Breaking

Fuga Fingerprints the Logical Prompt Across Arguments and Standard Input, Not the Raw Pipe

September 16, 2026 at 15:06 EDT

AI Agents
Software Dev & Coding

When a replay harness verifies that the system under test asked an external agent the same question the recording captured, the obvious implementation is to hash whatever arrived on standard input and compare. Fuga's replay transport did roughly that, and two commits in this window replaced it with something narrower: hash the logical prompt, wherever it happens to live, and never hash the transport envelope around it.

2026-09-17 · FUGA · LOCAL DEVELOPMENT

Fingerprint the prompt, not the pipe

A replay harness stopped hashing raw standard input and started extracting the logical prompt from wherever the command line actually put it, then hashing that.

places a prompt is recognised

distinct failure boundaries, up from one

identifier or path normalisation inferred

A mismatch used to be one answer. Now it is two.

Both columns count named missing-input boundaries the harness can report for a prompt check.

before: prompt mismatch

after: never observed, or observed and different

Where the harness looks, in order

1 · AFTER THE DELIMITER

Accepted only when it is the final argument and is not itself a flag

2 · PROMPT FLAG

Long or short form, and now required to be non-empty so a blank value cannot end the search

3 · STREAMING INPUT

Two exact flag spellings only; the user turn is read from the stream and text blocks joined

4 · PLAIN INPUT

Only if nothing above matched is the raw stream treated as the prompt

WHY THE ENVELOPE IS NOT THE QUESTION

A piped stream and a prompt argument can both be present at once, and they mean different things.

Hashing the stream would fingerprint the supporting context instead of the instruction
Hashing the streaming envelope would fingerprint the transport framing instead of the text
The observation record now carries the prompt digest and the place it was read from

WHAT IMPROVED

Prompt identity is derived from what the system under test actually wrote; a pre-seeded input value was removed from the check.

WHAT IS UNPROVEN

No measurement of how many recorded cases changed verdict, and no claim that four shapes cover every engine the orchestrator can launch.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.