# TODOS — gstack-auto ## P1 — High Priority ### Follow-up question validation When collecting pending-questions.md from worktree agents, validate the format before parsing. Malformed files (missing confidence tags, empty lines, binary content) should be caught and skipped with a warning in the round retrospective — not silently ignored. **Effort: S** **Context:** Identified during v2 eng review as a critical gap. The follow-up system depends on agents writing well-formatted questions, but LLMs can produce unexpected output. Without validation, bad questions are silently dropped, which looks like the pipeline working when it isn't. ### Deep-link to conductor.build Add an "Open in Conductor" button that launches conductor.build with the project workspace pre-loaded, eliminating copy-paste entirely. Blocked on knowing Conductor's URL scheme or web API for opening workspaces. The copy-to-clipboard prompt was shipped as the universal fallback. **Effort: S** (once URL scheme is known) **Depends on:** Conductor exposing a URL scheme or web API ### Conductor deep link / programmatic handoff Investigate Conductor's URL scheme or web API for opening workspaces programmatically. Currently the handoff is manual (copy prompt, open Conductor, paste). A deep link would eliminate copy-paste entirely. **Effort: S** (once URL scheme is known) **Depends on:** Conductor exposing a URL scheme or web API **Context:** Identified during gstack-auto-as-a-service CEO review. Existing "Deep-link to conductor.build" TODO above covers the same gap — merge when investigation completes. ## P2 — Medium Priority ### API key rotation for office hours Claude API The web service uses a single server-side Anthropic API key for office hours chat. Add rotation support: multiple keys with round-robin or failover, admin alerting on auth failures, and a config UI to add/remove keys without restarting the server. **Effort: S** **Context:** Identified during gstack-auto-as-a-service CEO review. Single key is acceptable for launch but becomes a single point of failure. ### Pipeline results webhook from Conductor Add a webhook endpoint that Conductor can call to push incremental build progress (phase completion, scores). Currently the pipeline POSTs final results only. Incremental updates would enable real-time SSE progress in Mission Control during builds. **Effort: M** **Depends on:** Conductor supporting outbound webhooks or the pipeline being modified to POST progress at each phase boundary. **Context:** Identified during gstack-auto-as-a-service CEO review. ### Chat keyboard accessibility for office hours The office hours chat is the core new UI pattern. Keyboard users need: focus management when new messages arrive, Escape to stop streaming, arrow keys to navigate message history, and screen reader announcements for assistant responses (aria-live region). Without this, keyboard-only and screen reader users can't use the primary product feature. **Effort: S** **Context:** Identified during gstack-auto-as-a-service design review. ### Mobile chat viewport management On mobile, the software keyboard pushes the viewport up and can hide the chat input field or cause scroll jumps during streaming responses. Needs explicit handling via the Visual Viewport API or fixed-bottom input with dynamic padding. Without it, mobile chat feels broken on iOS/Android. **Effort: S** **Context:** Identified during gstack-auto-as-a-service design review. ### Admin audit log Add an `admin_audit_log` table: `(admin_id, action, target_user_id, timestamp, details)`. Log every admin action (user approval, revocation, session viewing). Without this, a compromised admin account leaves no forensic trail. **Effort: S** **Context:** Identified during gstack-auto-as-a-service adversarial review. Both Codex and Claude subagent flagged this independently. ### Handoff conversion tracking Track how many users complete office hours vs. how many ever POST results back. If the ratio is below 30%, the manual Conductor handoff is the #1 priority to fix. Add metrics: `sessions_completed` count, `builds_with_results` count, conversion funnel in admin dashboard. **Effort: S** **Context:** Identified during gstack-auto-as-a-service adversarial review. Claude subagent flagged manual handoff as the highest product risk. ### Round-over-round early stopping When running multi-round (rounds > 1), detect if a round's winner scores LOWER than the previous round's winner. If so, stop early and keep the best round's output rather than wasting compute on regression rounds. Implementation: compare `round_results[-1].winner_score` with the new winner's score after Step 2e. If lower, skip to Step 3 (final report). **Effort: S** ### Cross-run scoring Currently each agent scores its own work (self-grading). A separate scoring agent that reads ALL runs' output/ and produces comparative scores would be more reliable. The self-grading bias means "best self-promoter wins" rather than "best code wins." **Effort: M** **Context:** Identified during v2 adversarial review. Fundamental architecture change — deferred to post-v2 stabilization. ### Auto-deployment with preview URLs (partially done) GitHub Pages deployment is now available via "Create Repo" in Mission Control. Remaining: support additional deploy targets (Vercel, Netlify, Railway) for projects that need a backend or custom build step. Requires detecting tech stack, choosing platform, configuring deploy, handling failures. **Effort: M** (reduced — GitHub Pages covers the common case) ### User email reply loop Monitor for user's email reply containing scope expansion instructions, then start a new pipeline cycle inheriting the winning run's codebase. Requires Gmail polling, reply parsing, cycle restart logic. **Effort: L** **Unblocked by:** email config (v0.1.1.0) — SMTP send works, polling is the remaining gap ### Live progress notifications Send email updates during pipeline run (not just at end). E.g., "Phase 9/13: Implementation complete." Prevents the 30+ minute silence that feels like it's broken. **Effort: S** **Unblocked by:** email config (v0.1.1.0) — SMTP infrastructure now available ### Persistent learning / knowledge base Store what worked (high-scoring patterns) and what didn't across runs. Future runs consult this history to bias toward successful