@launchchair/agent 0.4.11 → 0.5.0

package/AGENTS.md

@@ -6,7 +6,7 @@ This file is for fresh agent chats that need to use the LaunchChair Agent API or
 
 Use LaunchChair first for product, research, validation, positioning, pricing, MVP blueprint, and phase-loop work. Do not inspect the local repo first unless the user is in Stack Setup, Build MVP, Design and Landing Page implementation, or another code-changing phase.
 
-Fresh LaunchChair projects often do not have a local repo yet. Before Stack Setup, the configured runner folder may be empty on purpose.
+Fresh LaunchChair projects often do not have a local repo yet. Before Stack Setup, each project uses its own isolated workspace under `~/.launchchair/projects`.
 
 ## Public Docs To Read
 
@@ -35,6 +35,8 @@ mkdir -p "$HOME/launchchair-agent"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/local-bridge.mjs" -o "$HOME/launchchair-agent/local-bridge.mjs"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/launchchair-cli.mjs" -o "$HOME/launchchair-agent/launchchair-cli.mjs"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/mcp-bridge-server.mjs" -o "$HOME/launchchair-agent/mcp-bridge-server.mjs"
+curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/mcp-capabilities.mjs" -o "$HOME/launchchair-agent/mcp-capabilities.mjs"
+curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/project-workspace.mjs" -o "$HOME/launchchair-agent/project-workspace.mjs"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/runner-command.mjs" -o "$HOME/launchchair-agent/runner-command.mjs"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/codex-runner.mjs" -o "$HOME/launchchair-agent/codex-runner.mjs"
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/claude-code-runner.mjs" -o "$HOME/launchchair-agent/claude-code-runner.mjs"
@@ -42,21 +44,8 @@ curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/chatgpt-runner.mjs" -
 curl -fsSL "https://www.launchchair.io/api/agent/v1/bridge/claude-runner.mjs" -o "$HOME/launchchair-agent/claude-runner.mjs"
 curl -fsSL "https://www.launchchair.io/agent-artifacts/AGENTS.md" -o "$HOME/launchchair-agent/AGENTS.md"
 chmod +x "$HOME/launchchair-agent"/*.mjs
-LAUNCHCHAIR_AGENT_WORKDIR="${LC_AGENT_CODEX_CWD:-${LC_AGENT_CLAUDE_CODE_CWD:-$PWD}}"
-if [ -d "$LAUNCHCHAIR_AGENT_WORKDIR" ] && [ ! -f "$LAUNCHCHAIR_AGENT_WORKDIR/AGENTS.md" ]; then
-  printf '%s\n' \
-    '# LaunchChair Agent Bootstrap Pointer' \
-    '' \
-    'Before searching this repo, read ~/launchchair-agent/AGENTS.md.' \
-    'Use LaunchChair MCP/API first for product phases, project discovery, project creation, and phase loops.' \
-    'Do not ask for extra confirmation before using LaunchChair MCP/API when the user already requested the action and the token permits it.' \
-    'Only inspect this repo for Stack Setup, Build MVP, Design/Landing Page, SEO implementation, smoke testing, security review, or code remediation.' \
-    > "$LAUNCHCHAIR_AGENT_WORKDIR/AGENTS.md"
-fi
 ```
 
-The final block creates a tiny project-local `AGENTS.md` pointer only when the target folder does not already have one. It helps fresh Codex chats auto-load the LaunchChair instructions without overwriting a real project instruction file.
-
 ## CLI
 
 The npm CLI is the recommended setup path for Codex Desktop and Claude Desktop. It can log in through LaunchChair, save the token locally, print paste-ready MCP config, write Claude Desktop config, and start the local bridge.
@@ -64,8 +53,9 @@ The npm CLI is the recommended setup path for Codex Desktop and Claude Desktop.
 ```bash
 npm install -g @launchchair/agent
 launchchair login
-launchchair setup codex --cwd /absolute/path/to/repo-or-empty-folder
-launchchair setup claude --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup codex
+launchchair setup claude
+launchchair bind --project-id PROJECT_UUID --cwd /absolute/path/to/verified-clone
 launchchair projects --query LeafLore
 launchchair create --name SignalLift --one-liner "Distribution workspace for technical founders" --project-type b2b --idempotency-key signallift-v1
 launchchair run-once
@@ -76,7 +66,7 @@ launchchair mcp-config codex --ui-fields
 If Codex Desktop is being configured through the custom MCP form instead of `~/.codex/config.toml`, run:
 
 ```bash
-launchchair setup codex --cwd /absolute/path/to/repo-or-empty-folder --copy-config
+launchchair setup codex --copy-config
 ```
 
