Improve planning alternatives and pre-mortems

[warwick-bit]

Who is submitting this PR? (required)

Field	Value
Your model + version	Codex coding session, GPT-5 family. Nested pressure probes ran via Codex CLI with model gpt-5.5.
Harness + version	Codex CLI 0.142.5 on WSL2/Linux.
All plugins installed	atlassian-rovo 1.0.3; build-web-apps 0.1.2; github 0.1.5 and 0.1.6; google-drive 0.1.7; notion 0.1.5; slack 0.1.2; soph-accuracy 0.2.7; soph-braze 1.0.14; soph-hubspot 1.1.11; soph-investigate 1.2.48; soph-posthog 1.2.43; soph-systems-eng 0.6.11; stripe 1.0.2; superpowers 5.1.3.
Human partner who reviewed this diff	Warwick approved opening the upstream PR after the pushed fork diff link was provided on 2026-07-03.

What problem are you trying to solve?

In live Claude/Codex planning sessions, Warwick found that explicitly asking "what edge cases could this plan hit?" and "what alternative viable solutions are there?" produced better plans. Without a built-in planning gate, agents can still converge on the first plausible implementation, omit a pre-mortem, or comply with pressure such as "skip edge cases, just give tasks."

Observed failure modes during this PR:

• Pressure probe 1 asked for a Stripe/FastAPI idempotent webhook plan while saying to skip alternatives and edge cases. The patched local skill included alternatives and a pre-mortem, but only listed 3 alternatives for non-trivial payment work. That led to the explicit "exactly 5" count rule.
• Pressure probe 2 asked for a payment-provider webhook plan while saying to skip alternatives and edge cases. The agent produced a task-only list without the required plan header, alternatives, or pre-mortem. That led to the "Pressure To Skip Plan Quality Gates" section.

What does this PR change?

This PR makes writing-plans require risk-appropriate alternative viable solutions and a compact pre-mortem before task decomposition. It also updates the plan reviewer plus both execution paths (executing-plans and subagent-driven-development) to reject non-trivial plans that omit these sections or collapse into task-only lists.

Is this change appropriate for the core library?

Yes. This is general planning behavior, not project-specific configuration. It applies across implementation domains because it asks agents to compare plausible approaches and name likely failure modes before they create step-by-step work.

It does not add dependencies, tools, or third-party integrations.

What alternatives did you consider?

• Personal/local config only: useful for Warwick, but it leaves the Superpowers planning workflow with the same omission risk for everyone else.
• Add a new separate planning-review skill: heavier workflow surface and easier for agents to skip before writing a plan.
• Only add pre-mortem prompts: catches edge cases but not first-solution convergence.
• Only add alternatives: catches design convergence but not invariants, adjacent cases, or coverage gaps.
• Put checks only in executing-plans: too late and incomplete because generated plans recommend subagent-driven-development first.

Chosen approach: update writing-plans as the source of truth and add matching checks to the reviewer and execution surfaces.

Does this PR contain multiple unrelated changes?

No. All changes are one behavior-shaping planning improvement across the plan writer, plan reviewer prompt, and execution review gates.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: Add verifying-plan-coverage skill: prove a plan covers 100% of its spec before execution #1704, writing-plans: optional Invariants block for spec→test traceability #1831, Require evidence summary in writing plans #1808, feat(writing-plans): add Evaluator section + self-review check #1627, Improved Plan Review #1644, fix(writing-plans): write plan file incrementally to avoid stall watchdog #1865, feat(executing-plans): optional task graph for durable resume #1862, fix: namespace SDD task reports per-plan so multiple plans don't overwrite task-N-report.md #1858, fix(writing-plans): avoid imposing commit style examples #1813

Searches found no exact duplicate PR for "alternative viable", "pre-mortem", or writing-plan edge-case prompts. Related prior art touches plan coverage, invariants, evaluator/self-review sections, evidence summaries, plan review, execution-task durability, and SDD task reporting. This PR is narrower: compare alternatives before task decomposition, add pre-mortem coverage, and reject task-only outputs that are presented as complete Superpowers plans.

Environment tested

Harness (e.g. Claude Code, Cursor)	Harness version	Model	Model version/ID
Codex CLI	0.142.5	OpenAI Codex	gpt-5.5 for nested pressure probes

New harness support (required if this PR adds a new harness)

Not applicable. This PR does not add a new harness.

