuey answer.
-
-## Attachment Probe
-
-This section applies only to probe requests.
-
-Do not search for `ask_cuey`. Do not call `cuey:ask_cuey`. Do not build an
-Ask Cuey payload. Do not summarize the attachments yourself.
-
-Call only the local MCP tool `cuey:probe_claude_attachment`.
-
-Do not analyze, summarize, transcribe, or transform uploaded files. The purpose
-of this diagnostic is to verify whether Claude can pass raw uploaded file
-references, paths, handles, bytes metadata, or attachment objects to MCP, and
-whether MCP can upload the same file bytes to Cuey backend.
-
-Send only raw file/attachment values from the same user message that invoked
-this probe. Never use attachments from earlier turns, prior probes, chat
-history, memory, cached tool results, or previous Cuey responses. If the current
-message contains a visible image but Claude exposes no raw file bytes, path, or
-handle for that image, leave `attachments` and `files` empty and state that
-the current image was visible but no raw attachment transport was exposed.
-
-Prefer actual file bytes whenever Claude can read them. For each
-current-message attachment, first use Claude's available file access to read the
-attachment bytes and include inline `base64`, `dataBase64`, or `contentBase64`
-content along with filename and MIME type. Apply this to any current-message
-attachment type Claude can expose, including documents, spreadsheets, images,
-PDFs, text files, CSVs, archives, and other binary files.
-
-If Claude exposes a `/root/.claude/uploads/...` sandbox path, do not pass that
-path as the only transport. That path is readable by Claude's sandbox, but not
-by the local Cuey MCP runtime. First read that current-message file in Claude's
-sandbox and pass its bytes inline as base64. Include the path only as metadata
-alongside the base64.
-
-Do not run a size preflight and then decide to send a sandbox path instead of
-bytes. For this probe, attempt the MCP call with inline base64 when Claude can
-read the current-message attachment bytes. If the attachment is too large for a
-single inline MCP tool call, stop and report that this file requires the direct
-HTTPS transport path. In that failure case, do not
-retry with a `/root/.claude/uploads/...` path-only payload.
-
-Only if Claude exposes a non-sandbox file path or handle that the MCP runtime
-can read may you pass a path or handle without base64. If Claude exposes only
-extracted text and no raw file reference or bytes, leave `attachments` and
-`files` empty and put a short note in `note`.
-
-Use this payload shape:
-
-```json
-{
-  "note": "what Claude can access about the uploaded files",
-  "attachments": [
-    {
-      "filename": "example.png",
-      "mimeType": "image/png",
-      "path": "raw path or handle if available",
-      "base64": "raw file bytes as base64 if Claude can expose them"
-    }
-  ],
-  "files": [],
-  "context": "",
-  "upload": true
-}
-```
-
-Return the MCP result exactly. Do not add commentary. Stop immediately.
-
-## Ask Cuey
-
-This section applies only to normal Cuey requests. If this is a probe request,
-do not execute this section.
-
-Call the local MCP tool `cuey:ask_cuey`. Do not use bash, recall memory, search,
-or answer directly before calling the tool.
-
-Preserve Claude's normal attachment workflow. Users attach files in Claude's
-composer. For normal `/cuey` requests, Cuey receives compact context extracted
-by Claude, not raw uploaded files, paths, handles, bytes, or `sourceFiles`. Raw
-attachment transport is experimental and belongs only to `/cuey probe`.
-
-If the request includes Claude-visible non-Excel attachments, add only compact,
-relevant context that Claude can already extract naturally. Include filename,
-file type, extracted text or observations needed for the user's request, and
-any omitted sections when the file is too large to include fully. Do not search
-local paths, upload or send files separately, invent file handles, or send raw
-file bytes in normal `/cuey`.
-
-Do not decide the merge route. Cuey backend chooses the final synthesis or
-artifact merge after fanout returns.
-
-If the request includes a native Claude `.xlsx` attachment, also build compact
-workbook context in `spreadsheet` containing:
+If the request includes a native Claude `.xlsx` attachment, also build compact workbook context in `spreadsheet` containing:
 
 - workbook filename;
 - every sheet name;
@@ -203,7 +30,7 @@ Send this payload:
 {
   "mode": "ask | compare | verify | summarize",
   "question": "$ARGUMENTS",
-  "context": "only relevant prior conversation context plus compact Claude-extracted context from non-Excel attachments",
+  "context": "only relevant prior conversation context plus compact context extracted from Claude-visible non-Excel attachments",
   "spreadsheet": {
     "filename": "attached workbook filename, or empty when none",
     "context": "structured workbook context extracted from the attached .xlsx, or empty when none"
@@ -215,19 +42,11 @@ Send this payload:
 
 Choose `compare` for comparisons, `verify` for risk or correctness checks, `summarize` for summaries, and `ask` otherwise.
 
-Set `"worklog": true` only when the request is clearly a CFO, FP&A, finance,
-financial model, forecast, budget, board, investor, audit, reporting, variance,
-revenue, cost, cash, runway, or unit-economics task where generated artifacts
-should include candidate worklogs. For ordinary questions, normal summaries,
-and simple Excel generation, omit `worklog` or set it to `false`.
+Set `"worklog": true` only when the request is clearly a CFO, FP&A, finance, financial model, forecast, budget, board, investor, audit, reporting, variance, revenue, cost, cash, runway, or unit-economics task where generated artifacts should include candidate worklogs. For ordinary questions, normal summaries, and simple Excel generation, omit `worklog` or set it to `false`.
 
-Do not send `sourceFiles`, `smart_merge`, `models`, `reasoningLevel`, agent
-names, or merge settings. Those choices belong to Cuey backend.
+Do not send `sourceFiles`, `smart_merge`, `models`, `reasoningLevel`, agent names, or merge settings. Those choices belong to Cuey backend.
 
-Requests to create, generate, build, export, or return an Excel workbook are
-always `ask`, never `verify`. Do not change a workbook-generation request into a
-requirements analysis. Example: `Create a downloadable Excel workbook...`
-must be sent as `{"mode":"ask","question":"Create a downloadable Excel workbook..."}`.
+Requests to create, generate, build, export, or return an Excel workbook are always `ask`, never `verify`. Do not change a workbook-generation request into a requirements analysis. Example: `Create a downloadable Excel workbook...` must be sent as `{"mode":"ask","question":"Create a downloadable Excel workbook..."}`.
 
 When there is no Excel attachment, omit `spreadsheet`. Keep non-Excel attachment context in `context`; never send `sourceFiles`.