Spec author

/spec-author <feature-id> <title>

Author a hardened operational specification for a feature through multi-pass adversarial review: structured authoring, falsification, and automatic depth passes when critical findings indicate structural gaps. The output is a complete spec file ready for /spec-write registration.

This skill operates on design intent, not implementation. The only code read during authoring is pre-existing components (for prerequisite stubs). Checking the spec against the feature's implementation is /spec-verify's job, not this skill's.

Two entry paths:

• Default (fresh authoring): /spec-author <feature-id> <title> — runs Pass 1 (draft) → Pass 2 (falsification) → Pass 3 (depth pass) → register.
• Depth-pass-only: /spec-author <spec-id> --depth-pass-only [--lens <name>] — re-falsify an existing APPROVED spec under the current lens corpus (or a specific lens). Skips Pass 1 and Pass 2. See the Depth-pass-only mode section below.

---

Inputs expected in context

• The feature ID and title passed as arguments
• The feature brief (from .feature/<slug>/brief.md or user description)
• The domain analysis (from .feature/<slug>/domains.md if available)
• The work plan (from .feature/<slug>/work-plan.md if available)
• Optionally, a resolved spec context bundle from /spec-resolve

---

Depth-pass-only mode

When invoked with --depth-pass-only, this skill skips Pass 1 (drafting) and Pass 2 (initial falsification) and runs only the Pass 3 depth pass against an existing APPROVED spec. Use this to re-falsify a mature spec under the current lens corpus — typically because /curate flagged the spec as authored before a falsification lens shipped (e.g., specs from before the security lens added in v0.14.2).

Invocation:

/spec-author <spec-id> --depth-pass-only [--lens <lens_name>]

• <spec-id> — must resolve to an APPROVED spec via the manifest. DRAFT specs go through normal Pass 1/2/3; depth-pass-only refuses them.
• --lens <name> — optional. Bias the depth pass toward a specific lens reference. Without it, the full lens corpus applies. Lens references live at prompts/audit/lens-<name>.md (e.g., lens-security.md). See scripts/lens-registry.txt for the current set of lens names and their introduction dates.

Behavior:

1. Resolve the target spec. Look up <spec-id> via spec_file_for_id in the manifest. If state is not APPROVED, refuse and tell the user to use the default /spec-author flow instead.
2. Validate the lens (if --lens given). Check that .claude/prompts/audit/lens-<name>.md exists (installed layout) or prompts/audit/lens-<name>.md (dev-repo). If neither exists, list the available lenses (from scripts/lens-registry.txt) and stop.
3. Skip the fresh-authoring pre-flight steps (revival detection, work-group context, layered-spec family load) — they are scoped to new authoring. The only pre-flight needed for depth-pass-only is:
   • Read the spec body via the resolved file path.
   • If the spec has parent_spec set, run the layered-spec family load (Pre-flight Step 6) so siblings are visible to the depth pass.
   • Read any KB adversarial findings for the spec's domains (Pass 2 "KB adversarial findings" loader).
4. Skip Pass 1 and Pass 2.
5. Run Pass 3 with adapted inputs. Dispatch the depth-pass subagent with:
   • The full spec body as input (not "Pass 1 draft").
   • Empty prior-findings set (the original Pass 2 findings are not retrieved — too lossy across kit versions; the depth pass starts fresh against the spec as it stands).
   • If --lens <name> was given: include the named lens reference prompt as additional context, with framing: "Focus your falsification on attack categories surfaced by this lens. Other lenses may apply but are secondary."
   • If --lens was not given: the full standard lens corpus.
6. Standard arbitration UI. Findings flow through the Pass 2/3 arbitration grouping (confirmed gaps, implementation constraints, observability, uncertain). User accepts/declines per finding.
7. Apply accepted findings to the spec via the standard amendment flow (in-place edits + spec-write re-registration; the manifest's version: field bumps).

Signal-to-noise expectation. Lower than a fresh Pass 2 — specs in production have already encountered real-world inputs and survived many edge cases. A strong-model depth pass under a newer lens commonly surfaces 5-15 findings on mature specs that predate the lens; some will be already-mitigated in code (the spec just doesn't say so) and others will be genuine gaps. The arbitration UI lets the user judge.

