docs: codex-efficiency fix-cycle spec and plan (campaign record)

docs/superpowers/specs/2026-07-30-codex-efficiency-fixes-design.md

@@ -0,0 +1,252 @@
+# Codex Efficiency Fixes — Design
+
+Date: 2026-07-30
+Status: approved by Jesse (in-session)
+Branch: `codex-efficiency-fixes` off `dev`
+
+## Sources
+
+- Eval campaign closeout: `superpowers-autoresearch/reports/2026-07-codex-efficiency-campaign.md`
+  (treatment table §4; every treatment below has a scorer and a measured
+  `dev` baseline).
+- Codex source recon: `superpowers-autoresearch/docs/2026-07-29-codex-multiagent-v2-capabilities.md`
+  (file:line citations against the Codex CLI source; grounds T2, T3, T5).
+- Published experiment write-ups: `superpowers-evals/docs/experiments/`.
+- Drew's spinout stack (PRs #2036, #2035) is **evidence, not adopted text**:
+  Jesse wants to dig into those fixes in more detail before adopting any
+  of them; they inform the problem statements only.
+
+## Goal
+
+Ship the five evidence-strong treatments from the codex-efficiency eval
+campaign as superpowers skill/doc changes, each graded against its
+pre-registered criterion by the campaign's scorers before its PR is cut.
+Phase 2 (everything else in the closeout treatment table) follows, each
+item gated on new baseline work first.
+
+## Scope decisions (settled with Jesse)
+
+- **Phase 1 = the evidence-strong five** (T1–T5 below). Phase 2 items
+  each need a failing baseline before any fix ships (discrimination
+  rule: inconclusive-by-zero is a stop).
+- **One branch, PR per treatment.** Development and batteries happen on
+  `codex-efficiency-fixes`; when a treatment beats its criterion, it is
+  cut into its own PR against `dev` with its eval evidence. No merge
+  without Jesse's per-PR approval.
+- **T4 ships cross-harness with a global regression battery** (Claude
+  Code, Codex, Gemini), variant C shape: ceremony scales, approval never
+  does.
+
+## The five treatments
+
+### T1. SDD worker-review prohibition
+
+**Evidence:** 9/9 depth-2 spawns across 4 corpora were implementer-issued
+reviewers; all 9 were same-task duplicates of the review the controller
+dispatches anyway. The dispatch contract never says review is not the
+worker's job; "self-review" in the implementer prompt gets reified into a
+reviewer subagent on harnesses where children can spawn (Codex).
+
+**Changes:**
+- `skills/subagent-driven-development/implementer-prompt.md`: an explicit
+  "You do not dispatch subagents" clause — self-review means reading your
+  own diff; the controller owns all review dispatch; a reviewer you spawn
+  duplicates a review the process already provides.
+- `skills/subagent-driven-development/SKILL.md`: one dispatch-contract
+  line in the task loop, plus a Red Flags row: "An independent review
+  would strengthen my report" → review is the controller's next step;
+  your reviewer is a duplicate seat.
+- Harness-agnostic wording (no-op where children cannot spawn).
+
+**Graded by:** `score_e6.py` (depth-2 spawns by spawner role, duplicate
+review families); `score_e5.py` for the same-scope variant.
+**Baseline:** 9/9 worker-issued, 0 counter-examples.
+**Criterion:** 0 worker-issued depth-2 spawns AND review coverage
+preserved (every task still gets exactly one controller-dispatched task
+review).
+
+### T2. Event-driven waiting
+
+**Evidence:** 60–78% of `wait_agent` calls time out in every corpus
+(dev 67.1%, spinout 60.2%). Source recon: V2 waits are event
+subscriptions, not polls — one long wait has the same wake latency as a
+10s poll at ~1/90th the calls; a completed child's FINAL_ANSWER is pushed
+into the parent's mailbox and drained into the next model request with no
+wait at all.
+
+**Changes** (`skills/using-superpowers/references/codex-tools.md`):
+- Never short-timeout poll.
+- While local work remains, do not wait — child results arrive with your
+  next turn via the mailbox.
+- When genuinely idle, issue ONE `wait_agent` with a long `timeout_ms`
+  (900000+; harness max 3600000).
+- V2 caveat stated: completion mail carries `trigger_turn=false` and will
+  not wake an idle controller — that is the one job `wait_agent` has.
+
+**Graded by:** `score_e7.py` (timeout rate, inter-poll cadence,
+cache-rebill estimate — the rebill figure stays labeled as an estimate).