Clean-session transcript for "Let's make a react todo list"

Not applicable; no new harness support.

Evaluation

• Initial prompt that led to this change: Warwick asked whether codifying a planning habit of listing possible edge cases and alternative viable solutions would improve thoroughness, then asked to add it to Superpowers planning skills.
• Eval sessions after making the change: 3 pressure probes during the change, plus local static/package checks.
• Formal drill evals: 0. This checkout does not include an evals/ directory/submodule, so the drill harness was not available locally.

Pressure probe results:

1. Installed-cache Codex probe, Stripe/FastAPI idempotent webhook plan, with pressure to skip alternatives/edge cases:
   • Result: partial fail. It included alternatives and a pre-mortem, but only 3 alternatives for non-trivial payment work.
   • Follow-up change: added explicit "exactly 5" count rule and downshift guard.
2. Installed-cache Codex probe, payment-provider/FastAPI webhook plan, with pressure to skip alternatives/edge cases:
   • Result: fail. It produced a task-only list without required plan header, alternatives, or pre-mortem.
   • Follow-up change: added "Pressure To Skip Plan Quality Gates" and task-list/non-plan distinction.
3. Isolated Codex pressure probe using the final wording excerpt:
   • Result: pass for the targeted behavior. Output started with an implementation-plan header, included exactly 5 alternatives, and included a concrete pre-mortem despite the prompt saying to skip alternatives and edge cases.
4. Independent codex exec review --base upstream/dev review:
   • Result: found a P2 gap that the recommended subagent-driven-development path could bypass the new plan quality gates.
   • Follow-up change: added the same plan-contract, alternatives, and pre-mortem scan to SDD pre-flight review.

Local checks:

• git diff --check - passed after the SDD follow-up.
• bash tests/codex-plugin-sync/test-sync-to-codex-plugin.sh - passed.
• bash tests/codex/test-marketplace-manifest.sh - passed.
• bash tests/shell-lint/test-lint-shell.sh - passed.
• TZ=UTC PATH=/tmp/superpowers-local-zip/root/usr/bin:$PATH bash tests/codex/test-package-codex-plugin.sh - fails on the tar.gz half for a pre-existing base-branch issue. A local zip binary was downloaded/extracted without sudo and the zip archive assertions pass under UTC; the tar.gz assertions fail because this environment's GNU tar 1.35 does not support the repo script's --uid/--gid options. The same failure occurs in a clean shallow clone of upstream/dev, and this PR does not modify scripts/package-codex-plugin.sh or tests/codex/test-package-codex-plugin.sh.

Rigor

• If this is a skills change: I used superpowers:writing-skills and completed adversarial pressure testing (paste results below)
• This change was tested adversarially, not just on the happy path
• I did not modify carefully-tuned content (Red Flags table, rationalizations, "human partner" language) without extensive evals showing the change is an improvement

Pressure-test summary is in the Evaluation section above. The tests deliberately used prompts that asked the agent to skip the new quality gates.

Merge Coordination

This PR is not something I can merge. The authenticated account has READ permission on obra/superpowers.

Coordination check:

• Command: python3 /home/warwi/mcp_servers/scripts/pr_coordination_check.py --repo /home/warwi/mcp_servers/worktrees/superpowers-plan-quality --context-path /home/warwi/mcp_servers/worktrees/superpowers-plan-quality --base dev --include-base-advance --fail-on-risk file-overlap
• Last checked: 2026-07-03.
• Result: file-overlap.
• Latest detector action: competing PRs touch the same files, and several are dirty/conflicting; do not try gh pr merge without an upstream merge coordinator.
• Coordination mode: maintainer coordination required.
• Merge owner: upstream maintainers.

Notable overlapping planning/execution PRs include recent dirty/conflicting sibling PRs #1894, #1885, #1884, #1871, #1870, #1869, plus earlier planning/execution overlaps such as #1704, #1831, #1808, #1627, #1644, #1865, #1862, #1858, and #1813. The overlap is expected because this PR intentionally changes skills/writing-plans/SKILL.md, skills/writing-plans/plan-document-reviewer-prompt.md, skills/executing-plans/SKILL.md, and skills/subagent-driven-development/SKILL.md.

Human review

• A human has reviewed the COMPLETE proposed diff before submission