The CI biome check only covered packages/cli/src/, .claude/scripts/,
and .claude/skills/setup-spa/ — packages/shared/src/ was unchecked,
allowing 7 lint/format violations to accumulate in its test files.
- Auto-fix import ordering, formatting, and useNumberNamespace lint
across 3 test files in packages/shared/src/__tests__/
- Add packages/shared/src/ to the biome check in lint.yml so future
violations are caught in CI
Co-authored-by: B <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* fix: destroy orphaned Packer builder instances on workflow cancel
When a Packer Snapshots workflow is cancelled mid-build, Packer's process
is killed before it can clean up its temporary builder droplet/server.
This leaves orphaned packer-* instances running and costing money.
Add `if: cancelled()` cleanup steps for both DigitalOcean and Hetzner
that destroy any packer-* prefixed instances after cancellation.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* chore: remove Hetzner cleanup step — only DO needed
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: remove Hetzner from Packer snapshots, add cancel cleanup
Remove Hetzner from the Packer workflow entirely — only DigitalOcean
snapshots are built. Deletes packer/hetzner.pkr.hcl and simplifies the
workflow by removing all Hetzner-specific steps and cloud conditionals.
Also adds a cancelled() cleanup step that destroys orphaned packer-*
builder droplets when a workflow run is cancelled mid-build.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: add missing sprite-keep-running.sh script
The keep-alive install was 404ing because sh/shared/sprite-keep-running.sh
never existed in the repo. The TypeScript code downloaded it from the CDN
(which maps to sh/shared/) but the file was never created.
The script wraps a command and pings the sprite's own public URL every 30s
to prevent inactivity shutdown. It resolves the URL via sprite-env info
(available on all sprites) and falls back to exec without keep-alive if
the URL can't be determined.
Also removes Hetzner from the Packer snapshots workflow entirely — only
DigitalOcean snapshots are built.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: address security review — scope cleanup filter, fix JSON injection
1. Add `spawn-packer` tag to DO builder droplets in Packer template and
filter cleanup by tag instead of broad `packer-` name prefix. Prevents
accidentally destroying builder instances from other concurrent builds.
2. Use `jq --arg` for SINGLE_AGENT_INPUT instead of string interpolation
to prevent JSON injection via crafted agent names.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The original bunfig.toml used `line` and `function` (singular) which Bun
silently ignores. The correct field names are `lines` and `functions` (plural).
Changes:
- Fix field names: line→lines, function→functions
- Set thresholds: lines=0.35 (floor: digitalocean.ts 38.5%), functions=0.5
(floor: preload.ts 50%)
- Add coverageSkipTestFiles=true
- Keep --coverage in CI (bunfig thresholds enforce exit code on failure)
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: L <6723574+louisgv@users.noreply.github.com>
- Move bunfig.toml to repo root with valid coverageThreshold syntax
(line=80%, function=0 to avoid per-file false positives)
- Add --coverage flag to CI test step
- Delete packages/cli/bunfig.toml (superseded by root config)
- Add tests for packages/shared (type-guards, parse, result)
- Colocate billing config into each cloud directory (aws/billing.ts,
gcp/billing.ts, hetzner/billing.ts, digitalocean/billing.ts)
- Refactor billing-guidance.ts: BillingConfig interface replaces
cloud-string-keyed Record maps
- Bump CLI version to 0.25.1
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: L <6723574+louisgv@users.noreply.github.com>
* feat: never-give-up resilience layer — retry every failure instead of exiting
Add retryOrQuit() helper to shared/ui.ts that prompts "Try again? (Y/n)"
after any recoverable failure. Wrap all fatal exit points with retry loops:
- Cloud auth (Hetzner, DigitalOcean, AWS, GCP): retry after 3 failed tokens
- API key acquisition: retry after 3 failed OAuth+manual attempts
- Server creation: retry on any createServer failure (both fast & sequential)
- SSH readiness: retry on waitForReady timeout
- Agent install: retry on install failure
- Pre-launch hooks: retry on preLaunch failure
Non-interactive mode (SPAWN_NON_INTERACTIVE=1) still throws immediately.
Ctrl+C at any retry prompt exits cleanly.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(e2e): add AI-driven interactive test harness
Add --interactive mode to the E2E test framework. Instead of running spawn
in headless mode (SPAWN_NON_INTERACTIVE=1), this spawns the CLI in a real
PTY and uses Claude Haiku to respond to prompts like a human user would.
New files:
- sh/e2e/interactive-harness.ts — Bun script that drives the PTY + AI loop
- sh/e2e/lib/interactive.sh — Bash integration with the E2E framework
Usage:
e2e.sh --cloud hetzner claude --interactive
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(qa): wire interactive E2E into scheduled QA pipeline
- Add `e2e-interactive` option to workflow_dispatch in qa.yml
- Add `e2e-interactive` run mode to qa.sh (loads cloud creds + ANTHROPIC_API_KEY)
- Runs `e2e.sh --cloud hetzner claude --interactive` directly (no Claude Code needed)
- Defaults to hetzner (cheapest), overridable via E2E_INTERACTIVE_CLOUD/AGENT env vars
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat(qa): schedule interactive E2E daily at 6am UTC
Runs one agent (claude) on one cloud (hetzner) with AI-driven prompts.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(qa): offset soak cron to avoid GitHub Actions schedule dedup
GitHub Actions deduplicates overlapping cron schedules into one run,
making `github.event.schedule` unpredictable. The soak test at `0 3 * * 1`
was getting absorbed by the `0 */4 * * *` quality sweep and never firing
as reason=soak.
Move soak to `30 1 * * 1` (Monday 1:30am UTC) — safely between the
0am and 4am quality sweep slots. Interactive E2E at `0 6 * * *` is
already safe (between the 4am and 8am slots).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(qa): add e2e-interactive to trigger server valid reasons
The trigger server validates reason query params against an allowlist.
Without this, the `e2e-interactive` dispatch returns 400.