approaches. Risk of overfitting to past successes. **Effort: M** ### Custom inline style principles Allow `style: custom` with a `style_principles:` multi-line block in config.yml instead of referencing a built-in profile. For users whose engineering philosophy doesn't map to a named engineer. Requires parsing multi-line YAML values and handling the precedence when both `style:` and `style_principles:` are set. **Effort: S** **Depends on:** Style inspiration feature (v0.1.4.0) ### Style adherence scoring dimension Add a "style_adherence" scoring dimension that evaluates whether the implementation actually follows the legendary engineer's principles, not just whether the code is generically good. Currently style injection is pure vibes — a 7.0 with Carmack style and 7.0 with Marlinspike style are scored identically because the rubric doesn't differentiate. **Effort: S** **Context:** Identified during v2 adversarial review. ### Branching iterations (fork a build into two directions) Currently iteration is linear: Build 1 → Build 2 → Build 3. Users might want to fork: Build 1 → Build 2a (fix UX) AND Build 1 → Build 2b (add feature). The lineage model (parent_build_id + root_session_id) already supports this — the UI needs a way to pick which ancestor to iterate from and display the tree. **Effort: S** **Priority: P3** **Depends on:** Iteration feature (iterate-from-mission-control plan) **Context:** Identified during CEO review of iterate feature. Linear iteration covers 90% of use cases; branching adds UI complexity for a rare need. ### Iteration diff viewer (side-by-side code changes) The iterate feature includes a text diff summary (iteration_summary), but a real code diff between parent and child output/ directories would let users see exactly what changed. Could use a simple unified diff view rendered in the build detail page. **Effort: M** (human) → S (CC) **Priority: P2** **Depends on:** Iteration feature (iterate-from-mission-control plan) **Context:** Identified during CEO review. Text summary is sufficient for v1; code diff is a power-user feature. ## P3 — Low Priority ### Chat history retention policy Define and implement a retention policy for office hours chat messages. Options: auto-delete after N days, archive completed sessions, let users export/delete their own history. Without a policy, the messages table grows unbounded and stale conversations clutter the UI. **Effort: S** **Context:** Identified during gstack-auto-as-a-service CEO review. ### Visual references in Phase 01 Have the CEO persona name 2-3 real websites whose look and feel match the product's audience (e.g., "feel like Linear", "feel like Notion"). Gives Phase 03 and 05 concrete aesthetic anchors beyond abstract design style principles. Risk: LLM may hallucinate outdated sites, but directional guidance is still valuable. **Effort: S** ### Score history sparklines Add inline SVG sparkline charts to the dashboard results cards showing how each scoring dimension changed across rounds. Requires storing per-round per-run scores in results-history.json and rendering tiny line charts (no library needed — inline SVG path elements). Most useful when rounds > 1. **Effort: M** **Depends on:** results-history.json persistence ### Project history browser Let users see past projects (spec title, date, best score) and switch between them. `results-history.json` already accumulates this data. Once "New Project" exists, users will run multiple projects and want to revisit old ones. Needs a list view UI and archive/restore logic. **Effort: M** ### Keyboard shortcuts for dashboard Add keyboard navigation to Mission Control: `j`/`k` to move between run cards, `p` to toggle preview, `d` to jump to diff compare, `g` to toggle config panel, `?` to show shortcut overlay. Keep it opt-in (no conflicts with browser shortcuts). **Effort: S** ### Auto-probe email on page load When index.html (Mission Control) loads and .env already exists, silently POST /test-email in the background and show a green/red dot in the collapsed email status line. Removes the manual "Test Connection" click for returning users. Risk: SMTP probe adds ~2s latency to page load; transient failures could show false red. Mitigate by showing "checking..." state and only going red on definitive auth failures. **Effort: S** **Depends on:** Collapsible email section (done in setup.html launchpad redesign) ### One-click branch selection in email Include git checkout commands in email for each run's worktree branch. Zero friction between "I like this one" and "I'm working on it." **Effort: S** **Unblocked by:** email config (v0.1.1.0) — email delivery now works ### N>3 differentiation strategy Current differentiation (Run A=quality, B=UX, C=robustness) caps at 3 meaningful biases. Runs D+ get vacuous "pick a dimension" which leads to convergent solutions. Need structured biases for N=5 and N=7. **Effort: S** **Context:** Identified during v2 adversarial review. ## Completed ### ~~Build step for prompt composition (DRY)~~ ✓ DONE → SUPERSEDED Originally implemented as `pipeline/gen-phases.mjs` + `pipeline/phase-config.json`. Deleted in v2 pipeline rewrite — v2 uses explicit phase files with no generation. **Completed:** v0.1.12.0 (2026-03-17) | **Superseded:** v2 pipeline (2026-03-25) ### Design review pipeline (v1 phases 11-13 → v2 Phase 05) Design review moved from post-implementation audit to pre-implementation planning phase (Phase 05) in v2. Design principles are now defined before code is written, not evaluated after. ### Differentiation strategy for parallel runs Run A/B/C have distinct approach biases in phase 01 prompt. ### Multi-round pipeline with auto-accept winner Pipeline supports `rounds: N` in config.yml. Each round auto-selects the winner, commits it to git with a feature summary, and feeds it as input to the next round. Dashboard shows round progression. ### v2 Pipeline Rewrite Replaced 13-phase v1 pipeline with skill-based v2 pipeline featuring: - Office hours / design doc as entry point (replaces blank product-spec.md) - Dual adversarial review (Claude + Codex) at configurable phases - 4 planning phases before implementation (CEO → Eng → Design → Eng v2) - Mid-run follow-up question system (3 per round, shared) - Document release phase (Phase 12) **Completed:** 2026-03-25