 ## Environment Variables
@@ -86,15 +76,15 @@ LAUNCHCHAIR_BASE_URL="https://www.launchchair.io"
 LAUNCHCHAIR_PROJECT_ID="optional project UUID; omit or leave blank if creating or discovering a project"
 ```
 
-`launchchair login` is the recommended token path. It saves the token to `~/.launchchair/config.json`, and `launchchair mcp-server` reads that saved token. Use `LAUNCHCHAIR_AGENT_TOKEN` only as a manual fallback when browser login is skipped or a per-client token override is needed.
+`launchchair login` saves the token to `~/.launchchair/config.json`, shared by the CLI and MCP. Use `LAUNCHCHAIR_AGENT_TOKEN` as a manual fallback when browser login is skipped. A fresh saved login takes precedence over stale environment credentials; logout disables authentication until login. Use a separate `LAUNCHCHAIR_CONFIG_DIR` for a separate per-client account.
 The packaged MCP and local bridges also read the same saved CLI config as a fallback, so a desktop MCP launch should not require pasting a token into Codex or Claude config after browser login. Guided setup verifies standards-compliant MCP initialize and tool discovery before it reports success.
 
-For Codex Desktop local runs:
+For Codex Desktop local runs, choose the runner globally and bind each project clone independently:
 
 ```bash
 LC_AGENT_ROUTER_GPT_TARGET="codex"
 LC_AGENT_ROUTER_CODEX_TARGET="codex"
-LC_AGENT_CODEX_CWD="/absolute/path/to/the-folder-codex-should-build-into"
+launchchair bind --project-id PROJECT_UUID --cwd /absolute/path/to/verified-clone
 ```
 
 For Claude Desktop plus Claude Code local runs:
@@ -102,10 +92,10 @@ For Claude Desktop plus Claude Code local runs:
 ```bash
 LC_AGENT_ROUTER_GPT_TARGET="claude_code"
 LC_AGENT_ROUTER_CODEX_TARGET="claude_code"
-LC_AGENT_CLAUDE_CODE_CWD="/absolute/path/to/the-folder-claude-code-should-build-into"
+launchchair bind --project-id PROJECT_UUID --cwd /absolute/path/to/verified-clone
 ```
 
-The build folder can be an empty scratch folder until Stack Setup creates or records the GitHub repository. After Stack Setup, the user should clone the generated repo locally or update the runner CWD to the cloned repo path.
+Pre-code work uses an isolated per-project scratch folder. After Stack Setup records the GitHub repository, the bridge auto-detects matching clones in standard repo folders and verifies the Git `origin`. If needed, clone the repo and run `launchchair bind`; code-changing runs fail closed until that verification succeeds.
 
 ## MCP Server Config Example
 
@@ -124,22 +114,21 @@ LAUNCHCHAIR_CONFIG_DIR = "/absolute/path/to/.launchchair"
 LAUNCHCHAIR_PROJECT_ID = ""
 LC_AGENT_ROUTER_GPT_TARGET = "codex"
 LC_AGENT_ROUTER_CODEX_TARGET = "codex"
-LC_AGENT_CODEX_CWD = "/absolute/path/to/the-folder-codex-should-build-into"
 
 # Optional fallback if launchchair login was skipped:
 # LAUNCHCHAIR_AGENT_TOKEN = "<full lc_at token from Settings API Access>"
 ```
 
-Prefer `launchchair setup codex --cwd /absolute/path/to/build-folder` so the CLI writes native config and verifies MCP tool discovery. Use `launchchair mcp-config codex --ui-fields` only for manual custom-form setup.
+Prefer `launchchair setup codex` so the CLI writes native config and verifies MCP tool discovery. Use `launchchair bind --project-id PROJECT_UUID --cwd /absolute/path/to/verified-clone` after cloning when automatic discovery cannot locate the repo. Use `launchchair mcp-config codex --ui-fields` only for manual custom-form setup.
 
