fix(codex): bound SDD dispatch and final review#2036
Draft
arittr wants to merge 6 commits into
Draft
Conversation
…dation
Experiment branch — NOT for merge without eval evidence (writing-skills
discipline; RED/GREEN campaign to follow if the manual run validates).
Two targeted changes against the measured Codex 5.6-era spinout
(36% of SDD runs >8h, PRI-2672):
1. codex-tools.md — SDD dispatch rules: fork_turns "none" on every
spawn (the default "all" forks the whole transcript and refuses
model/effort overrides — S2's terminal 13-agent final wave was all
full-history forks at sol/xhigh); capability-check for 0.145+ spawn
params with an explicit terra/high role table (re-review terra/medium),
no sol seats, no effort escalation between fix rounds; honest
inheritance warning for <=0.144 where routing is impossible (T0-probed
on 0.144.4: schema is {task_name, message, fork_turns} only, role
TOMLs inert).
Role-table bet: census shows task seats already ran terra/high in the
spinout sessions — the surgical bet is fork_turns:none + no-sol-seats
+ wave termination, not task-seat downgrade. T1 cross-review matrix
will refine reviewer tiers (may support terra/medium or lower).
2. SKILL.md final review — wave closure is policy, not a verdict:
strong reviewers find real defects indefinitely, so "review until
clean" never terminates; new breakage in the final fix diff joins
residual adjudication instead of opening wave two; one-off human
review procedures (competing reviewers, scoring) never become
standing. Two matching rationalization-table rows.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…r PRI-2672 Run-2 dispatch-4 autopsy: after compaction, the codex-tools role table fell out of context and the controller reasoned itself into an inherited sol@max reviewer — the harness's own 'inherited parent model is preferred' plus SKILL.md's 'most capable model' outvoted the compaction summary's distilled 'terra/high' line. Its introspection named the only artifact it actually consults at the moment of dispatch: the review-package output. So the role tuple now rides the script output, reprinted fresh every round, immune to compaction by construction: - task-brief prints role=implementer terra/high fork_turns=none - review-package prints per role via a new leading --role flag (task-review default | re-review | final-review), and SKILL.md call sites pass the flag at the re-review and final-review sites - SKILL.md Model Selection now defers to platform role tables and the printed dispatch hints, closing the 'most capable model' loophole - codex-tools.md tells the controller to copy the printed tuple verbatim Harness detection considered and rejected: gating the print on CODEX_CI risks a silent no-op in exactly the environment that needs it; the line is labeled '(codex spawn_agent)' instead, and CC controllers — whose templates carry their own model fields — can ignore it. Experiment branch for PRI-2672 validation; not for merge without eval evidence. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…e termination Codex SDD runs at frontier tiers spin out: subagents inherit the session's model and effort, review seats find real-but-endless defects every round, and the final whole-branch review has no reachable termination state. Field forensics across multiple real sessions traced the mechanism: spawn_agent's fork_turns defaults to "all" (full-context forks that also refuse model overrides), omitted model/effort params inherit the frontier parent, and dispatch rules loaded at session start do not survive context compaction — a compacted controller reverts to inheritance exactly when the session is longest and most expensive. Three coordinated changes: - codex-tools.md: every SDD spawn sets fork_turns "none" plus explicit model/reasoning_effort when the schema supports them (Codex 0.145+), with an inheritance warning for older builds; reviewer tier never exceeds implementer tier and fix rounds never get effort bumps. - task-brief/review-package print a dispatch tuple with their output (review-package grows --role for re-review/final-review), putting the routing values in front of the controller at the moment of dispatch — reprinted every round, so they survive compaction by construction. - SKILL.md: the final-review wave closing is policy, not a verdict — one fix dispatch, one scoped re-review, residuals adjudicated to the ledger; one-off review procedures never become standing; Model Selection defers to platform role tables where one exists. Validated live: a previously spun-out 14-task run (8h, 6 tasks, died mid-loop) re-executed to completion under these changes — bounded fix loops, single-cycle final review, ~20 consecutive correctly-routed dispatches across two compactions. Reviewer-tier matrix (125 cells) showed no calibration cost: clean-diff approvals and planted-defect recall are identical across tiers. Eval scenarios to follow before any upstream submission. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…o Codex literals Responds to maintainer review of this branch: the dispatch-hint lines were hardcoded Codex strings inside shared skill tooling that every harness runs. The per-role hint lines now live in a platform-owned data file, skills/using-superpowers/references/codex-dispatch.hints, and the task-brief/review-package scripts only relay the current role's line. The relay is suppressed when CLAUDECODE is set (Claude Code's dispatch templates already carry model selection) and on any harness with no hints file; unknown harnesses fail toward printing, because a silent no-op in the environment that needs the hint is the failure mode this mechanism exists to prevent. Printing at the moment of dispatch is load-bearing: skill text loaded at session start does not survive context compaction, but script output reprints every round. codex-tools.md's SDD dispatch section shrinks to the behavioral rules (fork_turns: "none" always; copy the printed hint verbatim; reviewer tier never exceeds implementer tier; no effort bumps; the <=0.144 inheritance fallback) and points at the hints file for values. Tests cover the relay path, the per-role efforts, and the new CLAUDECODE suppression case. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Responds to maintainer review: "re-review" names the action badly — it
reads as "review again," which is the exact failure observed in the
field (fix reviewers re-running full reviews and package-wide suites
against explicit scope instructions). "Fix review" names the artifact
under review — the fix diff since the previous review — and makes the
scope self-enforcing.
