Session spec

Session-Spec Generator

Produces plans/<branch>-session-spec.md in step→verify format. Single-pass: one interview round, one conflict check, one output.

---

Adversarial Review Mandate

Primary job: critique requirements, not agree with them. Before accepting any requirement, actively search for:

1. Explicit conflicts — contradictory requirements, mutually exclusive acceptance criteria.
2. Implicit conflicts vs. codebase — contradictions with existing ADRs, CLAUDE.md patterns, domain model, CONTRIBUTING.md. Read CLAUDE.md for ADR locations, scan ADR dirs, search project source dirs before accepting any requirement.
3. Ambiguity-hidden conflicts — vague requirements that seem compatible but force contradictory impl choices.

Conflicts detected → push back via AskUserQuestion: state conflict, why matters, propose alternatives, block until resolved.

Seedling philosophy: Seedling doc gives head start — not sacred. Challenge it same as any input.

---

The Job

Step 1 — Environment Setup

1. Read $ARGUMENTS. Empty → AskUserQuestion to gather input before proceeding.
2. Detect input type:
   • $ARGUMENTS is path to existing .md file → seedling mode (use as baseline draft).
   • Otherwise → text description mode (generate from scratch).
3. Derive <slug> as short lowercase-kebab-case label from title or description.
4. Generate <branch> as feat-<slug>.
5. Detect GitHub issue references:
   • Scan $ARGUMENTS for URLs matching https://github.com/{owner}/{repo}/issues/{N}.
   • Matches → <source_issues>: list of owner/repo#N. No match → <source_issues> null.
   • Per issue in <source_issues>:

     gh issue edit <N> --repo <owner>/<repo> --add-assignee @me
     

     Fails → print ⚠ Warning: could not assign issue — <error>. Do NOT block spec creation.
6. Sync main from origin:
   1. Run git fetch origin main.
   2. git log main..origin/main --oneline — commits listed → main behind. AskUserQuestion to surface; stop until user confirms.
7. Create and checkout branch:

   git checkout -B <branch> main
   

   Verify git branch --show-current equals <branch>. Mismatch → AskUserQuestion and stop.
8. Create plans/ if absent: mkdir -p plans
9. Create symlink:

   ln -sf ~/.claude/tasks/<branch> plans/<branch>
   

   Verify test -L plans/<branch> and readlink ends in .claude/tasks/<branch>. Fails → AskUserQuestion and stop.

Step 2 — Conflict Scan

Read CLAUDE.md for ADR locations. Scan ADR dirs and relevant source dirs for existing code related to requested feature. Identify contradictions between requested and existing. Mandatory in both modes.

Versioning detection (same pass): Check for CHANGELOG.md + version manifest (plugin.json, package.json, pyproject.toml, Cargo.toml, etc.) in the repo. If both present → read CLAUDE.md for bump-type rules (major/minor/patch criteria) and determine <bump_type> from those rules + nature of changes being spec'd.

Step 3 — Single Focused Interview

Max 2–3 AskUserQuestion calls total. Rules:

• At least one question must address conflicts found in Step 2 (if any).
• No conflicts found → ask about goals, scope boundaries, success criteria.
• Do not re-ask content user already provided.

Seedling mode: ask 2–3 targeted questions on gaps, conflicts, ambiguities — not repeating seedling content.

Text description mode: ask 2–3 questions on problem/goal, scope/boundaries, success criteria.

Step 4 — Pre-Generation Checkpoint

1. Summarize in 1–3 sentences: goals gathered, scope boundaries, conflicts resolved in Steps 2–3.
2. AskUserQuestion: "Proceed to generate spec, or refine further?"
   • Proceed → continue to Step 5.
   • Refine → return to Step 3 for one additional targeted interview round (budget: 1–2 questions, separate from initial Step 3 budget). Cap: max 2 refinement iterations total — if Refine selected twice, proceed to Step 5 regardless on next pass.

Step 5 — Generate Spec

1. Draft spec content in step→verify format per references/example-session-spec.md — do not write to disk yet; Step 6 sweep runs first.

2. Spec structure:

   • <source_issues> non-null → write metadata table immediately after H1 and before ## Goal, one | Source Issue | row per issue.
   • <source_issues> null → omit table. H1 followed directly by ## Goal.
   • Steps framing: each deliverable = ### STEP-<slug>-NN: <name> with **Acceptance Criteria:** checklist.
   • One step = one commit. Scope each step so it can be implemented and committed independently (assuming prior steps already on branch). Steps that cannot be committed in isolation must be merged or re-scoped.
   • AC per step: - [ ] items — concrete, independently falsifiable checks. Include: artifact exists, behavior correct, project verification command passes. For new API surface, first step creates stubs with failing tests; subsequent steps implement against stable contracts.

