fix: time constraints cap scale and bind the whole /plan chain (#886)

[time-attack]

Problem

A /plan run on a hackathon-demo ask (SwipeMusic) burned 18 AskUserQuestion rounds / 28 questions despite the user stating a tight time budget:

• The time box was asked in round 2 and never consumed by anything downstream.
• Scale budgets from overlay GENERAL: Reduce Token Overhead garrytan/gstack#886 are keyed to /plan's Scale: header, which chained eng/design review invocations never see — 13 of the 18 rounds came from that chain running at full pressure.
• review, plan-design-review, and design-review were not GENERAL: Reduce Token Overhead garrytan/gstack#886 targets at all.
• Reviews spent rounds on ceremony: re-confirming a target the handoff already named (twice), one round per finding (6 consecutive single-question rounds in the design review), and a round offering to open blog links mid-timebox.
• The max-vector scale classifier ranks a room demo as audience=public, pushing a throwaway hackathon build toward product-grade treatment.

Evidence: ~/.claude/projects/-Users-sinabina-orca-projects-SwipeMusic/1b922b73-*.jsonl

Fix

• Time-constraint ceiling: an explicit user time constraint caps scale at session (one sitting) or hobby (day or two) regardless of higher vectors; printed with the scale.
• Scale binds the chain: every handoff carries scale + time box; the receiving specialist inherits them without re-asking. Overlay GENERAL: Reduce Token Overhead garrytan/gstack#886 now also targets review, plan-design-review, design-review.
• No ceremony rounds: unambiguous handoff targets are printed, not re-confirmed; obvious in-boundary fixes are applied and reported in a compact list; real decisions are batched (≤4/round); optional extras never get their own round at session/hobby.
• Shared judgment contract (all six dispatchers): infer-don't-ask + time-budget clauses.

Mechanics

• EXPECTED_PARITY_CHECKS 5027 → 5033 (+2 per newly targeted module); GENERAL: Reduce Token Overhead garrytan/gstack#886 regression fixture/evaluator extended to pin the ceiling, no-target-confirmation, and apply-not-ask behaviors.
• Regen commit included (check:gstack2-generated gate).

Testing

• bun run test:gstack2 — parity green at 5033 checks.
• bun test — free suite green (exit 0).
• bun run eval:select — 0/178 E2E, 0/28 LLM-judge tests affected by this diff.

🤖 Generated with Claude Code

---

Summary by cubic

Caps build scale by user-stated time boxes and binds scale/time across the entire /plan chain, including review modules, to cut wasted rounds. Extends garrytan#886 proportional planning so chained specialists inherit scale and act proportionally; parity now at 5,080 checks.

• Bug Fixes
  • Bind scale and time box in every handoff; downstream specialists inherit without re-asking.
  • Extend GENERAL: Reduce Token Overhead garrytan/gstack#886 to review, plan-design-review, and design-review.
  • Time ceiling: user-stated limits cap scale at session/hobby; print the constraint with the scale.
  • Remove ceremony rounds: don’t reconfirm named targets, batch decisions, apply obvious in-boundary fixes, and skip optional extras at session/hobby.
  • Shared judgment: add infer-don’t-ask and time-budget clauses across dispatchers.

^{Written for commit 420793d. Summary will update on new commits.}

[cursor]

Bugbot is not enabled for your account, so this pull request was not reviewed.

Enable Bugbot in the Cursor dashboard to get automatic reviews on future PRs.

[time-attack]

CI failure attribution (verified, not assumed):

The three failing checks are pre-existing repo-wide breakage, not this branch:

• windows-free-tests: same six gstack-paths GSTACK_HOME-resolution tests fail identically on every branch for 2+ days, including time-attack/planfixe (merged as fix: /plan empty-target fast path — stop burning ~5k tokens in empty repos #22) — compare this branch vs planfixe.
• GStack 2 Gate → dev-container + Native windows-latest: failing on every branch since 2026-07-22 (last success 2026-07-17), same two-job signature on merged branches planfixe/questions. Root cause of dev-container: the generated-helper-closure test runs gstack-gbrain-sync.ts --help through Node 20, which cannot execute .ts (ERR_UNKNOWN_FILE_EXTENSION). The same test passes locally on this branch (bun test test/gstack2-runtime-install.test.ts -t 'generated helper closure' → 1 pass).

This diff is judgment-text only (scripts/gstack2/ + regenerated markdown); it touches no path resolution or runtime helper code. The gate jobs that exercise this change (installer-matrix, Installed six-skill first use, Native macos/ubuntu, quality) are all green. The Windows/devcontainer breakage deserves its own fix branch.

