Clarifies that spawn agents use remote LLM APIs, not local inference,
which is why cheap CPU instances suffice and GPU clouds are unnecessary.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Spawn runs coding agents that call LLM APIs — they need affordable
CPU instances with SSH access, not expensive GPU VMs. Updated the
discovery prompts and CLAUDE.md to explicitly avoid GPU clouds and
prioritize budget VPS providers and container platforms.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rename the GitHub workflow, scripts, and service from "improve" to
"discovery" to better reflect what the automation does. Remove the
`spawn improve` CLI command entirely — the discovery/refactor loops
are internal automation, not user-facing CLI features.
File renames:
- .github/workflows/improve.yml → discovery.yml
- .claude/skills/.../improve.sh → discovery.sh
- .claude/skills/.../start-improve.sh → start-discovery.sh
- Service: improve-trigger → discovery-trigger
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Bumped CLI version from 0.2.4 to 0.2.5
- Added rule to CLAUDE.md: ANY change to cli/ requires a version bump
- Uses semantic versioning (patch for fixes, minor for features, major for breaking)
- Auto-update ensures users get latest version immediately
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
The 663-line bash CLI (spawn.sh) has drifted from the TypeScript CLI,
missing --prompt, security validation, download fallback, and other
features. Rather than maintaining two implementations, the installer
now auto-installs bun (~5 seconds) when it's not present, ensuring
every user gets the full-featured TypeScript CLI.
- Remove cli/spawn.sh (663 lines)
- Simplify install.sh: remove npm method, add bun auto-install
- Extract build_and_install() helper to deduplicate build logic
- Update cli/README.md and CLAUDE.md to reflect bun-only strategy
Co-authored-by: A <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Documents the dual-mode cycle system (issue vs refactor), concurrency
model, worktree isolation, and guidance for modifying the service.
Also adds trigger service files to the file structure convention.
Co-authored-by: A <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
PRs created by autonomous loops must always be either merged or
closed with a comment explaining why. Updates improve.sh, refactor.sh,
and CLAUDE.md to enforce this consistently.
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Remove testing documentation file per documentation policy
- Add rule to CLAUDE.md: docs must go in .docs/ directory (git-ignored)
- Only README.md, CLAUDE.md, and cloud READMEs allowed in repo
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
- Convert all test files to use bun:test instead of vitest
- Update CLAUDE.md to prohibit vitest, mandate bun:test
- Replace vi.fn() with mock() from bun:test
- Replace vi.spyOn with spyOn from bun:test
- Note: commands.test.ts needs module mocking refactor (TODO)
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
improve.sh:
- Cloud Scouts get 2 slots, Agent Scout gets 1 (and only if justified)
- Agent discovery requires evidence: 1000+ GH stars, 50+ HN points,
100+ Reddit upvotes — at least 2 of these
- New Issue Responder role checks gh issues for user requests
- Agents must be searched on HN Algolia API, Reddit, GitHub before adding
- Priority order: fill gaps > new clouds > new agents > issue triage
CLAUDE.md:
- Reordered: clouds are priority #2, agents are #3 with strict gate
- Added community buzz requirements with specific thresholds
- Added GitHub issue response as #4
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add gptme agent to spawn matrix
Add gptme (https://github.com/gptme/gptme) - a personal AI agent in the
terminal with tools for code editing, terminal commands, web browsing,
and more. Natively supports OpenRouter via OPENROUTER_API_KEY.
- Add gptme agent entry to manifest.json with OpenRouter env vars
- Implement sprite/gptme.sh deployment script
- Implement hetzner/gptme.sh deployment script
- Add "missing" matrix entries for remaining 8 clouds
- Update README.md with usage instructions for Sprite and Hetzner
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add Fly.io cloud provider with claude and aider agents
Add Fly.io as a new cloud provider using the Machines REST API for
provisioning and flyctl CLI for SSH access. Docker-based machines
with pay-per-second pricing.
