Commit graph

4 commits

Author SHA1 Message Date
7cba1ae57a feat: migrate orchestrator-llm to fai-module-sdk (v0.3.0)
Some checks failed
CI / Linux x86_64 (Forgejo) (push) Failing after 1m34s
The module now uses the #[fai_module] macro from
fai-module-sdk v0.1.2. The behavior is unchanged from v0.2.1 —
deterministic stub when llm_endpoint is empty, real Ollama-shaped
HTTP otherwise. The migration drops:

  - The hand-written wit_bindgen::generate! invocation
  - The exports::fai::platform::invoke::Guest impl boilerplate
  - The unsafe_op_in_unsafe_fn allow workaround
  - The manual export!() macro call
  - The text_input helper that matched Payload variants by hand
  - The wit/ directory mirror of the platform's frozen WIT

In their place, the module body is one annotated function:

  #[fai_module]
  pub fn invoke(_ctx: Context, inputs: Inputs)
      -> Result<Outputs, ModuleError>
  { ... }

Compared with v0.2.1, lib.rs lost ~70 lines of WASM plumbing.
The SDK macro emits the WIT glue only on target_arch = wasm32
so the host build now compiles cleanly without a manual gate.
That removed two #[cfg(target_arch = wasm32)] blocks in the
process.

The Cargo.toml dep on wit-bindgen is replaced by a single git
dep on fai-module-sdk (branch=main, transitively pinning v0.1.2).
The macros crate comes through transitively. waki stays as a
target-gated dep — the SDK does not (yet) abstract HTTP.

Output payload extends the v0.2.1 schema additively: alongside
the existing 'plan' (JSON Plan), the module now emits
'model_endpoint' and 'model_name' so consumers (and audit logs)
can identify which LLM produced which plan. Existing flows that
only consume 'plan' are unaffected.

Wasm artifact verified to build with v1.0 imports. Platform-side
integration tests in fai/platform :: orchestrator_llm_path.rs
and orchestrator_module.rs pass against the migrated wasm with
no platform-side changes. Total platform suite: 151 green.

The wit/ directory is removed because the SDK ships the WIT
inline via include_str! at proc-macro compile time. Module repos
no longer maintain their own mirror copy.

Bumps the module to 0.3.0 (minor — output schema additively
extended).

Signed-off-by: flemming-it <sf@flemming.it>
2026-05-02 11:49:10 +02:00
62792a20aa chore(wit): bump mirror to fai:platform@1.0
All checks were successful
CI / Linux x86_64 (Forgejo) (push) Successful in 2m42s
The platform repo froze its WIT contract at v1.0 and a
matching update is required here so the orchestrator-llm
component links against the same package version the hub
provides at runtime.

The wit/world.wit copy in this repo is a byte-for-byte
mirror of fai/platform :: wit/world.wit. Once the planned
fai-module-sdk crate ships, the mirror disappears — modules
will pull WIT bindings transitively through the SDK.

Bumps the module patch version 0.2.0 -> 0.2.1.

Signed-off-by: flemming-it <sf@flemming.it>
2026-05-01 00:45:49 +02:00
7d07937913 feat: real LLM call via wasi-http (Ollama adapter) — v0.2.0
The orchestrator now calls a configured Ollama-compatible LLM
endpoint over wasi-http and emits the LLM-generated plan back
to the hub. When llm_endpoint is empty, the v0.1.0 deterministic
stub is preserved as fallback (useful for tests and air-gapped
environments).

- module.yaml: bump to 0.2.0; new inputs llm_endpoint,
  llm_model, llm_api_key; net: permissions for OpenAI,
  Anthropic, localhost, 127.0.0.1 (Ollama defaults)
- src/llm.rs: LlmClient trait, OllamaParams, build_user_prompt,
  build_ollama_body, extract_ollama_content, generate_plan;
  9 unit tests for prompt-building and response-parsing
- src/plan.rs: tagged enum with Explain/SaveFlow/RunFlow/
  InstallModule variants; parse_and_validate; build_stub_plan
  (renamed from build_plan); validate(); 6 unit tests
- src/lib.rs: WakiClient implementing LlmClient over the waki
  WASI-0.2 HTTP crate (wasm32-only); empty endpoint → stub
- Cargo.toml: 0.2.0; thiserror dep; waki dep gated on wasm32

17 host-side tests pass. WASM bundle is 386KB (was 74KB; the
delta is wasi-http via waki, expected). The deny path is
exercised by the platform-side integration test in
fai_platform/crates/fai_hub/tests/orchestrator_llm_path.rs.

Signed-off-by: flemming-it <sf@flemming.it>
2026-04-29 21:27:07 +02:00
3bb8fc9e70 feat: deterministic stub orchestrator (v0.1.0)
The orchestrator-llm module turns a goal text into a Plan JSON
that the F∆I hub can apply. v0.1.0 is a deterministic stub —
it always emits a three-step plan (explain, explain-stub-notice,
save_flow with debug.echo) regardless of the goal.

Why ship a stub: it validates the module → plan → apply pipeline
end-to-end before adding LLM integration. The output is a
structurally valid Plan that round-trips cleanly through
fai_hub::plan::Plan deserialization.

v0.2.0 will replace the stub body with an actual LLM call once
the F∆I platform exposes outbound HTTP to permitted modules via
wasi-http.

Files:
- module.yaml — capability orchestrator.plan@0.1.0, no permissions
- wit/world.wit — copy of fai:platform@0.1.0
- src/lib.rs — wit_bindgen Guest impl that delegates to plan.rs
- src/plan.rs — pure plan-builder (host + WASM compatible)
- tests/plan_compatibility.rs — plan JSON shape assertions
- rust-toolchain.toml — pin to Rust 1.86 with wasm32-wasip2 target
- README.md — usage and status

Signed-off-by: flemming-it <sf@flemming.it>
2026-04-29 17:43:46 +02:00