fix(planning): the spec travels with the plan

[obra]

Who is submitting this PR? (required)

Field	Value
Your model + version	Claude Fable 5 (claude-fable-5)
Harness + version	Claude Code 2.1.220
All plugins installed	superpowers 6.2.0, episodic-memory, linear, context7, superpowers-chrome, agent-sdk-dev, code-simplifier, github-triage, plugin-dev
Human partner who reviewed this diff	Jesse Vincent (@obra) — directed the eval program and this submission; reviewing the diff here (opened as DRAFT pending that review)

What problem are you trying to solve?

Plans executed without the spec they were derived from are unauditable for cross-task coherence. In our eval program this failure recurred at every layer we instrumented: whole-branch reviewers detected 0/45 seeded cross-module inconsistencies (praising drift as design), a mechanical consistency scanner's findings were 100% dismissed at adjudication ("the plan explicitly requires both values"), and a preflight value-table with amendment authority enumerated the exact seeded conflicts and ruled them "internally explained." The root cause is structural: each side of a plan-internal contradiction is individually plan-mandated, so once the plan is the only authority in scope, incoherence is unfalsifiable — every instrument reaches the same "intentional per-module policy" conclusion.

What does this PR change?

Two navigational additions (+7/−1 across two files):

• skills/writing-plans/SKILL.md: the plan header template gains a **Spec:** line pointing at the spec/design doc the plan implements — the spec travels with the plan.
• skills/subagent-driven-development/SKILL.md: the setup step reads the plan's named spec ("the spec is the authority the plan argues from; conflicts inside the plan resolve against it"), and a plan with no reachable spec gets a ledger note that rulings are provisional.

Is this change appropriate for the core library?

Yes — core planning/execution plumbing, project-agnostic, no dependencies. The superpowers flow already produces specs (brainstorming → spec → plan); this closes the loop so execution can see them.

What alternatives did you consider?

• Stronger gate text instead of spec plumbing: tested — a preflight value-table with plan-amendment authority (an arm building on fix(sdd): preflight emits its checks as a ledger table and rules on what it surfaces #2080's evidence-bearing preflight) scored 0-1/5 on a seeded-incoherence plan when no spec was in scope. Text cannot substitute for ground truth.
• Mechanical consistency scanning at final review: built and tested — the scanner surfaced 14/14 candidates with file:line and adjudication dismissed 100% under plan authority. Same root cause.
• Anti-rationalization wording alone (naming the dismissal pattern in the skill text): under test in a running battery; interrogation of the dismissing sessions suggests the operative driver is "tests are green so the divergences must be fine," which ground-truth availability addresses directly.

Does this PR contain multiple unrelated changes?

No — one mechanism (the spec travels with the plan), two files that each carry half of it.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related: fix(sdd): rule and continue — non-catastrophic conflicts get ledgered rulings, not blocking questions #2077/fix(sdd): batch small same-shape tasks into one dispatch #2078/fix(sdd): preflight emits its checks as a ledger table and rules on what it surfaces #2080 (same eval program, disjoint sections — no textual overlap with this diff; all apply independently). No open or closed PR carries spec context into plan execution.

Environment tested

Harness (e.g. Claude Code, Cursor)	Harness version	Model	Model version/ID
Claude Code (authoring)	2.1.220	Claude Fable 5	claude-fable-5
Codex CLI (containerized evals, quorum harness)	0.46 lineage	GPT-5.6-Codex	gpt-5.6-codex

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

N/A.

Evaluation

• Method: containerized quorum evals; one seeded-incoherence fixture (a 6-task plan whose task briefs pin 5 deliberately incompatible shared values), run in two scenario variants: specless, and with the product spec the plan derives from present and named in the kickoff message. Pre-registered criteria; mechanical spec-resolution scoring + controller hand-reads; raw records public.
• Results: specless execution: 0-1/5 conflicts resolved correctly, with controllers ruling the divergences "deliberate module-local differences... internally explained" (3/3 reps). Spec present, unmodified stock skills (the exact configuration this PR ships): 4/5, 5/5, 4/5 — detection came free; controllers enumerated all five plan/spec conflicts immediately. Gate-text arms on the same spec'd scenario reached uniform 5/5 and resolved autonomously (no human ask), which is why the SDD half of this change also anchors preflight rulings to the spec.
• Cross-model (2026-08-05): the mechanism replicates on claude-sonnet-5 via the quorum claude adapter, stock skills: spec present 5/5 ×2, specless 0/5 ×2 (hand-verified divergent values in fully built trees). With the gpt-5.6 batteries this covers the two largest user segments; kimi/glm cells deferred with triggers.
• Disclosure/update: the SDD sentence in this diff has now ALSO run verbatim as a composed battery arm (with the reviewer-template paragraph from fix(sdd): reviewers re-read illegible evidence instead of re-running to regenerate it #2089 co-present): 3/3 reps resolved 5/5 seeded plan/spec conflicts (hand-verified), post-checks 19/19, no regression from the stock-with-spec baseline. The specless ablation also completed after this PR was opened: three further specless text arms (including a counter elicited from the dismissing sessions themselves) all scored 0/5 — the spec's presence, not any wording, is the active ingredient.
• Full records (public): https://github.com/prime-radiant-inc/superpowers-autoresearch — logs/2026-08-03-plan-decomposition-campaign.md (P2-on-cp-x10 FAIL verdict, P2' three-arm battery pre-registration + verdict, dismisser-interrogation entry) and the cp-x10-spec scenario + answer key under campaigns/cost-pathologies/scenarios/.

https://claude.ai/code/session_0185AJr98gHx5EmwqNeft4Sy

[obra]

@arittr Ready for your eval treatment — same program as the merged stack and #2089. Evidence summary: the specless wall (0-1/5 seeded plan-incoherence conflicts resolved, four different text arms) vs spec-present 4-5/5 with stock skills, on gpt-5.6 (terra-uniform battery); the exact texts in this diff ran verbatim in a composed battery (5/5 ×3, no regression); and the mechanism replicates on claude-sonnet-5 via the quorum claude adapter (spec 5/5 ×2, specless 0/5 ×2). Scenario fixtures (cp-x10-spec / cp-x10-consistency, seeded-value answer key) are public in prime-radiant-inc/superpowers-autoresearch if you want to run them directly — happy to walk you through the setup.

— Claude (Fable 5), on the eval program with @obra

[arittr]

LGTM

[obra]

Post-merge cross-model validation, for the record: the spec-in-the-room contrast was re-run on two more model families in the same containerized eval harness (superpowers at the merged dev base, full SDD sessions, same seeded 5-defect plan).

• kimi (Kimi Code CLI, oauth lane), 2 reps/condition: spec-present 5/5 + 5/5 seeded inconsistencies resolved to spec values; specless 0/5 + 0/5.
• glm-5.2 (pi harness via OpenRouter), 2 reps/condition: spec-present 5/5 + 5/5; specless 0/5 + 0/5.

That makes four model families with the same shape — gpt-5.6 (0-1/5 → 4-5/5), claude-sonnet-5 (0/5 → 5/5), kimi and glm above. All 8 new reps passed the LLM-verifier gate; the seeded-defect scans are tree-resolving and were hand-checked.

— Claude (Fable 5, Claude Code session; evidence log in superpowers-autoresearch logs/2026-08-05-adjudication-battery.md, 2026-08-07 entry)