solo-cto-agent 1.3.1 → 1.4.0

package/CHANGELOG.md

@@ -1,5 +1,52 @@
 # Changelog
 
+## v1.4.0 (2026-04-19)
+
+**Theme**: End-to-end automation complete. Install is fully automatic; review → rework → visual → merge runs without human copy-paste; Telegram / Discord carry the full operational loop.
+
+### Highlights
+* **Natural-language work orders** — `solo-cto-agent do "..."` CLI + Telegram `/do` route a plain-English instruction to the right product repo as a labeled, spec-rich issue the existing worker pipeline picks up.
+* **3-round agent consensus** — `cross-reviewer.js` runs an A/B debate (R1 propose → R2 agree/disagree/add → R3 verdict) with early-exit on agreement; non-consensus after R3 still dispatches rework with a distinguishable reason.
+* **Before/After visual report** — new `visual-report.yml` + `visual-reporter.js` capture screenshots of the Vercel preview at the pre- and post-rework SHA, compose side-by-side PNGs, commit them to the orchestrator, post to PR + Telegram `sendMediaGroup`.
+* **Opt-in GitHub auto-merge** — PR with `auto-merge-when-ready` label is merged by GitHub the moment all required checks pass (native `enablePullRequestAutoMerge` mutation; branch protection respected).
+* **Unified dispatcher** — `solo-cto-pipeline.yml` is now the single product-repo dispatcher with 7-layer anti-loop guards. Legacy `cross-review-dispatch.yml` + `rework-dispatch.yml` deleted; concurrency guards added on orchestrator receivers.
+* **Full install automation** — `setup.sh` now creates the orchestrator repo on GitHub, pushes it, and sets the `TRACKED_REPOS` variable itself instead of printing copy-paste commands.
+* **Telegram CTO command surface** — `/status`, `/list`, `/rework`, `/approve`, `/do`, `/digest`, `/merge` (admin-gated). Every review / rework / report message includes inline buttons for ✅ Approve · ❌ Reject · 🔧 Rework · 🔀 Merge.
+* **Discord mirror** — set `DISCORD_WEBHOOK_URL` and visual-change screenshots / auto-diagnose reports mirror to Discord as file attachments.
+* **Repo auto-discovery** — `init --wizard` shells out to `gh api` and offers a multi-select of the user's repos; saved selection auto-fills `--repos` on every subsequent command.
+
+### Pipeline fixes
+* `review-request` dispatch no longer orphaned — solo-cto-pipeline now emits `cross-review` to match the existing orchestrator listener.
+* Anti-loop guards in solo-cto-pipeline extended to recognise new comment formats: `## 🔍 Consensus Review`, `## Visual Report — Before / After`, `[visual-report-skipped:…]`, circuit-breaker comments, auto-merge-enabled comments, `<!-- cross-reviewer:consensus -->` machine tag.
+* Claude model IDs unified to `claude-sonnet-4-6` across rework-agent, claude-reviewer, claude-worker (was mixed 4.0 / 4.6).
+* Hardcoded `seunghunbae-3svs` owner in `solo-cto-pipeline.yml` replaced with `{{GITHUB_OWNER}}` placeholder — every non-maintainer user was hitting silent dispatch failures.
+* OpenAI call in rework-agent was passing `system` as a top-level parameter (wrong shape); moved into `messages[]` as a system-role message.
+
+### New notification paths
+* **`pr-merge-notify.yml`** — fires on PR closed (merged or not); posts consolidated Telegram + Discord summary with rework round count.
+* **`combined-pr-with-uiux.yml` rewired** — now triggers on the real workflow name (`Visual Report (Before/After)`), posts the single "all agent checks passed" message exactly once per PR.
+* `visual-check.yml` fires on `workflow_run: Auto Rework on Review completed` → fresh preview screenshots after every rework.
+* `auto-diagnose.yml` fires on rework-auto **failure** → Telegram-attached JSON diagnostic.
+* Skip paths (`[visual-report-skipped:…]`) now notify Telegram/Discord too so silent failures aren't silent.
+
+### Docs + hygiene
+* README front section rewritten to describe the full pipeline (consensus, rework, visual, Telegram) instead of just dual-agent review.
+* New `docs/user-journey.md` — install → trigger → review → rework → visual → merge, with ASCII flow diagram, common scenarios, troubleshooting table.
+* `docs/hero-banner-prompt.md` (new) — regeneration prompt for the README hero image aligned with the expanded surface.
+* `.env.example` covers Discord / Vercel / Browserless / admin Telegram chat IDs.
+* PAT scope guidance expanded for classic vs fine-grained tokens.
+* `require.main === module` guards on `visual-check.js` + `auto-diagnose.js` so tests can import them without firing `main()`.
+* Audit report from 2026-04-19 published as a gist: https://gist.github.com/seunghunbae-3svs/4f3da08f149fdb2b2451b43751f6f35c
+
+### Merged PRs
+* #106 — vision batch (repo-discovery / consensus / visual-report / NL orders / Telegram CTO)
+* #107 — pipeline consolidation
+* #108 — docs sync with #106/#107
+* #109 — end-to-end loop closure (install automation, merge notifications, D1/D2 wiring, README)
+* (this release) — `require.main` guards + hero banner prompt + de-dupe combined-pr-with-uiux trigger
+
+---
+
 ## v1.3.0 (2026-04-17)
 