 ## What To Do First In A Fresh Chat
 
-1. If LaunchChair MCP tools are available, call `launchchair_heartbeat`.
-2. Call `launchchair_status` to confirm token, bridge, runner, and project context.
+1. Call `launchchair_status`. If login is required, call `launchchair_login`, present the approval URL, then poll `launchchair_login_status` at its returned interval.
+2. If paid access is required, present the returned upgrade URL and wait for the user to upgrade before resuming the blocked operation.
 3. If `LAUNCHCHAIR_PROJECT_ID` is blank and the user named a project, call `launchchair_list_projects` and select the matching project.
 4. If the user asks to create a new LaunchChair project, call `launchchair_create_project`.
-5. If the user clicked Run with agent in LaunchChair, call `launchchair_run_once`.
-6. If the user asks to advance the product workflow, call `launchchair_continue_project`.
+5. To pick up queued work or advance the workflow, call `launchchair_next_step`, execute its claimed prompt in the current host agent, and submit the result with `launchchair_complete_run`.
+6. Use `launchchair_run_once` or `executionMode=local_runner` only when the user explicitly requests a separate local runner.
 7. If LaunchChair asks for a user-choice bucket such as wedge, pricing, or feature direction, call `launchchair_select_choice`. Do not use generic SPEC_UPDATE apply for user-choice buckets.
 8. When the user asks for validation scores or wants to compare ideas, call `launchchair_get_scorecards`. Report validation scorecards separately from generated direction-option scores.
 
@@ -230,7 +219,7 @@ Avoid long explanations of local file searching when LaunchChair MCP tools are a
 Run a full project loop:
 
 ```text
-Call launchchair_continue_project with projectId, maxSteps, and optional stopPhaseKey.
+Call launchchair_next_step with projectId and optional stopPhaseKey, execute the claimed prompt with the host's tools, and call launchchair_complete_run. Repeat until the requested target or a genuine checkpoint.
 ```
 
 Create a project from a new idea:
@@ -245,18 +234,20 @@ Find a project by name:
 Call launchchair_list_projects with query set to the user-provided project name.
 ```
 
-Run one queued UI job:
+Claim one queued UI job for execution by the current agent:
 
 ```text
-Call launchchair_run_once.
+Call launchchair_next_step with projectId.
 ```
 
-Run one queued UI job from CLI:
+Run one queued UI job through an explicitly requested local runner:
 
 ```bash
 launchchair run-once
 ```
 
+Every MCP tool is also accessible through `launchchair call TOOL --json '{...}'` or `--input FILE`. Use `launchchair tools --json` for schemas. Tool results are JSON, tool errors exit nonzero, and token secrets remain in private files. Prefer host execution for agents with connected GitHub/Vercel/Supabase tools.
+
 Select a generated wedge:
 
 ```text
@@ -271,3 +262,15 @@ Call launchchair_select_choice with domain="wedge" and optionTitle or optionId.
 - Do not run production SQL without explicit user authorization and a known project ref.
 - Do not mark QA complete without matching evidence. For Feature projects, produce that evidence by running the repository-native or browser checks whenever possible.
 - Do not search the whole filesystem for LaunchChair source when the public docs and MCP tools are available.
+
+## MCP login and host execution
+
+The MCP can start without a token. Call `launchchair_login`, show the returned LaunchChair approval URL, and call `launchchair_login_status` after approval. Respect its poll interval. Tokens are stored privately in the local config and never returned to the conversation. Existing tokens need a fresh owner-approved login for `tokens.manage`. Use `launchchair_create_token` to save a scoped credential file, `launchchair_list_tokens` for redacted metadata, and `launchchair_revoke_token`/`launchchair_logout` for cleanup.
+
+Prefer `launchchair_next_step`, execute the claimed prompt in the current agent with its already-connected plugins, call `launchchair_complete_run`, and repeat. Do not spawn another model or local CLI unless explicitly requested (`executionMode=local_runner`). A waiting run is already claimed; poll it instead of executing it twice. Inspect forms/checklists with `launchchair_get_bucket`, save allowed patches with `launchchair_update_bucket`, and supply factual `executionEvidence` for setup and verification. Use `launchchair_get_starter` after validation to read the starter files and the host's GitHub/Vercel/Supabase tools to create and deploy the user's app. Bind the verified clone with `launchchair_bind_repository`. Use `stopPhaseKey=launch_sales` for the complete Product workflow.
+
+Preserve a supplied idea and specified requirements, even with low validation scores. `ideaIntent=preserve` is the default. Report risks and continue the requested work when no score gate was requested; never pivot merely to improve a score. Choose a generated direction consistent with the supplied idea. Only use `ideaIntent=explore` for explicit discovery or pivot requests.
+
+For “do not build until VERY high validation,” create candidates with `validationGate=very_high` (pain >=85, opportunity >=85, saturation <=30), validate through `market_validation`, and compare with `launchchair_compare_projects`. The gate is enforced on LaunchChair build/setup actions and cannot be weakened through generic mutations. Do not provision external resources until it passes. If no candidate qualifies, report the unmet thresholds honestly. Scores are research estimates, not a guarantee of market demand. Do not fabricate evidence or replace validation scores with direction-option scores.
+
+Research sources, competitor notes, Pexels assets, logo upload, form/checklist data, prompts, build cards, run management, and markdown exports are available as MCP tools. Tool descriptions distinguish data writes from actual external execution. Existing user authorization carries forward; ask only at a genuine missing authorization or information boundary.

