[feature] /goal should enforce HIGH-95% confidence gate (not just document it)

[BaseInfinity]

Feedback Type

Feature request — strengthen existing guidance to actual enforcement

Description

PR #355 (v1.77.0) added skill-text guidance to skills/sdlc/SKILL.md:

Confidence gate — NEVER invoke below HIGH 95%; below that the evaluator rubber-stamps flailing as progress.
DLC binding — condition MUST name the DLC (/sdlc, /gdlc, /ldlc, etc.) so the evaluator anchors on "doing it right."

This is documentation, not enforcement. Nothing prevents (or even warns) when a user types /goal <condition> without first reaching HIGH 95% confidence — exactly the failure mode the guidance is meant to prevent (Haiku evaluator rubber-stamping flailing).

This mirrors the plan-mode auto-approval gate pattern already in /sdlc (skip plan approval ONLY when confidence ≥ 95% AND task is single-file/trivial). We have that gate for plans. We should have an analogous gate for /goal.

The gap

Layer	/effort low	/goal low-confidence
Skill-text guidance	✅	✅ (PR #355)
Hook warning	✅ hooks/model-effort-check.sh (LOUD warning when below xhigh)	❌ none
Programmatic block	❌ (intentional — soft nudge)	❌

model-effort-check.sh is the proven pattern: SessionStart hook reads settings, emits a LOUD warning when effort < recommended. /goal deserves the same treatment as a per-invocation guard.

Proposed scope

1. New hook hooks/goal-confidence-check.sh — UserPromptSubmit matcher on prompts starting with /goal (excluding /goal status and /goal clear).
2. Detection signal: scan the prior assistant turn in transcript for an explicit confidence statement (HIGH (95%, HIGH 95%, Confidence: HIGH, etc. — same patterns the SDLC scoring rubric already greps for).
3. If prior turn has HIGH 95%+ confidence stated → silent, exit 0.
4. If prior turn does NOT → emit LOUD warning block:

   /goal invoked without HIGH-95% confidence statement in prior turn.
   The Haiku evaluator can rubber-stamp flailing as progress.
   Recommendation: cancel with /goal clear, do research first, restate
   confidence as HIGH (≥95%), then re-issue /goal.
   See: skills/sdlc/SKILL.md (Long-Running Goals section)
   

   Non-blocking (exit 0) — same soft-warn pattern as model-effort-check.sh.
5. DLC-binding co-check: also warn when the /goal condition string does not name a DLC (/sdlc, /gdlc, /ldlc). Both checks in one hook.
6. Tests (quality, not existence): fixture transcripts where prior turn DID/DID NOT state HIGH 95%; fixture conditions WITH/WITHOUT DLC binding; assert warnings fire/don't fire correctly. Mirror the test structure in tests/test-hooks.sh for model-effort-check.

Alternatives considered

• Statusline nudge — too easy to ignore.
• Native CC primitive request — feature-request to Anthropic to bake confidence-gating into /goal itself. Could be filed in parallel but doesn't help wizard users until shipped.
• Skill-only (status quo) — what we have. Doesn't survive the failure mode it's trying to prevent (LLM ignores documentation under pressure).

Enforcement layer is the missing piece.

Prerequisites / blast radius

• CC version: requires UserPromptSubmit hook event (already required by sdlc-prompt-check.sh, so no version bump).
• No id-token / network / package.json changes.
• Plugin manifest needs the new hook registered.
• CLI distributes the hook file (8th → 9th distributed hook).

Why now

/goal is new enough (CC v2.1.139, ~mid-May 2026) that wizard users are forming habits NOW. The earlier the gate ships, the less ingrained the "/goal everything" reflex becomes.

Context

• Wizard version: 1.77.0
• Prior art: PR feat(cc-features): document new CC features + ship /goal /sdlc wrapper (closes #347) #351 (wrapper section), PR feat(sdlc): /goal SDLC-discipline gates — 95% confidence floor + DLC binding (PR-D) #355 (95%/DLC text guidance), PR docs(sdlc): promote 5 portable lessons from v1.76.0+v1.77.0 release arc #357 (lesson promoted to SDLC.md)
• Originating discussion: 2026-05-25 session — /feedback invocation while pausing the v1.76+v1.77 release arc
• Pattern reference: hooks/model-effort-check.sh (already in repo) is the closest existing analogue

---

Submitted via `/feedback` skill