+**Baseline:** dev 67.1% timeout rate.
+**Criterion:** timeout rate < 25% with no loss of task completion.
+
+### T3. codex-tools.md corrections
+
+**Evidence:** five claims in the current guidance are contradicted by the
+Codex source (all file:line-cited in the capabilities doc):
+1. `close_agent` does not exist in multi-agent V2 (V1-only). V2 LRU-evicts
+   finished children automatically; not closing costs nothing;
+   `followup_task` transparently reloads an evicted child.
+2. Fix rounds can always resume the implementer via `followup_task` —
+   dev's "if your harness cannot send another message to a spawned agent,
+   dispatch each fix round as a fresh implementer" branch is dead on V2.
+3. Role files (`~/.codex/agents/**.toml`) DO attach to spawns via
+   `agent_type` on isolated forks (0.145+).
+4. Full-history forks accept `model`/`reasoning_effort` overrides; only
+   `agent_type` is refused. (Isolated forks remain the SDD guidance for
+   context-hygiene reasons, stated accurately.)
+5. Dispatch guidance must never name non-V2 model presets — the V2 spawn
+   allowlist is v2 presets only; others hard-error.
+
+**Changes:** rewrite the multi-agent paragraph of
+`skills/using-superpowers/references/codex-tools.md` to be
+version-honest (V1 vs V2 behavior labeled where they differ).
+
+**Graded by:** source citation (already verified); no scorer regressions
+on the shared battery. `score_e8.py` is retained as a V1/V2 schema
+detector, not a hygiene grader — no `close_agent` checklist ships.
+
+### T4. Brainstorming three-path router (variant C: approval always)
+
+**Evidence:** micro — the current HARD-GATE text pushes a bounded task to
+FULL ceremony 5/5, while Z-null (no guidance) and a three-path router
+both differentiate 5/5: the absolute wording suppresses discrimination
+the model draws natively. FULL battery — ceremony volume scales
+moderately (16.7 vs 24.0 tool calls, bounded vs arch), but the
+two-document ritual (spec file → plan file) ran unconditionally in every
+rep. The measured waste is the unconditional artifact ritual, not the
+approval gate.
+
+**Design (variant C):** three paths scale the ARTIFACT; every path keeps
+human approval before implementation:
+- **Spike** (feasibility question, explicitly throwaway): present the
+  question and the intended probe in 2–3 sentences, get a nod, go. No
+  docs. Findings return as a recommendation; anything built stays labeled
+  throwaway.
+- **Bounded** (well-scoped change to an existing, understood flow):
+  present a short design in chat, get approval, implement. No spec file,
+  no writing-plans invocation.
+- **Architectural** (restructures components, new subsystem, public
+  interface change): the full current flow — spec doc, review,
+  writing-plans.
+
+**Guards (all ship with the router):**
+- Classification is said out loud ("this looks bounded, so I'll present a
+  short design here rather than write a spec") so the human can override.
+- When in doubt between two paths, take the heavier one.
+- One-way ratchet: hidden complexity discovered mid-path upgrades the
+  path; never downgrade mid-task.
+- New Red Flags rows targeting classification-as-escape-hatch ("I'll call
+  it bounded to skip the doc").
+
+**Changes** (`skills/brainstorming/SKILL.md`): HARD-GATE keeps "no
+implementation before approval" and drops "regardless of perceived
+simplicity" as the ceremony driver; anti-pattern section reframed (the
+sin is skipping approval, not skipping documents); checklist steps 6–9
+become the architectural path; process-flow graph gains the router; Red
+Flags rows added. This is carefully-tuned content — the edit follows
+writing-skills methodology and ships only with the full eval evidence
+below.
+
+**Graded by (three layers):**
+1. **Micro** (`ceremony-path-micro.py`, adapted): variant C literal text,
+   plus adversarially ambiguous briefs the campaign never tested (a task
+   that pattern-matches bounded but hides a public interface change).
+   Criteria: spike/bounded/arch differentiate (≥4/5 per cell); ambiguous
+   briefs escalate to FULL (≥4/5); arch never downgrades (5/5).
+2. **Codex ceremony battery:** `cx-ceremony-{spike,bounded,arch}` on the
+   fix arm, 3 reps each, `score_e4.py` census. Criteria: bounded reps