Also note: `soak` is already in VALID_REASONS in the repo but the running
service on the QA VM is stale — needs a restart to pick up both soak and
e2e-interactive reasons.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The old-asset cleanup pipeline `gh release view | grep | while` fails
when grep finds no matches (exit 1) and pipefail is set. This kills
the entire step before gh release upload runs.
Fix: wrap grep in `{ grep ... || true; }` so no-match is not fatal.
This caused all arm64 builds and some x86_64 builds to fail nightly.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The nested comprehension `[($agents[] | . as $a) | ...]` is invalid jq.
Use `[$agents[] as $a | $clouds[] as $c | ...]` instead.
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
CLI changes:
- Add findSpawnSnapshot() to query Hetzner /images?type=snapshot API
for pre-built spawn-{agent}-* images (matches by description prefix)
- Add waitForSshOnly() for snapshot boots (skips cloud-init polling)
- Update createServer() to accept optional snapshotId — boots from
snapshot instead of ubuntu-24.04, skips cloud-init userdata
- Wire up orchestrator with skipAgentInstall flag
Packer changes:
- Add packer/hetzner.pkr.hcl using hcloud plugin, mirroring the DO
template (tier scripts, agent install, cleanup, manifest)
- Unify packer-snapshots.yml to build both DO and Hetzner in a single
workflow with cloud×agent matrix and per-cloud cleanup steps
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: increase packer snapshot transfer timeout to 60m
The default 30m timeout is too short for transferring snapshots to
distant DO regions (blr1, sgp1, syd1). This caused zeroclaw and
kilocode builds to fail despite successful provisioning.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* revert: remove batch splitting from packer workflow
DO droplet cap is no longer an issue — revert to single parallel build
job for all agents.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Splits the 8 agents into 2 sequential batches of 4 so we stay under
DigitalOcean's concurrent droplet creation limit. Batch 2 waits for
batch 1 to finish before starting. Single-agent builds are unaffected.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: A <258483684+la14-1@users.noreply.github.com>
The junie.Dockerfile was added in PR #2601 but the docker.yml workflow
matrix was not updated, so no Docker image for junie was ever being built.
Add junie to the agent list so ghcr.io/openrouterteam/spawn-junie gets
built alongside all other agents.
Co-authored-by: spawn-qa-bot <qa@openrouter.ai>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
- soak.sh: SOAK_CLOUD env var makes cloud configurable (default: sprite)
- qa.sh: load TELEGRAM_BOT_TOKEN, TELEGRAM_TEST_CHAT_ID, SOAK_CLOUD from
/etc/spawn-qa-auth.env in soak mode
- qa.yml: add weekly Monday 3am UTC scheduled soak trigger
- fix: bun eval → bun -e across soak.sh, key-request.sh, github-auth.sh
(bun eval is not a valid subcommand in bun 1.3.9)
- fix: export _TOKEN via env prefix so process.env._TOKEN works in bun -e
- docs: update shell-scripts.md rule to say bun -e (not bun eval)
Verified: 3/4 Telegram tests pass in smoke test on DigitalOcean (120s wait)
getMe ✓ sendMessage ✓ getWebhookInfo ✓; cron test needs full 55-min window.
Co-authored-by: spawn-qa-bot <qa@openrouter.ai>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Add a soak test that provisions OpenClaw on Sprite, waits 1 hour for
stabilization, injects a Telegram bot token, and runs integration tests
against the Telegram Bot API (getMe, sendMessage, getWebhookInfo).
- New: sh/e2e/lib/soak.sh — soak test library with all Telegram-specific logic
- Modified: sh/e2e/e2e.sh — add --soak flag to arg parser
- Modified: qa.sh — add soak run mode (bypasses Claude, runs e2e.sh directly)
- Modified: trigger-server.ts — add "soak" to VALID_REASONS
- Modified: qa.yml — add soak to workflow_dispatch options
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: A <258483684+la14-1@users.noreply.github.com>
The cli-release workflow was deleting releases before recreating them,
leaving a window where users downloading cloud bundles (gcp.js, aws.js,
etc.) would get a 404. This affected all clouds on every push to main.
Switch to gh release upload --clobber which atomically replaces assets
without removing the release, and only create releases if they don't
already exist.
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use Google Chrome .deb instead of Playwright for OpenClaw browser
Snap Chromium on Ubuntu 24.04 fails because AppArmor confinement blocks
CDP control. OpenClaw's own docs recommend installing Google Chrome via
.deb package which bypasses snap entirely.
Also adds browser.noSandbox and browser.executablePath to the OpenClaw
config so the browser tool works out of the box on Linux VMs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: remove unnecessary confirmation prompt when OAuth fails
If OAuth didn't complete, the user obviously wants to paste a key.
The "Paste your API key manually? (Y/n)" prompt was pointless friction.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: remove unnecessary "Continue anyway?" credential confirmation
If the user selected a cloud, they obviously want to continue.
The warning + setup guidance is sufficient — no need to block on a confirm.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: move Chrome install to configure step so it runs after tarball
The tarball path skips agent.install() entirely, so Chrome never got
installed. Moving it to configure() (setupOpenclawConfig) ensures it
always runs regardless of install method.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: bundle Google Chrome in openclaw tarball
Add Chrome .deb install to openclaw's tarball build so it ships
pre-installed. Capture /usr/bin/google-chrome and /opt/google/chrome/
in the tarball. Add dl.google.com to the workflow domain allowlist.