package/README.md

@@ -4,24 +4,19 @@ LaunchChair Agent packages the LaunchChair CLI, MCP bridge, local bridge, and ru
 
 Use it when you want an agent to create or discover LaunchChair projects, run product phases, continue project loops, pick up `Run with agent` jobs from LaunchChair, or expose LaunchChair tools through MCP.
 
-## Install Locally For Testing
-
-From this package directory:
+## Install or update
 
 ```bash
-npm pack
-npm install -g ./launchchair-agent-*.tgz
-launchchair help
+npm install -g @launchchair/agent@latest
+launchchair --version
 launchchair doctor
 ```
 
-After publishing:
+Version 0.5.0 includes the expanded MCP, browser login, private token management, host-agent execution, idea preservation, and optional validation gates. Existing installations must update the package and restart their MCP connection to load the new tools.
 
-```bash
-npm install -g @launchchair/agent
-launchchair help
-launchchair doctor
-```
+After updating, run `launchchair setup codex --skip-login` or `launchchair setup claude --skip-login` to refresh the installed path and verify tool discovery. Then restart/reload that client. Your saved account login and project bindings remain in `~/.launchchair/config.json`. Older tokens still work within their existing scopes; run `launchchair login` again with workspace-owner approval if you need the new `tokens.manage` scope.
+
+For local release testing, run `npm pack ./packages/launchchair-agent --pack-destination /tmp` from the repository root and install the resulting tarball into an isolated npm prefix before publishing.
 
 ## Login
 
@@ -49,59 +44,60 @@ Tokens are masked in CLI output. Do not paste real tokens into public docs, scre
 
 ## Guided Setup
 
-Use `setup` as the easiest path. It saves the build folder, writes a small `AGENTS.md` pointer into that folder, starts browser login when no token is saved yet, writes the MCP config for your agent, and then proves the server can complete MCP `initialize` and `tools/list` before reporting success.
+Use `setup` as the easiest path. It starts browser login when no token is saved yet, writes the global MCP config for your agent, creates isolated per-project workspaces, and then verifies MCP discovery and an authenticated status call before reporting authenticated success. Unauthenticated discovery is reported separately as login required.
 
 Codex Desktop:
 
 ```bash
-launchchair setup codex --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup codex
 ```
 
 If you prefer Codex Desktop's custom MCP form, copy paste-ready fields instead:
 
 ```bash
-launchchair setup codex --cwd /absolute/path/to/repo-or-empty-folder --copy-config
+launchchair setup codex --copy-config
 ```
 
 Claude Desktop + Claude Code (Claude Code is registered at user scope when its CLI is installed):
 
 ```bash
-launchchair setup claude --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup claude
 ```
 
 Configure only one Claude client when needed:
 
 ```bash
-launchchair setup claude-desktop --cwd /absolute/path/to/repo-or-empty-folder
-launchchair setup claude-code --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup claude-desktop
+launchchair setup claude-code
 ```
 
 Terminal / CLI agents:
 
 ```bash
-launchchair setup terminal --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup terminal
 ```
 
 If you do not want browser login during setup, add `--skip-login`. You can also save an existing token during setup:
 
 ```bash
-launchchair setup codex --cwd . --skip-login
-launchchair setup codex --cwd . --token "<full lc_at token from LaunchChair Settings API Access>"
+launchchair setup codex --skip-login
+launchchair setup codex --token "<full lc_at token from LaunchChair Settings API Access>"
 ```
 
 Setup writes client configuration by default. Use `--copy-config`, `--ui-fields`, or `--print-config` only when you intentionally want manual configuration. Authenticated setup exits nonzero if the MCP handshake or required tool discovery fails; `--skip-mcp-check` is an explicit diagnostic escape hatch, not the recommended first-run path.
 