+   show an approval turn but zero committed spec files and zero
+   writing-plans ritual; arch reps keep the full two-doc flow; spike reps
+   stay minimal.
+3. **Global regression battery:** the same three ceremony scenarios on
+   Claude Code and Gemini (rig work: those scenarios are currently
+   codex-gated), 3 reps each; plus the triggering acceptance check
+   ("Let's make a react todo list" auto-triggers brainstorming into the
+   full/architectural path) on all three harnesses.
+
+### T5. Explicit model on child-issued spawns
+
+**Evidence:** root spawns are 100% explicit-model at CLI 0.146 (dev
+14/14); the live gap is depth-2 — 2/2 child-issued spawns omitted
+`model`. Source recon: `model` without `reasoning_effort` resets effort
+to the MODEL's default, not the parent's.
+
+**Changes** (`skills/using-superpowers/references/codex-tools.md`):
+- Every spawn you issue — including as a child — sets `model` AND
+  `reasoning_effort`; the effort-reset trap is named.
+- Advise `[agents].default_subagent_model` and
+  `[agents].default_subagent_reasoning_effort` in `~/.codex/config.toml`
+  as the machine-level backstop for anything that slips through.
+
+**Graded by:** `score_e1.py` (per-spawn explicit-model rate, by depth) on
+the shared battery.
+**Baseline:** depth-2: 0/2 explicit.
+**Criterion:** every spawn at every depth carries explicit model +
+effort. Pre-registered caveat: if T1 eliminates depth-2 spawns entirely,
+T5 grades as root-spawn regression (hold 100%) plus doc correctness and
+is recorded inconclusive-by-zero at depth-2 — the config backstop is then
+the operative mechanism.
+
+## Grading plan
+
+- **Shared SDD battery** carries T1, T2, T5: `cx-sdd-small`, fix-branch
+  arm (`/tmp/sp-arm-fix`), 8 reps across both container lanes. Dev
+  baselines are already measured; no baseline re-runs.
+- **T4 batteries** as listed above (micro + codex ceremony + global
+  regression).
+- **Pre-registration:** every battery gets a hypothesis-log entry
+  (prediction, scorer, criterion) in
+  `superpowers-autoresearch/logs/2026-07-30-codex-efficiency-fixes.md`
+  BEFORE it runs. Standing rules carry over: append-only log, manual
+  inspection of scorer matches on fix-arm runs (non-circular
+  verification), no raw rollouts committed, correctness rides beside
+  cost in every verdict.
+- **Attribution:** orthogonal scorers on one combined branch; unexpected
+  regressions bisect by treatment commit.
+- **Budget:** shared battery ~$40, codex ceremony ~$40, global
+  regression ~$40–80, micros ~$5 → phase 1 ≈ $150–200 of the ~$850
+  remaining from the campaign's $1000.
+
+## Process
+
+- Work happens in the `codex-efficiency-fixes` worktree (branched off
+  `dev`); execution via subagent-driven-development from a written plan.
+- Skill-text changes follow writing-skills methodology.
+- Scenario/rig changes (un-gating ceremony scenarios for Claude
+  Code/Gemini, adversarial micro briefs) land in `superpowers-evals`
+  main, as authorized.
+- PR-per-treatment against `dev`, each with its eval evidence and the
+  standard identification block; merges only on Jesse's per-PR approval.
+
+## Phase 2 queue (baseline-first; not in this plan's tasks)
+
+Each item requires a failing baseline before any fix ships:
+1. **Dispatch routing / long-session drift** — needs a long-session
+   elicitation rig (fresh sessions don't reproduce the pathology at CLI
+   0.146). Drew's stack informs the treatment shape.
+2. **Verification leases / evidence receipts** — needs the
+   substring-aware duplicate counter added to `score_e3.py` first
+   (current baseline 1/23 exact-string pairs is too weak).
+3. **Remediation cap** — small-n baseline (2/3 reps) needs more reps.
+4. **Cross-task-race probe redesign** — `score_e5.py`'s probe is
+   inconclusive-by-zero by design tradeoff; needs a stronger probe.
+5. **E5 D4 shell-command parser** — fix-review-scope classifier cannot
+   parse compound commands; scorer work, not skill work.
+
+## Out of scope
+
+- Adopting Drew's spinout stack (#2036/#2035) or its text.
+- RoboRev, Codex token telemetry (separate codebases).
+- A `close_agent` hygiene checklist (V2 has no such tool — closed as
+  do-not-ship in the campaign).
+- Claude Code/Gemini-specific efficiency treatments beyond the T4
+  regression battery.

