Repository navigation
Description
Situation
This came from user feedback. An agent installs custom apps and discovers their actions, but cannot execute those actions through the normal action tool. Calling an app's own API provides a workaround.
Symptom
The reporter observed successful source installs and awaited runtime reloads. Both existing and newly installed custom actions appeared in search_actions and read_action, but execute_action rejected the same canonical IDs.
The report quotes Action not found and Unknown action, without complete parameterized responses. A workflow's documented /run API still invoked its run action successfully.
How to reproduce
These steps combine the reporter's sequence with code-derived preconditions. They were not executed during this read-only investigation.
- Derived from code: Start an isolated profile and open a main-agent conversation before installing the test app.
- Derived from code: Prepare a format-version-2 workflow app with a public, schema-bearing
runaction. Make its workflow a no-op for this test. - Reporter-provided: Install the app from source and await completion of its runtime reload.
- Reporter-provided: Find its canonical action ID through
search_actions, then inspect that exact ID withread_action. - Reporter-provided: Call
execute_actionwith that ID andjson_args: {"dryRun":true}. Record the complete response. - Reporter-provided: Compare the result with POST
/api/apps/<app-id>/run, using{"dryRun":true}. Also check an existing custom action and repeat after reinstalling the test app.
Initial triage
- Discovery does not use a separate static catalog. Search and read consult a live getter. Execution delegates to the session callback (facade). Both session discovery and its permission gate use the same registry and captured allow-list (session resolution). A stable permission mismatch does not explain successful discovery in that same session.
- The two reported error families identify different stages.
Unknown action: <name>comes from session resolution or permission rejection (execution callback). The engine instead throwsAction "<name>" not foundwhen its registry lookup fails (engine lookup, run). Full responses are necessary to distinguish them. - A reload exposes a real temporary absence. The subscriber unregisters an app's actions before awaiting their replacement factories. Factory failures leave actions absent (reload subscriber). However, installation awaits its final catalog refresh. Subscriber exceptions are logged and swallowed (install completion, subscriber handling). A successful install alone does not prove every runtime subscriber succeeded.
- Workers have independent registries. Each worker hydrates the lockfile and loads its own action records once (worker initialization). Catalog changes already invalidate the worker pool (wiring). Missing invalidation cannot be assumed.
- The standard workflow API does not bypass the engine. Its
/runhandler calls the canonical action throughctx.runAction(workflow API), which reaches the action engine (app context). The affected app's implementation was not supplied, so this establishes the template behavior, not its exact implementation. - Existing coverage misses the production process boundary. The install-after-session-open test checks discovery and execution, but constructs its engine with
processRole: "worker"over the same registry. It does not exercise a separately hydrated subprocess registry (test setup, test). - History: Searches covered open and closed issues and PRs, including user-feedback reports. apps: a legacy bare-name clash drops a format version 1 app's whole action, agent, or skill, and routines bound to it fail on every fire #965 concerns legacy-name clashes that drop artifacts before discovery, not this established discrepancy. codex: a resumed chat keeps its first tool list, so subagents added later never reach the agent #1301 concerns added subagent tools on resumed Codex threads. Neither is an established duplicate. Comparing the reporter commit with main found no change that resolves the relevant lookup, reload-publication, or worker-invalidation behavior.
Suspected root cause
Confidence: medium — the report points to runtime registration state changing between discovery and execution, or differing between main and a worker. The source does not support a permanently stale discovery snapshot: discovery and the session execution gate share the same live registry. Non-atomic action replacement can explain a transient Unknown action, while a worker-local omission can explain an engine not-found error. Neither is proven to explain the post-reload persistence and successful API comparison. The missing evidence is the complete failing response, exact canonical ID, reload subscriber outcome, and executing process or worker generation.
Possible fixes
- Publish an app's replacement actions atomically after successful instantiation, rather than removing them before asynchronous loading.
- Surface runtime-subscriber failures separately from successful bundle installation, with action-registration readiness in the install result.
- Add registry-generation and lookup-stage diagnostics across the session gate and worker engine. If divergence is confirmed, validate worker readiness against the dispatched generation.
- Extend the lifecycle regression test to real subprocess workers, covering existing and newly installed actions, reinstalls, and the workflow API comparison. Preserve permission checks and avoid automatic retries of actions with uncertain outcomes.
Environment
- Agent-filed report from the main agent, not a browser feedback dialog.
- Rome 1.1.166, commit
90d32d8d3f4a717c2bf1e650d1cb909d78d69990, 53 commits behind inspected main815cff04109bdb1f622363738415edd4f48f1ab2. - Build timestamp: 2026-10-26 15:49:40 UTC.
- Uptime: 49,572 seconds, about 13.8 hours. Upgrade flag true, previous version 1.1.165.
- Database probe healthy. 59 apps, zero failed and one broken. No evidence ties the broken app to the affected actions.
- Relay configured, including its deposit URL. Email and webchat channels connected.
- Route, browser family, viewport class, theme, mode, and language were not supplied.
Triage confidence: medium — Discovery and the session execution gate demonstrably share one live registry, while subprocess workers hydrate independent registries and reload replacement is non-atomic. Complete failing responses and runtime generation data are missing, so the investigation cannot distinguish a transient main-registry gap from worker divergence or confirm either as the persistent cause. Suggested priority: P2. Estimated size: medium.
Source: Rome Cloud user feedback, submitted 2026-10-27T23:58:12.905065Z, build 1.1.166. Analyzed against 815cff0410 (main); reporter build commit 90d32d8, 53 commits behind main. Filed automatically by Feedback Triage.
Activity