Skip to content

Latest commit

 

History

History
149 lines (122 loc) · 8.1 KB

ASTRA.md

File metadata and controls

149 lines (122 loc) · 8.1 KB

Astra configuration and instruction audit

Checked against official OpenAI documentation on 2026-10-06. This is a quality-to-credit baseline, not a measured optimum.

Model defaults

Setting Choice Reason
model gpt-6-astra Preserve the requested model
model_reasoning_effort medium Existing live setting and local catalog default
plan_mode_reasoning_effort high Retain the existing planning setting
model_verbosity low Concise output
features.fast_mode false Disable Fast-tier selection in the TUI
service_tier unset Do not request priority processing

The repo previously installed Sol/high over a live Astra/medium configuration. The updated defaults resolve that drift. The local model catalog describes medium as balancing speed and reasoning depth. OpenAI's Astra migration guidance recommends retaining effective effort when migrating from supported levels.

The pricing page distinguishes included subscription usage from purchased credits and Enterprise pay-as-you-go usage:

Speed mode Included usage multiplier Purchased-credit / pay-as-you-go multiplier
Fast 2.5x 2x
Astra Ultrafast 8x 6x

These are billing multipliers relative to Standard for the same model, not speed or quality guarantees. Availability depends on plan and workspace. API dollar prices are separate and do not establish subscription usage saved per task. Recheck dated rates before making cost decisions.

The configuration reference distinguishes Fast-tier selection from service_tier. An explicit task, profile, project, or app override can supersede global defaults. Existing tasks may retain their model settings; start a new task and check its model and effort.

For simple work or difficult debugging, use an explicit per-session override:

codex -c model_reasoning_effort='"low"'
codex -c model_reasoning_effort='"high"'

Keep medium as the starting point. Compare accepted results, rework, elapsed time, and actual account usage on representative tasks before changing the default further. No credit-savings benchmark was run during this audit.

October 2026 follow-up

The python-ai, rust-cli, and find-skills skills were removed on 2026-10-27. Their mentions below record the earlier audit, not the current skill inventory.

The current model guide recommends GPT-6.1 Sol for complex coding and agentic work when available, with Astra for the most demanding work and Luna for focused tasks. Keep the explicitly chosen Astra baseline; compare accepted results and actual usage before switching. GPT-5.5 retires from ChatGPT and Codex on October 14, 2026, but not from the API. The former tui.model_availability_nux entries were notification state, not model selections. They were removed from the portable configuration along with personal project paths during the portability cleanup.

The September 11 skills guidance emphasizes focused triggers, conditional references, and clear completion boundaries. The follow-up changes apply it as follows:

  • Global instructions authorize validated local commits as work progresses; external publication still needs authorization.
  • pr-readiness selects uncommitted, branch, or commit review scope.
  • hugo and mdbook inspect context relevant to the requested change.
  • python-ai validates affected components and routes API migrations to a dated migration reference (removed with the skill on 2026-10-27).
  • autofix, code-review, and find-skills are now repository-managed. Local fixes do not imply publication; generic review and how-to requests no longer select CodeRabbit or skill discovery indiscriminately.
  • Skill authoring guidance covers invocation policy and behavioral evaluation cases. These cases are proposed checks, not claimed passing runs.

All 17 skill entrypoints are in scope; bundled domain manuals are not an exhaustive current-version audit. Model defaults and security settings remain unchanged. Plugin-owned skills remain separate and may still overlap; loading a plugin does not supersede repository review policy.

September 2026 findings and changes

All 14 repository skill entrypoints, global and root instructions, and the planning instruction template were inspected. Conditional references were read where needed to reconcile a finding; bundled upstream manuals were not rewritten.

  • Global instructions now resolve skill precedence explicitly, preserve existing authorization, scope diagnostic examples, and reuse passing evidence.
  • python-ai removes an environment dump that could disclose API keys and stops requiring speculative feature flags or provider fallbacks.
  • homelab-admin keeps TLS verification enabled and scopes disruptive testing.
  • linux-sysadmin, forgejo-maintainer, and podman-operator distinguish diagnosis from mutation, target affected paths, and preserve deployment modes. Container inspection must select fields rather than expose credentials.
  • quickshell and its build reference preserve required features when build dependencies are missing.
  • ai-project-manager scales planning to the request and cannot mark skipped required checks complete. Its template reuses prior authorization.
  • pr-readiness distinguishes local validation from remote merge gates and accepts existing independent review evidence when repository policy permits.
  • bash-scripting, hugo, mdbook, rust-cli, wayfinder, and youtube-thumbnail retain their domain guidance. Planning-only boundaries, shell correctness, publication examples, and photographed-person preservation remain intact.

OpenAI's Astra instruction guidance calls for auditing conflicting skills. Its verification guidance favors relevant checks over repeated unchanged tests. The changes apply those principles while retaining required gates.

Global scope and security

During the September audit, standalone CLI 0.149.0 was rejected by the server for Astra requests. The then-installed app bundled CLI 0.153.0-alpha.5, which successfully started Astra review. The local ~/.local/bin/codex launcher was linked to /usr/lib/chatgpt/resources/codex; its former symlink is backed up under ~/.codex/backups/astra-cli-20260908-190153/. This machine-specific repair is not part of the portable installer. If the app is removed, install a current standalone CLI before replacing the launcher; restoring the old launcher also restores its Astra incompatibility.

The installer links global AGENTS.md and managed skills into this repository, so edits to linked files apply globally. The live config is a regular file; update only the intended model settings when applying this audit to an existing installation. The installer now initializes defaults and trust entries only when no config exists; subsequent installs preserve the installed config. Adopt later defaults by editing that file explicitly.

This audit preserves the existing approval_policy = "never" and sandbox_mode = "danger-full-access", trust entries, command rules, disabled memory features, and machine-local integrations. These are existing full-access preferences, not a sandboxed security baseline. Written authorization rules do not create OS isolation. No permission or security control was relaxed for cost.

The October follow-up brings the previously unmanaged autofix, code-review, and find-skills into the portable skill collection. Preserve existing local copies in backups before replacing them with managed links. Plugin-owned caches remain untouched; future upstream updates need re-auditing.

Credentials, session data, runtime caches, and plugin settings are not copied into this repository. Required verification remains ./scripts/validate.sh, skill frontmatter validation, and codex review --uncommitted after local gates.