[cubic-dev-ai]

5 issues found across 36 files

Prompt for AI agents (unresolved issues)

Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="scripts/gstack2/bug-fix-overlays.ts">

<violation number="1" location="scripts/gstack2/bug-fix-overlays.ts:355">
P2: Direct review/design runs may infer a capped scale without printing the scale or time box, so users cannot verify the constraint being enforced. The shared overlay should explicitly print both after resolving them, including for non-`/plan` entry points.</violation>

<violation number="2" location="scripts/gstack2/bug-fix-overlays.ts:677">
P2: An empty or whitespace-only handoff target is treated as unambiguous and skips the target gate. Checking for a non-empty string preserves the no-ceremony behavior only when a target was actually supplied.</violation>

<violation number="3" location="scripts/gstack2/bug-fix-overlays.ts:678">
P2: Session/hobby regressions claim minor findings were applied even when `minor_findings` is zero or absent. Include the finding count in this result so the executable contract describes an action that occurred.</violation>

<violation number="4" location="scripts/gstack2/bug-fix-overlays.ts:678">
P2: The regression marks all findings as safe to apply solely because the scale is session/hobby. That does not test the overlay's required obvious-fix and authorized-mutation conditions, and can hide a workflow that applies unsafe findings without asking. The fixture/evaluator should model those conditions explicitly and only expect automatic application when both are true.</violation>
</file>

<file name="scripts/gstack2/generate-skill-tree.ts">

<violation number="1" location="scripts/gstack2/generate-skill-tree.ts:483">
P2: At product and venture scale, the shared contract now suppresses optional-extra question rounds even though the proportional-planning overlay only removes those rounds at session/hobby and explicitly preserves full question pressure at larger scales. The global rule should keep the no-ceremony restriction scoped to session/hobby (or otherwise defer to the specialist workflow at larger scales).</violation>
</file>

<sub>Tip: instead of fixing issues one by one fix them all with cubic

Re-trigger cubic</sub>

[cubic-dev-ai]

P2: Session/hobby regressions claim minor findings were applied even when minor_findings is zero or absent. Include the finding count in this result so the executable contract describes an action that occurred.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/gstack2/bug-fix-overlays.ts, line 678:

<comment>Session/hobby regressions claim minor findings were applied even when `minor_findings` is zero or absent. Include the finding count in this result so the executable contract describes an action that occurred.</comment>

<file context>
@@ -647,20 +655,27 @@ export function evaluateBugFixRegression(pr: number, rawInput: unknown): Record<
         review_iterations_max: small ? 1 : 3,
         step_unit: ['hours', 'days', 'weeks', 'stage-appropriate', 'stage-appropriate'][rank],
+        target_confirmation_round: input.handoff_target == null,
+        minor_findings_applied_not_asked: small,
       };
     }
</file context>

Suggested change

	minor_findings_applied_not_asked: small,
	minor_findings_applied_not_asked: small && Number(input.minor_findings ?? 0) > 0,

[cubic-dev-ai]

P2: An empty or whitespace-only handoff target is treated as unambiguous and skips the target gate. Checking for a non-empty string preserves the no-ceremony behavior only when a target was actually supplied.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/gstack2/bug-fix-overlays.ts, line 677:

<comment>An empty or whitespace-only handoff target is treated as unambiguous and skips the target gate. Checking for a non-empty string preserves the no-ceremony behavior only when a target was actually supplied.</comment>

<file context>
@@ -647,20 +655,27 @@ export function evaluateBugFixRegression(pr: number, rawInput: unknown): Record<
         outside_voices: small ? 'on-request' : 'offer',
         review_iterations_max: small ? 1 : 3,
         step_unit: ['hours', 'days', 'weeks', 'stage-appropriate', 'stage-appropriate'][rank],
+        target_confirmation_round: input.handoff_target == null,
+        minor_findings_applied_not_asked: small,
       };
</file context>

Suggested change

	target_confirmation_round: input.handoff_target == null,
	target_confirmation_round: typeof input.handoff_target !== 'string' || input.handoff_target.trim().length === 0,

[cubic-dev-ai]

P2: Direct review/design runs may infer a capped scale without printing the scale or time box, so users cannot verify the constraint being enforced. The shared overlay should explicitly print both after resolving them, including for non-/plan entry points.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/gstack2/bug-fix-overlays.ts, line 355:

<comment>Direct review/design runs may infer a capped scale without printing the scale or time box, so users cannot verify the constraint being enforced. The shared overlay should explicitly print both after resolving them, including for non-`/plan` entry points.</comment>

<file context>
@@ -346,21 +346,29 @@ At the Phase 5 approval gate, print the full design-doc body as direct assistant
-The /plan dispatcher prints a \`Scale:\` header line classified from fifteen build-scale vectors (its Build scale section). This rule authorizes every planning specialist to size its machinery to that scale while keeping every STOP gate and approval boundary; a polite user answering every question is not evidence the full machinery is wanted, and the user can always ask for the complete treatment.
+The /plan dispatcher prints a \`Scale:\` header line classified from fifteen build-scale vectors (its Build scale section). This rule authorizes every planning and review specialist to size its machinery to that scale while keeping every STOP gate and approval boundary; a polite user answering every question is not evidence the full machinery is wanted, and the user can always ask for the complete treatment.
+
+The binding scale comes from the first available source: the printed \`Scale:\` header, a chain handoff that names a scale or time box, or on-the-spot classification from the prompt and cheap repository evidence. An explicit user time constraint is a ceiling, not one vector among fifteen: work that must fit one sitting caps the scale at \`session\`, and a day-or-two deadline caps it at \`hobby\`, regardless of higher vectors. A chained invocation inherits the upstream scale and time box without re-asking, and every handoff it emits carries them forward.
 