The configure() step still has a fallback install with idempotency
check (command -v google-chrome) for non-tarball installs.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use openclaw config set for browser setup + correct binary name
- Use `google-chrome-stable` (actual .deb binary name) not `google-chrome`
- Set browser config via `openclaw config set` CLI (the supported way)
instead of writing JSON directly which wasn't being picked up
- Remove browser section from JSON config to avoid conflicts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Packer template:
- Match official 90-cleanup.sh: remove SSH host keys, create
revoked_keys, remove cloud-init instances, zero-fill free space,
use --force-confold for upgrades, autoremove/autoclean
- Add Packer manifest post-processor for snapshot ID extraction
- Remove PACKER_LOG=1 (debug logging not needed in production)
Workflow:
- Add "Submit to DO Marketplace" step after successful build
- Reads agent→app_id mapping from MARKETPLACE_APP_IDS secret (JSON)
- Extracts snapshot ID from Packer manifest, PATCHes Vendor API
- Gracefully handles 400 (app already pending review)
- Skips silently if no MARKETPLACE_APP_IDS secret is configured
Setup: add MARKETPLACE_APP_IDS secret as JSON, e.g.:
{"claude":"60089fc6...", "codex":"60089fc7..."}
App IDs come from the DO Vendor Portal after initial approval.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* fix: claude snapshot build — remove npm fallback from install command
The native install (curl | bash) succeeds but exits non-zero due to a
PATH warning. The || fallback then tries `npm install` which doesn't
exist on the "minimal" tier → exit 127.
Fix: replace npm fallback with binary existence check (same pattern
as hermes agent). If install exits non-zero but ~/.local/bin/claude
exists, the build succeeds.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: snapshot cleanup and lookup — use name prefix instead of tags
DO Packer builder `tags` only apply to the temporary build droplet,
not the resulting snapshot image. Both the workflow cleanup step and
the CLI's findSpawnSnapshot() were querying by `tag_name` which
returned nothing — old snapshots piled up and the CLI couldn't find
existing snapshots.
Fix: filter by snapshot name prefix (`spawn-{agent}-`) instead of
tags, in both the workflow and the CLI. Remove misleading `tags`
from the Packer template. Add test cases for name-prefix filtering.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* refactor: remove packages/shared, deduplicate with packages/cli/src/shared
packages/shared duplicated packages/cli/src/shared (parse.ts, result.ts,
type-guards.ts) with the CLI never importing from the shared package.
The only consumer was .claude/skills/setup-spa, which now imports directly
from packages/cli/src/shared via relative paths.
- Delete packages/shared entirely
- Update setup-spa imports to use relative paths to CLI shared
- Remove @openrouter/spawn-shared workspace dependency from setup-spa
- Update CLAUDE.md and type-safety.md references
Agent: complexity-hunter
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* fix: remove packages/shared from lint workflow, fix import sorting
The Biome Lint CI step referenced packages/shared/src/ which no longer
exists after this PR removes the package. Also fix import ordering in
setup-spa files to satisfy Biome's organizeImports rule.
Agent: pr-maintainer
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* fix: address Devin review — update stale packages/shared references
- Update type-safety.md line 67: packages/shared/src/parse.ts → packages/cli/src/shared/parse.ts
- Update install.ps1 sparse-checkout: remove packages/shared reference
Agent: pr-maintainer
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
---------
Co-authored-by: B <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
Restores the nightly Packer snapshot build pipeline (reverted in #2205)
that pre-bakes agent images as DigitalOcean snapshots. When a snapshot
exists on the user's account, droplet boot skips cloud-init and tarball
install entirely — cutting provisioning from ~10min to ~2min.
- Add `packer/digitalocean.pkr.hcl` HCL2 template with multi-region
distribution, apt-lock wait, and snapshot marker
- Add `.github/workflows/packer-snapshots.yml` nightly build with
matrix strategy, auto-cleanup of old snapshots, and injection-safe
env var handling
- Add `findSpawnSnapshot()` to query DO API for pre-built snapshots
- Add `waitForSshOnly()` for snapshot boots (skip cloud-init wait)
- Modify `createServer()` to accept optional `snapshotId` param
- Wire snapshot detection in DO `main.ts` orchestrator
- Add `skipAgentInstall` to `CloudOrchestrator` interface to skip
tarball + install steps when booting from snapshot
- Add 5 unit tests for snapshot lookup (happy path, empty, error,
invalid ID, network failure)
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: ARM tarball builds + arch-aware download
- Add ARM64 matrix entries for native binary agents (zeroclaw, opencode,
hermes, claude) in agent-tarballs.yml workflow
- Update agent-tarball.ts to detect remote VM arch via uname -m and
download the correct tarball (x86_64 or arm64)
- Change release strategy to support multiple arch assets per tag
- Document ARM build requirements in discovery.md for future agents
- Bump CLI version to 0.15.2
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use sudo for tarball extraction on non-root SSH clouds
On AWS Lightsail, SSH connects as 'ubuntu' (not root), but tarballs
extract to /root/. Without sudo, tar fails with "Permission denied".
Conditionally use sudo when not running as root (id -u != 0).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
- Use sudo mv + chown for tarball in release step (root-owned from capture)
- Skip swapfile creation if /swapfile already exists (GitHub Actions runners)
- Tolerate hermes setup wizard failure when /dev/tty unavailable in CI
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: pre-built agent tarballs on GitHub Releases for fast install
Adds a nightly GitHub Actions workflow that builds and uploads agent
tarballs to rolling GitHub Releases. During provisioning, the CLI now
attempts to download and extract a tarball before falling back to live
install. Priority chain: snapshot > tarball > live install.
