v1.4.0 — forward cloud creds into box env + lease-TTL leak fix

Forward cloud credentials into the box env (#39, closes #19)

Box-CI roles that run inside the box and drive the cloud REST API now get credentials. The always-needed trio CONTAINARIUM_SERVER / CONTAINARIUM_TOKEN / CONTAINARIUM_ORG (org from the existing org-id input) is forwarded into the box's setup/test command env, plus a new optional multiline env: input (KEY=VALUE per line) for extras like a second org's CONTAINARIUM_TOKEN_PEER / CONTAINARIUM_ORG_PEER.

Done securely: secrets are written to a mode-0600 file (umask 077), scp'd over the existing channel, and chmod 600 on the box — never on the ssh argv (which would leak into the box process list and Action logs). Every value is ::add-mask::'d, the file is sourced before the command runs, and it's stripped on the keep-on-failure debug path so a debugger can't inherit the CI token.

Unblocks arming the Containarium-cloud actuation verifier (#192) and the Layer-4 per-org network-isolation sentry (#191) — the two remaining MVP verification gates.

Consumer usage

- uses: FootprintAI/containarium-run@v1.4.0
  with:
    server: ${{ secrets.CONTAINARIUM_SERVER }}
    token:  ${{ secrets.CONTAINARIUM_TOKEN }}
    org-id: ${{ vars.CONTAINARIUM_ORG }}
    env: |
      CONTAINARIUM_TOKEN_PEER=${{ secrets.CONTAINARIUM_TOKEN_PEER }}
      CONTAINARIUM_ORG_PEER=${{ vars.CONTAINARIUM_ORG_PEER }}

Lease TTL + heartbeat — no box ever leaks (#37, Containarium#526)

Every box is now born with a short birth TTL (default 20m) stamped right after create, renewed by a background heartbeat while the job runs, so a CANCELLED job or a runner that dies mid-run can no longer leak a box — it self-reaps within one window. Success still deletes immediately; failure still keeps for the debug window.

Full changelog: v1.3.0...v1.4.0

v1.0.0 — first public release

First public release of containarium-run — the GitHub Action for running CI in a Containarium box.

What you get

• Spawns a fresh Containarium container per job — real Linux, your code goes in via rsync, your tests run there, the box is torn down on success.
• Failure-debug surface: when a job fails with `keep-on-failure: true` (default), the box stays alive for an hour and a comment is posted on the PR with an SSH command + an `agent-box` MCP URL. Hand the URL to Claude Code / Cursor / any MCP-aware agent and they can read the `containarium://ci-context` resource (PR + commit + failing test + log tail) before they connect — so the agent doesn't waste its first turn asking "what am I looking at?".
• Three-field `containarium.yml` in your repo: `image`, `setup`, `test`. Plus an optional `serve:` block for PR-preview-environment use.

Tiered setup (pick yours)

Tier	Setup	Where your code runs
0 — Evaluator	This Action + a token from cloud.containarium.dev	Code → GHA-hosted runner → cloud → Containarium box
1 — Self-host	This Action + your own Containarium daemon	Code → GHA-hosted (orchestration only) → your boxes
2 — Enterprise	Tier 1 + audit log + SSO	Nothing crosses to FootprintAI
3 — Air-gapped	Tier 2 + offline install bundle	Fully contained in your VPC

See https://containarium.dev/security for the full data-residency breakdown.

Pin to this release

```yaml

• uses: FootprintAI/containarium-run@v1
  with:
  server: https://cloud.containarium.dev
  token: ${{ secrets.CONTAINARIUM_TOKEN }}
  ```

What's in v1

• Action scaffold with full `action.yml` contract (inputs, outputs, branding)
• `ci-context.json` writer that drops PR + commit + failing-test context into `/workspace/.containarium/` inside the box, for the agent-box MCP resource to surface
• Default `server` set to `https://cloud.containarium.dev\` (was previously a placeholder that didn't resolve)

What's next

• v0.1: cache-key implementation (warm volumes across runs)
• v0.1: sticky PR comments (edit existing instead of post new)
• v1.1: composite → JS Action when complexity warrants

Companion projects

• FootprintAI/containarium — the OSS daemon + CLI + agent-box (Tier 1+ self-host path)
• cloud.containarium.dev — the hosted control plane (Tier 0)
• hacks/runner/ — alternative: run CI workloads on Containarium-hosted self-hosted runners (zero GHA-hosted minutes)

v1.2.0 — per-role box sizing (box-cpu/box-memory/box-disk)

Highlights

Per-role box sizing (#33). Adds box-cpu / box-memory / box-disk
action inputs (defaults unchanged: 4 / 4GB / 50GB), plus a
resources: block in containarium.yml that overrides them per role.
Resolution order: resources: in the job config wins, else the action
inputs, else the defaults. Lets memory-heavy jobs (e.g. go test -race ./...)
get a larger box instead of OOM-killing in the default 4GB one.

This is v1.1.1 + #33 — no other changes. Existing consumers that set no
resources: block and pass no box-* inputs behave exactly as v1.1.1.

Inherited from v1.1.1

• create + delete via curl HTTP/1.1, not the CLI (FootprintAI/Containarium#422)
• SSH by the cloud-assigned identity (cld-<id>), not the box name
• source push via tar-over-ssh into ~/work
• debug boxes self-reap via a 1h auto-delete TTL on keep-on-failure
• idempotent create + job-unique box name; retry on transient edge failures

🤖 Generated with Claude Code

v1.1.1 — auto-delete kept-alive debug boxes + accumulated CI fixes

Highlights

Debug boxes now self-reap. On a failed CI run with keep-on-failure, the box is kept alive for debugging but now also gets a 1h auto-delete TTL (SetContainerTTL → daemon ttlsweeper), so it cleans itself up instead of leaking until someone deletes it by hand. Best-effort: a server predating the TTL handler returns 501 and the step warns rather than failing the build.

@v1 now ships the accumulated fixes. The floating v1 tag was still pointing at v1.1.0 and missing everything merged since. This release moves v1 → v1.1.1, finally delivering to @v1 consumers:

• create + delete via curl HTTP/1.1, not the CLI (FootprintAI/Containarium#422)
• SSH by the cloud-assigned identity (cld-<id>), not the box name
• source push via containarium sync / tar-over-ssh into ~/work
• idempotent create + job-unique box name; retry on transient edge failures
• curl exit-code surfacing + empty-server guard

Known gap

Cancelled jobs (neither success nor failure) still skip both teardown and the failure-path TTL — tracked for a follow-up (always() reaper or create-time TTL). See STATUS.md.

Not included: per-role box sizing (box-cpu/box-memory/box-disk, #33) — that's a feature, landing in a future v1.2.0.

🤖 Generated with Claude Code

v1.1.0 — self-hosted-runner fixes + dashboard debug-box reporting

Six fixes for failures observed against self-hosted runners during the 2026-05-24/25 cloud-CI dogfood (see STATUS.md and #14 for the full chain), plus one additive feature.

Fixes

• CLI version subcommand (#8) — containarium --version was unsupported; the action now uses the version subcommand. Fixes Error: unknown flag: --version.
• gh→curl fallback for PR comments (#9) — self-hosted runners often lack the gh CLI; the failure-debug PR comment now falls back to a curl POST. Fixes gh: command not found.
• PyYAML parser instead of yq (#10) — yq isn't preinstalled on self-hosted runners; switched to python3 + PyYAML. Includes edge-case handling (missing image, scalar setup/test, missing serve block).
• Per-run SSH keypair (#11) — containarium create requires --ssh-key. The action now mints a fresh ed25519 keypair under \$RUNNER_TEMP per run; subsequent ssh/scp/rsync use the privkey directly.
• Box-name truncation (#12) — jump-server validator caps usernames at 32 chars. Switched box-name template from ci-<repo>-<run-id>-<attempt> to ci-<run-id>-<attempt> (GITHUB_RUN_ID is globally unique).

Features

• Debug-box info to cloud dashboard (#16) — new optional input `org-id`. When set, the runner POSTs debug-box endpoint info (SSH host/port/user, MCP URL, keep-alive deadline) to the cloud's `ReportCIJobRuntime` endpoint after a test failure, so the `/ci/runs/` FailureDebugSurface panel renders real connection info instead of the "Coming soon" stub. Backward-compatible: default empty → no behavior change.

Internal

• Self-test workflow to catch `action.yml` load failures before consumers do (#6, #7).
• STATUS.md records the 2026-05-24/25 cloud-CI silence chain (#14).

Upgrading

The floating `v1` tag has been moved to v1.1.0. Repos pinning to `@v1.0.1` should bump to `@v1.1.0` (or `@v1` if they accept moving aliases) to pick up these fixes.

v1.0.1 — action.yml load-time fix

Single-commit patch release fixing two action.yml parse errors that blocked the action from loading at all (downstream CI runs failed with Failed to load action.yml).

What's fixed

1. cache-key input description no longer contains a ${{ hashFiles('go.sum') }} example. GHA evaluates ${{ }} expressions in input descriptions at action-load time and hashFiles() is workflow-only. Now describes the recommended caller usage in plain prose.
2. Expose port step name is now quoted and rephrased to avoid an unquoted : inside parentheses that was tripping GHA's YAML parser.

Pinning

• @v1 (moving) — gets this fix automatically; recommended for most users
• @v1.0.1 — pin to exactly this release if you want immutability
• @v1.0.0 — still broken; no reason to stay pinned here, please bump

Upgrade path

If you saw any of these errors:

Unrecognized function: 'hashFiles'
Mapping values are not allowed in this context
Failed to load FootprintAI/containarium-run/main/action.yml

Bumping the pin to @v1 (or @v1.0.1) resolves them with no other workflow changes needed.

PR: #5