 - \`session\` and \`hobby\`: batch every question the initial prompt left unanswered into one AskUserQuestion round (two rounds for hobby); skip web or landscape research, outside voices, second opinions, and visual sketches unless the user asks (privacy gates are unchanged whenever they run); cap any adversarial or spec review loop at one iteration; keep the decision artifact near one page with next steps sized in hours (session) or days (hobby), never a phased multi-week roadmap or a distribution plan the user did not ask for.
</file context>

Suggested change

The binding scale comes from the first available source: the printed \`Scale:\` header, a chain handoff that names a scale or time box, or on-the-spot classification from the prompt and cheap repository evidence. An explicit user time constraint is a ceiling, not one vector among fifteen: work that must fit one sitting caps the scale at \`session\`, and a day-or-two deadline caps it at \`hobby\`, regardless of higher vectors. A chained invocation inherits the upstream scale and time box without re-asking, and every handoff it emits carries them forward.

The binding scale comes from the first available source: the printed \`Scale:\` header, a chain handoff that names a scale or time box, or on-the-spot classification from the prompt and cheap repository evidence. An explicit user time constraint is a ceiling, not one vector among fifteen: work that must fit one sitting caps the scale at \`session\`, and a day-or-two deadline caps it at \`hobby\`, regardless of higher vectors. A chained invocation inherits the upstream scale and time box without re-asking, and every handoff it emits carries them forward. After resolving them, print the binding \`Scale:\` and \`Time box:\` before proceeding, including when classification happened on the spot.

[cubic-dev-ai]

P2: At product and venture scale, the shared contract now suppresses optional-extra question rounds even though the proportional-planning overlay only removes those rounds at session/hobby and explicitly preserves full question pressure at larger scales. The global rule should keep the no-ceremony restriction scoped to session/hobby (or otherwise defer to the specialist workflow at larger scales).

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/gstack2/generate-skill-tree.ts, line 483:

<comment>At product and venture scale, the shared contract now suppresses optional-extra question rounds even though the proportional-planning overlay only removes those rounds at session/hobby and explicitly preserves full question pressure at larger scales. The global rule should keep the no-ceremony restriction scoped to session/hobby (or otherwise defer to the specialist workflow at larger scales).</comment>

<file context>
@@ -476,7 +480,9 @@ function sharedJudgmentContract(): string {
     '8. Match the user language. Empty or contradictory evidence blocks confident success.',
     '9. Recommendations remain traceable downstream, including what evidence would change them.',
-    '10. The user makes the final decision.',
+    '10. Ask only what cannot be inferred. Question rounds are for decisions that are consequential and still open after the prompt, the repository, and platform convention are consulted; infer the rest, state each inferred default in one line, and batch what remains. Never spend a round confirming what a handoff already names or offering optional extras.',
+    '11. A user-stated time constraint binds every phase and every chained skill. Skip or compress optional phases that do not fit it, noting each skip in one line.',
+    '12. The user makes the final decision.',
</file context>

[cubic-dev-ai]

P2: The regression marks all findings as safe to apply solely because the scale is session/hobby. That does not test the overlay's required obvious-fix and authorized-mutation conditions, and can hide a workflow that applies unsafe findings without asking. The fixture/evaluator should model those conditions explicitly and only expect automatic application when both are true.

Prompt for AI agents

Check if this issue is valid — if so, understand the root cause and fix it. At scripts/gstack2/bug-fix-overlays.ts, line 678:

<comment>The regression marks all findings as safe to apply solely because the scale is session/hobby. That does not test the overlay's required obvious-fix and authorized-mutation conditions, and can hide a workflow that applies unsafe findings without asking. The fixture/evaluator should model those conditions explicitly and only expect automatic application when both are true.</comment>

<file context>
@@ -647,20 +655,27 @@ export function evaluateBugFixRegression(pr: number, rawInput: unknown): Record<
         review_iterations_max: small ? 1 : 3,
         step_unit: ['hours', 'days', 'weeks', 'stage-appropriate', 'stage-appropriate'][rank],
+        target_confirmation_round: input.handoff_target == null,
+        minor_findings_applied_not_asked: small,
       };
     }
</file context>