Decline routing back to /curate. If the user's arbitration declines all findings as not-applicable, that's a signal the lens really doesn't apply to this spec. Append a falsification-decline row to .curate/curation-state.md keyed by <spec-id>:<lens> so future /curate scans show the prior decline. This step only runs when invoked via /curate --verify routing — manual invocations don't record declines (the user can still surface them via /curate).

---

Pre-flight

1. Run /spec-resolve if no context bundle is in your context:

   bash .claude/scripts/spec-resolve.sh "<feature description>" 25000
   

2. Read .spec/CLAUDE.md for the domain taxonomy.

3. Read any relevant KB entries and ADRs referenced in the domain analysis or context bundle.

4. Revival detection: Check for INVALIDATED specs in matching domains:

   INCLUDE_INVALIDATED=true \
     bash .claude/scripts/spec-resolve.sh "<feature description>" 25000
   

   If the bundle contains a ## INVALIDATED Specs (historical reference) section, check whether any INVALIDATED spec's requirements overlap with the feature brief's intent (shared subject tokens, related domain).

   If overlap is found, present the INVALIDATED spec(s):

   ── Revival candidate ──────────────────────────
     <spec_id> — "<spec title>"
     Displaced by: <displacing spec id>
     Reason: <displacement_reason>
     Requirements: <count> (R1-R<N>)
   

   Use AskUserQuestion:

   • "Use as reference" — load the INVALIDATED spec as additional context for Pass 1. The spec is a reference, not a template — author the new spec fresh through the normal two-pass flow. After registration (via spec-write), set revives: ["<old_spec_id>"] in the new spec's frontmatter and update the old spec's revived_by field.
   • "Author from scratch" — proceed normally without the INVALIDATED spec as context.

   If no INVALIDATED specs overlap or the section is absent, skip silently.

5. Work group context: If .work/ exists, run:

   bash .claude/scripts/work-context.sh --domains "<domains from spec-resolve bundle>"
   

   If output is non-empty, read it and note:

   • Downstream consumers: Which work definitions depend on specs in this domain. Surface: "WD-03 and WD-05 depend on specs in this domain — the spec you author will affect their readiness."
   • Interface contracts: If any work definition expects an interface contract (kind: interface-contract) in this domain, note it: "WD-02 expects an interface contract at <path> — consider whether this spec should be authored as an interface contract."
   • This information does not change the authoring flow — it adds awareness of who will consume the spec being authored.

6. Layered-spec family load (if amending a child spec): If the target spec already exists in the manifest AND its frontmatter has a non-empty parent_spec, this is a child in a layered family. Adversarial review needs the parent (cross-cutting requirements) AND all siblings (cross- sibling contradictions) loaded as full files, not summaries. Re-run the resolver with sibling inclusion:

   EXPLICIT_SPEC_IDS="<target-spec-id>" INCLUDE_SIBLINGS=true \
     bash .claude/scripts/spec-resolve.sh "<feature description>" 40000
   

   spec-resolve.sh auto-includes the parent chain when a child is selected; INCLUDE_SIBLINGS=true adds every other child of the same parent. The resulting bundle becomes Pass 2 / Pass 3's "resolved context bundle" instead of the pre-flight default — falsification subagents need this richer family view to spot contradictions across the family.

   Also re-run the work-group context query against the family's domains (Step 5) so consumers of any sibling are surfaced.

   If the target is a top-level spec (no parent_spec), skip this step.

---

Pass 1 — Structured authoring

Mode: Author. Fast, exploratory, high-trust. Produce the draft.

Step 1a — Extract surface requirements

From the feature brief, domain analysis, and work plan, extract every explicit requirement. These are the "what to build" requirements — behavioral contracts, API signatures, storage formats, configuration options.

Write each as a numbered requirement following the requirement rules:

• One falsifiable claim per requirement
• Explicit subject ("The vector index must..." not "Must...")
• Measurable condition where applicable
• Present tense, active voice
• Behavioral, not structural: every requirement must be verifiable by observing inputs and outputs, not by reading source code. Never reference specific class names, method names, file paths, or call chains. The spec describes what must be true; the work planner decides which constructs deliver it.
  • BAD: "The LsmVectorIndex.encodeFloat16s method must produce bytes"
  • GOOD: "Float16 encoding must produce big-endian bytes with length exactly dimensions * 2"
  • BAD: "IvfFlat.index() must quantize before calling assignCentroid"
  • GOOD: "Centroid assignment must use the quantized vector value, not the original float32 input"

Step 1b — Identify prerequisites

For each surface requirement, ask: "What must already be true in the existing system for this requirement to hold?"

