docs(roadmap): add #453 — slash-command guard fires only on bare token; any extra arg falls through to prompt dispatch

[Yeachan-Heo]

ROADMAP pinpoint #453 — slash-command guard is shape-blind to subcommand args

Dogfooded for the 2026-05-24 06:00 Clawhip pinpoint nudge (message 1507986539538419863).

Repro (clean isolated env, ANTHROPIC_* unset)

Bare form caught by guard (good):

$ claw cost
[error-kind: unknown] error: `claw cost` is a slash command. Use `claw --resume SESSION.jsonl /cost` ...

Add ONE argument → guard breaks, prompt dispatched to Claude (bug):

$ claw cost list
[error-kind: missing_credentials] error: missing Anthropic credentials; export ANTHROPIC_AUTH_TOKEN ...

The missing_credentials envelope proves "cost list" was sent to the LLM as a user prompt.

Affected spellings (all currently fall through to CliAction::Prompt)

Spelling	Bare form	With extra arg
cost	✅ guard	❌ cost list → prompt
tokens	✅ guard	❌ tokens list → prompt
model	✅ guard	❌ model openai/gpt-4 → prompt
permissions	✅ guard	❌ permissions show → prompt
context	✅ guard	❌ context show → prompt
providers	✅ guard	❌ providers list → prompt
history	✅ guard	❌ history list → prompt
cache	✅ guard	❌ cache list → prompt
models (plural)	❌ falls through	❌ falls through
aliases	❌ falls through	❌ falls through
profiles	❌ falls through	❌ falls through
logs	❌ falls through	❌ falls through
settings	❌ falls through	❌ falls through

Code location

rust/crates/rusty-claude-cli/src/main.rs:1120-1149:

fn bare_slash_command_guidance(command_name: &str) -> Option<String> {
    if matches!(command_name, "dump-manifests" | "bootstrap-plan" | ... | "export") {
        return None;
    }
    let slash_command = slash_command_specs()
        .iter()
        .find(|spec| spec.name == command_name)?;
    ...
}

The guard is only consulted from the bare parse arm — never from the subcommand-args path. Adding any positional arg routes the entire argv to CliAction::Prompt.

Why distinct from #78 / #145 / #452

• 见证历史 #78 / 合影 #145 filed claw plugins as a single missing CliAction variant.
• avicii #452 (PR docs(roadmap): add #452 — claw models/list/help/--help falls through to prompt dispatch (prompt misdelivery, sibling of #78) #3065) filed claw models* as a single missing CliAction variant.
• This pinpoint is the guard itself being shape-blind to subcommand args — even commands whose bare form is correctly handled regress to prompt misdelivery the moment an operator types one more word. Includes the dangerous claw model openai/gpt-4 typo (natural given the existing claw --model openai/gpt-4 prompt … flag).

Why it matters

Prompt misdelivery is the Clawhip top-of-list category. With creds set, every operator/claw typo in this cluster burns provider tokens on a meaningless completion. Without creds, the wrong-shaped missing_credentials envelope is even more confusing — operators see auth error when they meant local CLI dispatch.

Required fix shape (full detail in ROADMAP entry)

(a) Widen bare_slash_command_guidance to fire from the subcommand-args parse arm too. (b) Extend known-slash-command set to include natural CLI discovery spellings (models, aliases, profiles, logs, settings) and emit did_you_mean suggestions. (c) Structured kind:"argv_misroute_prevented" JSON envelope. (d) Regression coverage for every spelling in the table above, asserting never CliAction::Prompt.

Acceptance check (one-liner)

env -u ANTHROPIC_API_KEY -u ANTHROPIC_AUTH_TOKEN claw model openai/gpt-4 should NOT exit with missing_credentials; it should exit with a typed CLI dispatch error mentioning --model openai/gpt-4 prompt … or /model.

—
[repo owner's gaebal-gajae (clawdbot) 🦞]

[code-yeongyu]

Content merged to main via batch commit c881069. Closing PR.