 **Theme**: Tier 3 deep integration features — plugin registry search, setup automation, type system enhancements.
@@ -121,6 +168,26 @@ non-interactive verify in CI, and tear it all down with one command.
 
 ## Unreleased
 
+* Merge pull request #109 from seunghunbae-3svs/claude/e2e-audit
+
+* Merge pull request #108 from seunghunbae-3svs/claude/docs-sync
+
+* Merge pull request #107 from seunghunbae-3svs/claude/pipeline-consolidation
+
+* Merge pull request #106 from seunghunbae-3svs/claude/vision-batch
+
+* Merge pull request #105 from seunghunbae-3svs/claude/friendly-black-efec58
+
+* docs: sharpen README intro — feature-first, install in 2 lines
+
+* chore: v1.3.2 — clean up README slop, sync versions, update metrics
+
+* ci: add VS Code extension auto-publish to release workflow
+
+* docs: add hero banner to README, update test badge to 996
+
+* fix: rewrite managedAgentReview to match real Managed Agents API
+
 * chore: vscode extension packaging verified (icon, license, gitignore)
 
 * fix: routine.js readTier import from personalization (not core)

package/README.md

@@ -1,45 +1,106 @@
+<p align="center">
+  <img src="docs/hero-banner.png" alt="solo-cto-agent — AI code review for solo founders" width="720" />
+</p>
+
 # solo-cto-agent
 
-**Your AI coding agent already writes code. This makes it think like a CTO.**
+**The full CTO loop for solo founders — dual-agent review, multi-turn consensus, auto-rework, before/after visual reports, and a Telegram/Discord control surface.**
 
 [![npm](https://img.shields.io/npm/v/solo-cto-agent)](https://www.npmjs.com/package/solo-cto-agent)
-[![Test](https://img.shields.io/badge/tests-894%20passing-brightgreen)](https://github.com/seunghunbae-3svs/solo-cto-agent/actions/workflows/test.yml)
+[![Test](https://img.shields.io/badge/tests-996%20passing-brightgreen)](https://github.com/seunghunbae-3svs/solo-cto-agent/actions/workflows/test.yml)
 [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 [![PRs Welcome](https://img.shields.io/badge/PRs-welcome-brightgreen.svg)](CONTRIBUTING.md)
 
-> Stop babysitting your AI agent. `solo-cto-agent` adds circuit breakers for error loops, dual-agent code review, design quality gates, session memory, and deployment checklists — so you can focus on building instead of supervising.
+You push code. Two AI agents review it, debate for up to three rounds until they reach consensus, auto-push fixes for any blockers, shoot before/after screenshots of your Vercel preview, and ping your phone on Telegram with action buttons. When you tap ✅ Merge (or set the `auto-merge-when-ready` label), GitHub merges it once CI is green. You stay in the loop from your phone; you never touch YAML.
 