-For a fresh project before a repo exists, point `--cwd` at an empty folder. LaunchChair can run ideation, validation, positioning, pricing, and MVP blueprint before code exists; Stack Setup and Build MVP can later use that folder as the build target.
+Fresh projects automatically use an isolated folder under `~/.launchchair/projects/<project-id>/workspace` before a repo exists. After Stack Setup, LaunchChair looks for the repo by its recorded Git URL in standard local repo folders and verifies the `origin` before switching. If it cannot find the clone, code work stops with an exact clone and `launchchair bind` instruction instead of using another project.
 
-Use `setup terminal` when the agent runs inside a normal shell instead of Codex Desktop or Claude Desktop MCP. It writes the same project-local `AGENTS.md` pointer, saves the build folder, and prints the CLI commands the agent should use: `launchchair projects`, `launchchair create`, `launchchair continue`, `launchchair run-once`, and `launchchair bridge`.
+Use `setup terminal` when the agent runs inside a normal shell instead of Codex Desktop or Claude Desktop MCP. It installs the user-scoped bootstrap and prints the CLI commands the agent should use: `launchchair projects`, `launchchair create`, `launchchair continue`, `launchchair run-once`, and `launchchair bridge`.
 
 ## Common Commands
 
 ```bash
-launchchair setup codex --cwd /absolute/path/to/repo-or-empty-folder
-launchchair setup claude --cwd /absolute/path/to/repo-or-empty-folder
-launchchair setup claude-code --cwd /absolute/path/to/repo-or-empty-folder
-launchchair setup terminal --cwd /absolute/path/to/repo-or-empty-folder
+launchchair setup codex
+launchchair setup claude
+launchchair setup claude-code
+launchchair setup terminal
+launchchair bind --project-id <project-id> --cwd /absolute/path/to/verified-clone
 launchchair projects
 launchchair projects --query LeafLore
 launchchair scorecards --project-id <project-id>
@@ -114,9 +110,32 @@ launchchair bridge
 launchchair mcp-config codex
 launchchair mcp-config codex --ui-fields --copy-config
 launchchair mcp-config claude
-launchchair install-agent-md --cwd /absolute/path/to/repo-or-empty-folder
+launchchair install-agent-md --cwd /absolute/path/to/repo
+```
+
+## Use any MCP tool from the terminal
+
+`tools` discovers the same catalog as your MCP connection (47 tools when authenticated). `call` uses the bundled MCP implementation, including its login, permission checks, private credential storage, and host-execution defaults. Tool errors return a nonzero exit status and preserve details such as an upgrade URL.
+
+```bash
+launchchair tools --json
+launchchair call status
+launchchair call next_step --json '{"projectId":"PROJECT_UUID"}'
+launchchair call complete_run --input completed-run.json
+launchchair call list_tokens
+launchchair call create_token --json '{"name":"Build agent","scopes":["spec.read"]}'
+```
+
+Tool names work with or without the `launchchair_` prefix. Use `--input -` to read JSON arguments from stdin. `next_step` returns a claimed prompt for your current agent to execute with its connected plugins; it does not start another agent. `complete_run` submits the actual result and execution evidence. Token creation returns a private credential-file path, never the secret in terminal output.
+
+Project creation preserves a supplied idea by default, even when validation is low. Only request exploration and a score gate when that matches the user's intent:
+
+```bash
+launchchair create --name "Groomer candidate" --one-liner "Reduce missed appointments" --idea-intent explore --validation-gate very_high --save-project
 ```
 
