Roll out Open Pstack 1.5.0 to Claude Code and Codex on the Mac and studio2

[ericlitman]

Goal

Move both of Eric's hosts (this Mac and studio2) to the released Open Pstack 1.5.0 in Claude Code and Codex. Bring both model sheets in line with it. Eric answers the setup prompts.

Depends on

#88 must be merged, with v1.5.0 tagged and the release published. Tracked in Linear as PSTACK-3.

Installed state

Read it back on each host before step 1: the Claude Code and Codex pstack versions, the Codex marketplace ref, and which sheets exist, with their sha256. The 2026-09-24 version of this section was wrong for pro14. On pro14 on 2026-09-30:

• Claude Code runs pstack@open-pstack 1.2.1 (f466256), last updated from the marketplace on 2026-09-01. The update to 1.5.0 therefore also brings the 1.3.0-1.4.1 changes: forge-neutral shipping, the retired How critics, and the evidence-label rule.
• Legacy plugin. pstack@pstack-claude 0.9.12 is installed and set to false in ~/.claude/settings.json.
• Codex tracks [marketplaces.open-pstack] ref = "main" in ~/.codex/config.toml, with cached plugin 1.4.1.
• Sheets. ~/.claude/pstack-models.md, ~/.codex/pstack-models.md and the bounded <!-- pstack:models:begin --> … <!-- pstack:models:end --> block of ~/.codex/AGENTS.md are byte-identical. The rows use all four families at xhigh (fable, sol, grok:grok-4.6, opus), Sol xhigh on bug-fix/perf-issue/hillclimb, and a retired how critics row. Because a sheet exists in both harnesses, the update keeps Eric's roles. Only setup's role question moves them to the new defaults.

Steps (this Mac first, then studio2)

1. Update Claude Code. Run claude plugin marketplace update open-pstack, then update pstack@open-pstack (with claude plugin update pstack@open-pstack or /plugin). Restart and confirm it reports 1.5.0.
   • On studio2, check that pstack@pstack-claude is false in ~/.claude/settings.json. If it is true, ask Eric before disabling it.
   • If the host has no Claude sheet, write none unless Eric chose to keep Fable. If it has one, copy it to ~/.claude/pstack-models.md.bak-pre-1.5.0 and run /pstack:setup-pstack. Setup lists how critics as dropped and asks for replacements for the grok:grok-4.6 rows (answer Grok family). Eric answers the role question, as in step 5. Confirm, and setup writes the sheet.
2. Back up Codex. Copy ~/.codex/config.toml, ~/.codex/AGENTS.md and, if it exists, ~/.codex/pstack-models.md to *.bak-pre-1.5.0 alongside the originals.
3. Seed the sheet file only if it is missing. An existing ~/.codex/pstack-models.md is setup's source of truth, as on pro14; leave it. If it is missing, write the exact bytes between the begin and end marker lines of ~/.codex/AGENTS.md (markers excluded) to it.
4. Upgrade Codex. In ~/.codex/config.toml, change the ref under [marketplaces.open-pstack] (main on pro14) to ref = "v1.5.0". Run codex plugin marketplace upgrade open-pstack, then codex plugin add pstack@open-pstack, and confirm that diff -rq between ~/.codex/plugins/cache/open-pstack/pstack/1.5.0 and plugins/pstack in a checkout of v1.5.0 prints nothing. (codex plugin marketplace add only adds sources; it does not re-point an existing one.)
5. Rerun setup in Codex. Run pstack:setup-pstack from a Codex session that can reach the Grok CLI (see PR docs(dispatch): note that Grok lanes need an unsandboxed Codex parent #82).
   • It lists how critics as dropped.
   • It reports the grok:grok-4.6 rows for replacement. Answer Grok family, which gives grok:grok-4.7.
   • The role question follows Eric's decision: keep the four-family panel, or change the panels and the judgment/prose, hardest-tasks and How-explainer roles to the new defaults.
   • It asks one effort question per assigned family. The probes pass.
   • Confirm. Setup writes the sheet and the bounded block and runs its smoke.
6. Smoke Claude Code. Run /pstack:interrogate in Claude Code on a one-file scratch diff and note the lanes. Codex needs no extra smoke; setup's own step 9 already ran a mixed panel there.
7. Record the results as one issue comment per host: installed versions, the Claude and Codex sheet diffs against their backups (for Codex, diff <(sed -n '/pstack:models:begin/,/pstack:models:end/p' ~/.codex/AGENTS.md.bak-pre-1.5.0 | sed '1d;$d') ~/.codex/pstack-models.md), and the smoke results.

