perf(codex): trim context and tune parallelism #159

Merged
claudiomaradonna merged 2 commits from perf/codex-reviewer-context into main 2026-10-21 16:57:32 +02:00
Collaborator

Why

Local Codex reviewers inherit app, plugin and skill context unrelated to
their repository review. Parallel implementation uses CPU cores as its
only capacity limit, even when the Codex session has fewer agent slots.

What changed

The three local reviewer adapters disable apps, plugins, bundled skills
and automatic skill instructions, retaining their project briefs, shell
access, models and effort. Configuration checks guard these settings.

Codex dispatch now accounts for available session slots, queues excess
tasks, reuses compatible implementers and falls back to parent execution
when no usable slot can be obtained. Build and test resource contention
can further reduce concurrency. Task isolation and landing order remain.

How to verify

Run the Maintenance commands in README.md. The configuration suite passes
297 tests, including 36 new regression cases that fail on main and pass
on this branch. Work-folder tests, ShellCheck, Markdown formatting, static
configuration validation and diff whitespace checks pass.

Review the scheduling rule with occupied slots, unknown capacity and more
tasks than slots. Independent change, security and documentation reviews
completed without unresolved findings.

Considerations

The settings follow the official Codex custom-agent schema. An offline
Codex 0.163.0 prompt probe confirms that the equivalent root overrides
remove automatic skill instructions while retaining repository rules;
it does not exercise a spawned custom role or measure speed or cost.
The README's billed behavioral evals remain for the maintainer before
merge. No new tools or dependencies were installed.

## Why Local Codex reviewers inherit app, plugin and skill context unrelated to their repository review. Parallel implementation uses CPU cores as its only capacity limit, even when the Codex session has fewer agent slots. ## What changed The three local reviewer adapters disable apps, plugins, bundled skills and automatic skill instructions, retaining their project briefs, shell access, models and effort. Configuration checks guard these settings. Codex dispatch now accounts for available session slots, queues excess tasks, reuses compatible implementers and falls back to parent execution when no usable slot can be obtained. Build and test resource contention can further reduce concurrency. Task isolation and landing order remain. ## How to verify Run the Maintenance commands in README.md. The configuration suite passes 297 tests, including 36 new regression cases that fail on main and pass on this branch. Work-folder tests, ShellCheck, Markdown formatting, static configuration validation and diff whitespace checks pass. Review the scheduling rule with occupied slots, unknown capacity and more tasks than slots. Independent change, security and documentation reviews completed without unresolved findings. ## Considerations The settings follow the official Codex custom-agent schema. An offline Codex 0.163.0 prompt probe confirms that the equivalent root overrides remove automatic skill instructions while retaining repository rules; it does not exercise a spawned custom role or measure speed or cost. The README's billed behavioral evals remain for the maintainer before merge. No new tools or dependencies were installed.
Sign in to join this conversation.
No reviewers
No labels
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
claudiomaradonna/agent-foundry!159
No description provided.