diff --git a/.gitignore b/.gitignore
index 895dc97..615e362 100644
--- a/.gitignore
+++ b/.gitignore
@@ -12,6 +12,7 @@ plugins/*/skills/
 # any transitive dist/ dependencies the HUD imports — currently scripts/utils/).
 scripts/hud/
 scripts/utils/
+scripts/package.json
 
 # Generated shared agents (copied from shared/agents/ at build time)
 # Note: plugin-specific agents (catch-up.md, devlog.md) are committed in their plugins
diff --git a/package.json b/package.json
index 5249430..d1cee43 100644
--- a/package.json
+++ b/package.json
@@ -14,6 +14,8 @@
     "scripts/hooks/",
     "scripts/hud.sh",
     "scripts/hud/",
+    "scripts/utils/",
+    "scripts/package.json",
     "src/templates/",
     "README.md",
     "LICENSE",
diff --git a/plugins/devflow-plan/commands/plan.md b/plugins/devflow-plan/commands/plan.md
index 51fab42..7aefeaa 100644
--- a/plugins/devflow-plan/commands/plan.md
+++ b/plugins/devflow-plan/commands/plan.md
@@ -33,7 +33,7 @@ For **multi-issue** mode: collect all `#N` tokens from `$ARGUMENTS` as `ISSUE_NU
 
 | Gate | Phase | Purpose |
 |------|-------|---------|
-| Gate 0 | Phase 1 | Confirm understanding before exploration |
+| Gate 0 | Phase 1 | Requirements discovery before exploration |
 | Gate 1 | Phase 7 | Validate scope + gap analysis results |
 | Gate 2 | Phase 13 | Confirm final plan + design review |
 
@@ -45,15 +45,24 @@ No gate may be skipped. If user says "proceed" or "whatever you think", state re
 
 ### Block 1: Requirements Discovery
 
-#### Phase 1: Gate 0 — Confirm Understanding
+#### Phase 1: Gate 0 — Requirements Discovery
 
-Present interpretation using AskUserQuestion:
-- Core problem this solves
-- Target users
-- Expected outcome
-- Key assumptions
+Explore the user's intent through focused Socratic questioning before spawning agents.
 
-For multi-issue: present unified scope across all issues.
+**Skip discovery when** (semantic assessment, not word count):
+- User has specified WHAT to build, HOW it should behave, and WHERE it integrates
+- User input references an existing design document or detailed issue
+
+**Process:**
+
+1. **First question**: Confirm your understanding of the core problem and expected outcome. Frame as multiple choice when 2-3 interpretations exist.
+2. **Follow-up questions** (if ambiguity remains): Probe constraints, scope boundaries, or tradeoffs via AskUserQuestion.
+3. **Present approaches**: When multiple valid approaches exist, present 2-3 options with explicit tradeoffs. Lead with your recommendation and why.
+4. **Confirm scope**: Summarize understanding including: core problem, target users, expected outcome, key assumptions, chosen approach (if applicable).
+
+For multi-issue: present unified scope across all issues after individual discovery.
+
+If the user says "skip" or "just proceed" — skip remaining questions, present inferred understanding (core problem, users, outcome, assumptions, recommended approach) in one message for confirmation, then proceed. Gate 0 is satisfied by the confirmation, not by the discovery questions.
 
 **MANDATORY**: Do not spawn any agents until Gate 0 is confirmed.
 
@@ -350,7 +359,7 @@ Display completion summary:
 /plan (orchestrator - spawns agents only)
 │
 ├─ Block 1: Requirements Discovery
-│  ├─ Phase 1: GATE 0 - Confirm Understanding ⛔ MANDATORY
+│  ├─ Phase 1: GATE 0 - Requirements Discovery ⛔ MANDATORY
 │  │  └─ AskUserQuestion: Validate interpretation
 │  ├─ Phase 2: Orient + Load Knowledge
 │  │  ├─ Skimmer agent (codebase context)
diff --git a/scripts/build-hud.js b/scripts/build-hud.js
index cd9441b..9b5c51c 100755
--- a/scripts/build-hud.js
+++ b/scripts/build-hud.js
@@ -80,4 +80,11 @@ while (queue.length > 0) {
   }
 }
 
+// All compiled .js files use ESM syntax — Node needs this to resolve imports
+// across sibling directories (e.g. hud/ importing from ../utils/).
+const pkgJsonPath = path.join(scriptsRoot, 'package.json');
+if (!fs.existsSync(pkgJsonPath)) {
+  fs.writeFileSync(pkgJsonPath, '{"type": "module"}\n');
+}
+
 console.log(`\u2713 HUD distribution: ${visited.size} files copied to ${scriptsRoot}`);
diff --git a/shared/skills/plan:orch/SKILL.md b/shared/skills/plan:orch/SKILL.md
index 6e30239..1f38e36 100644
--- a/shared/skills/plan:orch/SKILL.md
+++ b/shared/skills/plan:orch/SKILL.md
@@ -24,6 +24,7 @@ This is a focused variant of the `/plan` command pipeline for ambient ORCHESTRAT
 