Rollback

• Codex: restore the .bak-pre-1.5.0 files and delete ~/.codex/pstack-models.md if step 3 created it. A restored ref = "main" would reinstall 1.5.0, so set ref to the tag of the recorded version (v1.4.1 on pro14). Then run codex plugin marketplace upgrade open-pstack and codex plugin add pstack@open-pstack.
• Claude Code: restore ~/.claude/pstack-models.md.bak-pre-1.5.0 if step 1 made one. Then ship a fix release, or load a checkout of the recorded version (v1.2.1 on pro14) with claude --plugin-dir until one ships.

Out of scope

• No repository changes and no new tooling.
• Do not uninstall or disable any plugin without Eric's OK.

Acceptance

On both hosts:

• Claude Code reports pstack@open-pstack 1.5.0, and interrogate runs the trio (or Eric's chosen roles).
• Codex reports 1.5.0. Its sheet and AGENTS.md block match and contain no how critics or grok-4.6 rows. Setup's probes and smoke passed.
• A Claude sheet, where the host has one, contains no how critics or grok-4.6 rows.
• The evidence is recorded on this issue.

[ericlitman]

Edited in place on 2026-09-30. Tracked in Linear as PSTACK-3.

The installed-state section was wrong for pro14. Claude Code runs 1.2.1 (not 1.4.1). Codex tracks ref = "main" with 1.4.1 cached (not v1.2.1). ~/.claude/pstack-models.md and ~/.codex/pstack-models.md both exist and match the AGENTS.md block. The section now starts with a per-host read-back, because studio2 was not checked.

Consequences in the steps:

• Both sheets exist, so the update keeps Eric's roles. Step 1 backs up the Claude sheet and reruns setup in Claude Code as well.
• Step 3 seeds the Codex sheet only when it is missing. It never overwrites an existing sheet.
• Step 4 re-points the recorded ref and checks the installed cache against the v1.5.0 tree with diff -rq.
• Rollback pins Codex to the recorded release tag (v1.4.1 on pro14), because a restored main would reinstall 1.5.0. Claude Code rolls back to the recorded version (v1.2.1 on pro14).

[ericlitman]

pro14 rolled out to Open Pstack 1.5.0 (2026-10-01). Eric answered the role question: new defaults (the Opus, Sol, Grok trio panel; Opus max for judgment, prose, hardest tasks and the How explainer).

Before:

• Claude Code: pstack@open-pstack 1.2.1. The legacy pstack@pstack-claude 0.9.12 is still disabled.
• Codex: ref = "main", 1.4.1 cached.
• Sheets: all three byte-identical (e01bece8…), including the stale how critics row and grok:grok-4.6.

Steps:

1. Claude Code updated from 1.2.1 to 1.5.0 through claude plugin marketplace update and claude plugin update. ~/.claude/pstack-models.md was backed up to .bak-pre-1.5.0.
2. Codex config.toml, AGENTS.md and pstack-models.md were backed up to *.bak-pre-1.5.0.
3. No seeding was needed: ~/.codex/pstack-models.md already existed.
4. The Codex ref changed from main to v1.5.0, followed by the marketplace upgrade and plugin add. 1.5.0 is cached, and diff -rq against git archive v1.5.0 plugins/pstack is clean.
5. Setup ran in both harnesses, Codex with --sandbox danger-full-access. Each dropped how critics as retired and reported the grok:grok-4.6 rows, which were answered with Grok (grok:grok-4.7). The role question applied the new defaults. Efforts were asked for Opus (max), Sol (max) and Grok (xhigh); Fable is unassigned and got no question. The probes passed: Opus claude-opus-5-5, Grok grok-4.7-build and Sol (Codex native, or the runner from Claude Code). Readback was byte-exact, and the mixed smoke and cross-judge passed in both harnesses.

Smoke (step 6): /pstack:interrogate on a one-file scratch diff in a fresh Claude Code 1.5.0 session ran the trio: A was native pstack-opus-max, B the Sol runner (complete, pinned-argv) and C the Grok runner (reportedModel: grok-4.7-build). All three flagged the injected bug.

Result:

• Claude Code reports 1.5.0. Codex reports 1.5.0 (v1.5.0, 1c91e65).
• ~/.codex/pstack-models.md equals the AGENTS.md block (sha256 0445556d…). Everything outside the block is unchanged, and the Claude sheet is byte-identical.
• No sheet has a how critics or grok-4.6 row.