skills/subagent-driven-development/SKILL.md

@@ -14,21 +14,7 @@ Execute plan by dispatching a fresh implementer subagent per task, a task review
 **Narration:** between tool calls, narrate at most one short line — the
 ledger and the tool results carry the record.
 
-**Continuous execution:** Do not pause to check in with your human partner between tasks. Execute all tasks from the plan without stopping. The only reasons to stop are the four named below, or all tasks complete. "Should I continue?" prompts and progress summaries waste their time — they asked you to execute the plan, so execute it.
-
-**Rulings, not stalls.** A running plan does not wait on a human. Conflicts,
-ambiguities, plan defects, a cap you would have asked to exceed — decide
-them. The spec is the binding authority, the plan is its argument, and your
-judgment settles what neither answers. Record every decision in the ledger as
-`Ruling: <what you decided> — <why> — <what it costs if wrong>`, and keep
-going. A wrong ruling costs rework your human partner can see and undo; a
-session parked on a question costs their whole day and buys nothing.
-
-Four things stop you, and only these: an irreversible or destructive
-operation; a security-sensitive action; a side effect outside this worktree
-that norms say you ask about first (a merge, a push to a shared branch, a
-publish); and a plan so broken that every path forward is a guess. For those,
-stop and ask.
+**Continuous execution:** Do not pause to check in with your human partner between tasks. Execute all tasks from the plan without stopping. The only reasons to stop are: BLOCKED status you cannot resolve, ambiguity that genuinely prevents progress, or all tasks complete. "Should I continue?" prompts and progress summaries waste their time — they asked you to execute the plan, so execute it.
 
 ## When to Use
 
@@ -71,14 +57,14 @@ digraph process {
         "Generate review package, dispatch task reviewer (./task-reviewer-prompt.md)" [shape=box];
         "Spec ✅ and quality approved?" [shape=diamond];
         "Finding conflicts with plan text?" [shape=diamond];
-        "Rule on the conflict, ledger the ruling" [shape=box];
+        "Ask human partner which governs" [shape=box];
         "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model" [shape=box];
         "Dispatch scoped re-review (./re-review-prompt.md)" [shape=box];
         "All findings addressed?" [shape=diamond];
         "R = 5?" [shape=diamond];
         "Adjudicate each open finding" [shape=box];
         "Any load-bearing finding?" [shape=diamond];
-        "Rule and continue; stop only if every path forward is a guess" [shape=box];
+        "STOP: report BLOCKED to human partner" [shape=box];
         "Park findings in ledger with rulings" [shape=box];
         "Append completion to ledger, mark todo complete" [shape=box];
     }
@@ -99,8 +85,8 @@ digraph process {
     "Generate review package, dispatch task reviewer (./task-reviewer-prompt.md)" -> "Spec ✅ and quality approved?";
     "Spec ✅ and quality approved?" -> "Append completion to ledger, mark todo complete" [label="yes"];
     "Spec ✅ and quality approved?" -> "Finding conflicts with plan text?" [label="no"];
-    "Finding conflicts with plan text?" -> "Rule on the conflict, ledger the ruling" [label="yes"];
-    "Rule on the conflict, ledger the ruling" -> "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model";
+    "Finding conflicts with plan text?" -> "Ask human partner which governs" [label="yes"];
+    "Ask human partner which governs" -> "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model";
     "Finding conflicts with plan text?" -> "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model" [label="no"];
     "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model" -> "Dispatch scoped re-review (./re-review-prompt.md)";
     "Dispatch scoped re-review (./re-review-prompt.md)" -> "All findings addressed?";
@@ -109,7 +95,7 @@ digraph process {
     "R = 5?" -> "Fix round R of 5: R≤3 resume implementer; R≥4 fresh implementer, more capable model" [label="no - next round"];
     "R = 5?" -> "Adjudicate each open finding" [label="yes - breaker trips"];
     "Adjudicate each open finding" -> "Any load-bearing finding?";
-    "Any load-bearing finding?" -> "Rule and continue; stop only if every path forward is a guess" [label="yes"];
+    "Any load-bearing finding?" -> "STOP: report BLOCKED to human partner" [label="yes"];
     "Any load-bearing finding?" -> "Park findings in ledger with rulings" [label="no"];
     "Park findings in ledger with rulings" -> "Append completion to ledger, mark todo complete";
     "Append completion to ledger, mark todo complete" -> "More tasks remain?";
@@ -162,8 +148,9 @@ Before dispatching Task 1, scan the plan once for conflicts:
 - anything the plan explicitly mandates that the review rubric treats as a
   defect (a test that asserts nothing, verbatim duplication of a logic block)
 
-Rule on everything you find before execution begins — each finding against
-the plan text that mandates it — and record each ruling in the ledger. If the
+Present everything you find to your human partner as one batched question —
+each finding beside the plan text that mandates it, asking which governs —
+before execution begins, not one interrupt per discovery mid-plan. If the
 scan is clean, proceed without comment. The review loop remains the net for
 conflicts that only emerge from implementation.
 
