spryng-agent 0.1.10

Run Claude Code or Codex on your machine against your Spryng board (the Spryng Local Agent / claude-code-runner).

claude-code-runner

Out-of-process sidecar that runs the claude CLI on behalf of Spryng AgentRuns.

Spryng's apps/agent_runs/runtimes/claude_code.py adapter speaks the protocol; this service owns the subprocess and posts progress back. Deploy it once per environment (one container per runtime config), point Spryng at it via Settings → AI → Agent Runtimes, and agents start driving themselves end-to-end.

---

Quick start

# 1. Provision the auth token both sides will share.
export RUNNER_AUTH_TOKEN=$(python -c "import secrets; print(secrets.token_urlsafe(32))")
echo "Runner token: $RUNNER_AUTH_TOKEN"

# 2. Provision the Anthropic API key the `claude` CLI uses.
export ANTHROPIC_API_KEY=sk-ant-...

# 3. Build + run the container.
docker build -t claude-code-runner .
docker run --rm -p 8088:8088 \
    -e RUNNER_AUTH_TOKEN \
    -e ANTHROPIC_API_KEY \
    -e RUNNER_ALLOWED_CALLBACK_HOSTS=app.spryng.io \
    claude-code-runner

# 4. In Spryng, AI Settings → Agent Runtimes → Add:
#    type:        claude_code
#    endpoint:    http://<runner-host>:8088
#    credential:  $RUNNER_AUTH_TOKEN

Local dev without Docker:

pip install -e .[dev]
cp .env.example .env
# edit RUNNER_AUTH_TOKEN
uvicorn runner.main:app --reload --host 0.0.0.0 --port 8088

---

Protocol

The runner exposes four endpoints. All except /healthz require Authorization: Bearer $RUNNER_AUTH_TOKEN.

POST /jobs

Spryng's adapter calls this to enqueue a job. Returns 202 with a job_id immediately; the work happens in the background.

Request body (see models.py for the full schema):

{
  "run_id": 4242,
  "card_ref": "ON-914",
  "card_id": 12345,
  "project_slug": "onboarding",
  "org_slug": "spryng-internal",
  "agent": {
    "username": "pilot-claude",
    "display_name": "Pilot Claude"
  },
  "spec": {
    "content": "---\nfeature_identity: …\n---\n\n## Context\n…",
    "format": "md",
    "complexity_score": 12
  },
  "replan_context": null,
  "spryng_callback": {
    "base_url": "https://app.spryng.io",
    "bearer_token": "agent-run-sub-token-xxxxx",
    "progress_path": "/api/scrumdo/organizations/spryng-internal/agent-runs/4242/progress/",
    "agent_run_header_name": "X-Spryng-Agent-Run",
    "agent_run_header_value": "4242"
  },
  "runtime": {
    "model_slug": "claude-sonnet-4-7",
    "max_wallclock_seconds": 1800,
    "max_cost_cents": 500
  }
}

Response:

{ "job_id": "1f8c…", "url": "/jobs/1f8c…" }

GET /jobs/<job_id>

Returns the runner's view of the job — runner-local state, plan text (once generated), outcome (once executed), exit code, cumulative cost in cents, and the live PID if a subprocess is currently running.

POST /jobs/<job_id>/cancel

Cooperative cancel. Sets cancel_requested, cancels the orchestration task, SIGTERMs any live subprocess (then SIGKILL after 10s), and posts state=cancelled back to Spryng.

Idempotent — calling cancel on a terminal job returns ok with note: "already terminal".

POST /jobs/by-run/<run_id>/cancel

Same as above but looked up by Spryng's run_id rather than the runner-local job_id. Used by Spryng's adapter when it doesn't have the runner job id (e.g. when the original start callback failed).

GET /healthz

Liveness probe. No auth required. Returns { "ok": true, "version": "0.1.0" }.

---

Orchestration loop

Per job, the runner walks through these states:

                              ┌──────────────────────┐
queued (Spryng) ──► planning ─┤                      │
                              │   claude `generate   │
                              │   plan` subprocess   │
                              │                      │
                              └────────┬─────────────┘
                                       │ plan text
                                       ▼
                              ┌──────────────────────┐
                              │ POST progress to     │
                              │ Spryng with plan     │
                              └────────┬─────────────┘
                                       │
                                       ▼
                              awaiting_approval ──┐
                                                  │ poll Spryng's
                                                  │ /agent-runs/<id>/
                                                  │ every 15s
                                                  ▼
                                          state=executing observed
                                                  │
                                                  ▼
                              ┌──────────────────────┐
                              │   claude `execute    │
                              │   plan` subprocess   │
                              └────────┬─────────────┘
                                       │ outcome text
                                       ▼
                              ┌──────────────────────┐
                              │ finalize: stage /    │
                              │ commit / push branch │
                              │ / open-or-reuse PR   │
                              └────────┬─────────────┘
                                       │ result {pr_url, diffstat, …}
                                       ▼
                              ┌──────────────────────┐
                              │ POST progress to     │
                              │ Spryng: completed +  │
                              │ outcome + cost +     │
                              │ result              │
                              └──────────────────────┘

Any failure / timeout / cancel along the way posts the appropriate state=failed | cancelled callback so Spryng's audit log stays correct.