-**For solo founders, indie hackers, and small teams using Claude Cowork + OpenAI Codex.**
+```bash
+npm i -g solo-cto-agent
+solo-cto-agent init --wizard         # pick your repos, pick your tier
+solo-cto-agent do "fix the auth bug in tribo"   # natural-language work order
+# ...PR opens → review → rework → merge, all visible in Telegram.
+```
+
+**Surfaces**: CLI · [GitHub Action](action.yml) · [VS Code Extension](https://marketplace.visualstudio.com/items?itemName=seunghunbae-3svs.solo-cto-agent) · Telegram bot · optional Discord mirror.
 
 > **Languages**: English (primary) - [한국어 안내](#한국어-안내) below.
 
-## Quickstart (5 minutes)
+## Quickstart
 
 ```bash
-# 1. Install
+# 1. Install + wizard (wizard auto-discovers your repos via gh CLI,
+#    creates the orchestrator repo on GitHub, sets TRACKED_REPOS,
+#    and installs workflows on every product repo you pick.)
 npm install -g solo-cto-agent
-
-# 2. Initialize (recommended: choose mode during wizard)
 npx solo-cto-agent init --wizard
 
-# 3. Set your Anthropic API key (required for reviews)
-#    Get one at: https://console.anthropic.com/settings/keys
-export ANTHROPIC_API_KEY="sk-ant-..."
+# 2. Keys (all set in your shell, then copied to repo secrets by setup)
+export ANTHROPIC_API_KEY="sk-ant-..."     # required — Claude review + NL orders
+export OPENAI_API_KEY="sk-..."            # required for dual-agent / CTO tier
+export ORCHESTRATOR_PAT="ghp_..."         # required — cross-repo dispatch + write
+export TELEGRAM_BOT_TOKEN="..."           # optional — PR notifications + /commands
+export TELEGRAM_CHAT_ID="..."             # optional
+export DISCORD_WEBHOOK_URL="https://..."  # optional — mirror of Telegram output
+export VERCEL_TOKEN="..."                 # optional — before/after visual reports
 
-# 4. (Optional) Set OpenAI key for dual-review mode
-#    Get one at: https://platform.openai.com/api-keys
-export OPENAI_API_KEY="sk-..."
+# 3. Run setup-pipeline (reads saved wizard selection; no manual --repos needed)
+solo-cto-agent setup-pipeline --org <your-github-org>
 
-# 5. Verify everything is ready
-solo-cto-agent doctor --quick
+# 4. Verify
+solo-cto-agent doctor
 
-# 6. Run your first review (inside a git repo with staged changes)
-solo-cto-agent review
+# 5. Kick off a real work order
+solo-cto-agent do "add a monthly ARPU chart to the tribo admin dashboard"
+#   → LLM picks target repo, drafts a spec issue, labels agent-claude
+#   → claude worker opens a PR
+#   → cross-reviewer.js runs 3-round consensus
+#   → rework-agent.js pushes fixes if needed
+#   → visual-report.yml posts before/after screenshots
+#   → combined-pr-gate.yml sends "all checks passed" to Telegram
+#   → auto-merge-when-ready label makes GitHub merge on CI green
+#   → pr-merge-notify.yml sends final "✅ merged" to Telegram/Discord
 ```
 
-That is it. `doctor --quick` will tell you what is missing, where to get it, and the next command to run.
+Every step above ships end-to-end today. The `doctor` subcommand tells you anything missing with the exact command to run.
+
+### Telegram bot — the phone-first control surface
+
+After you set `TELEGRAM_BOT_TOKEN` and `TELEGRAM_CHAT_ID`, the bot gives you:
+
+| Command | What it does |
+|---|---|
+| `/status [repo]` | Open, non-draft PRs + review state across tracked repos |
+| `/list [repo]` | Last 10 PRs, one-line summary each |
+| `/rework <pr>` | Force a rework cycle on an existing PR |
+| `/approve <pr>` | Approve the PR |
+| `/do "<instruction>"` | Natural-language work order (same as CLI `do`) |
+| `/digest` | Today's PR activity summary |
+| `/merge <pr>` | Immediate merge (admin-only: `TELEGRAM_ADMIN_CHAT_IDS`) |
 
