Elixir runner continues after failed workspace setup and stalls on Codex MCP elicitation

[stavarc]

Summary

The Elixir Symphony runner can leave a Linear issue permanently In Progress when workspace setup fails and the agent later reaches a Codex MCP elicitation request. In our run, Symphony kept retrying the same Linear issue every ~5 minutes without surfacing a terminal blocker or moving the issue to the configured handoff state.

Environment

• Symphony commit: 58cf97da06d556c019ccea20c67f4f77da124bf3
• Runner: elixir/bin/symphony in a Podman container
• Tracker: Linear
• Codex CLI: 0.128.0
• Codex model config: gpt-5.5, model_reasoning_effort=xhigh
• Workspace root: external volume mounted at /var/lib/symphony/workspaces

Workflow shape

The workflow uses an after_create hook similar to:

workspace:
  root: $SYMPHONY_WORKSPACE_ROOT
hooks:
  after_create: |
    git clone --depth 1 "$SOURCE_REPO_URL" .
    git config --global --add safe.directory "$PWD"
    if command -v bun >/dev/null 2>&1; then
      bun install --frozen-lockfile
    fi

The issue was a normal Linear issue in an active state. Symphony created /var/lib/symphony/workspaces/SKI-5 and dispatched Codex.

Observed behavior

The workspace ended up in a bad/incomplete state:

/var/lib/symphony/workspaces/SKI-5/.git exists but is empty
git status --short --branch
fatal: not a git repository (or any of the parent directories): .git

Symphony logs showed the hook starting, but the worker continued into Codex anyway:

info: Running workspace hook hook=after_create issue_identifier=SKI-5 workspace=/var/lib/symphony/workspaces/SKI-5
info: Codex session started ... cwd=/var/lib/symphony/workspaces/SKI-5

After that, Codex attempted fallback work through GitHub connector operations. The latest session got as far as trying to create a branch, then Symphony observed:

debug: Codex notification: "mcpServer/elicitation/request"

No further useful Codex activity occurred. Roughly five minutes later Symphony restarted the same issue:

warning: Issue stalled: issue_identifier=SKI-5 ... elapsed_ms=306292; restarting with backoff
warning: Retrying issue_identifier=SKI-5 in 10000ms (attempt 1) error=stalled for 306292ms without codex activity

This repeated many times. The /api/v1/state endpoint still reported the issue as running/In Progress, and Linear stayed In Progress, but no PR or branch was produced.

Expected behavior

One or more of these would be safer:

1. If after_create exits non-zero or leaves the workspace without a valid Git repo, mark the issue blocked/failed and stop retrying.
2. Surface the hook stderr/stdout clearly in Symphony logs and state API.
3. Treat mcpServer/elicitation/request as a hard blocker in unattended mode (approval_policy: never) rather than repeatedly restarting the same issue.
4. Move the issue to the configured handoff state or otherwise expose a terminal operator action.

Impact

A single issue can consume a large amount of Codex tokens while making no progress.

Notes

This may be partly a deployment/configuration problem on our side, but the runner behavior still seems unsafe: it continued after a broken workspace setup and retried indefinitely after an unattended elicitation request.