- New workflow: .github/workflows/agent-tarballs.yml
- New capture script: packer/scripts/capture-agent.sh
- New module: packages/cli/src/shared/agent-tarball.ts
- Orchestrate tries tarball first on non-local clouds
- Skip tarball when using DO snapshot (skipTarball flag)
- Tests for tarball install + orchestration integration
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use global.fetch mock pattern and address security review
- Use `global.fetch = mock(...)` instead of `spyOn(globalThis, "fetch")`
to match codebase convention and fix CI mock interception
- Add URL validation regex to reject shell metacharacters (CRITICAL)
- Add agent name validation in workflow input (MEDIUM)
- Add `jq has()` check before executing install commands (CRITICAL)
- Use `tar -T` instead of unquoted word-splitting in capture-agent.sh (MEDIUM)
- Resolve merge conflicts with upstream/main (keep Docker fields, adapt
to simplified DO flow, bump version to 0.15.0)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use globalThis.fetch for testability in CI
Bun's native fetch binding doesn't go through global.fetch property
lookup, so global.fetch = mock(...) doesn't intercept it. Using
globalThis.fetch explicitly ensures the mock interception works.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add missing packer dependencies and harden install command safety
- Add packer/agents.json (agent tier + install command definitions)
- Add packer/scripts/tier-{minimal,node,bun,full}.sh (dependency scripts)
- Add basic command safety check rejecting suspicious patterns
- Document packer/agents.json as a trust boundary requiring PR review
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): fix npm prefix mismatch, add apt-get update, cleanup
- Add apt-get update -y before apt-get install in all tier scripts
- Add --prefix ~/.npm-global to npm install commands in agents.json
so installed packages land where capture-agent.sh expects them
- Rename misleading MARKER_DIR → MARKER_FILE in capture-agent.sh
- Remove stale comment referencing packer snapshots in workflow
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): detect empty agent installs in capture script
The "no files found" check was dead code — the marker file is always
created before filtering, so FILTERED_FILE always had at least one
entry. Now we count non-marker entries to catch cases where the agent
install silently fails and no actual files are on disk.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): use bare fetch() for Bun mock compatibility in CI
In Bun, global.fetch = mock(...) overrides bare fetch() calls but NOT
globalThis.fetch() calls. Every other source file in the codebase uses
bare fetch() and their mocks work fine in CI. Switch to match.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): use dependency injection for fetch in tests
Bun's global.fetch mock doesn't reliably intercept bare fetch() calls
across all Bun versions in CI. Instead of fighting the runtime, accept
an optional fetchFn parameter (defaults to fetch) and pass mock fetch
directly in tests.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): bypass mock.module bleed in agent-tarball tests
orchestrate.test.ts uses mock.module("../shared/agent-tarball", ...)
which is process-global in Bun and bleeds into agent-tarball.test.ts.
Import via URL (import.meta.url resolution) to bypass the specifier-
based mock matching and get the real module.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tarballs): eliminate mock.module bleed between test files
Bun's mock.module is process-global — orchestrate.test.ts mocking
agent-tarball poisoned agent-tarball.test.ts (the mock function
ignored the fetchFn parameter and always returned false).
Fix: make tryTarballInstall injectable via OrchestrationOptions.
orchestrate.test.ts passes the mock directly via options instead
of using mock.module. agent-tarball.test.ts imports the real module.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(tests): mock Bun.which in credential priority tests
Tests assumed no cloud CLIs were installed, but machines with hcloud/
doctl would get "CLI installed" hint overrides, failing the assertion.
Spy on Bun.which to return null so tests are environment-independent.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: fix import ordering after rebase
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* security: add curl domain allowlist and expand command blocklist
Addresses security review findings:
- Add domain allowlist for curl/wget targets (claude.ai, opencode.ai,
raw.githubusercontent.com, registry.npmjs.org, crates.io, github.com)
- Expand suspicious command blocklist (python -c, perl -e, ruby -e, dd, /dev/)
- Document 4-layer security model in workflow comments
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* security: add rm -rf to command blocklist
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Signed-off-by: Ahmed Abushagur <ahmed@abushagur.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(docker): replace Packer snapshots with Docker-based agent delivery
Docker images on GHCR are public and cross-account, unlike DO snapshots
which are private/account-scoped. Cloud-init installs Docker + pulls the
agent image during boot. The install step extracts pre-built binaries via
`docker cp` and falls back to normal install if unavailable.
- Add Dockerfiles for all 7 agents (claude, codex, openclaw, opencode,
kilocode, zeroclaw, hermes)
- Convert docker.yml to matrix build for all agents
- Add tryInstallFromDocker() shared helper with Docker-first install
- Add Docker pull to DigitalOcean cloud-init userdata
- Remove Packer snapshot pipeline, lookup, and SSH-only wait
- Remove packer/ directory (HCL templates, tier scripts, agents.json)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* security: address review findings in docker agent delivery
- Add agentName validation regex (/^[a-z0-9-]+$/) in digitalocean.ts
before interpolation into cloud-init script
- Quote dockerImage variable in all docker command strings in
agent-setup.ts to prevent command injection
- Restrict docker cp to specific known directories (.claude, .bun,
.local, .npm, .cargo, .opencode) instead of blanket /root/.
Agent: pr-maintainer
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: B <6723574+louisgv@users.noreply.github.com>
DO snapshots are private and account-scoped — users on different
accounts cannot see snapshots built by the CI token. Docker images
are the better approach for cross-account pre-built agents.
Removes: packer/, packer-snapshots workflow, snapshot lookup code,
and snapshot test. Reverts DO CLI to plain cloud-init flow.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Packer wasn't auto-loading build.auto.pkrvars.json, causing
"Unset variable" errors. Pass it explicitly with -var-file.
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat(digitalocean): Packer nightly snapshot pipeline for fast boot
Add pre-built Packer snapshots for DigitalOcean droplets. Instead of
10-20 min cloud-init + agent install on every boot, snapshot-based
droplets boot in ~2-3 min (SSH only, agent pre-installed).
- Packer HCL2 template with parametrized agent/tier builds
- Agent build matrix (packer/agents.json) for all 7 agents
- Tier scripts mirroring cloud-init.ts package tiers
- Nightly GitHub Actions workflow (4 AM UTC, max-parallel: 3)
- Automatic cleanup: keeps only latest snapshot per agent
- CLI: findSpawnSnapshot() looks up pre-built images via DO API
- CLI: waitForSshOnly() skips cloud-init when using snapshots
- CLI: createServer() accepts optional snapshotId, skips user_data
- CLI: main.ts routes to fast path when snapshot detected
- Tests for findSpawnSnapshot() (5 cases, all passing)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(packer): use var-file for install_commands to avoid shell quoting issues
The previous approach passed install_commands as `-var` inline, but
GitHub Actions expands `${{ }}` before shell evaluation — JSON arrays
with `|`, `&&`, and `"` characters break shell quoting.