-### Platform-specific setup
+Every review / rework / visual-report message includes inline buttons — ✅ Approve · ❌ Reject · 🔧 Rework · 🔀 Merge. Tap to act without leaving Telegram.
+
+If `DISCORD_WEBHOOK_URL` is set, visual-change screenshots and auto-diagnose reports mirror to Discord as attachments.
+
+### External services
+
+| Service | Used for | Required? | Setup |
+|---|---|---|---|
+| **GitHub** | orchestrator repo + product repo workflows | ✅ required | wizard auto-creates orchestrator repo via `gh repo create` |
+| **Anthropic** | Claude consensus review, NL orders, rework | ✅ required | `ANTHROPIC_API_KEY` env var |
+| **OpenAI** | Codex counter-review (dual-agent) | CTO tier | `OPENAI_API_KEY` env var |
+| **Vercel** | preview URLs for before/after visual-report | optional | `VERCEL_TOKEN` + `VERCEL_PROJECT_ID`. Works with **Netlify / Cloudflare Pages / Render / Railway previews** too if their `deployment_status` webhooks fire — the visual stage resolves the URL from SHA and shoots whichever host serves it. |
+| **Telegram** | notifications + CTO commands | optional | `/telegram wizard` command + bot token |
+| **Discord** | optional mirror of Telegram output | optional | `DISCORD_WEBHOOK_URL` on orchestrator secrets |
+| **Browserless** | alternate screenshot provider (skips Playwright install cost) | optional | `VISUAL_REVIEW_PROVIDER=browserless` + `BROWSERLESS_API_KEY` |
+
+### Compatibility
+
+- **Stack-agnostic.** The toolkit never touches your application code directly. Agents produce patches that land on your PR branch; your repo's existing CI / build tools verify them. Works with Next.js, Vite, Remix, SvelteKit, FastAPI, Rails — anything with a PR workflow.
+- **Hosting-agnostic.** Vercel is the default for the visual-report preview URL resolver, but any host that ships preview URLs tied to commit SHAs works. For hosts without that (plain Docker, bare-metal, self-hosted): set `VISUAL_REVIEW_PROVIDER=off` and the pipeline just skips the visual stage — everything else still runs.
+- **Database-agnostic.** The toolkit doesn't read or write your database. Postgres (Supabase / Neon / PlanetScale-Postgres), MySQL, SQLite, MongoDB — all fine. Agent workers DO see your schema files if they're in the repo (prisma/schema.prisma, supabase/schema.sql, etc.) so suggested fixes can be schema-aware.
+- **Docker.** If your product repo is dockerized, nothing changes — GitHub Actions runners handle the build per your existing Dockerfile. The agents commit to the PR branch, your CI rebuilds the container, the auto-merge gate waits on that CI.
+- **Windows / macOS / Linux** for the CLI. GitHub Actions runners are Linux for all automation paths.
+
+### Platform-specific CLI setup
 
 **macOS / Linux**
 
@@ -60,6 +121,8 @@ solo-cto-agent doctor
 If you choose `codex-main` during the wizard, also install:
 - GitHub CLI: [cli.github.com](https://cli.github.com/)
 - GitHub PAT for cross-repo dispatch: [github.com/settings/personal-access-tokens/new](https://github.com/settings/personal-access-tokens/new)
+  - **Classic PAT**: check `repo` + `workflow` scopes.
+  - **Fine-grained PAT**: grant `Contents: write`, `Issues: write`, `Pull requests: write`, `Actions: write` on every product repo listed in `setup-pipeline --repos`. The orchestrator pushes fix commits and posts comments on those repos on your behalf.
 
 If you choose `codex-main`, template drift audit is enabled by default:
 - local check: `solo-cto-agent template-audit`
@@ -90,9 +153,9 @@ The point is simple:
 
 * less repetitive setup work
 * less context loss between sessions
-* less AI slop in code and design
-* more useful criticism before you commit to bad ideas
-* more initiative from the agent on low-risk work
+* two models cross-checking each other's review (not one model's opinion)
+* actual criticism before you commit to bad ideas
+* secrets caught before they leave your machine
 
 ## What changes in practice
 
@@ -116,12 +179,12 @@ This is running on three private repos (Next.js + Supabase, Vite + React, Next.j
 | PRs opened | 53 |
 | PRs merged | 48 |
 | Mean time to merge | 0.64 hours |
-| Test suite | 894 tests, 48 files, all passing |
+| Test suite | 996 tests, 57 files, all passing |
 | CLI commands | 25 subcommands |
 | Skills | 8 (44 reference docs) |
-| npm version | 1.2.0 |
+| npm version | 1.3.2 |
 
-What is not there yet: dual-agent cross-review metrics are still accumulating (rate shows 0 because the structured scoring pipeline was deployed recently). Decision tracking is wired but the decision queue has not produced enough data for meaningful stats. Those will fill in over the next few weeks of normal use.
+Dual-agent cross-review and Managed Agents deep review are live and tested against real diffs. Decision tracking is wired but the decision queue has not produced enough data for meaningful stats yet.
 