Finalize — commit / push / PR (deterministic)

After a successful execute, runner/finalize.py::finalize_execution makes the artifacts a property of the runner, not of the agent's behaviour:

1. git add -A, then detect a no-op (no staged diff and no new commits vs. the checkout base) → report result.no_changes and open no PR.
2. Commit any uncommitted edits on expected_branch with a Fixes <card-ref> message so Spryng's webhook always links it.
3. git push the branch using the write-scoped token embedded in the clone URL.
4. Open (or reuse, idempotently) a PR via the GitHub REST API.
5. Return a structured result (pr_url, pr_number, diffstat, commits) on the completed callback. Spryng folds it into AgentRun.evidence['result'] and links the PR immediately — the card shows it without waiting on the webhook. On a push/PR failure the run still completes, with result.finalize_error set so the failure is visible.

Idempotency & restart behaviour

• POST /jobs is idempotent on run_id. Spryng mints a unique run_id per real attempt, so a second POST for a run_id already in the store (active or recently terminal, within RUNNER_JOB_TTL_SECONDS) is a duplicate at-least-once delivery — it returns the existing job and never starts a second claude run or double-charges.
• No resume across restart. The job store is in-memory (runner/jobs.py), so a container restart loses in-flight jobs. This is intentional — the monolith reaper is the agreed backstop: it sweeps runs stuck in planning / executing to failed (wallclock_timeout) once they pass max_wallclock_seconds + 5 min grace (apps/agent_runs/reaper.py), freeing the card and revoking the run token. Keep max_wallclock_seconds realistic so abandoned runs free their card promptly. To survive restarts without the reaper delay, swap JobStore for a Redis/SQLite-backed implementation.

What the runner still does NOT do

• Auto-merge. The human reviews and merges the PR. Spryng's large-diff approval gate still applies.
• PR discovery from a webhook. Finalize opens the PR itself; the webhook remains the reconciler/source of truth for review + check state.

---

Configuration

All env vars; defaults shown.

Var	Default	Purpose
RUNNER_HOST	0.0.0.0	Bind address
RUNNER_PORT	8088	Bind port
RUNNER_AUTH_TOKEN	required	Bearer token shared with Spryng's AgentRuntimeConfig
CLAUDE_CLI_BINARY	claude	Path to the Claude Code CLI binary
ANTHROPIC_API_KEY	inherited	Read by the claude CLI from env
RUNNER_WORKDIR_ROOT	/tmp/spryng-runs	Per-job working directories
RUNNER_JOB_TTL_SECONDS	3600	How long terminal jobs stay in memory (also the idempotency window)
RUNNER_SPRYNG_POLL_INTERVAL_SECONDS	15	Approval-polling cadence
RUNNER_GITHUB_API_BASE	https://api.github.com	GitHub REST base for the finalize PR open/reuse (override for GHE)
RUNNER_ALLOWED_CALLBACK_HOSTS	*	Comma-separated hostname allow-list for callbacks (SSRF guard)
LOG_LEVEL	info	uvicorn / app logging

The allow-list is the most important security knob:

# Prod
RUNNER_ALLOWED_CALLBACK_HOSTS=app.spryng.io,qa.spryng.io

# Multi-tenant: limit to your own Spryng deployments only.

When set to *, the runner posts callbacks to whatever URL Spryng sends. Fine for dev; dangerous in prod if a malicious payload could ever reach the POST /jobs endpoint.

---

Cost tracking

The runner pulls token counts off the claude --output-format json metadata and converts to cents using the same price table as Spryng's Anthropic provider adapter. The cost_cents_delta field on the progress callback aggregates planning + execution costs.

Update claude_runner.py _PRICE_TABLE when Anthropic pricing changes; mirror the same edit on Spryng's anthropic_adapter.py to keep the two sides honest.

---

Tests

pip install -e .[dev]
pytest -v

The test suite mocks both the claude subprocess and Spryng's callbacks so it runs hermetically. Coverage:

• /healthz doesn't require auth, returns version
• All other endpoints reject missing / wrong tokens
• POST /jobs accepts the payload shape and starts the orchestration
• GET /jobs/<id> returns the correct state shape
• POST /jobs/<id>/cancel and /jobs/by-run/<run_id>/cancel both work
• Idempotent cancel on terminal jobs
• Pydantic rejects malformed payloads with 422

---

Production checklist

1. Generate the auth token (32+ bytes random) and store both ends — RUNNER_AUTH_TOKEN in the runner's env, same value in Spryng's AgentRuntimeConfig.api_credential.
2. Provision ANTHROPIC_API_KEY in the runner's env (Secrets Manager / SSM SecureString — same pattern as Spryng's keys).
3. Lock the callback allow-list to your Spryng hostnames.
4. Set max_wallclock_seconds on the AgentRuntimeConfig to match your tolerance for stuck runs (default 1800s).
5. Set max_cost_cents_per_run on the AgentRuntimeConfig — the runner reports cumulative cost and Spryng's dispatcher fails the run when it exceeds this cap.
6. Run behind a reverse proxy (nginx / ALB) with TLS termination. The runner trusts X-Forwarded-For via --proxy-headers.
7. Health checks — point your orchestrator's liveness probe at GET /healthz.
8. Logs — uvicorn writes structured access logs to stdout; forward them to your usual aggregator.

---

Spec deviation note

The unified spec at BOARD_AI_AGENTS_UNIFIED_SPEC.md §E.4 calls for a strategy-pattern driver layer in Spryng plus an out-of-process compute that owns the subprocess. The driver layer lives at runtimes/claude_code.py; this repo IS the out-of-process compute. The HTTP protocol is the contract — both sides version it together.

When Spryng's adapter changes (new fields on the payload, new endpoint expectations), update models.py here in lockstep and bump the version in pyproject.toml.