Fix: generate a `.auto.pkrvars.json` file (auto-loaded by Packer)
using jq with --argjson for safe JSON handling. Also route all
`${{ inputs }}` and `${{ matrix }}` values through env vars to
prevent script injection.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
* feat: migrate shell script URLs to openrouter.ai/labs/spawn CDN
Users on older CLI versions can't auto-update because the repo was restructured
(cli/ → packages/cli/), so old version-check URLs 404. This decouples the CLI
from the repo's internal directory structure:
- Shell script URLs (install, agent scripts, github-auth) now use
openrouter.ai/labs/spawn/* as primary with GitHub raw as fallback
- Version checks now use GitHub release artifact (cli-latest/version)
as primary — a static URL that never changes regardless of repo layout
- CI workflow updated to publish a `version` file alongside cli.js
- Remove GITHUB_RAW_URL_PATTERN validation (no longer needed since
install URL is now a hardcoded CDN string, not interpolated)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* style: fix biome formatting in update-check test
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: CLAUDE.md says biome lint but should say biome check
biome lint only checks lint rules, not formatting. biome check does both.
The hooks and CI already run biome check — the docs were out of sync.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(hooks): PostToolUse hook wasn't running biome on CLI source files
Two bugs in validate-file.ts:
1. Config search only checked 1-2 levels up from the edited file, but
biome.json is at packages/cli/ — 3 levels above src/__tests__/*.ts.
Fix: walk up directories until biome.json is found (or hit root).
2. Ran `biome format` (prints formatted output, always exits 0) instead
of `biome format --check` (exits non-zero if file needs formatting).
Fix: use `biome check` which does lint + format check in one pass.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor: extract inline hook commands to TypeScript scripts in .claude/scripts/
Replace long inline `bash -c '...'` one-liners in .claude/settings.json with
standalone TypeScript scripts that are easier to read, debug, and maintain:
- enforce-worktree.ts: PreToolUse hook ensuring edits happen in worktrees
- validate-file.ts: PostToolUse hook for .sh/.ts file validation
- pre-merge-check.ts: PreToolUse hook running biome + tests before merge
Add .claude/scripts as a bun workspace package (@spawn/hooks).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor: replace manual typeguards with valibot schemas in hook scripts
- Extract shared schemas (FilePathInput, CommandInput, parseStdin) to schemas.ts
- Replace inline multi-level typeof/in checks with v.safeParse() calls
- Add valibot dependency to @spawn/hooks package
- Add CLAUDE.md rule: always prefer valibot over manual typeguards, share schemas
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor: split CLAUDE.md into modular .claude/rules/ files
Split the 437-line monolithic CLAUDE.md into a lean 89-line project overview
plus 9 focused rules files in .claude/rules/ (auto-loaded by Claude Code):
- culture.md — embrace bold changes, parallelize, verify exhaustively
- shell-scripts.md — curl|bash compat, macOS bash 3.x, ESM only, bun not python
- type-safety.md — no `as` assertions, ALWAYS use valibot (never manual typeguards)
- testing.md — bun:test only, no vitest, no subprocess spawning
- git-workflow.md — worktree-first mandatory workflow
- autonomous-loops.md — discovery/refactor service architecture
- discovery.md — how to fill matrix gaps, add clouds/agents
- documentation.md — never commit docs, use .docs/
- cli-version.md — bump version on every CLI change
The type-safety rule now explicitly mandates valibot schemas over manual
typeguard chains in all cases beyond single-primitive narrowing.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix(lint): run biome check across all packages in CI
The lint workflow only checked packages/cli/src/. Now it checks all
TypeScript locations in a single biome check command:
- packages/cli/src/ (with GritQL plugins)
- packages/shared/src/ (new biome.json)
- .claude/scripts/ (new biome.json)
- .claude/skills/setup-spa/
Fixed all pre-existing lint/format errors:
- node: protocol on all Node.js built-in imports in hook scripts
- useBlockStatements in packages/shared/src/type-guards.ts
- expand formatting in .claude/skills/setup-spa/main.ts and spa.test.ts
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* feat!: remove Fly.io cloud provider support
Drop Fly.io as a supported cloud provider. Sprite (which uses Fly.io
infrastructure internally) is retained.
- Delete packages/cli/src/fly/ module, sh/fly/ scripts, fixtures/fly/
- Remove fly cloud entry and 6 fly matrix entries from manifest.json
- Remove fly imports, destroy cases, and connection handlers from commands.ts
- Remove fly-ssh sentinel from security.ts
- Port E2E test suite from Fly.io to AWS Lightsail (fly-e2e.sh → aws-e2e.sh)
- Update README (7 clouds, 42 combinations), CLAUDE.md, and skill prompts
- Clean up fly references in build config, gitignore, icon sources
- Bump CLI version to 0.11.0
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* chore: restore Docker image build under sh/docker/
Move openclaw Dockerfile from sh/fly/docker/ to sh/docker/ and rename
workflow from fly-docker.yml to docker.yml with updated paths.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* style: fix extra blank lines in commands.ts
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: spawn-bot <spawn-bot@openrouter.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: L <6723574+louisgv@users.noreply.github.com>
* ci: add Mock Tests job to satisfy required status check
Split the unit-tests job into mock-tests (runs bun test) and unit-tests
(verifies cloud bundles build). The repo ruleset requires "Mock Tests",
"Unit Tests", and "Biome Lint" checks — the missing "Mock Tests" job was
blocking all PR merges.
Fixes#1901
Agent: issue-fixer
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
* style: fix pre-existing Biome format issues in 9 files
Auto-applied Biome formatter to src/ to resolve failing "Biome Lint"
required status check. No logic changes — formatting only.
