consult

Files (1)

• SKILL.md

SKILL.md (first lines)

## Argument Parsing

Parse the user's input for:
- First positional argument → `question`
- `--focus`: one of `general`, `architecture`, `security`, `performance`, `style`
- `--depth`: one of `quick`, `balanced`, `deep`
- `--file` / `-f`: file path(s) to include as context

## Attachment verification (Part 2.1 — fail-loud)

Before calling `mori-consult_advisor`, scan the question text for attachment references:
- Explicit `(attached: <name>)` patterns
- Instructions matching a read-the-source intent: "attached", "READ the", "read the actual",
  "source read", "vs pin", "primary source", "primary-source"

For each referenced file: verify it exists on disk. If **any** referenced file is absent, **abort —
do not call `mori-consult_advisor`** — and report:

```
CONSULT ABORTED: attached file(s) missing from payload:
  - <path>  (referenced in question but not found on disk)
Fix: pass --file <path> or ensure the file exists before consulting.
```

After verifying (or if no attachment references), append a manifest line to the context:

```
ATTACHED FILES: <name1> (<bytes>B), <name2> (<bytes>B)
```

or `ATTACHED FILES: none` if no files were supplied.

## Source-dependence classification (Part 2.2 — mode-integrity)

A consult is **source-dependent** if the question contains any attachment reference OR `--file`
arguments were supplied.

## Depth enforcement

If `--file` arguments were supplied, `--depth` MUST be `deep`. If the user passed `quick`
or `balanced` (or omitted `--depth`), silently coerce to `deep` and prepend a one-line note
to your response:

```
Note: depth coerced to deep — --file requires full analysis budget.
```

## File content injection (client-side)

If `--file` arguments were supplied and all files passed verification:

For each path, use the CC `Read` tool to read the file content **client-side**. Build a list of
objects for the `file_contents` parameter:

```json
[{"name": "path/to/file.py", "content": "<file text>"}, ...]
```

Record each file's name and byte count (bytes = `len(content)`) for the `ATTACHED FILES:` manifest.
Keep the manifest (and any other non-file background) in the `context` parameter — **do not**
paste full file bodies into `context`.

**Do NOT pass raw paths in `files=`** — the MCP server runs on remote GCE; local filesystem
paths are inaccessible there. Always pass `files=[]`; file bodies arrive via `file_contents`.

## Execution (async submit + poll)

`consult_advisor` is **non-blocking**: it returns a `job_id` immediately. Deep consults can take
up to ~300s of Bifrost inference (round-trip often 400–450s). Poll until done.

1. Run attachment verification. Abort loudly if any referenced file is missing.
2. Classify source-dependence.
3. Read each `--file` path client-side; build `file_contents`; append `ATTACHED FILES:` to `context`.
4. Call `mori-consult_advisor` with `question`, `context`, `focus`, `depth`, `files=[]`, and
   `file_contents=[...]` (or omit / pass `[]` when no files).
5. Parse the JSON response: `{"job_id": "...", "status": "pending"}`.
6. Poll `mori-consult_status` with that `job_id` every **30–60 seconds** (minimum 30s —
   even small-file consults commonly take ~2 minutes of Bifrost latency; polling faster
   just burns tool calls) until:
   - `{"status": "done", "result": "<advice>"}` → present `result`
   - `{"status": "error", "error": "..."}` → report the error loudly
   - **Deadline** (~**10 minutes** for deep; ~4 minutes for quick/balanced) → fail with timeout
7. Present the advice (or the failure).

**On job not found:** in-memory jobs are wiped on server restart/CD — resubmit; do not invent advice.

**On error or empty response:**
- **Not source-dependent**: retry once with `--depth quick` (new submit + poll).
- **Source-dependent**: do NOT fall back. Return:

```
CONSULT UNAVAILABLE (source-dependent; fallback would be source-blind).
Options: retry manually, postpone, or proceed without consultation (disclose this).
```

Empty deep-mode responses must surface as `EMPTY_RESPONSE (elapsed: Xs)` — not silent success.

From fjwood69/mori, AGPL-3.0.

More from fjwood69/mori