writing-plans: self-review misses task-state/DAG closure in executable plans

What problem does this solve?

writing-plans currently requires each task to produce a self-contained, independently testable deliverable and gives tasks explicit Consumes / Produces interfaces. It also asks for concrete test code, commands, implementation snippets, and frequent commits.

At current main (b36e0829c6d0140e93cfef2ca599b1b07d4a7797), the final Self-Review checks three things: spec coverage, placeholders, and type/name consistency:

https://github.com/obra/superpowers/blob/b36e0829c6d0140e93cfef2ca599b1b07d4a7797/skills/writing-plans/SKILL.md

What it does not check is whether the repository state at each task boundary can actually satisfy the plan: whether consumers exist before use, deleted inputs have no surviving consumers, promised RED tests are discoverable, workflow/runtime values exist in the phase where they are consumed, and each commit is runnable under its stated verification.

A real planning session produced a plan-only PR where this missing class repeated many times before implementation began:

https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7

The PR changes one Markdown plan file only. Its review history contains 24 top-level Codex findings. Representative task-state/DAG failures:

• Matrix rows before their consumers exist: a workflow task enabled LongGaps/PrimeGaps rows before their replay scripts were created in later tasks. https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4033330349

• Delete input before replacing consumer: producer/zeta23.json was deleted in Task 2 while the legacy workflow that reads it survived until a later task; deleting it would itself trigger that workflow and fail. https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4033399751

• Runtime value consumed before transport makes it visible: the plan wrote SOURCE_ROOT to $GITHUB_ENV and then read it in the same Actions step. https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4028363646

• Promised RED not actually in the test graph: def test_*(self) appeared at module scope, so unittest skipped the intended cases. https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4033112363 https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4033399759

• Workflow replacement not tested while legacy state still existed: the structure assertion was added only after replacement, so it started GREEN instead of proving it detected the old workflow. https://github.com/TeaShaman-cyber/theseus-repo-search-lab/pull/7#discussion_r4033330353

The repo-wide audit grouped these under a recurring invariant: state N must remain valid until state N+1 is independently ready; a task/commit/step may consume only state already produced and verified in an earlier valid phase.

Full public checkpoint: https://github.com/TeaShaman-cyber/theseus-repo-search-lab/issues/10#issuecomment-5709497603

This is consistent with model error being part of the cause, but the recurrence suggests a framework-level review blind spot: the plan format explicitly models task interfaces and independent testability, while the inline self-review never validates those properties.

Suggested change

Not another reviewer subagent. The history in #229 / #1962 / #1130 already documents why the heavier review loop was removed after evals.

A cheap fourth inline self-review lens would be enough to test the hypothesis:

4. Execution-state closure

For every task boundary, simulate the repository state at task start and end:
- every consumed file/interface/runtime value exists before it is used;
- every produced output precedes all consumers;
- deleting or renaming an input leaves no surviving consumer;
- every promised RED command actually discovers the intended test and fails for the intended reason;
- the task's commit is independently runnable under its stated verification;
- cross-step environment/runtime values name the producer, transport, and consumer phase.

This is deliberately narrow. It does not add a new skill, subagent, plan format, or generalized workflow engine.

Why this is distinct from nearby issues

I searched open and closed issues before filing.

• #895 is about how much implementation detail belongs in plans. This issue assumes the current detailed-plan design and asks whether its existing independently testable / Consumes / Produces contract is actually self-reviewed.
• #229 / #1962 / #1511 cover the old plan-document reviewer subagent loop, which was deliberately removed after evals. This proposal does not restore it.
• #1130 proposes an adversarial reviewer. Again, this is only one cheap deterministic review lens.
• #642 discusses loose skill wiring generally. This is a focused observed failure in writing-plans itself with a public transcript/artifact trail.

Duplicate searches for writing-plans with task-boundary / independently-runnable / consumer-producer / DAG-closure terms did not find a focused issue for this class.

Reproduction / environment

• Upstream skill inspected at exact obra/superpowers main commit b36e0829c6d0140e93cfef2ca599b1b07d4a7797 (v6.3.0 release commit).
• Planning/review harness: ChatGPT project runtime using the bundled Superpowers skills, with GitHub Codex review on the resulting public plan PR.
• Exact bundled-plugin package commit is not exposed by this runtime, so no claim is made that it is byte-identical to upstream main; the missing self-review lens described above is independently observable in upstream main at the exact commit linked above.
• Public plan/review corpus: TeaShaman-cyber/theseus-repo-search-lab#7.
• Consolidated causal audit: TeaShaman-cyber/theseus-repo-search-lab#10.

If a before/after eval is required for behavior-shaping skill text, the PR #7 history provides concrete planted/replayed failure shapes rather than a synthetic prompt: missing consumer, premature deletion, same-step env lifetime, undiscovered RED, and post-change-only assertion.