Agent: issue-fixer
Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>
---------
Co-authored-by: B <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Sonnet 4.5 <noreply@anthropic.com>
E2E tests now run as a 4th teammate alongside test-runner,
dedup-scanner, and code-quality-reviewer during schedule-triggered
QA cycles. The standalone e2e mode is preserved for on-demand use.
- Add e2e-tester teammate to qa-quality-prompt.md
- Increase quality mode timeout from 35 to 40 min
- Add "e2e" to trigger-server valid reasons
- Re-enable daily schedule in qa.yml, default to "schedule"
Co-authored-by: spawn-bot <spawn-bot@openrouter.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The linter was running in CI with --warn-only, meaning it never
blocked anything — effectively vaporware. This removes --warn-only
to make it a real gate.
Also adds rules for bash 4.0+ features that were documented in
CLAUDE.md but not enforced:
- MC014: readarray/mapfile (bash 4.0+)
- MC015: coproc (bash 4.0+)
- MC016: &>> redirect (bash 4.0+)
- MC017: relative source paths (breaks curl|bash)
- MC018: wait -n (bash 4.3+)
- MC019: declare -g (bash 4.2+)
Excludes .claude/worktrees/ from scanning (temp copies, not
committed code).
Co-authored-by: spawn-bot <spawn-bot@openrouter.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Restructure the repo as a Bun workspace monorepo:
- Move cli/ → packages/cli/
- Create packages/shared/ (@openrouter/spawn-shared) with type-guards and parse utilities
- Add root package.json with workspace configuration
- Update all CLI imports to use @openrouter/spawn-shared
- Deduplicate toRecord/toObjectArray helpers from 4 cloud modules
- Update SPA (slack-bot) to use shared package instead of local toObj()
- Update 48 agent shell scripts for new packages/cli/ path
- Update install.sh, install.ps1, e2e, and test scripts
- Update all GitHub workflows, .gitignore, pre-commit hooks
- Update CLAUDE.md, README.md, and skill prompt references
- Pin all dependency versions (no ^ ranges)
- Bump CLI version 0.9.1 → 0.10.0
All 1908 tests pass. Lint clean. All 8 cloud bundles build.
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Reorganizes the project so all shell scripts live under a dedicated
/sh directory, enabling the OpenRouter rewrite URL to point at /sh/
instead of the repository root.
Moves:
- cli/install.sh → sh/cli/install.sh
- shared/*.sh → sh/shared/*.sh
- {cloud}/{agent}.sh → sh/{cloud}/{agent}.sh (48 scripts)
- {cloud}/README.md → sh/{cloud}/README.md
- e2e/*.sh → sh/e2e/*.sh
- test/macos-compat.sh → sh/test/macos-compat.sh
- test/fixtures/**/*.sh → sh/test/fixtures/**/*.sh
Updates all references:
- RAW_BASE path construction in commands.ts, update-check.ts
- GitHub auth URL in agent-setup.ts
- Self-referencing URLs in install.sh, github-auth.sh
- CI workflow paths in lint.yml, cli-release.yml
- Test file paths in install-script-validation, manifest-integrity
- Documentation in README.md, cli/README.md, CLAUDE.md
- QA scripts in .claude/skills/
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add e2e/ directory with fly-e2e.sh orchestrator and lib/ helpers
(provision, verify, teardown, cleanup) that provision real Fly.io VMs,
verify agent installation, and tear everything down
- Fix openclaw E2E failure by setting MODEL_ID=openrouter/auto to bypass
interactive model selection prompt in headless mode
- Add e2e mode to qa.sh (reason=e2e) that launches a Claude agent to run
the E2E suite and investigate/fix any failures
- Update qa.yml with reason dropdown (e2e/schedule/fixtures), kept disabled
Co-authored-by: spawn-bot <spawn-bot@openrouter.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
The race condition: two PRs merged 3 seconds apart both triggered the
CLI Release workflow. The second run (v0.7.12) finished last and
overwrote the release with a stale binary, even though the repo HEAD
was at v0.8.0.
- Add concurrency group so concurrent releases cancel the older one
- Add workflow_dispatch trigger for manual re-runs
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Run `biome format --write` on all 98 source files (38 needed fixes).
The main change: object literals and long argument lists are now expanded
onto separate lines per Biome's `"expand": "always"` setting, making
code much easier to scan on narrow screens.
Add `biome format` check step to CI lint workflow so formatting
regressions are caught on every PR.
Co-authored-by: Claude <claude@anthropic.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: Result monad for retry logic — prevent duplicate server creation
SSH exit 255 after an interactive session caused runWithRetries to retry
the entire bash script, creating duplicate servers. The old withRetry
also blindly retried all errors including timeouts where the remote
command may have already completed.
Introduces a Result<T> monad (Ok/Err) so callers explicitly signal
whether a failure is retryable (return Err) or fatal (throw). Adds
wrapSshCall() that classifies SSH errors: transient connection failures
are retryable, timeouts are not. Removes retry loop from the top-level
script runner entirely since it spans server creation + interactive
session.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* docs: mandate draft-PR-first workflow for all changes
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: add biome lint to CI and pre-commit hook, fix lint violations
- Add Biome lint job to .github/workflows/lint.yml
- Add TypeScript lint check to .githooks/pre-commit
- Fix useBlockStatements violations in ui.ts and tests
- Add biome lint to CLAUDE.md "After Each Change" checklist
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* refactor: rename Result.value to Result.data
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: clean up stale pre-commit hook
- Remove dead check for deleted functions (write_oauth_response_file,
create_oauth_response_html) — they no longer exist in the codebase
- Fix early exit skipping Biome lint when no .sh files are staged
- Replace echo -e with printf (the hook was using the pattern it bans)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: resolve biome lint errors blocking CI
- Fix useImportType: import { type Result } → import type { Result }
- Fix noUnusedImports: remove unused KNOWN_FLAGS import
- Fix noUnusedTemplateLiteral: template literal → string literal
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
shared/common.sh (3852 lines) was dead code — the entire architecture
was rewritten to TypeScript in cli/src/. No agent scripts source it
anymore. The only consumer was github-auth.sh which just needed 4
log functions (now inlined).