For each assumption identified:

• Check the resolved context bundle for an existing spec that covers it
• If covered: add to the requires array
• If not covered: create a prerequisite stub (see /spec-write Step 2)
• Read the relevant existing code ONLY for prerequisite verification — do not read the feature's own implementation

Step 1c — Expand by operational sequence

For each requirement that describes a temporal relationship ("X before Y", "X uses the result of Y", "X must happen during Z"):

Trace the operational sequence step by step:

1. What is the input to each step?
2. What transformation happens?
3. What is the output?
4. Who consumes the output?
5. At which step does the requirement's constraint apply?

If the trace reveals that the requirement is ambiguous about when in the sequence the constraint applies, split into separate requirements that specify the exact step.

Example: "Centroid assignment must use the quantized vector" is ambiguous. The operational trace reveals: input arrives as float32 → encoding quantizes to float16 bytes → decoding produces float32 from float16. The requirement must specify: "centroid assignment must use the quantized vector value (the result of encoding to the configured precision then decoding back to float32), not the original float32 input." Note: the requirement describes the behavioral constraint (use quantized value), not the implementation mechanism (which methods to call in which order).

Step 1d — Expand by failure mode

For each component or technology the feature touches, search for known failure modes BEFORE reading the design:

• IEEE 754 edge cases for floating point
• Concurrency hazards for shared data structures
• Encoding format ambiguities for serialization
• Input validation gaps for API boundaries
• Precision/rounding issues for numerical computation

Build a checklist from what you find. Then review each surface requirement against the checklist. For each failure mode that isn't addressed by an existing requirement, add a requirement.

Step 1e — Declare concurrency contracts

For every construct introduced by this spec, add a concurrency requirement. This is not optional — see the concurrency contract probe in Pass 2a for the full rationale.

For each construct, ask the user (or determine from the domain analysis):

• Will this be used from multiple threads?
• If shared: what concurrency model? (immutable, synchronized, confined)
• If not shared: state it explicitly as a requirement

If the answer is unknown, write the requirement as thread-confined by default: "This type is not thread-safe. Callers must not share instances across threads without external synchronization." The user can upgrade to thread-safe during review if the use case demands it.

This step ensures that Pass 2 falsification has a concurrency contract to verify against, rather than discovering the gap and having to add it.

Step 1f — Collapse user decisions, expand requirements

Review the requirements for clusters that stem from the same conceptual decision. For each cluster:

• State the conceptual decision once in the Design Narrative
• Keep individual requirements separate (they evolve independently)
• Ensure the user would only need to make one decision, not N

Step 1f — Write the draft

Assemble the requirements into a spec file following the format in .spec/CLAUDE.md. Include:

• All numbered requirements grouped by category
• Design Narrative (Intent, Why this approach, What was ruled out)
• Front matter with domains, requires, decision_refs, kb_refs
• [UNVERIFIED] annotations on any requirement you couldn't confirm

Present the draft to the user. Do NOT proceed to Pass 2 until the user confirms the draft direction is correct. The user may add, remove, or adjust requirements at this point.

---

Pass 2 — Adversarial falsification

Mode: Adversarial. Slow, zero-deference, exhaustive. Treat the Pass 1 output as a hypothesis to be falsified.

KB adversarial findings (load before launching subagent)

If .kb/CLAUDE.md exists, scan the Topic Map for categories relevant to this feature's domains. Read any type: adversarial-finding entries in matching categories. These are real bug patterns from prior audits — concrete attacks that broke real code in similar domains.

Pass these findings to the subagent as additional input. They provide concrete attack vectors that the falsification pass should check against each requirement. A KB entry like nan-score-ordering-corruption turns "what about NaN?" from an abstract question into a proven failure pattern with specific test guidance.

Implementation grounding (load before launching subagent)

Read the source files that the feature will modify or interact with. Use the domain analysis (domains.md) file list and the work plan stubs (if available) to identify which files exist. Read method signatures, type declarations, and API boundaries — not full implementations.

The purpose is NOT to spec implementation details. It is to discover which concrete technology decisions the implementation will encounter:

• Does the code use floating point? → IEEE 754 edge cases are real, not abstract
• Does it manage resources (streams, connections, handles)? → lifecycle bugs are real
• Does it parse external input? → encoding/validation gaps are real
• Does it share state across threads? → concurrency hazards are real

Without this grounding, the falsification pass reasons about spec text in isolation and misses gaps that are obvious when you see the actual types involved. A spec that says "numbers stored as text" doesn't trigger "what about NaN?" until you see that the implementation uses double.

Launch a subagent for this pass. The subagent receives:

• The complete draft spec from Pass 1
• The feature brief and domain analysis
• The resolved context bundle
• The list of prerequisite stubs created in Pass 1
• KB adversarial findings for the feature's domains (if any)
• Key source file signatures from the feature's domain (types, method signatures, API boundaries — NOT full implementation bodies)

The subagent's prompt must include:

2.0 — Subagent contract (load-bearing preamble)

Subagent contract: Honor rules/completeness-contract.md (load-bearing — no silent deferrals; trigger phrases = escalation signals, not completion modes). If you cannot complete assigned scope, escalate via AskUserQuestion with user-validatable proof. A return claiming COMPLETE alongside deferred items is a contract violation.

2a — Requirement-level falsification

For each requirement in the draft, begin with the assertion: "This requirement is complete and unfalsifiable." The adversary's job is to disprove this by constructing a concrete attack — a specific input, sequence of operations, or initial state that produces wrong behavior while technically satisfying the requirement text.

For each successful disproof, produce:

1. Requirement: the specific requirement being challenged (by number)
2. Attack: the concrete scenario — exact input values, operation sequence, or initial state that triggers the failure
3. Expected wrong behavior: what breaks, and how an observer would detect it
4. Gap in requirement text: which specific words are ambiguous, missing, or overly broad, and why they permit the attack
5. Tightened requirement: suggested replacement text that closes the gap

KB adversarial pattern check (mandatory if findings were provided). For each adversarial-finding entry provided, check whether any requirement in the draft is vulnerable to the same pattern. The entry's domain field tells you which requirements to focus on. The ## Test guidance section describes the specific attack. The ## What happens section describes the failure mode.

For each match: construct a concrete attack against the specific requirement (not just "this pattern might apply" — show exactly which requirement, which input, what breaks). If the requirement already handles the case, note it and move on. If it doesn't, produce a finding.

Degenerate value checklist (mandatory). For every requirement that involves a typed value (numeric, string, collection, or nullable), check these cases mechanically. These are language-agnostic — they apply regardless of implementation language.

Numeric types:

• NaN — not-a-number. Can the value be NaN? What happens downstream?
• Positive and negative infinity — same questions.
• Negative zero — does -0.0 vs 0.0 matter for equality, ordering, display?
• Boundary values — if the spec declares a precision or range, what happens at and beyond the boundary? (overflow, underflow, truncation)
• Zero — does zero have special semantics? Division, indexing, sizing?

String/text types:

• Empty string — vs null vs absent. Are these three cases distinguished?
• Null bytes — embedded \0 in strings. Truncation risk.
• Unicode boundaries — surrogate pairs, BOM, RTL markers, combining chars.

Collection types:

• Empty collection — zero elements. Does downstream code assume non-empty?
• Single element — boundary between "one" and "many" behavior.
• Null elements — can the collection contain nulls? What happens?

Nullable/optional types:

• Null vs absent vs default — are these three states distinguished? A field that is null, a field that is missing, and a field that has its default value may need different behavior.

Size/capacity:

• Maximum size — if the spec declares a limit, what happens at the limit and one beyond it? Is the error behavior specified?

For each case that produces wrong or unspecified behavior: produce a finding using the standard format above.

Boundary validation probe (mandatory). For every requirement that uses "reject", "validate", "must not accept", or "throw on invalid":

• Is the validation unconditional? Could it be bypassed by configuration, optimization, debug flags, or assertion settings?
• Are ALL entry points that could send invalid data enumerated? For each entry point, does validation happen there or does a requirement on the caller guarantee validity?
• Is the validation complete? If the requirement says "reject invalid X", does it enumerate what "invalid" means, or could an implementation interpret it narrowly?
• For mutable inputs: is the validation on the value at call time, or could the caller mutate the input after validation but before use?

Standards compliance probe (mandatory). For every requirement that references a standard, specification, or protocol (RFC, IEEE, Unicode, HTTP, SQL, etc.):

A compliance claim is a contract. "Compliant with RFC 8259" means the implementation MUST accept everything the RFC says is valid and MUST reject everything the RFC says is invalid. No exceptions, no "we're stricter." If the spec rejects something the standard allows, the spec is non-compliant — period.

• Identify every OTHER requirement in the spec that restricts or extends the standard's behavior. For each one: does the standard allow what the spec rejects? If yes, the compliance claim is false. The spec must either remove the restriction, or replace the compliance claim with an explicit deviation list: "Implements RFC 8259 with the following deviations: (1) duplicate keys rejected, (2) trailing content rejected." These are not stricter compliance — they are deviations from the standard that must be documented as requirements.
• For every degenerate case in the standard (empty values, maximum sizes, optional features, deprecated behaviors): does the spec handle it? If the standard allows empty string keys and the spec says "non-blank keys", that is a contradiction — produce a finding.
• Every deviation from the standard must be its own numbered requirement so it can be tested independently. "Stricter than RFC 8259" is not a requirement — it's a vague claim. "Rejects duplicate keys (deviation from RFC 8259 Section 4)" is a testable requirement.

Resource lifecycle probe (mandatory). For every requirement that introduces a closeable, disposable, or acquirable resource — or any construct that wraps streams, connections, handles, buffers, or locks:

• What happens to objects derived from this resource (iterators, streams, cursors, views) after the resource is closed? Must they throw, or can they silently return stale data?
• What happens if creation/construction fails partway? Are partially allocated sub-resources released?
• What happens if close/dispose itself fails? Is the resource left in a usable state, an unusable state, or an undefined state?
• What happens if close is called twice? Must it be idempotent?

Cross-construct atomicity probe (mandatory). Review the full requirement set for groups of requirements that describe the same logical operation across multiple constructs (e.g., "insert updates index A" + "insert updates index B"):

• If one succeeds and the other fails, what state is the system in?
• Is partial completion acceptable, or must the operation be atomic?
• If atomic: is the rollback mechanism specified? What state do observers see during the operation?
• If not atomic: is the partial-failure state documented? Can the system recover, or is manual intervention required?

Error propagation probe (mandatory). For every requirement that specifies an error, exception, or failure condition:

• After the error occurs, what is the state of the object that threw? Is it reusable without re-initialization?
• If the operation was mid-stream (writing, parsing, iterating), what happens to output written so far? Is it visible to observers? Is it valid?
• If the object holds shared resources (streams, buffers, connections), what state are those resources in after the error? Must the caller close/dispose them explicitly?
• Can the error propagate to a context where it's unexpected? (e.g., a checked exception thrown through an interface that doesn't declare it)

Identity and equality probe (mandatory). For every data type introduced by the spec that will participate in comparison, lookup, deduplication, caching, or sorting:

• What does equality mean? Reference identity, structural equality, or domain-specific equivalence?
• If the type contains fields with non-obvious equality semantics (floating point, opaque handles, lazy-loaded values), does the spec define how those fields affect equality?
• If two instances are "equal", can they be substituted for each other in all contexts? If not, the spec needs a weaker equivalence relation.

Concurrency contract probe (mandatory). Every construct introduced by the spec must declare its thread-safety model. This is not optional — the absence of a concurrency statement is a spec gap, because it leaves the implementer guessing and the test writer unable to verify.

For every construct (type, interface, resource, service) in the spec:

• Is it designed to be called from multiple threads? If yes: what is the concurrency model? (immutable, internally synchronized, requires external synchronization, thread-confined)
• If NOT designed for concurrent use: the spec must say so explicitly. "This type is not thread-safe — callers must not share instances across threads without external synchronization" is a requirement.
• For constructs that manage shared resources (caches, pools, registries, indexes): concurrent access is likely regardless of intent. The spec must declare whether concurrent operations are safe, and if so, what guarantees hold (atomicity of individual operations, consistency of compound operations, visibility of mutations).
• For compound operations (check-then-act, get-or-create, position-then- read): are they atomic? If not, the spec must state that callers are responsible for atomicity.

A spec that says nothing about concurrency is incomplete. The test writer cannot write concurrency tests without knowing the contract. The implementer cannot choose between synchronized and unsynchronized without knowing the intent. Declare it.

Trust boundary probe (mandatory). For every requirement that describes a predicate, status check, or query consumed by other constructs:

• Does the predicate have sub-states within "true"? (e.g., "is member" could mean active, joining, or leaving.) Do all consumers agree on which sub-states count as true?
• If the input comes from another construct (not user input), does the spec explicitly trust it ("assumes X is valid, validated by Y") or validate it? Implicit trust is a spec gap.

For each case that produces wrong or unspecified behavior: produce a finding using the standard format above.

If the adversary cannot construct a concrete attack for a requirement, the requirement stands as written. No changes.

Bare assertions that a requirement "could be" violated without a concrete attack are not findings. The burden of proof is on disproof, not on the requirement.

2b — Enforcement path tracing

For each NEW requirement proposed in this pass:

• If it adds validation ("must reject X"): identify all callers that could send X. For each caller, either add a filtering requirement or confirm the caller already prevents X.
• If it changes behavior ("must do X instead of Y"): identify all consumers that depend on old behavior Y.
• A requirement that creates a cascading failure when enforced is a spec defect, not a finding.

2c — Cross-requirement interaction

Review the full requirement set (Pass 1 + new additions) for:

• Contradictions: two requirements that cannot both be satisfied
• Cascading failures: requirement A, when enforced, causes requirement B to fail
• Ordering dependencies: requirement A must be implemented before B but this ordering isn't stated
• Atomicity gaps: multiple requirements that describe the same logical operation on different constructs. If one can succeed while another fails, the partial-failure state must be specified.
• Ambiguous alternatives: a requirement that uses "either X or Y", "may", or conditional language allowing two mutually exclusive implementations. If two test writers could independently write tests that contradict each other while both satisfying the requirement, the requirement is ambiguous and must be tightened. Replace "either X or Y" with a concrete choice. The spec makes the decision — the implementer should not have to choose.

2d — Cross-module boundaries

Search the codebase for other implementations of the same concepts:

• Are there multiple encode/decode paths for the same data?
• Are there utility classes in other modules that duplicate functionality?
• Could a maintainer accidentally use the wrong implementation?

For each found: add a requirement that specifies which implementation to use and prohibits cross-use.

2e — Observability of silent failures

For each failure mode in the spec that produces no error (silent degradation, invisible data, suppressed exceptions):

• Is there a way for an operator to detect it?
• If not, add a requirement that the behavior is documented in the public API, or add a diagnostic mechanism.

2f — Uncertainty disclosure

For each finding where you are uncertain whether it's a real gap:

• State what you're uncertain about
• State what would need to be true for it to be a real gap
• Let the user decide — don't suppress uncertain findings

---

Arbitration

Mode: Arbitration. User has design authority, agent has technical accuracy.

Present the Pass 2 findings to the user, grouped by:

1. Confirmed gaps — requirements that are definitely missing
2. Implementation constraints — requirements that clarify how something must be done (operational sequence, enforcement path)
3. Observability — silent failure modes that need documentation
4. Uncertain — findings where you couldn't determine if it's a real gap

For each finding, provide:

• What's missing (the unspecified behavior)
• What breaks (the concrete failure scenario)
• Suggested requirement text

The user decides what to include, adjust, or drop. Apply their decisions to the spec.

Just-in-time subdivision signal (after arbitration applied)

After applying the user's arbitration decisions, evaluate the spec's post-amendment shape against the natural-progression heuristic. The goal is a non-blocking heads-up — if this amendment has tipped the spec into "multiple distinct concerns at scale" territory, /spec-split is now an option.

Mechanical check (apply mentally — no script needed):

1. Count R-numbered requirements (lines matching ^R[0-9]+).

2. Count distinct concern-style section headers in the machine section (## <Concern> or ### <Concern> — exclude headers named "Notes", "Narrative", "Design", "History", "Appendix").

