aliatx2017/reasonix-hermes/complete-step-evidence

Complete-Step Evidence Protocol

The problem

complete_step evidence citations are checked against your actual command history for the turn. If you cite a command you didn't literally run this turn, the step is REJECTED. This causes loops: you write a plausible grep/test command in the evidence, it gets rejected, you run it again, cite it again.

The rule

Run the exact command BEFORE calling complete_step. Never fabricate a citation from memory. The command string in the evidence must match a bash invocation from this exact turn — character for character.

Workflow

1. Decide what verification you need (grep for code, run a test, etc.)
2. Run the bash command now — in this turn, don't skip it
3. Call complete_step and cite that exact command (copy-paste the

command string you just typed into bash)

1. If multiple verifications are needed, run each one as a separate

bash call first, then cite all of them

Common failure patterns

• Citing grep -n WorkshopThreshold internal/boot/boot.go but you only

ran a combined grep -n "WorkshopThreshold|..." internal/boot/boot.go — these are different command strings → rejected

• Citing a test command from a previous turn that's now cached
• Citing a command you planned to run but didn't

Anti-pattern

complete_step(evidence=[{command: "grep -n X file.go", ...}])
# REJECTED — grep -n X file.go was never run this turn

Correct pattern

bash: grep -n X file.go
# output...
complete_step(evidence=[{command: "grep -n X file.go", ...}])
# ACCEPTED — command was literally run 2 seconds ago