 ## Who this is for
 
@@ -246,7 +309,7 @@ solo-cto-agent/
     product-repo/             # product repo scaffold (workflows, STATE.md, .env.example)
     builder-defaults/         # routing-policy.json, agent-scores.json
     workflows/                # solo-cto-review.yml (3-pass auto-review)
-  tests/                      # 894 tests across 48 files
+  tests/                      # 996 tests across 57 files
   benchmarks/                 # effectiveness reports, metrics
   docs/                       # claude.md, tier-matrix, configuration, policies
   examples/                   # real-world flows: build, ship, review, founder-workflow
@@ -255,18 +318,26 @@ solo-cto-agent/
 
 ## Three Axes: Tier / Agent / Mode
 
+At a glance:
+
+|          | Cowork (semi-auto) | Codex (full-auto)     |
+|----------|--------------------|-----------------------|
+| Builder  | local + manual review | CI dispatch + auto-fix |
+| CTO      | local + dual-agent cross-review | CI dispatch + dual + cross-review + scoring |
+
+Cowork runs in your terminal with you in the loop. Codex runs in CI and reworks itself until the PR passes.
+
 `solo-cto-agent` is configured across three independent axes. You choose each based on your workflow.
 
 | Axis | Decision | Options |
 |---|---|---|
-| Tier | Scope of capability | Maker / Builder / CTO |
+| Tier | Scope of capability | Builder / CTO |
 | Agent | Who reviews | Cowork (Claude) / Cowork + Codex |
 | Mode | Automation depth | Semi-auto (cowork-main) / Full-auto (codex-main) |
 
 Quick pick if you are unsure:
-- Start with Maker + Cowork + Semi-auto.
-- Move to Builder when you are shipping real features.
-- Move to CTO + Full-auto when you want always-on CI/CD and multi-agent routing.
+- Start with Builder + Cowork (single Claude agent, semi-auto, optional Telegram bot).
+- Move to CTO + Full-auto when you want dual-agent cross-review and always-on CI/CD across repos.
 
 ### Agents (summary)
 
@@ -303,15 +374,13 @@ Full-auto adds:
 ### Tiers (summary)
 
 **Not sure which tier? One question:**
-- Are you shipping code to production? → **Builder** (default, recommended for most users)
-- Only doing idea validation / reviews? → **Maker**
-- Running multi-repo CI/CD with full automation? → **CTO**
+- Solo dev shipping code with one Claude agent? → **Builder** (default, recommended for most users)
+- Want dual-agent cross-review (Claude + Codex) and multi-repo CI/CD? → **CTO**
 
-| Tier | Includes | Recommended for |
-|---|---|---|
-| Maker | spark + review + memory + craft | idea and validation loops |
-| Builder | Maker + build + ship | solo dev shipping |
-| CTO | Builder + orchestrate | multi-agent + routing |
+| Tier | Includes | Agents | Extras | Recommended for |
+|---|---|---|---|---|
+| Builder | spark + review + memory + craft + build + ship | solo Claude | optional Telegram bot for PR notify/approve | solo dev shipping |
+| CTO | Builder + orchestrate | Claude + Codex (dual-agent cross-review) | agent scoring, routing, decision queue, daily briefing | multi-agent CI/CD across repos |
 
 Details: `docs/tier-matrix.md`, `docs/tier-examples.md`, `docs/cto-policy.md`, `docs/cowork-main-install.md`, `docs/configuration.md`.
 

