[Bug] /review-code delegation to Oracle fails silently — Sisyphus reviews its own code

[ermmak]

Problem

The /review-code command specifies delegate_task(subagent_type="oracle", ...) but Oracle does not appear to be invoked. Instead, Sisyphus performs the review itself.

Impact: Self-review bias. The same model that wrote the code reviews it → blind spots to its own bugs.

Evidence

Observation	Expected	Actual
/coding-workflow REVIEW verdict	Mixed (Oracle/GPT finds bugs)	Always OKAY
Manual "delegate to Oracle for review"	Oracle finds issues	Oracle finds issues
Review file writing style	Oracle style (terse, direct)	Sisyphus style (✅ lists, self-congratulatory tone)

Style Comparison

Suspected Sisyphus-generated review (from /coding-workflow):

• Heavy use of ✅ checkmark lists
• Phrases like "Quality observations", "Test quality: Excellent", "Efficient implementation"
• Self-congratulatory framing

Confirmed Oracle-generated review (from explicit delegate_task):

• More factual, direct prose
• Explains why changes are correct rather than listing what was checked
• Calls out behavioral changes with risk context

Reproduction Steps

1. Configure oh-my-opencode.json with different models for sisyphus and oracle:

   {
     "agents": {
       "sisyphus": { "model": "anthropic/claude-opus-4-5" },
       "oracle": { "model": "openai/gpt-5.2-codex" }
     }
   }

2. Run /coding-workflow <task> through REVIEW stage
3. Examine .sisyphus/review-*.md for writing style
4. Compare with explicit delegate_task(subagent_type="oracle", ...) output
5. Observe style mismatch suggesting Sisyphus wrote both

Root Cause Hypothesis

The delegate_task(subagent_type="oracle", ...) call in /review-code either:

1. Fails silently and falls back to Sisyphus executing locally
2. Is not being executed as a real delegation (Sisyphus interprets the instructions and does the work itself)
3. Has a bug in the delegation mechanism that doesn't properly spawn the configured Oracle model

Requested Fixes

Priority	Fix
P0	Surface errors when delegate_task fails to spawn subagent
P0	Prevent silent fallback to parent agent
P1	Log which model actually executes (add to review file header)
P2	Add verification that subagent actually spawned

Suggested Review File Header

Adding model metadata would make this auditable:

VERDICT: OKAY
MODEL: openai/gpt-5.2-codex
AGENT: oracle
DELEGATED_BY: sisyphus
TIMESTAMP: 2026-01-30T14:41:57+05:00
---

Environment

• oh-my-opencode config: Custom oh-my-opencode.json with separate models per agent
• Sisyphus model: anthropic/claude-opus-4-5
• Oracle model: openai/gpt-5.2-codex

Workaround

Manually invoke Oracle review after IMPLEMENT stage instead of relying on /review-code:

"Delegate to Oracle to review the changes in this branch"

This bypasses the slash command and forces actual Oracle delegation.

[ermmak]

Update: Root Cause Identified + Workaround Found

Finding

The subagent_type parameter in delegate_task() does not respect the model configuration in oh-my-opencode.json.

Delegation Method	Expected Model (per config)	Actual Model
subagent_type="oracle"	openai/gpt-5.2-codex	claude-opus-4-5 ❌
category="ultrabrain"	openai/gpt-5.2-codex	openai/gpt-5.2-codex ✅

Evidence

Review file headers from test runs:

Using subagent_type="oracle":

REVIEWER: oracle
MODEL: claude-opus-4-5  ← Wrong! Config says gpt-5.2-codex

Using category="ultrabrain":

REVIEWER: ultrabrain
MODEL: openai/gpt-5.2-codex  ← Correct!

Workaround

Modified /review-code command to use category="ultrabrain" instead of subagent_type="oracle". This correctly invokes the configured model.

Suggested Fix

The subagent_type parameter should respect the model configuration in oh-my-opencode.json just like category does. Currently it appears to either:

1. Ignore the config entirely
2. Fall back to the parent agent's model
3. Have a hardcoded default

This is a significant issue because users expect subagent_type="oracle" to use Oracle's configured model, not the calling agent's model.

[zxsleebu]

could be related to #1255