+The gate requires pain >=85, opportunity >=85, and saturation <=30 before build/setup. For custom thresholds or the full Feature input schema, use `launchchair call create_project` or `launchchair call create_feature_project` with JSON arguments.
+
 Feature projects can be created and planned from the CLI for every authenticated LaunchChair workspace:
 
 ```bash
@@ -157,7 +176,7 @@ launchchair mcp-config claude
 
 The recommended config launches the MCP server through the exact Node executable and CLI file path generated by `launchchair setup` or `launchchair mcp-config`. It also pins `LAUNCHCHAIR_CONFIG_DIR` to the exact directory used by `launchchair login`, so Codex and Claude users do not need to paste tokens into MCP config after browser login—even when the desktop app starts with a different home or shell environment. The server uses standard newline-delimited JSON-RPC over stdio, as required by MCP clients.
 
-Use `LAUNCHCHAIR_AGENT_TOKEN` only as a manual fallback when you skip `launchchair login`, need a per-client token override, or your desktop app cannot read the saved CLI config. If you use a desktop custom MCP form, run `launchchair mcp-config codex --ui-fields` or `launchchair mcp-config claude --ui-fields` and paste the printed command, each argument, env vars, and working directory exactly as shown.
+Use `LAUNCHCHAIR_AGENT_TOKEN` as a manual fallback when you skip `launchchair login` or your desktop app cannot read the saved config. A saved browser login takes precedence over stale environment credentials. Logout disables both CLI and MCP authentication until another login. For a separate per-client account, use a separate `LAUNCHCHAIR_CONFIG_DIR`. If you use a desktop custom MCP form, run `launchchair mcp-config codex --ui-fields` or `launchchair mcp-config claude --ui-fields` and paste the printed command, each argument, env vars, and working directory exactly as shown.
 
 ## Local Bridge
 
@@ -175,17 +194,17 @@ launchchair run-once
 
 ## Runner Targets
 
-Codex:
+Codex routes are selected globally, while repository paths are bound per LaunchChair project:
 
 ```bash
-launchchair config set codex-cwd /absolute/path/to/repo-or-empty-folder
+launchchair bind --project-id <project-id> --cwd /absolute/path/to/verified-clone
 launchchair bridge
 ```
 
 Claude Code:
 
 ```bash
-launchchair config set claude-code-cwd /absolute/path/to/repo-or-empty-folder
+launchchair bind --project-id <project-id> --cwd /absolute/path/to/verified-clone
 LC_AGENT_ROUTER_GPT_TARGET=claude_code LC_AGENT_ROUTER_CODEX_TARGET=claude_code launchchair bridge
 ```
 
@@ -206,3 +225,15 @@ This package includes `AGENTS.md`. Fresh Codex, Claude, and Hermes chats should
 - MCP docs: https://www.launchchair.io/docs/mcp
 - Agent API docs: https://www.launchchair.io/docs/agent-api
 - Machine-readable catalog: https://www.launchchair.io/api/agent/v1/docs
+
+## MCP login and host execution
+
+The MCP can start without a token. Call `launchchair_login`, show the returned LaunchChair approval URL, and call `launchchair_login_status` after approval. Respect its poll interval. Tokens are stored privately in the local config and never returned to the conversation. Existing tokens need a fresh owner-approved login for `tokens.manage`. Use `launchchair_create_token` to save a scoped credential file, `launchchair_list_tokens` for redacted metadata, and `launchchair_revoke_token`/`launchchair_logout` for cleanup.
+
+Prefer `launchchair_next_step`, execute the claimed prompt in the current agent with its already-connected plugins, call `launchchair_complete_run`, and repeat. Do not spawn another model or local CLI unless explicitly requested (`executionMode=local_runner`). A waiting run is already claimed; poll it instead of executing it twice. Inspect forms/checklists with `launchchair_get_bucket`, save allowed patches with `launchchair_update_bucket`, and supply factual `executionEvidence` for setup and verification. Use `launchchair_get_starter` after validation to read the starter files and the host's GitHub/Vercel/Supabase tools to create and deploy the user's app. Bind the verified clone with `launchchair_bind_repository`. Use `stopPhaseKey=launch_sales` for the complete Product workflow.
+
+Preserve a supplied idea and specified requirements, even with low validation scores. `ideaIntent=preserve` is the default. Report risks and continue the requested work when no score gate was requested; never pivot merely to improve a score. Choose a generated direction consistent with the supplied idea. Only use `ideaIntent=explore` for explicit discovery or pivot requests.
+
+For “do not build until VERY high validation,” create candidates with `validationGate=very_high` (pain >=85, opportunity >=85, saturation <=30), validate through `market_validation`, and compare with `launchchair_compare_projects`. The gate is enforced on LaunchChair build/setup actions and cannot be weakened through generic mutations. Do not provision external resources until it passes. If no candidate qualifies, report the unmet thresholds honestly. Scores are research estimates, not a guarantee of market demand. Do not fabricate evidence or replace validation scores with direction-option scores.
+
+Research sources, competitor notes, Pexels assets, logo upload, form/checklist data, prompts, build cards, run management, and markdown exports are available as MCP tools. Tool descriptions distinguish data writes from actual external execution. Existing user authorization carries forward; ask only at a genuine missing authorization or information boundary.