Pure vocabulary substitution: re-review-prompt.md becomes
fix-review-prompt.md, SKILL.md and the review-package --role value
follow, and the term is introduced once as "a scoped fix review — a
review of the fix diff, not a fresh review." No behavioral rules
changed. Historical records (docs/superpowers/plans, specs,
RELEASE-NOTES) keep the old vocabulary; other skills' generic verb
usage ("no need to re-review") is untouched.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Responds to maintainer review asking for better implementer self-reflection. The report format already demands command+output for TDD evidence and fix-round covering tests, but the full-suite claim asked only for prose — and that is the claim that rotted in the field: three consecutive fix rounds shipped a "full suite passes" claim the fix reviewer found unreproducible, each time because the suite had run before later edits made the result stale. Two changes, both at the moment the report is written: every claimed gate needs its exact command and the tail of fresh output — fresh meaning after the final edit, otherwise rerun or report the gate as unverified — and a missing pasted output is itself a defect for the reviewer to flag, which puts enforcement at the consumption side the same way the dispatch hints do. Fix reports claiming a full-suite pass need a fresh run, not the pre-findings one. This is verification-before-completion's gate function relocated into the one prompt subagents actually receive. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
5 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Who is submitting this PR? (required)
claude-fable-5). Stack preparation/submission: OpenAIgpt-5.6-solat max effort.The commits identify Claude Fable 5 as co-author. Codex performed the current-
devrebase, deterministic verification, duplicate search, stack split, and PR submission.This is PR 1 of 2. The Codex compaction-hook work follows in #2035.
What problem are you trying to solve?
Drew reported a real Codex/Superpowers reliability failure: SDD implementations that normally finished in 1–2 hours in Claude Code were taking 8+ hours in Codex, often without finishing. Agents looped through review and minor test fixes, expanded beyond the approved plan, or turned a bounded final review into repeated competing review waves.
The deduplicated session corpus showed 8/22 GPT-5.6-era runs spinning out versus 2/36 GPT-5.5-era runs. One 14-task run spent eight hours completing only six tasks; another reached a terminal final-review loop that continued for more than three hours without a reachable stop state.
This layer addresses three measured causes:
spawn_agentdefaults to full-history forks, and full-history forks refuse per-call model/effort overrides.The post-compaction bootstrap loss that can reintroduce this drift is handled separately in the stacked follow-up, #2035.
What does this PR change?
This PR pins Codex SDD dispatches to
fork_turns: "none"with platform-owned per-role model/effort hints, bounds final review to one fix dispatch plus one scoped fix review, and renames “re-review” to “fix review” so the target is the fix diff. It also requires exact commands and fresh output tails for implementer gate claims, with deterministic coverage for hint routing, role validation, Claude Code suppression, and the bounded review workflow.Is this change appropriate for the core library?
Yes. Codex is an existing first-class Superpowers harness, and SDD is a general-purpose core workflow. The failure occurs across projects and customers; the changes are not Prime Radiant-specific, add no dependency, and keep Codex-specific routing values in the existing platform reference rather than hard-coding them into shared scripts.
What alternatives did you consider?
task-briefandreview-packageimmediately before dispatch; those are the durable chokepoints for the routing tuple.Does this PR contain multiple unrelated changes?
No. Dispatch routing, bounded final review, the scoped fix-review contract, and evidence-locked completion claims all address the same SDD spinout loop. The separable post-compaction hook and packaging work is intentionally isolated in stacked PR #2035.
Existing PRs
No duplicate PR was found for Codex SDD spinout,
fork_turnspinning, role-bound review dispatch, or bounded final-review termination.Environment tested
gpt-5.6-sol;gpt-5.6-terrasuperpowers-evalsRED/GREEN campaigncodex-spinout-scenarios@3c0a2f7, Codex 0.145.0gpt-5.6-sol;gpt-5.6-terrasuperpowers-evalsregression guards3c0a2f7claude-opus-4-8tests/claude-code/test-sdd-workspace.shNew harness support (required if this PR adds a new harness)
Not applicable. Codex is already supported; this changes SDD routing and review-loop behavior in that existing harness.
Clean-session transcript for "Let's make a react todo list"
Evaluation
Recorded campaign results in
prime-radiant-inc/superpowers-evalsbranchcodex-spinout-scenarios:sdd-codex-dispatch-pinning: RED was bimodal/unpinned; GREEN 3/3 gauntlet passes.sdd-codex-no-tier-escalation: RED 3/3 inherited or escalated; GREEN 3/3 passes.sdd-final-review-contaminated-resume: RED met the planned probabilistic threshold; GREEN 3/3 passes.sdd-implementer-evidence-locked-report: improved from 1/3 to 2/3 judge passes, so this remains an explicit partial GREEN, not a claim of full binding.sdd-round4-escalates-modelwas 2/2 clean. No regression-guard failure was attributable to this branch; unrelated anomalies and infrastructure/judge failures remain documented in the campaign record.No new behavioral eval was run merely to split the already-reviewed branch. Fresh split verification was deterministic:
git diff --check origin/dev...codex/sdd-review-loop-fix bash tests/claude-code/test-sdd-workspace.sh # 19/19 checks passedRigor
superpowers:writing-skillsand completed adversarial pressure testing (results above)The pressure cases combine frontier-parent inheritance, round-four escalation pressure, contaminated final-review history, stale evidence claims, and a Claude Code regression column. The branch adds final-wave rationalization counters because those exact rationalizations appeared in the failed field sessions; it does not broadly rewrite the skill's voice or philosophy.
Human review
Drew reviewed the complete combined GitHub compare, explicitly approved submission on 2026-07-24, and then requested this exact two-PR split. This lower PR is an unchanged subset of that reviewed diff.