 For GUIDED depth, the main session performs planning directly:
 
+0. **Discover** — If the planning question is open-ended, ask clarifying questions via AskUserQuestion and present 2-3 approaches with tradeoffs before orienting. Skip if the user's prompt is already specific. If the user says "skip" or "just proceed": skip remaining questions, present inferred scope for confirmation.
 1. **Spawn Skimmer** — `Agent(subagent_type="Skimmer")` targeting the area of interest. Use orientation output to ground design decisions in real file structures and patterns.
 2. **Design** — Using Skimmer findings + loaded pattern/design skills, design the approach directly in main session. Apply `devflow:design-review` skill inline to check the plan for anti-patterns before presenting.
 3. **Present** — Deliver structured plan using the Output format below. Use AskUserQuestion for ambiguous design choices.
@@ -44,6 +45,41 @@ KNOWLEDGE_CONTEXT=$(node scripts/hooks/lib/knowledge-context.cjs index "{worktre
 
 This produces a compact index of active ADR/PF entries. Pass `KNOWLEDGE_CONTEXT` to Explorer and Designer agents — prior decisions constrain design, known pitfalls inform gap analysis. Agents use `devflow:apply-knowledge` to Read full entry bodies on demand.
 
+## Phase 0.5: Requirements Discovery
+
+Before committing to an approach, surface ambiguity through focused Socratic questioning.
+
+**Skip when** (semantic assessment, not word count):
+- User has specified WHAT to build, HOW it should behave, and WHERE it integrates — regardless of prompt length
+- Invoked from within another pipeline (pipeline:orch, implement:orch)
+- Single clear approach exists with no meaningful alternatives
+
+**Skip examples** (proceed directly to Phase 1):
+- "Add retry with exponential backoff to HttpClient in src/http.ts, max 3 retries, configurable timeout" — specific files, clear behavior, defined parameters
+- "Implement the design from .docs/design/caching.md" — pre-existing specification
+
+**Discover examples** (run Phase 0.5):
+- "Add a caching layer" — open-ended, multiple valid approaches
+- "Improve the auth flow" — vague scope, unclear what aspects need improvement
+- "Design a notification system" — system-level, many architectural choices
+
+**Process:**
+
+1. **Assess** — Does the request have meaningful ambiguity or multiple valid approaches? If not, skip to Phase 1.
+2. **Question** — Ask clarifying questions via AskUserQuestion. Prefer multiple choice (2-4 options) when tradeoffs exist.
+3. **Propose approaches** — Present 2-3 options with explicit tradeoffs:
+   - Lead with your recommended approach and why
+   - Each option: 2-3 sentences + key tradeoff (complexity, performance, maintenance)
+   - Final option: "Other — describe your preferred approach"
+4. **Confirm** — Get user's choice, then proceed to Phase 1 with a constrained problem.
+
+If the user says "skip", "just proceed", or signals impatience — skip remaining questions, present your inferred understanding (problem, scope, recommended approach) in one message for confirmation, then proceed to Phase 1 after confirmation. This matches /plan Gate 0 behavior.
+
+**Question design:**
+- Ask about constraints and goals, not implementation details
+- Surface hidden assumptions ("Does this need to handle concurrent writes?")
+- Reveal scope boundaries ("Just the API layer, or the UI as well?")
+
 ## Phase 1: Orient
 
 Spawn `Agent(subagent_type="Skimmer")` to get codebase overview relevant to the planning question:
diff --git a/shared/skills/router/classification-rules.md b/shared/skills/router/classification-rules.md
index 4e56e57..bd6f3ca 100644
--- a/shared/skills/router/classification-rules.md
+++ b/shared/skills/router/classification-rules.md
@@ -21,7 +21,10 @@ Classify each prompt by **intent** and **depth** before responding.
 
 Default to ORCHESTRATED for substantive work — it produces better results.
 Reserve GUIDED for small focused changes where orchestration adds no value.
-Prefer GUIDED over QUICK for any prompt involving code changes.
+Prefer GUIDED over QUICK for any prompt involving code changes — UNLESS the
+change AND target are both explicit and trivial (e.g., "rename foo to bar in
+utils.ts", "change color from red to blue in header.tsx"). Such single-site,
+obvious edits stay QUICK.
 
 ## Action
 
diff --git a/shared/skills/test-driven-development/SKILL.md b/shared/skills/test-driven-development/SKILL.md
index ad5c2c8..106e4ed 100644
--- a/shared/skills/test-driven-development/SKILL.md
+++ b/shared/skills/test-driven-development/SKILL.md
@@ -72,21 +72,81 @@ Apply DRY, extract patterns, improve readability.
 
