Review what shipped in the OpenAI Codex CLI since the last adoption analysis and file an umbrella issue (☂️) + one subissue per adoptable feature, each grounded in how Sybra actually drives Codex headless. Use when asked to "check what's new in Codex", "adopt new Codex features", or to refresh the Codex CLI adoption tracker. For Claude Code, use claude-feature-adoption instead.

2026-07-257

adversarial-review

Automaat/sybra

Fast adversarial code review. Red-teams a diff or PR to find the concrete bug, then refutes its own findings to cut false positives. Works with Claude, Codex, or Copilot — uses subagents when the runtime has them, falls back to sequential local passes. Use for routine changes where a full staff review is overkill.

2026-07-187

codex-feature-adoption

Non-interactive PR review fix — fetch unresolved review threads, auto-apply valid fixes, reply on every thread (applied/questionable/invalid). Use when given a PR URL and asked to address review feedback without prompts.

2026-07-187

sybra-audit

Automaat/sybra

Analyze Sybra health findings + audit summary to explain workflow issues and propose grounded fixes. Use when asked to review how work is flowing or suggest process improvements.

2026-07-187

sybra-eval-report

Automaat/sybra

Analyze the Sybra evaluation scorecard, breakdowns, and weaknesses, then recommend concrete improvements toward an autonomous mid-level engineering team. Use when asked how well/efficiently Sybra is working, to review the evaluation data, or to act on the fleet scorecard.

2026-07-187

name	sybra-evaluate
description	Evaluate completed Sybra tasks — determine status transition and link PRs. Use when asked to evaluate task completion.
allowed-tools	Bash
user-invocable	true
disable-model-invocation	true

Sybra Task Evaluation

Decide what happens to a task after an agent finishes. Do NOT read source code, review diffs, or explore the codebase.

Constraints

Do NOT use Read tool or explore any files
Do NOT review code quality, style, or correctness
Only analyze the agent result text provided in the prompt
Keep total cost under $0.02 per evaluation

Process

1. Read the task

sybra-cli --json get <id>

2. Link PR if created

Search the agent result for PR references. Look for:

gh pr create output containing a URL like https://github.com/.../pull/N
Mentions of "PR #N" or "pull request #N"
Branch names pushed (for branch linking)

If found, link to task:

# Link PR number
sybra-cli --json update <id> --pr <number>

# Link branch if known and not already set
sybra-cli --json update <id> --branch <branch-name>

3. Decide status transition

Based ONLY on the agent result text:

Condition	New Status
Agent completed work, PR created or code pushed	in-review
Agent completed but no PR/push (partial work)	human-required
Agent failed, hit errors, looped	human-required
Agent blocked, needs input	human-required

Rules

Never set done — only humans do that
Never set todo — triggers auto-dispatch, creates duplicate agents
Default to in-review when uncertain
Set human-required if agent output shows errors, loops, or incomplete work

4. Update status

sybra-cli --json update <id> --status <new-status>

Input: Agent result text contains `https://github.com/acme/repo/pull/42` and "Successfully created PR".

Actions:

sybra-cli --json update task-abc --pr 42
sybra-cli --json update task-abc --status in-review

Input: Agent result text contains "Error: rate limit exceeded" with no PR reference.

Actions:

sybra-cli --json update task-abc --status human-required

Input: Agent result says "Refactored auth.go, all tests pass" but no branch/PR reference.

Actions: status=human-required (partial work, no push detected).