feat(automation): normalize headless runtime output

[luokerenx4]

What changed

• normalize Claude Code, Codex, OpenCode, and Pi JSONL into one reply/message-block/tool/error contract
• persist a debounced structured snapshot for live Automation polling and keep bounded historical-tail parsing
• serialize concurrent task-registry writes, expose the global 8-worker capacity, and prune structured logs with old task records
• compact Pi diagnostics before disk by discarding documented cumulative message_update / tool_execution_update events; keep a 16 MB per-stream cap as the second guard
• redesign Automation Runs around compact run cards, final Markdown replies, tool activity/status/input/output, and collapsed runtime diagnostics
• treat runtime-reported failures as failed runs even when a CLI exits zero
• cover Codex stream/turn/error items plus command, file, MCP, web-search, and collaboration tool blocks
• extend packaged Workspace acceptance to exercise the real async Automation API and require a decoded Pi reply, completed tool block, CLI side effect, and compacted diagnostics
• document each runtime's native stream contract and normalization policy

Runtime contract checks

• Pi 0.80.6: official JSON mode emits all session events. Packaged async smoke confirms transient update events do not reach diagnostics; deterministic diagnostic tail fell from about 15.8 KB to 7.2 KB.
• OpenCode 1.17.13: official source and real mock runs emit terminal text/tool events plus step boundaries, not token deltas. Text run: 948 bytes / 3 events. Tool run: 2.3 KB / 6 events with one completed tool snapshot. No filter needed.
• Codex 0.142.5: official JSONL contract has no token-delta stream; item.updated currently carries todo-list updates. Parser now covers top-level and item errors, failed/declined commands, failed file/MCP calls, web search, and collaboration. Local success sampling is blocked by the installed CLI/model compatibility (gpt-5.6-sol requires a newer Codex), while its real failure stream and the version-matched official event schema are covered.

Verification

• pnpm test — 2512 passed, 6 skipped
• npx tsc --noEmit
• cd ui && npx tsc -b
• pnpm build
• real browser at /automation/runs: 69 historical runs, bounded 84.5 px prompt headers, structured reply/tool display, renamed runtime diagnostics
• pnpm electron:smoke:workspace — packaged Electron Chat Workspace, Git/CLI manifests, managed Pi async task, structured tool output, compacted diagnostics, CLI side effect, and cleanup all passed

Operational cleanup

Removed legacy oversized files only from ~/.openalice/workspaces/state/headless-logs: eight 77 MB–3.0 GB files plus six 17–45 MB files. The directory dropped from roughly 5 GB to 115 MB; Workspace contents were untouched.

This PR is intentionally left open for product review; do not auto-merge yet.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
openalice-demo	Ready	Preview, Comment	Jul 11, 2026 8:44am