3. Skill assignment per step. Every step must declare skills. Parsed by hyperteam Phase 1 → skills array in task YAML front-matter.

   Apply all rules that match.

   • Python code involved → add hyperwork-python
   • TypeScript code involved → add hyperwork-typescript
   • Step implements logic against existing contracts (not pure scaffolding) → add hyperwork-tdd
   • Step is docs, README, changelog, ADR, or user-facing writing → add hyperwork-tech-writing
   • hyperwork-api-scaffold as a standalone step only when BOTH hold: (a) No existing structure — target modules/files absent and shape indeterminate from existing code. (b) Parallelism unlock — scaffolded stubs allow ≥2 workers to proceed independently in next wave; if work remains serial after scaffolding, skip standalone step. Gate not met → bundle structure-definition into first FEAT task requiring it: add hyperwork-api-scaffold to that task's skills: array + note in description that worker creates structure inline before proceeding.

   No matches → skills: none (explicit sentinel; worker skips loading).

   Each step annotation (embedded in spec body, parsed by hyperteam Phase 1):

   > skills: hyperwork-tdd, hyperwork-python
   

   Example step with skill annotation:

   ### STEP-auth-01: Implement JWT validation middleware
   
   > skills: hyperwork-tdd, hyperwork-python
   
   **Acceptance Criteria:**
   - [ ] `src/auth/middleware.py` exists with `validate_jwt(token: str) -> Claims` function
   - [ ] Unit tests in `tests/test_middleware.py` cover valid, expired, malformed token cases
   - [ ] `pre-commit run --all-files` exits 0
   

   The hyperteam skill (Phase 2, Step 3) reads these annotations and produces the native task YAML:

   ---
   id: FEAT-auth-01
   type: FEAT
   skills:
     - hyperwork-tdd
     - hyperwork-python
   blocked_by: []
   ---
   

   Reading note for agents: Metadata table (if present) appears immediately after H1 and before first ## section. Parsers: locate H1, scan forward collecting all | Source Issue | rows before ##; none found → source_issues is null.

4. Version bump step (conditional). CHANGELOG.md + version manifest found in Step 2 → append final step after all FEAT/DOC steps:

   ### STEP-<slug>-NN: Bump version and update CHANGELOG
   
   > skills: hyperwork-tech-writing
   
   **Acceptance Criteria:**
   - [ ] Version manifest (`plugin.json` / `package.json` / etc.) bumped to next <bump_type> version per CLAUDE.md rules
   - [ ] `CHANGELOG.md` entry added under correct version header covering all changes in this spec
   - [ ] Project verification command exits 0
   

   <bump_type> from Step 2 detection. Step always last — blocked by all preceding FEAT steps.

Step 6 — Final Conflict Sweep

Verify: no step contradicts another; no step conflicts with codebase findings from Step 2. Conflict found → raise with user via AskUserQuestion and resolve before saving.

Open-questions gate: Every item in ## Open Questions must be: answered inline, explicitly deferred (note rationale in item), or removed. Any unresolved item remaining → raise via AskUserQuestion and resolve before saving.

All sweep checks pass → (1) if plans/<branch>-session-spec.md exists, move to plans/archive/<branch>-session-spec.md; (2) write drafted spec to disk.

Do NOT start implementing. Create spec only.

---

Before Saving

• Step 1 complete: <branch> chosen, main synced, branch checked out and verified, plans/ exists, symlink created and validated
• GitHub issue URLs scanned: <source_issues> set to ["owner/repo#N", ...] or null; gh issue edit assignment attempted per issue (warning on failure, spec not blocked)
• Input mode detected: seedling or text description
• CLAUDE.md ADR locations, ADR dirs, project source dirs searched for conflicts (Step 2)
• Interview complete: ≤3 AskUserQuestion calls; ≥1 addressed Step 2 conflicts (if any)
• Pre-generation checkpoint (Step 4) presented; user selected Proceed, or refinement cap (2 iterations) reached
• <source_issues> non-null → metadata table present immediately after H1 and before ## Goal; null → no table
• Spec uses ## Goal / ## Context / ## Non-goals / ## Steps / ## Open Questions structure
• Each step has **Acceptance Criteria:** checklist with ≥1 concrete, independently falsifiable - [ ] item
• Each step has > skills: annotation (skills: none if no rules match)
• Old plans/<branch>-session-spec.md archived to plans/archive/ if existed
• Final conflict sweep complete — no intra-spec contradictions, no codebase conflicts
• All ## Open Questions items answered, deferred with rationale, or removed
• CHANGELOG.md + version manifest present → final step is version bump + CHANGELOG; <bump_type> matches CLAUDE.md rules; step blocked by all preceding FEAT steps