package/bin/cli.js

@@ -30,6 +30,8 @@ let telegramBot;
 try { telegramBot = require("./telegram-bot"); } catch (_) { telegramBot = null; }
 let selfEvolve;
 try { selfEvolve = require("./self-evolve"); } catch (_) { selfEvolve = null; }
+let repoDiscovery;
+try { repoDiscovery = require("./repo-discovery"); } catch (_) { repoDiscovery = null; }
 
 const ROOT = path.resolve(__dirname, "..");
 const DEFAULT_CATALOG = path.join(ROOT, "failure-catalog.json");
@@ -65,6 +67,8 @@ function printHelp() {
 Usage:
   solo-cto-agent init [--force] [--preset maker|builder|cto] [--wizard]
   solo-cto-agent setup-pipeline --org <github-org> [--tier builder|cto] [--repos <repo1,repo2,...>]
+  solo-cto-agent repos list [--org <github-org>]      # show/re-pick the saved repo selection
+  solo-cto-agent do "<instruction>" [--dry-run] [--repo owner/name] [--agent claude|codex]
   solo-cto-agent setup-repo <repo-path> --org <github-org> [--tier builder|cto]
   solo-cto-agent auto-setup                 # Install solo-cto-pipeline.yml to your repos (centralized)
   solo-cto-agent setup --central --org <owner> [--orchestrator <repo>] [--repos <r1,r2,...>] [--dry-run]
@@ -93,6 +97,8 @@ Usage:
 Commands:
   init              Install skills to ~/.claude/skills/, then run doctor to verify setup
   setup-pipeline    Full pipeline setup: create orchestrator repo + install workflows to product repos
+  repos list        Print current saved repo selection (from init wizard) and re-pick interactively
+  do                Natural-language work order: LLM parses intent → creates labeled issue → worker runs
   setup-repo        Install dual-agent workflows to a single product repo
   auto-setup        Install solo-cto-pipeline.yml (centralized thin workflow) to selected repos
   setup --central   Centralize cross-repo workflows (digest, bot-runner) to orchestrator repo
@@ -401,6 +407,82 @@ Style: {{YOUR_STYLE}}
   doctorCommand({ exitOnError: false, quick: true });
 }
 
+// ─── repos: show/re-pick saved selection ────────────────────
+
+async function reposCommand(args) {
+  if (!repoDiscovery) {
+    console.error("❌ repo-discovery module not available in this install.");
+    process.exit(1);
+  }
+  const sub = args[1] || "list";
+  if (sub !== "list") {
+    console.error(`Unknown repos subcommand: ${sub}`);
+    console.error("Usage: solo-cto-agent repos list [--org <github-org>]");
+    process.exit(1);
+  }
+
+  const orgIndex = args.indexOf("--org");
+  let org = orgIndex >= 0 ? args[orgIndex + 1] : null;
+
+  const saved = repoDiscovery.loadSelection();
+  if (saved) {
+    console.log(`Saved selection (${repoDiscovery.selectionPath()}):`);
+    console.log(`  org:     ${saved.org || "(user-scoped)"}`);
+    console.log(`  updated: ${saved.updatedAt || "—"}`);
+    console.log(`  repos:   ${saved.selected.length ? saved.selected.join(", ") : "(none)"}`);
+    if (!org && saved.org) org = saved.org;
+  } else {
+    console.log("No saved selection yet. Run `solo-cto-agent init --wizard` first, or re-pick below.");
+  }
+
+  // Non-TTY callers (CI) just get the print-out.
+  const { isTTY, createRl } = require("./prompt-utils");
+  const { ask } = require("./prompt-utils");
+  if (!isTTY()) {
+    console.log("\nℹ️  Non-interactive terminal — skipping re-pick prompt.");
+    return;
+  }
+
+  const rl = createRl();
+  try {
+    const again = await ask(rl, "\nRe-pick repos now?", "n");
+    if (!/^y(es)?$/i.test(again.trim())) {
+      rl.close();
+      return;
+    }
+
+    let repos = null;
+    try {
+      repos = repoDiscovery.fetchRepos({ org });
+    } catch (err) {
+      console.log(`⚠️  ${err.message}`);
+    }
+
+    if (repos == null) {
+      console.log("`gh` CLI not found. Install from https://cli.github.com/ then `gh auth login`.");
+      const manual = await ask(rl, "Paste repo slugs manually (comma-separated, or blank to cancel)", "");
+      const selected = manual.split(",").map((s) => s.trim()).filter(Boolean);
+      if (selected.length) {
+        const file = repoDiscovery.saveSelection({ org, selected, discovered: [] });
+        console.log(`✅ Saved ${selected.length} repo(s) to ${file}`);
+      } else {
+        console.log("No changes.");
+      }
+    } else if (repos.length === 0) {
+      console.log("No repositories returned from gh.");
+    } else {
+      const preselected = saved && Array.isArray(saved.selected) && saved.selected.length
+        ? saved.selected
+        : repoDiscovery.defaultPreselect(repos);
+      const selected = await repoDiscovery.pickReposInteractive(rl, ask, repos, preselected);
+      const file = repoDiscovery.saveSelection({ org, selected, discovered: repos });
+      console.log(`✅ Saved ${selected.length} repo(s) to ${file}`);
+    }
+  } finally {
+    rl.close();
+  }
+}
+
 // ─── setup-pipeline: Full Pipeline Deploy ───────────────────
 
