Skip to content

feat: gpt-6-astra flagship, codex exec guards, 0.153 epoch (0.0.55) - #55

Merged
okisdev merged 6 commits into
mainfrom
feat/gpt-6-astra-epoch
Sep 5, 2026
Merged

feat: gpt-6-astra flagship, codex exec guards, 0.153 epoch (0.0.55)#55
okisdev merged 6 commits into
mainfrom
feat/gpt-6-astra-epoch

Conversation

@okisdev

@okisdev okisdev commented Sep 5, 2026

Copy link
Copy Markdown
Owner

summary

  • gpt-6-astra becomes the codex flagship: the capability table row (5/4/2, measured shape in its notes), the spec grade routing row with an xhigh effort requirement, the gpt-5.6-sol row removed under the single flagship precedent, the lane reminder, and the doctor checks for a codex default without a table row or a persistent reasoning effort. the dated codex listing enumeration in the rules becomes a rule so it stops rotting with every listing change, and gpt-reserve is documented as the Luna Reserve fallback.
  • the codex companion guards exec runs for the new model: every invocation also disables sleep_tool (Astra advertises clock, so codex would register a sleep tool that spends the flight budget) and memories (privacy parity with the grok lane); the persistent reasoning effort is rejected when explicit or inherited through a resume, and a run whose rollout resolved to persistent terminalizes as policy through the generalized policyViolation; the foreground write warning retargets from gpt-5.6-sol to gpt-6-astra; a job whose resolved model differs from its explicit --model records modelDrift and renders a warning, which makes a Luna Reserve handoff visible at collection.
  • the codex tested interval moves to [0.152.0, 0.154.0) after live verification on 0.153.2 and 0.153.4; the never populated appliedServiceTier field and its dead session.configured probe are removed; guardian routing and the skills context cost are settled by reading upstream and documented in the contract.
  • the grok runtime socket preflight names the engine behind a symlinked /var/run/docker.sock and prints the exact repair (OrbStack: orb config set setup.use_admin false plus removing the symlink; Docker Desktop: the default socket option) in the setup report, the rendered next step, and the managed run sandbox failure, without downgrading strict.
  • reviewable logic lives in plugins/codex/scripts/codex-companion.mjs, plugins/codex/scripts/lib/codex-exec.mjs, plugins/grok/scripts/lib/grok-exec.mjs, and plugins/fusion/rules/orchestration.md; plugins/fusion/rules-manifest.json is regenerated and the seven version fields plus CHANGELOG carry the 0.0.55 release.

breaking changes

  • codex-cli below 0.152.0 now fails task and review preflight with failureKind: "setup" (the minimum was 0.147.0); the sleep_tool feature the companion disables is absent from older registries.
  • appliedServiceTier disappears from codex job records; records that still carry it are normalized on read. the requested serviceTier stays.
  • delegated codex runs no longer read or write the account's codex memory store and cannot use the sleep tool or the persistent reasoning effort.
  • the routing table loses its gpt-5.6-sol row; explicit --model gpt-5.6-sol still works, but no routing row names it.

test plan

already verified

  • npm test on the merged tree with nothing else running -> 1148 tests, 1147 pass, 0 fail, 1 standing skip
  • claude plugin validate for codex, grok, fusion, and the marketplace -> all pass
  • live probes through the companion on codex-cli 0.153.4: structured output via --output-schema, thread resume by stored thread id, native review, and a task with no effort flag -> all exact; the same set plus a live web consult on gpt-6-astra passed on 0.153.2 earlier the same day
  • two adversarial reviews on gpt-6-astra over the diff (150s and about 2.5 min wall clock); both findings (inherited persistent effort on resume, and a config.toml preflight scan that misread profile overrides) were fixed or withdrawn before this PR
  • /grok:setup on an OrbStack host renders the orb config set setup.use_admin false remedy under Next steps

reviewer should verify

  • after the plugin caches refresh to 0.0.55, /fusion:setup reconciles ~/.claude/rules/orchestration.md to the new prose and keeps the scored table, and /codex:setup reports ready with compatibility tested on codex-cli 0.153.x
  • /codex:task --effort persistent -- anything fails before job creation with failure: input naming the persistent mode
  • /fusion:smoke passes the codex and worker probes; the grok probe fails fast with the engine specific remedy while a Docker engine keeps the socket symlink

notes

  • issue codex 0.153 and gpt-6-astra follow-ups #54 tracked the follow-ups from the first pass of this change and was closed once they landed here.
  • the grok lane stays down on hosts where OrbStack runs with admin privileges until the host is repaired or upstream stops rejecting symlinked endpoints; the host change and the upstream report are deliberately left to the user.
  • release cadence after merge: refresh the installed plugin caches, run /fusion:setup, /fusion:smoke, /fusion:doctor, and /fusion:stats --prune-dead --yes through the raw args transport.

the spec grade routing row names gpt-6-astra with an xhigh effort requirement, the example capability table mirrors the re-scored live table (astra 5/4/2 added, gpt-5.6-sol removed), the dated codex listing enumeration becomes the rule that a listed model without a scored row holds no seat and gpt-reserve is documented as the Luna Reserve fallback, the session lane reminder names astra for spec grade and deep review, and the doctor flags a codex default without a table row or a persistent reasoning effort. rules manifest regenerated.
…rval to 0.152.0 through 0.153.4

every exec invocation also disables sleep_tool and memories; the persistent reasoning effort is rejected before job creation when explicit or inherited through a resume, and a run whose rollout resolved to persistent terminalizes as policy through the generalized policyViolation; the foreground write warning retargets from gpt-5.6-sol to the flagship gpt-6-astra; a job whose resolved model differs from its explicit --model request records modelDrift with requestedModel and resolvedModel and renders a warning; the never populated appliedServiceTier field and its dead session.configured probe are removed; the tested interval becomes [0.152.0, 0.154.0) with fake-codex and fixtures on 0.153.4.
verified versions move to codex-cli 0.153.4 on 2026-09-05 with the README pin, and the codex contract documents the sleep_tool and memories disables, the persistent effort rejection and rollout fail-closed, explicit model drift, the guardian routing that keeps auto review out of companion runs, the skills context cost, and the new tested interval; the rules sync regex follows the interval sentence.
…repair advice

the symlink target identifies OrbStack or Docker Desktop, the setup report carries hostEnvironment.runtimeSocketEngine and hostEnvironment.remedy, and the rendered next step and the managed run sandbox failure both print the engine specific repair (orb config set setup.use_admin false plus removing the symlink for OrbStack, the default socket option for Docker Desktop) without downgrading strict.
the codex companion suite exceeded the 120 second file level budget on the macOS Node 22 runner while every individual test passed; node applies --test-timeout to each test file as a top level test, so the whole file must fit the budget.
@okisdev
okisdev merged commit 6198045 into main Sep 5, 2026
6 checks passed
@okisdev
okisdev deleted the feat/gpt-6-astra-epoch branch September 5, 2026 05:44
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant