paulnsorensen/read-mode-probe

read-mode-probe — Five probes, one stance

Reader-first. No edits. Pick the probe closest to the user's question; if ambiguous, ask. Output is a numbered list of findings tagged with confidence (high / med / low) and a file:line citation, followed by a "next probes I would run" closer.

The five probes

1. Invariant probe

List the things that must always be true in this code, and cite where they're enforced. Catalog what would happen if each invariant were violated (silent corruption? hard panic? rejected request?).

Tool order:

1. mcp__serena__find_symbol for the central type/function.
2. mcp__tilth__tilth_grok to pull its callers + tests in one shot.
3. mcp__serena__find_referencing_symbols to find every check site.

2. Data-flow probe

Trace a value — a user-id, a request body, an env var, a feature flag — from its entry point to its persistence or external use. Name every file:line where the value is read, transformed, or written.

Tool order:

1. mcp__tilth__tilth_search (kind=content) for the literal name.
2. mcp__serena__find_referencing_symbols on each definition site.
3. mcp__code-review-graph__get_affected_flows_tool for the downstream graph.

3. Error-path probe

For a given function or module, enumerate every throw, return Err, panic, reject, or unhandled rejection. For each: where it can be triggered, who catches it (if anyone), and whether it's recoverable or terminal.

Tool order:

1. mcp__tilth__tilth_grok(target=<function>) for body + callers.
2. mcp__tilth__tilth_search(query="catch,try,Result,Err,panic,throw,reject") to find the surrounding error machinery.

4. Hot-path probe

Find loops, N+1 queries, sync I/O on request paths, and unbounded allocations. Sort findings by likely impact — flag the worst offender first.

Tool order:

1. mcp__code-review-graph__get_hub_nodes_tool for high-fan-in functions.
2. mcp__code-review-graph__find_large_functions_tool for >150-LOC bodies.
3. mcp__tilth__tilth_search(query="forEach,for.*await,await.*for,for.*\\.length") for common loop antipatterns.

5. Security probe

Audit input validation, authz checks, secret handling, deserialization sinks, SQL/shell injection vectors, SSRF, prototype pollution.

Tool order:

1. mcp__tilth__tilth_search(query="JSON.parse,eval,exec,Function,vm,child_process,spawn") for deserialization + exec sinks.
2. mcp__tilth__tilth_search(query="req.body,req.params,req.query,process.argv") for untrusted-input entry points.
3. mcp__serena__find_referencing_symbols to walk from each input to where it's used.

Output template

## Findings (<probe-name> probe on <target>)

1. [high] <finding> — `path/to/file.ts:42`
   <one-sentence why this matters>

2. [med] <finding> — `path/to/file.ts:88`
   <one-sentence why this matters>

...

## Next probes I would run

- <name>: <one-sentence reason>
- <name>: <one-sentence reason>

Hard rules

1. Every finding has a file:line citation. No "looks like" or "appears to" — quote the line if needed.
2. Confidence tags are mandatory. high = verified by reading the exact code. med = pattern-matched but not fully traced. low = suspicious, worth checking but not confirmed.
3. No edits, no refactor suggestions. If the user asks "how should we fix this?", reply: "Switch out of read-mode-probe. Want me to propose a fix?" — and wait.
4. Reader-first verbs only. read_file, codebase_search, find_symbol, mcp__serena__*, mcp__tilth__* (no tilth_edit), mcp__code-review-graph__*. No edit_file, no run_terminal_cmd except git log|status|diff|show, ls, wc.