- Create fly/lib/common.sh with Fly.io Machines API integration
- Implement fly/claude.sh for Claude Code deployment
- Implement fly/aider.sh for Aider deployment
- Update README.md with Fly.io usage instructions and env vars
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add gemini, amazonq, cline, gptme to Fly.io
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add openclaw, nanoclaw, goose, codex, interpreter to Fly.io
Implements 5 new agent scripts for the Fly.io cloud provider:
- fly/openclaw.sh: OpenClaw with gateway + TUI, model selection, config
- fly/nanoclaw.sh: NanoClaw WhatsApp agent with .env configuration
- fly/goose.sh: Block's Goose agent with OpenRouter provider
- fly/codex.sh: OpenAI Codex CLI with OpenRouter base URL override
- fly/interpreter.sh: Open Interpreter with OpenRouter base URL override
All scripts follow the Fly.io pattern (flyctl-based, no IP args for
run_server/interactive_session) and use upload_file for env injection.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat: add gptme agent to 8 remaining clouds
Implement gptme agent scripts for digitalocean, vultr, linode, lambda,
aws-lightsail, gcp, e2b, and modal. Each script follows the exact
pattern of that cloud's existing aider.sh, adapted for gptme's install
and launch commands. Updates manifest.json matrix entries from "missing"
to "implemented".
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Add guardrails from insights: CLAUDE.md rules, hooks, pre-commit
Based on usage insights analysis:
CLAUDE.md:
- Shell script rules: curl|bash compat, macOS bash 3.x compat
- Autonomous loop rules: test after each iteration, never revert fixes
- Git workflow rules: always use feature branches
.claude/settings.json:
- PostToolUse hook validates .sh files on every Write/Edit:
syntax check, no relative source, no echo -e, no set -u
.githooks/pre-commit:
- Blocks commits with: syntax errors, relative sources, echo -e,
set -euo, references to deleted functions
- Install: git config core.hooksPath .githooks
README.md:
- Added developer setup section with hook installation
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The autonomous refactoring added `set -euo pipefail` but the scripts
check optional env vars with `[[ -n "$VAR" ]]` which is a fatal error
under nounset when the var isn't set (e.g. SPRITE_NAME, OPENROUTER_API_KEY).
Fix: downgrade to `set -eo pipefail` across all 42 affected files.
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Three issues broke the OAuth callback server on macOS:
1. echo -e doesn't work in bash 3.x — \r\n appears as literal text
in the HTTP response, browser gets malformed headers.
Fix: pre-write response with printf to a file before the subshell.
2. local variables inside ( ... ) & subshell — undefined behavior in
bash 3.x since subshells aren't function scope.
Fix: use plain variables in subshells.
3. ((elapsed++)) when elapsed=0 evaluates to falsy — set -e kills
the script on the first iteration of the timeout loop.
Fix: use elapsed=$((elapsed + 1)) instead.
Also simplified nc_listen detection to only check for BusyBox
(the -p flag check could misfire on macOS nc).
Applied to all 10 lib/common.sh files.
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Rewrites improve.sh to use the experimental Agent Teams feature:
- Lead coordinates in delegate mode (never touches code)
- Teammates work in parallel: Gap Fillers, Agent Scouts, Cloud Scouts
- Shared task list for self-claiming work
- Plan approval required for cloud provider work (lib/common.sh)
CLAUDE.md updated with team role definitions and coordination rules.
.claude/settings.json enables CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS.
Usage:
./improve.sh # one team cycle
./improve.sh --loop # continuous team cycles
./improve.sh --single # old single-agent mode (fallback)
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- manifest.json: tracks agents x clouds matrix with metadata
- CLAUDE.md: instructions for Claude Code to fill gaps and discover new agents/clouds
- improve.sh: loop script that launches Claude Code to expand the matrix
Current matrix: 3 agents (claude, openclaw, nanoclaw) x 2 clouds (sprite, hetzner)
with 2 gaps remaining (hetzner/openclaw, hetzner/nanoclaw).
Co-authored-by: Sprite <noreply@sprite.dev>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