 ---
 
+## Cycle Verification
+
+After each RED-GREEN-REFACTOR cycle, ALL must hold:
+
+- [ ] Test existed BEFORE production code (not concurrent, not after)
+- [ ] Test failed for the RIGHT reason (expected behavior absent, not syntax/import error)
+- [ ] Production code is minimal — no speculative additions beyond what the test demands
+- [ ] ALL tests pass, not just the new one
+- [ ] Refactoring happened in Step 3 (or code is already clean — state explicitly)
+- [ ] No untested production code remains
+
+If any fails: you skipped TDD. Back up and redo the cycle correctly.
+
+---
+
 ## Rationalization Prevention
 
 These are the excuses developers use to skip TDD. Recognize and reject them.
 
 | Excuse | Why It Feels Right | Why It's Wrong | Correct Action |
-|--------|-------------------|---------------|----------------|
+|--------|-------------------|----------------|----------------|
 | "I'll write tests after" | Need to see the shape first | Tests ARE the shape — they define the interface before implementation exists | Write the test first |
 | "Too simple to test" | It's just a getter/setter | Getters break, defaults change, edge cases hide in "simple" code | Write it — takes 30 seconds |
 | "I'll refactor later" | Just get it working now | "Later" never comes; technical debt compounds silently | Refactor now in Step 3 |
 | "Test is too hard to write" | Setup is complex, mocking is painful | Hard-to-test code = bad design; the test is telling you the interface is wrong | Simplify the interface first |
 | "Need to see the whole picture" | Can't test what I haven't designed yet | TDD IS design; each test reveals the next piece of the interface | Let the test guide the design |
 | "Tests slow me down" | Faster to just write the code | Faster until the first regression; TDD is faster for anything > 50 lines | Trust the cycle |
+| "Framework is hard to set up" | Setup is complex | One-time cost vs recurring regression cost; untested code compounds debt | Set up the framework first — that IS the work |
+| "Need the architecture first" | Can't test without structure | Tests DEFINE the architecture; they reveal what interfaces are needed | Let tests drive the structure |
+| "This is infrastructure, not logic" | Plumbing doesn't need tests | Infrastructure carries data; broken plumbing floods everything downstream | Test the contract, not the internals |
+| "Deadline is tight, no time for tests" | Ship now, test later | Untested code ships bugs; fixing bugs under deadline is slower than TDD | TDD is faster under pressure, not slower |
 
 See `references/rationalization-prevention.md` for extended examples with code.
 
+### Red Flags — STOP Immediately
+
+The rationalization table above catches excuses you make *before starting*. These red flags catch you *mid-work* — thoughts that signal you are about to skip a step.
+
+| Thought | Correction |
+|---------|------------|
+| "Let me write the implementation first, tests after" | Delete the code. Write the test. Watch it fail. Then rewrite. |
+| "I need to see the shape before I can test" | The test IS the shape. It defines the interface before implementation. |
+| "The test setup is too complex for this" | Complex setup = too much coupling. Simplify the design first. |
+| "I'll just spike this and add tests later" | Unless the user said "spike" — you're rationalizing, not prototyping. |
+| "Let me get it working, then lock it with tests" | Those tests verify your implementation, not the requirement. Backwards. |
+| "I know this works, I've written it before" | Past code passed past tests. This code needs its own failing test. |
+| "One more function, then I'll write the test" | STOP. Write the test for what you have now. Increments, not batches. |
+
+### Test-First vs Code-First
+
+**Code-first** (wrong):
+1. Write `parseConfig()` — split lines, filter comments, build map
+2. Write test: `parseConfig("key=val")` passes
+3. Ship. Undiscovered: empty input, malformed lines, multi-value keys, whitespace
+> Test mirrors implementation. Edge cases stay hidden until production.
+
+**Test-first** (correct):
+1. Test: "ignores comment lines" — `parseConfig("# comment\nkey=val")` → `{key: val}`
+2. Test: "handles empty input" — `parseConfig("")` → empty result
+3. Test: "rejects malformed lines" — `parseConfig("no-equals")` → error
+4. Implement `parseConfig()` to satisfy all three
+> Tests define the contract. Implementation forced to handle edges from the start.
+
+The difference: code-first tests verify what you *happened to build*. Test-first tests specify what *should exist*.
+
+### When Stuck
+
+| Blocker | Solution |
+|---------|----------|
+| Don't know what to test first | Test the simplest input/output pair. "Given X, expect Y." Start there. |
+| Test needs complex setup/state | Extract the logic into a pure function. Test that. Integrate after. |
+| Behavior depends on external state (DB, API, clock) | Inject the dependency as a parameter. Pass a fake in tests. |
+| Multiple behaviors tangled together | Decompose. One test = one behavior. If you can't isolate it, the design needs splitting. |
+| Modifying legacy code with no tests | Write a characterization test first — a test that captures the current behavior, even if wrong. Then make your change and see what breaks. |
+
 ---
 
 ## Process Enforcement
@@ -109,7 +169,7 @@ When implementing any feature under ambient IMPLEMENT/GUIDED or IMPLEMENT/ORCHES
 ### What to Test
 
 | Test | Don't Test |
-|------|-----------|
+|------|------------|
 | Public API behavior | Private implementation details |
 | Error conditions and edge cases | Framework internals |
 | Integration points (boundaries) | Third-party library correctness |