Remove 27 test files that spawned ~800+ real bash/bun subprocesses per
run (the root cause of slow bun test). Every shared-common-*.test.ts
file forked a real bash shell per test case to source shared/common.sh.
CLI subprocess tests spawned `bun run index.ts` per assertion. These
were integration tests, not unit tests.
Also removes:
- mock-tests CI job from test.yml (ran test/mock.sh which opens browser)
- Stale plan files referencing deleted infrastructure
- All CLAUDE.md/README.md references to the old lib/common.sh pattern
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Keep workflow_dispatch for manual testing. Re-enable cron when the
QA VM is back online.
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Eliminates the slow waitForCloudInit() + bun install phase by booting
a pre-built image with Node.js, bun, and openclaw already installed.
The image is rebuilt daily via GitHub Actions to pick up new releases.
Other agents are unaffected — they still use ubuntu:24.04 + cloud-init.
Co-authored-by: spawn-bot <spawn-bot@openrouter.ai>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Move all fly TypeScript files from fly/lib/*.ts and fly/main.ts into
cli/src/fly/. This gives them access to cli/node_modules (@clack/prompts),
biome linting, and the existing bun:test infrastructure — no symlinks or
NODE_PATH hacks needed.
The org picker now uses @clack/prompts select() directly (static import,
bundled at build time).
New: cli/build-clouds.sh — auto-discovers cli/src/*/main.ts and bundles
each into {cloud}.js. Scalable to future cloud TS migrations:
bash cli/build-clouds.sh # build all
bash cli/build-clouds.sh fly # build one
Shims now check for cli/src/fly/main.ts (local) or download fly.js from
GitHub releases (remote curl|bash).
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: use uv --upgrade to ensure Python 3.13-compatible Pillow across all clouds
aider-chat on Python 3.13 fails with `ImportError: cannot import name
'_imaging' from 'PIL'` when an old Pillow version (pre-10.4) is resolved
— those releases have no Python 3.13 binary wheels, so the C extension
is missing at runtime.
Replace `--with 'Pillow>=10.2.0'` (which was silently broken — the `>`
and single quotes get mangled by `printf '%q'` in run_server before the
command reaches the remote machine) with `--upgrade`, which forces all
transitive deps including Pillow to their latest compatible versions.
Also adds a plain-text echo before the install so users see progress
instead of a silent hang during the 2-4 minute install.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* test: update aider/gptme/interpreter assertions from pip to uv
The install method for aider, gptme, and open-interpreter was changed
from pip to `uv tool install` across all clouds. The mock test
assertions still checked for the old `pip.*install.*` patterns, causing
9 failures (3 agents × 3 clouds).
Update patterns to match the actual `uv tool install` commands now used
in all cloud scripts.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* ci: trigger test run for uv assertion fix
* fix: prevent SSH hangs, restore stderr, fix command escaping across clouds
- Add < /dev/null to ssh_run_server and generic_ssh_wait to prevent SSH
stdin theft causing sequential install/verify/configure steps to hang
- Add ServerAliveInterval, ServerAliveCountMax, ConnectTimeout to default
SSH_OPTS so long-running installs don't silently drop on flaky networks
- Remove 2>/dev/null from Fly.io run_server so remote command errors are
no longer silently swallowed (--quiet flag still suppresses flyctl noise)
- Fix Fly.io printf '%q' double-quoting: remove extra quotes around
$escaped_cmd that prevented the remote shell from consuming escapes,
breaking && || | operators in commands
- Remove broken printf '%q' from Daytona run_server and interactive_session
where it escaped shell operators into literal characters since daytona exec
has no intermediate shell layer
- Pin aider to --python 3.12 instead of --with audioop-lts across all clouds
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add --pty to fly ssh console for interactive sessions
fly ssh console -C does not allocate a pseudo-terminal by default,
causing interactive TUI agents (aider, claude) to fail with
"Input is not a terminal (fd=0)" or completely unresponsive input.
Adding --pty forces PTY allocation, matching how other clouds handle
interactive sessions (SSH uses -t, Sprite uses -tty).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: prepend ~/.local/bin to PATH in ssh_run_server
After uv installs to ~/.local/bin, the current shell session doesn't
have it in PATH, causing "uv: command not found" on DigitalOcean and
all other SSH-based clouds (Hetzner, AWS, GCP, OVH).
Fly.io's run_server already prepends this PATH — now the shared
ssh_run_server does the same, fixing all SSH-based clouds at once.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add Node.js to cloud-init for all cloud providers
npm-based agents (codex, kilocode, etc.) fail with "npm: command not
found" because Node.js isn't installed during cloud-init. Fly.io was
the only provider installing Node.js (in wait_for_cloud_init).
Now all cloud-init scripts install Node.js v22 LTS from nodesource,
matching Fly.io's setup. Also adds ~/.local/bin to PATH in AWS and
GCP cloud-init (was already in shared/DigitalOcean/Hetzner).
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: use apt packages for nodejs/npm instead of nodesource
The nodesource setup script (setup_22.x) runs its own apt-get update
and repository configuration, nearly doubling cloud-init time and
causing hangs on DigitalOcean. Ubuntu 24.04 includes nodejs and npm
in its default repos — just add them to the packages list.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: add timeouts and better error handling to Daytona CLI commands
Daytona CLI commands (login, list, create) can hang indefinitely when
the API is slow or unreachable. This causes:
- "Failed to create sandbox: timeout" with no recovery
- Token validation timeouts misreported as "invalid token"
- Users re-entering valid tokens that also timeout
Fixes:
- Wrap all daytona CLI calls with timeout (30s for auth, 120s for create)
- Detect timeout errors separately from auth errors
- Show actionable "try again / check status" messages for timeouts
- Add nodejs/npm to Daytona wait_for_cloud_init
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: set DAYTONA_API_URL to Daytona Cloud by default
The Daytona CLI may default to connecting to a local self-hosted
server instead of Daytona Cloud. Without DAYTONA_API_URL set to
https://app.daytona.io/api, every CLI command (login, list, create)
hangs trying to reach a non-existent local server and times out.