The sheet diff against the backup is the same for Claude and Codex:

5,13c5,12
< feature, refactoring: grok:grok-4.6@xhigh
< bug-fix: codex:gpt-5.6-sol@xhigh
< perf-issue: codex:gpt-5.6-sol@xhigh
< hillclimb: codex:gpt-5.6-sol@xhigh
< judgment and prose: claude:fable@xhigh
< hardest tasks: claude:fable@xhigh
< how explorer: grok:grok-4.6@xhigh
< how explainer: claude:fable@xhigh
< how critics: claude:fable@xhigh, codex:gpt-5.6-sol@xhigh, grok:grok-4.6@xhigh, claude:opus@xhigh
---
> feature, refactoring: grok:grok-4.7@xhigh
> bug-fix: codex:gpt-5.6-sol@max
> perf-issue: codex:gpt-5.6-sol@max
> hillclimb: codex:gpt-5.6-sol@max
> judgment and prose: claude:opus@max
> hardest tasks: claude:opus@max
> how explorer: grok:grok-4.7@xhigh
> how explainer: claude:opus@max
16,20c15,19
< arena runners: claude:fable@xhigh, codex:gpt-5.6-sol@xhigh, grok:grok-4.6@xhigh, claude:opus@xhigh
< arena cross-judge pool: claude:fable@xhigh, codex:gpt-5.6-sol@xhigh, grok:grok-4.6@xhigh, claude:opus@xhigh
< swarm workers: grok:grok-4.6@xhigh
< architect runners: claude:fable@xhigh, codex:gpt-5.6-sol@xhigh, grok:grok-4.6@xhigh, claude:opus@xhigh
< interrogate reviewers: claude:fable@xhigh, codex:gpt-5.6-sol@xhigh, grok:grok-4.6@xhigh, claude:opus@xhigh
---
> arena runners: claude:opus@max, codex:gpt-5.6-sol@max, grok:grok-4.7@xhigh
> arena cross-judge pool: claude:opus@max, codex:gpt-5.6-sol@max, grok:grok-4.7@xhigh
> swarm workers: grok:grok-4.7@xhigh
> architect runners: claude:opus@max, codex:gpt-5.6-sol@max, grok:grok-4.7@xhigh
> interrogate reviewers: claude:opus@max, codex:gpt-5.6-sol@max, grok:grok-4.7@xhigh

The headless Codex setup run also filed a Linear issue on its own (MOB-1465), because of a rule in ~/.codex/AGENTS.md. I cancelled it as a duplicate of PSTACK-3.

[ericlitman]

studio2 rolled out to Open Pstack 1.5.0 (2026-10-01), with the same role answer as pro14 (Eric: "Yes, same answer").

Before:

• Claude Code: pstack@open-pstack 1.4.1. pstack@pstack-claude 0.9.11 is still false.
• Codex 0.159.0: ref = "v1.2.1", 1.2.1 cached.
• Sheets: there was no ~/.claude/pstack-models.md and no ~/.codex/pstack-models.md. The ~/.codex/AGENTS.md block held the e01bece8… sheet, the same one pro14 had.

Steps:

1. Claude Code updated from 1.4.1 to 1.5.0. The host had no Claude sheet, and Eric did not keep Fable, so none was written; the 1.5.0 defaults apply.
2. Codex config.toml and AGENTS.md were backed up to *.bak-pre-1.5.0.
3. ~/.codex/pstack-models.md was seeded from the AGENTS.md block (e01bece8…). This rollout created that file, so a rollback deletes it.
4. The Codex ref changed from v1.2.1 to v1.5.0. My first edit wrote escaped quotes and broke config.toml parsing for about two minutes. After the fix, the only diff from the backup is the ref line. Then the upgrade and add ran: 1.5.0 is cached, the marketplace root is at 1c91e65, and diff -rq against that tree is clean.
5. Codex setup (--sandbox danger-full-access) dropped how critics and replaced grok:grok-4.6 with Grok 4.7. It applied the new-default role changes. Efforts were Opus max, Sol max and Grok xhigh, with no question for Fable. The probes passed (Opus claude-opus-5-5, Grok grok-4.7-build, Sol native), and the readback was byte-exact. All four smoke lanes returned the expected answer, and the Opus cross-judge passed.

Smoke (step 6): /pstack:interrogate in a fresh Claude Code 1.5.0 session ran the default trio: A was native pstack-opus-max, B the Sol runner (complete, pinned-argv) and C the Grok runner (grok-4.7-build, provider-report). There were no dropouts, and all three flagged the injected bug.

Result:

• Claude Code reports 1.5.0, with no sheet; the defaults are the trio.
• Codex reports 1.5.0. Its sheet equals the AGENTS.md block (sha256 0445556d…, identical to pro14), and everything outside the block is unchanged.
• No how critics or grok-4.6 rows remain.

The Codex sheet diff against the backed-up block is identical to pro14's diff above. Rollback ref for this host: v1.2.1.