3. For each section, count the requirements under it. Track the largest section's count.

4. Apply the same thresholds /curate uses (so detection here matches what /curate would flag at the next housekeeping pass):

   • Reqs ≥ 50 OR machine-section size ≥ ~15K tokens (~60 KB)
   • Section count ≥ 2
   • Largest section's share < 90% of total reqs (otherwise the spec is large-but-singular, not a real subdivision candidate)
   • Spec does NOT already have parent_spec set (skip — already part of a layered family)

If all thresholds match, surface alongside the Pass 2 findings with this exact framing — note its non-blocking nature explicitly so the user knows it's information, not a request:

── Just-in-time subdivision signal ───────────────
This spec now sits in subdivision territory:
  N requirements across M concern sections
  Largest section: <section name>  (K reqs, P% of total)

After this amendment lands, consider:
  /spec-split <spec-id>

This is a heads-up, not a blocker. Proceed to Pass 3 normally. The
candidate will also surface at the next /curate run.

If the thresholds do NOT match, do not surface anything — silence is the default when subdivision isn't applicable. Don't draw attention to the absence of the signal.

This signal does NOT block Pass 3. It does NOT modify any file. It just informs the user that natural subdivision has become an option.

---

Pass 3 — Depth pass (mandatory)

After applying the user's arbitration decisions from Pass 2, automatically run a depth pass. Do not prompt — this is mandatory.

(Pass 3 is also the entry point for depth-pass-only mode — see the top-of-file section. When entered via --depth-pass-only, the depth pass runs against the existing spec body with empty prior-findings, optionally biased toward a single lens reference.)

Round 2 fixes create new attack surface that round 1 could not see. Empirical data across multiple specs shows round 2 consistently finds critical/high issues that are consequences of round 1 fixes: deadlocks from new thread models, budget fictions from abstraction changes, duplicate data from retry semantics. These are invisible to a single falsification pass.

Depth pass execution

The depth pass is structurally identical to Pass 2 (same subagent setup, same falsification lenses, same arbitration mode) with two additions:

1. Prior findings as input. The subagent receives all accepted findings from Pass 2 as additional context. These are attack vectors — knowing "credits shifted from physical slabs to logical budget" tells the depth pass to probe whether the budget actually bounds what it claims to.

2. Duplicate suppression. The subagent must not re-report findings already accepted in Pass 2. If a finding overlaps with an existing requirement added during arbitration, skip it.

After the depth pass, present findings in the same arbitration format. Apply the user's decisions to the spec.

Pass 4+ — Prompted on severity

After the depth pass, evaluate the accepted findings:

If any critical or high findings were found: use AskUserQuestion:

• "Run another pass" (description: "Critical/high findings suggest more may remain")
• "Done" (description: "Proceed to output")

If only medium/low findings: proceed to output. Diminishing returns are likely.

There is no limit on passes, but each subsequent offer beyond pass 3 should note diminishing returns. Empirically: pass 3 captures the fix-consequence bugs, pass 4+ is rarely productive.

---

Output — Register the spec

After all adversarial passes are complete and arbitration decisions applied, register the spec by invoking /spec-write "<id>" "<title>" directly.

This is mandatory. /spec-author owns the full lifecycle: draft → falsify → arbitrate → register. The caller should never need to call /spec-write separately. This prevents the shortcut where an unfalsified draft gets registered without adversarial review.

After /spec-write completes, verify the spec is in APPROVED state in .spec/registry/manifest.json. If it registered as DRAFT (due to unresolved conflicts), report the issue — do not silently return.

---

Hard constraints

• Never read the feature's implementation during authoring — only pre-existing code for prerequisite stubs
• Never skip Pass 2 — nothing advances past a draft without a falsification pass
• Never skip Pass 3 — the depth pass is mandatory, not severity-gated
• Never suppress uncertain findings — surface them for arbitration
• Never add a validation requirement without tracing its enforcement path
• Always present the draft to the user between Pass 1 and Pass 2
• The subagent for Pass 2 must receive the complete Pass 1 output, not a summary — summaries lose the detail that adversarial review needs
• Always register the spec via /spec-write at the end — the caller must never need to call /spec-write separately