 function setupPipelineCommand(tier, org, repos, orchName, force) {
@@ -2061,6 +2143,17 @@ async function main() {
     return false;
   }
 
+  if (cmd === "repos") {
+    await reposCommand(args);
+    return;
+  }
+
+  if (cmd === "do") {
+    const doModule = require("./do");
+    await doModule.main();
+    return;
+  }
+
   if (cmd === "setup-pipeline") {
     if (checkCoworkMainMode()) {
       console.log("ℹ️  Not needed in cowork-main mode. Use `review`, `knowledge`, and `sync` commands instead.");
@@ -2071,7 +2164,15 @@ async function main() {
     const orgIndex = args.indexOf("--org");
     const org = orgIndex >= 0 ? args[orgIndex + 1] : null;
     const reposIndex = args.indexOf("--repos");
-    const repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    let repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    // Fall back to the selection persisted by `init --wizard` / `repos list`.
+    if (!repos && repoDiscovery) {
+      const saved = repoDiscovery.loadSelection();
+      if (saved && Array.isArray(saved.selected) && saved.selected.length) {
+        repos = saved.selected.join(",");
+        console.log(`ℹ️  Using saved repo selection (${saved.selected.length} repos from ${repoDiscovery.selectionPath()}).`);
+      }
+    }
     const orchIndex = args.indexOf("--orchestrator-name");
     const orchName = orchIndex >= 0 ? args[orchIndex + 1] : null;
     setupPipelineCommand(tier, org, repos, orchName, force);
@@ -2136,7 +2237,14 @@ async function main() {
     const orgIndex = args.indexOf("--org");
     const org = orgIndex >= 0 ? args[orgIndex + 1] : null;
     const reposIndex = args.indexOf("--repos");
-    const repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    let repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    if (!repos && repoDiscovery) {
+      const saved = repoDiscovery.loadSelection();
+      if (saved && Array.isArray(saved.selected) && saved.selected.length) {
+        repos = saved.selected.join(",");
+        console.log(`ℹ️  Using saved repo selection (${saved.selected.length} repos).`);
+      }
+    }
     const orchIndex = args.indexOf("--orchestrator-name");
     const orchName = orchIndex >= 0 ? args[orchIndex + 1] : null;
     upgradeCommand(org, repos, orchName);
@@ -2147,7 +2255,14 @@ async function main() {
     const orgIndex = args.indexOf("--org");
     const org = orgIndex >= 0 ? args[orgIndex + 1] : null;
     const reposIndex = args.indexOf("--repos");
-    const repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    let repos = reposIndex >= 0 ? args[reposIndex + 1] : null;
+    if (!repos && repoDiscovery) {
+      const saved = repoDiscovery.loadSelection();
+      if (saved && Array.isArray(saved.selected) && saved.selected.length) {
+        repos = saved.selected.join(",");
+        console.log(`ℹ️  Using saved repo selection (${saved.selected.length} repos).`);
+      }
+    }
     const orchIndex = args.indexOf("--orchestrator-name");
     const orchName = orchIndex >= 0 ? args[orchIndex + 1] : "dual-agent-orchestrator";
     const token = process.env.GITHUB_TOKEN || process.env.GH_TOKEN || process.env.ORCHESTRATOR_PAT;