feat(pstack): support optional Sonnet Astra Luna and Terra families

[tlmader]

Closes #54

What changed

The model matrix now supports opt-in Sonnet, GPT-6 Astra, GPT-5.6 Luna, and GPT-5.6 Terra families at every selectable effort. The existing Fable, Sol, Grok, and Opus first-run panel stays byte-identical.

Setup derives active families from the final role map. It probes and smokes only that set, rejects mismatches before writing, and keeps unchanged reruns byte-identical. Sonnet uses rolling aliases and five native Claude agent definitions. The existing runner handles all cross-provider routes without fallback.

Astra expands the upstream issue for PAT-3792.

Verification

• Bun tests, strict typecheck, static invariants, and plugin validation pass.
• The exact candidate is installed in every affected harness.
• The changed behavior passes from each real user surface.
• The installed version, action, and observed result appear below.

Live evidence:

• Commit ae37a16 passed 160 Bun tests, strict TypeScript checks, manifest parsing, static invariants, and git diff --check.
• Claude Code 2.1.261 loaded the exact plugin with --plugin-dir. Native pstack-sonnet-high returned PSTACK_NATIVE_SONNET_OK from claude-sonnet-5.
• Codex CLI 0.153.4 installed the exact local marketplace candidate. pstack:setup-pstack returned PSTACK_CODEX_PLUGIN_OK with all four optional descriptors in matrix order.
• The external runner returned PSTACK_MODEL_PROBE_OK for Sonnet, Astra, Luna, and Terra. Sonnet verified the provider-reported claude-sonnet-5; all three Codex receipts preserved the requested model and high effort in exact argv.
• Native Codex probes returned the expected markers for Astra, Luna, and Terra at high effort.
• Independent Fable and Opus reviews returned PASS after the accepted findings were fixed. The no-comments pass reported no remaining refactor flags.

[mergify]

Tick the box to add this pull request to the merge queue (same as @mergifyio queue).

• Queue this pull request

[greptile-apps]

Greptile Summary

This PR expands pstack's model registry with opt-in Sonnet, Astra, Luna, and Terra families while retaining the existing first-run panel.

• Adds five native Sonnet effort profiles and rolling-alias normalization.
• Generalizes setup around the active families derived from the final role map.
• Extends runner and matrix coverage for all newly supported model families.
• Updates package documentation and static panel invariants.
• One setup edge case remains: removing every active family allows configuration writes that cannot pass the subsequent behavioral smoke.

Confidence Score: 4/5

The PR should not merge until setup rejects an empty active-family set or supports completing its behavioral smoke for an alias-only configuration.

Removing all active families can pass validation and write the model sheet, after which the required smoke has no active descriptor or cross-judge to run.

Files Needing Attention: plugins/pstack/skills/setup-pstack/SKILL.md

Important Files Changed

Filename	Overview
plugins/pstack/skills/setup-pstack/SKILL.md	Generalizes setup to configurable active-family sets, but does not prevent an empty set that makes post-write smoke impossible.
plugins/pstack/skills/poteto-mode/references/provider-dispatch.md	Expands the authoritative model matrix and routing contract while preserving the original first-run panel.
plugins/pstack/skills/poteto-mode/scripts/runner/model-aliases.ts	Adds Sonnet to the shared rolling-Claude-alias boundary and concrete revision normalization.
plugins/pstack/skills/poteto-mode/scripts/runner/model-matrix.test.ts	Updates matrix parsing and assertions for eight families, explicit first-run membership, and native profiles.
tests/skill-collision-repro.sh	Derives the default panel from active matrix rows and broadens checks against pinned Claude revisions.
plugins/pstack/agents/pstack-sonnet-high.md	Adds a representative bounded native Sonnet profile; the other four effort variants follow the same contract.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Load and normalize role map] --> B[Derive current active families]
  B --> C[Operator adds or removes families]
  C --> D[Build final role assignments]
  D --> E[Select effort per active family]
  E --> F[Probe active families]
  F --> G[Render and validate family-set equality]
  G --> H[Confirm and write configuration]
  H --> I[Run mixed-panel behavioral smoke]
  C -->|All families removed| Z[Empty active set]
  Z --> F
  Z -->|No descriptor or cross-judge available| I

Loading

_{Reviews (1): Last reviewed commit: "feat(pstack): support optional Sonnet As..." | Re-trigger Greptile}

[ericlitman]

Gavel verdict: Approve

0 blocking findings; 1 unconfirmed plausible finding; 0 killed by adjudication.

Quorum: 4/3 successful reviewers.
Continuity: 0 open · 0 resolved · 0 reopened.
Scope: initial full review.

Top findings

No blocking findings.

Non-blocking findings

• MEDIUM plugins/pstack/skills/setup-pstack/SKILL.md:123 RIGHT Batch native smoke lanes within Codex's child limit — On a four-slot Codex parent with Sol, Astra, Luna, and Terra active, this launches four native spawn_agent smoke lanes before draining any, but the parent leaves only three child slots. The fourth launch is rejected after step 9 has already written the configuration, so setup cannot complete for a newly supported target set. Run native smoke lanes in capacity-bounded waves and retain their outputs for the subsequent judge.
• LOW plugins/pstack/skills/setup-pstack/SKILL.md:46 RIGHT Live evidence does not cover the reviewed head — The reviewed head adds the alias-only/empty-active-set rejection here, but the only commit identified in the recorded verification is ae37a16, while the candidate is 8ef15f8. No installed version or real-surface result demonstrates this new behavior from the current head in either Claude Code or Codex, so the non-draft PR does not satisfy the repository's exact-candidate merge gate. Install 8ef15f8, exercise this rejection from both affected harnesses, and record the installed commit, action, and result.

---

Gavel run panel:8ef15f8b93bb5123b97c547d9238ecda4e44bb28:kvLSXLRl85t:yPAZicdhlC8 · Commit 8ef15f8b93bb5123b97c547d9238ecda4e44bb28
Judge: native (primary) openai/gpt-6-astra · tools: yes
Full record: GET /gavel/run/panel:8ef15f8b93bb5123b97c547d9238ecda4e44bb28:kvLSXLRl85t:yPAZicdhlC8