Bound command-syntax recovery and pivot after repeated invalid calls

[justsml]

Parent

#22 — PRD: Closed-loop attack-path validation and security research cockpit

What to build

Make command-syntax recovery converge. After an invalid shell/search shape, the controller should expose the exact validation error, permit at most one materially different correction, and then pivot to another approved tool or stop with a clear blocker. Repeated malformed calls must not consume the run budget or create fake execution evidence.

The DeepSeek V4 Flash hosted matrix recorded 2 syntax errors and 1 repeated similar mistake in command-syntax-recovery; the GPT OSS 120B control still recorded one syntax correction.

Acceptance criteria

• Invalid command payloads are classified before execution and persisted as invalid attempts.
• The recovery policy allows at most one materially different correction per error class before pivot/stop.
• Three-repeat DeepSeek V4 Flash and GPT OSS 120B runs produce no repeated syntax mistake and at most one syntax error per scenario.
• The final summary distinguishes invalid attempts, successful tool calls, and the remaining blocker with toolCalls/maxToolCalls.

Blocked by

None - can start immediately.

[justsml]

Implemented in e082141c. Syntax recovery now permits one retryable malformed call and at most one materially different correction, then emits a terminal pivot/stop with no fake bash execution; run-scoped state resets between candidates and final text is forced after terminal recovery. Focused regression tests and typecheck pass. Full-suite verification: 1,336/1,338 tests passed; the two failures are unrelated timeout rows. Live DeepSeek V4 Flash/GPT OSS 120B three-repeat acceptance remains pending provider credentials, so this issue stays open.