The SDK documents this as the default, but the CLI doesn't always
pick it up — now we export it explicitly.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: symlink n-installed Node.js v22 over apt v18 to prevent shadowing
n installs Node.js v22 to /usr/local/bin/node but apt's v18 at
/usr/bin/node can shadow it in non-interactive SSH sessions. After
n 22, symlink the new binaries over the apt ones so v22 is always
resolved. Also fix hcloud CLI token extraction for new TOML format.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix: address security review, add curl timeouts to trigger workflows
- Fix ssh_run_server command injection concern: use single-quoted
path_prefix so $HOME/$PATH expand remotely, not locally
- Add --connect-timeout 15 --max-time 30 to trigger workflows to
prevent 5-min hangs when server streams responses
- Handle 409 (dedup) as success — expected when cron fires every 15min
but cycles take 35min
- Reduce workflow timeout-minutes from 5 to 2
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
* feat: add QA upgrade — macOS compat linter, per-agent mock assertions
Layer 1: macOS compat linter (test/macos-compat.sh)
- 12 rules (MC001–MC012) catching bash 3.2 incompatibilities
- Detects: base64 -w0 file args, non-portable echo flags, source <(),
((var++)), read -d, nounset flag, sed -i, date %N, local -n,
declare -A, ${var,,}, and |&
- Added to CI lint.yml in warn-only mode for burn-in
- Integrated as Phase 0.5 in qa-dry-run.sh
Layer 2: Per-agent mock assertions
- test/fixtures/_shared_agent_assertions.sh with install checks
for all 15 agents (claude, openclaw, aider, goose, etc.)
- Integrated into test/mock.sh via _run_agent_assertions()
Also includes branch fixes:
- Fix base64 -w0 to use stdin redirect (aws, daytona, fly)
- Fix fly/openclaw to use npm install instead of broken curl|bash
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* feat: add E2E test harness and integrate into QA pipeline
Add test/e2e.sh — a full E2E test harness that provisions real servers,
installs agents, and verifies setup across all clouds. Features:
- Smoke test (one canary agent per cloud) and full matrix modes
- Credential auto-detection for 8 clouds
- Per-cloud preflight validation (sequential) then parallel agent tests
- Stale server cleanup, timing history, cross-cloud comparison
- Auto-fix and optimization phases via Claude agents
- macOS bash 3.2 compatible
Integrate E2E as Phase 5 in both qa-cycle.sh and qa-dry-run.sh:
- Runs after mock tests pass, gated on cloud credentials
- Phase 5b auto-fixes failures using per-agent worktree branches
- Parses results and includes in QA summary
Also fixes:
- shared/common.sh: honour SPAWN_NON_INTERACTIVE=1 in safe_read()
- aws/lib/common.sh: fix SSH key import (use cat instead of base64,
handle race condition on concurrent imports)
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Replace the complex claude launch pattern (subshell + PID file + tee
pipe + stream-json + 50-line watchdog monitoring log file growth +
session-end detection) with a simple direct launch:
claude -p "..." >> "${LOG_FILE}" 2>&1 &
The watchdog is now just a wall-clock timeout. The idle-output detection,
stream-json result parsing, and tee piping are all removed.
Also remove GitHub Actions concurrency groups — the trigger server
already handles dedup (409 for same issue, 409 for same reason), making
the GH Actions concurrency groups redundant queuing.
Changes:
- refactor.sh: simple launch + wall-clock-only watchdog
- security.sh: same simplification
- discovery.sh: same (refactored _kill_claude_process and
_run_watchdog_loop to simpler signatures)
- All 4 workflows: remove concurrency groups
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The trigger server streamed script stdout back to GitHub Actions via a
long-lived HTTP response, requiring --http1.1, heartbeat injection,
server.timeout(req, 0), createEnqueuer, drainStreamOutput, and 90-min
GH Actions timeouts. In practice GitHub Actions is just a dumb trigger
— the real state lives on the VM (log files, journalctl). Simplify to
fire-and-forget: spawn script, return 200 JSON immediately.
Also fix the refactor and discovery team lead monitoring loops. The
prompts buried the loop in a single compressed line that the model
ignored (doing Bash("sleep 10") repeatedly without calling TaskList).
Replace with a dedicated "Monitor Loop (CRITICAL)" section with numbered
steps, matching the security.sh pattern that actually works.
Changes:
- trigger-server.ts: remove ~150 lines of streaming code (createEnqueuer,
drainStreamOutput, startStreamingRun, heartbeat, ReadableStream),
replace with startFireAndForgetRun (stdout: "inherit", immediate JSON)
- All 4 workflows: simple curl POST, timeout-minutes 90→5, remove
--http1.1/-N/--max-time/exit-code handling
- refactor.sh: add Monitor Loop (CRITICAL) section with numbered steps
- discovery-team-prompt.txt: same Monitor Loop fix
- SKILL.md: update architecture docs, remove streaming sections
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Add team lead pre-approval gate: teammates spawn in plan mode and must
get approval before creating any PR (hard gate, not just prompt rules)
- Add diminishing returns rule: default posture is "code is good, shut down"
- Add dedup rule: check for existing open/closed PRs before creating new ones
- Require concrete PR justification (what breaks without this change)
- Add off-limits files list (.github/workflows, .claude/skills, CLAUDE.md)
- Use git pathspec exclusions in refactor.sh to never stage protected files
- Constrain pr-maintainer to only act on approved or feedback PRs
- Reduce refactor cron from every 5 minutes to every 2 hours
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Previously only org members were allowed. Now checks both org membership
and repo collaborator status, so invited collaborators can open issues
and PRs without being blocked.
Co-authored-by: lab <6723574+louisgv@users.noreply.github.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