@@ -258,7 +245,7 @@ Implementer subagents report one of four statuses. Handle each appropriately:
 1. If it's a context problem, provide more context and re-dispatch with the same model
 2. If the task requires more reasoning, re-dispatch with a more capable model
 3. If the task is too large, break it into smaller pieces
-4. If the plan itself is wrong, rule on the correction, ledger it, and re-dispatch with the ruling carried in the dispatch
+4. If the plan itself is wrong, escalate to the human
 
 **Never** ignore an escalation or force the same model to retry without changes. If the implementer said it's stuck, something needs to change.
 
@@ -325,11 +312,10 @@ Before the loop starts, two routes leave it immediately:
   before merge. A roll-up nobody reads is a silent discard. Minor findings
   never enter the loop.
 - A finding labeled plan-mandated — or any finding that conflicts with
-  what the plan's text requires — is yours to rule on: weigh the finding
-  against the plan text, decide with the spec as the binding authority, and
-  ledger the ruling before you act on it. Do not dismiss the finding because
-  the plan mandates it, and do not dispatch a fix that contradicts the plan
-  without a recorded ruling.
+  what the plan's text requires — is the human's decision, like any plan
+  contradiction: present the finding and the plan text, ask which governs.
+  Do not dismiss the finding because the plan mandates it, and do not
+  dispatch a fix that contradicts the plan without asking.
 Everything else enters the loop. A fix round is one fix dispatch plus one
 scoped re-review. Five rounds maximum per task:
 
@@ -374,16 +360,15 @@ dispatching. Adjudicate each open finding yourself — you hold the plan and
 the cross-task context the reviewer lacks:
 
 - **The reviewer is wrong, or the point is contestable:** park it —
-  `Task <N>: parked — <finding> — Ruling: <why the code stands>`. The final
+  `Task <N>: parked — <finding> — ruling: <why the code stands>`. The final
   review sees both sides.
 - **Real, but nothing downstream builds on it:** park it the same way, with
   a ruling that says it's real and deferred.
 - **Real and load-bearing** — a later task builds on it, or it reveals a
-  plan defect: rule on the smallest change that unblocks the dependent work,
-  ledger it as `Task <N>: Ruling: <finding> — <what you decided and why>`,
-  and carry it into the next task's dispatch. Parking a structural failure
-  silently lets every dependent task build on it. Stop only when the defect
-  leaves every path forward a guess.
+  plan defect: STOP. Append `Task <N>: BLOCKED — <reason>` and report to
+  your human partner with the finding, the plan text it collides with, and
+  the fix history. Parking a structural failure lets every dependent task
+  build on it and hands the final review a problem it cannot fix either.
 
 Adjudicate only at the cap. Adjudicating earlier to end a loop is
 pre-judging with a different name. Every adjudication is a ledger entry —
@@ -424,22 +409,12 @@ Then run exactly one scoped re-review of the fix wave
 (`scripts/review-package PLAN_FILE FIX_BASE HEAD` over the fix range,
 [re-review-prompt.md](re-review-prompt.md)).
 Adjudicate any residual findings as in the task loop's breaker: park with
-rulings, or rule on the load-bearing ones and ledger what you decided. Only
-the four classes above stop you here. There is no second fix wave —
+rulings, or stop on load-bearing ones. There is no second fix wave —
 residual load-bearing findings surface to your human partner when
 finishing-a-development-branch presents the options.
 
 ## Finish
 
-Before you delete anything, collect every ledger line containing `Ruling:` —
-preflight rulings, parked findings, breaker adjudications, all of them — into
-your final message under "Rulings I made", in the order you made them, each
-with what it costs if wrong. The list is exhaustive: if the ledger holds a
-ruling, the list holds it. That list is the only place the decisions you
-took on your human partner's behalf reach them — they read it and rework
-whatever you got wrong. A ruling that dies with the workspace was a decision
-made in secret.
-
 When the final whole-branch review is clean and its fixes are merged,
 delete this plan's workspace (`rm -rf <workspace>`) — the git history is
 the record now. Sibling directories belong to other plans; leave them