Ralphex depends on subagents feature, and there is a universal fully compatible alternative

[francescoalemanno]

Hello,
First of all thank you for the amazing tool!

Context

I know that the repo philosophy is to be Claude-only and to support other agentic tools via adapters.
This works well for most other tools, but there is Pi for example which is extremely well-made (many argue that it is better designed than Claude) and which does not offer sub-agents natively (altough it has many subagents extension packages).

Proposal

In order not to lose the immensely useful workflow of making sub-agents' findings "fight" with each other for priority (I observed myself that your pattern works really well, even more so with weaker models, not at Opus level).
We could add a compatibility mode that, instead of asking the LLM to run sub-agents, Ralphex runs the cli agent in parallel with the prompts set equal to each of the sub-agents prompt.

Each parallel agent will then emit signals like:
e.g. Quality agent: <<<RALPHEX:SA:QUALITY>>> ... REVIEW CONTENT ... <<<RALPHEX:SA:END>>>

These signals are gathered by the Ralphex host in order to prepare a package for the actual reviewer agent, which is run after the parallel batch finishes. The reviewer will then perform its job of deduplicating and dismissing conflicting observations, and will apply its fixes and commit.

I'd be happy to attempt a PR at this compat-mode.

as a final note: Anthropic is version after version pushing to make Claude incompatible with external harnessing, and most workflows that customise its behaviours, first class multi-agent support would become much easier if Ralphex does not rely on the Claude specific Task tool semantics.

[umputun]

thx for the detailed write-up, but I think there's a misconception about how ralphex uses subagents, and the proposed solution is solving a problem that doesn't really exist.

ralphex doesn't depend on subagents

subagent usage is confined to two review prompt files (review_first.txt, review_second.txt) that instruct Claude to launch parallel Task tool agents. the task execution phase, external review, plan creation, finalize - none of these use subagents at all.

more importantly, these prompts are plain text files that the user fully controls. you can replace them with anything - a prompt that asks the model to review from 5 perspectives sequentially in a single session works just as well. no code changes needed, the customization mechanism already exists. that's the whole point of making prompts configurable.

almost everything is user-configurable text

ralphex is deliberately designed so that almost all communication between the orchestrator and the AI agents is controlled via a set of text files and config params:

• ~/.config/ralphex/prompts/*.txt - all prompts (task, review, codex, plan creation, finalize)
• ~/.config/ralphex/agents/*.txt - review agent definitions
• claude_command + claude_args config options - swap the entire CLI backend to any tool that produces compatible output
• external_review_tool = custom + custom_review_script - plug in any external reviewer
• per-project overrides via .ralphex/ directory

if your tool doesn't support subagents, just write review prompts that don't use {{agent:name}} references. the orchestrator doesn't care - it passes whatever text you put in those files.

the signal-based parallel approach is over-engineering

running N parallel CLI sessions, parsing custom <<<RALPHEX:SA:QUALITY>>> signals from stdout, collecting results, then feeding them into a separate "reviewer" session - this is reimplementing the Task tool at the orchestrator level. it adds a new signal protocol, new parsing logic, new execution flow, all to replicate what the LLM tool already does natively.

the simpler path for tools without subagents: sequential review in a single session. you lose parallelism but the review quality is the same, and it requires zero code changes - just different prompt text.

re: Anthropic compatibility

I understand the concern about Anthropic potentially discouraging external harnessing, but I don't see this as a real risk for how ralphex works. the integration surface is the claude -p invocation and --output-format stream-json protocol, both documented and stable. and even this is fully configurable via claude_command/claude_args config options, so you can swap Claude for any CLI that speaks the same protocol (there's a codex wrapper included, and a gemini wrapper in the repo).

---

btw, proposals like this are better suited for Discussions, not Issues. moving it there.