luigidemasi opened a new issue, #92:
URL: https://github.com/apache/camel-kit/issues/92

   > Imported from 
[luigidemasi/camel-kit#221](https://github.com/luigidemasi/camel-kit/issues/221),
 opened by [luigidemasi](https://github.com/luigidemasi) on 2026-09-09.
   
   ---
   
   ## Summary
   
   Camel Ship currently runs its AI stages through a Pi worker, even when 
another supported agent invokes the Ship entry point. Add a native execution 
path that lets the active agent use its own subagents where supported, while 
preserving Pi and agents with a specific execution model.
   
   The controller remains responsible for stage sequencing, durable state, 
oversight, deterministic validation, and publication. Agent integrations 
provide task execution and return results to that shared controller.
   
   ## Expected behavior
   
   - Native-capable integrations request the next permitted task, dispatch 
their own subagent, and submit its structured result and candidate artifacts to 
Ship.
   - Pi and agents without suitable native subagents, or with an established 
execution model that should be preserved, keep their existing behavior.
   - Select and persist the execution mode for each run. Do not silently switch 
backends during recovery or duplicate an in-flight attempt.
   - Existing Pi runs, registered standalone/plugin commands, oversight 
policies, and supported project contracts remain compatible.
   - Native capability must be verified on the actual host; a fresh 
conversation alone does not establish workspace or tool isolation.
   
   ## Shared controller work
   
   - [ ] Define a versioned task/result handoff bound to run, stage, attempt, 
input identity, allowed workspace, and expected result schema.
   - [ ] Generalize Pi-specific execution, result recovery, and provenance only 
where needed to support the native path; retain Pi transport/session handling.
   - [ ] Add controller operations that return pending work to the active agent 
and accept completed results without moving state-machine decisions into 
prompts.
   - [ ] Make repeated work requests and identical result submissions 
idempotent. Reject stale inputs, wrong attempts, conflicting results, and late 
submissions after abort.
   - [ ] Specify and implement parent-session loss, native-child loss, timeout, 
cancellation, and resume. Stop workers or establish isolation/quiescence before 
retry or publication.
   - [ ] Preserve tool/workspace restrictions, immutable validation inputs, 
artifact policy checks, and guarded publication on both paths.
   - [ ] Keep old Pi run records and retained results readable and resumable; 
retain the existing standalone path when no native host is present.
   
   ## Agent subtasks
   
   - [x] https://github.com/luigidemasi/camel-kit/issues/222
   - [x] https://github.com/luigidemasi/camel-kit/issues/223
   - [ ] https://github.com/luigidemasi/camel-kit/issues/224
   - [x] https://github.com/luigidemasi/camel-kit/issues/225
   - [x] https://github.com/luigidemasi/camel-kit/issues/226
   - [ ] https://github.com/luigidemasi/camel-kit/issues/227
   - [ ] https://github.com/luigidemasi/camel-kit/issues/228
   - [ ] https://github.com/luigidemasi/camel-kit/issues/229
   
   ## Affected repositories and website impact
   
   **Website impact: required.** Native execution changes the documented 
architecture, prerequisites, supported targets, invocation, and recovery 
behavior.
   
   | Repository | Work | Validation |
   |---|---|---|
   | luigidemasi/camel-kit | Shared controller handoff, agent integrations, 
compatibility, tests, CLI help, core docs, changelog | Focused regressions, 
full wrapper build, sourcecheck, generated-resource and real 
registered-entry-point checks |
   | luigidemasi/camel-kit-web | Execution model, per-agent support, 
prerequisites, invocation and recovery documentation | npm ci, npm run build, 
rendered-page/link checks |
   
   - [ ] https://github.com/luigidemasi/camel-kit-web/issues/45 — complete 
companion website documentation before closing this parent.
   
   ## Implementation and validation
   
   Implement the shared handoff with one complete native adapter first; Claude 
Code is a suggested pilot. Exercise a second host with a different 
dispatch/permission model, then complete every eligible agent subtask. Preserve 
specific execution models rather than forcing a universal migration.
   
   Use isolated branches/worktrees. Link each implementation PR and the 
matching website changes to this tracker. Shared controller changes must be 
available before dependent native adapters are enabled; publish documentation 
in step with available behavior.
   
   Completion requires real native-child runs through the generated, registered 
entry point, including a complete workflow, pause/resume, abort/recovery, and 
independent checks of the resulting artifacts. Cover crash windows, wrong/stale 
attempts, duplicate submissions, workspace boundaries, and legacy Pi recovery 
with focused regressions. Generated files or mocked responses alone do not 
establish native support.
   
   Run the required sequential Maven wrapper gates (`./mvnw -B clean install` 
and sourcecheck), shipped-resource checks for changed assets, packaged 
standalone/JBang-plugin checks, and exact-head hosted CI. Retain Pi's existing 
authenticated certification path. Record a concrete capability or 
execution-model reason for any target whose existing path must be preserved; 
missing validation is unfinished work.
   
   This work changes AI dispatch. It does not by itself broaden Ship's 
supported operating systems, Camel runtimes, or artifact formats.
   
   ## Context
   
   This extends the controller introduced by luigidemasi/camel-kit#149 and its 
product-boundary work in luigidemasi/camel-kit#182.
   
   Current implementation: 
[ShipCoordinator](https://github.com/luigidemasi/camel-kit/blob/9fa9771e15b3648b999883c4005de9f6383061b4/camel-kit-core/src/main/java/io/github/luigidemasi/camelkit/ship/controller/ShipCoordinator.java).
 Current per-agent integration surfaces are in [the agent 
architecture](https://github.com/luigidemasi/camel-kit/blob/9fa9771e15b3648b999883c4005de9f6383061b4/docs/architecture.md#sub-agent-driven-execution).
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]

Reply via email to