fix(hooks): pass failed commands through verbatim, never distill a non-zero exit into success

[fajarhide]

Closes #120

Problem

A command that failed could be distilled into output that reads as success. That is the worst failure mode: a fabricated success terminates investigation, while a fabricated error only costs a retry. In the filed case a vault call that failed exit=2 on a network timeout surfaced a clean, plausible, fictional ["n8n"].

Root cause (code-visible, not the unreproducible ["n8n"] itself): OMNI's normalize layer parsed each agent's failure signal and threw it away.

agent	failure signal	state before
Codex CLI	exit_code	never read into CodexInput
Pi	toolResponse.isError	#[allow(dead_code)]
Generic MCP	result.isError	in a doc comment, never deserialized
Claude Code	tool_response is a bare string "Error: Exit code N…"	parse-failed to None (safe by accident)

So failed commands from Codex/Pi/MCP ran the full distiller.

Fix

The invariant from the issue: a non-zero exit passes through verbatim, never distilled. Implemented at the one place format knowledge already lives:

• NormalizedInput gains failed, set from each agent's own signal (exit_code != 0, isError).
• One guard in post_tool::process_payload: if normalized.failed { return None } — passthrough, host keeps the raw bytes at zero marker cost.
• No is_command_error text heuristic: the guard keys on exit status only, so a successful build log full of the word "error" still distills. Net savings on passing commands are untouched.

Claude Code needs no code change (its failure string already passes through); a regression test locks that in so a future lenient parser can't silently reintroduce the fabrication.

Verified end-to-end (real binary, omni --hook)

input	before	after
docker build exit_code 1, 9207 B	distilled → 6090 B	0 B → passthrough (host keeps raw)
docker build exit_code 0, 9207 B	distilled → 6090 B	distilled → 6090 B (unchanged)

Teeth: disabling the guard turns the Codex/Pi/MCP tests red; restoring it makes them green.

Gates: cargo fmt / clippy -D warnings / cargo test (1027 passed, 0 failed) — run on local Homebrew 1.94.0, not CI's pinned 1.97.0, so CI clippy is the source of truth.

Out of scope (follow-up)

The omni exec/pipe.rs path reads piped stdout only and never sees the child exit code; hardening it needs exec.rs to capture and forward the code — a separate change, not this one.

PR Auto Describe

Summary

Added failed field to NormalizedInput to track non-zero command exits or error signals per-agent. Failed commands now pass through verbatim instead of being distilled — preventing fabricated success summaries that would terminate investigation early. Fixes #120 with full test coverage across all affected parsers.

Key Changes

• New field: NormalizedInput.failed: bool tracks command failure per agent format
• Passthrough guard: process_payload returns None (verbatim pass) when failed == true
• Per-agent parsing: exit codes / isError flags wired into failed where available
• Tests: 5 new cases covering Codex, Pi, MCP, and Claude Code failure paths

Detailed Breakdown

• normalize_claude_code: always sets failed: false — failures arrive as bare strings, not objects, so never reach the normalizer
• normalize_pi: reads isError boolean into failed
• normalize_opencode / normalize_vscode_continue / normalize_aider: set failed: false — no exit/error signal in payload
• normalize_codex: reads exit_code field; failed = exit_code.is_some_and(|c| c != 0)
• normalize_generic_mcp: reads isError from McpResultContent into failed
• process_payload: early return None when normalized.failed — host keeps original bytes; no risk of a distilled success hiding a real failure

Notes

• Design decision: "fabricated success terminates investigation; fabricated error only costs a retry" — skewed toward preserving fidelity of failed commands.
• Claude Code bare-string failures already passthrough via passthrough guard; the failed: false here documents intent and guards future parser changes.

Breaking Changes

None.

Last updated: 2026-07-21 00:51:52