Execute provider actions through a real transport and executor #17

Merged

Breadway merged 1 commit from feature/ram-48-provider-execution into main

2026-10-11 16:18:01 +08:00

Breadway commented

2026-10-11 16:17:56 +08:00

Owner

Closes RAM-48.

Provider relief was the biggest memory win breadram had and it could not happen: the adapters could describe an operation and the coordinator could execute one, but there was no transport to send it and no executor to turn an AdapterOperation into a write.

What this adds

adapter::HttpTransport, the real bounded loopback transport: one connection per request, connect/read/write timeouts, Connection: close, content-length and chunked parsing, a response size cap, credentials never rendered, and every resolved address tried in turn. Transport carries probes and OperationTransport carries writes, so a read-only path cannot be handed a write request. 13 tests.
adapter::AdapterExecutor, an Executor for the provider-backed actions: it revalidates the capability and the workload generation, refuses a release that needs proven idleness when activity is unknown, active, or too recent, checks every URL against the configured endpoints before anything is sent, applies, confirms with a read-only probe, and reports the memory the provider holds before and after so the effect is measured. 17 tests.
coordinator::ChainedExecutor, so host controls and provider actions run in one pass, dispatched by what each executor claims, with the first real measurement winning on either memory axis. 4 tests.
Adapter observation, so a provider is read before it is reasoned about: WorkloadAdapter::observe with ProbeOutcome, AdapterRegistry::observe, adapter::refresh_adapters, and inventory_is_current. The daemon refreshes every configured endpoint once per tick, records a failed read as a failure rather than leaving the previous success in place, and an empty current reading is a measured zero rather than a missing reading.
A journal entry now carries the consequence: when an executor returns a note it is recorded with the entry, so the cold start a model release costs is in the journal, not only in the pass that reported it.
Wire-level tests (crates/breadram-core/tests/provider_execution.rs) against an in-process fake Ollama service on a real socket: an idle release planned, performed, confirmed, and journalled with a measured effect; a backend with a live request never written to; an unreachable provider reported and unable to release; a provider that ignores the unload journalled as unconfirmed; monitor mode writing nothing.

No provider is installed on this host, so the end-to-end release is proven over a real socket against the fake, and the daemon's own read path is covered by a unit test against a closed port.

Verification

cargo fmt --all --check clean, cargo clippy --workspace --all-targets -- -D warnings clean, cargo test --workspace 376 passed, 0 failed (up from 349).

Deviations, stated rather than dropped

unix: endpoints are not supported, and say so: the hook protocol is a process protocol rather than HTTP, and needs a framed message format and a permission check that does not exist yet (RAM-15).
The requested host delta on a provider action is the plan's estimate from the workload's memory, while the provider knows its own figure. Both are now recorded in the entry rather than one being silently preferred.
With the default TrackingCompleteness::Unknown a resident model reports Active, so nothing is planned against it. Correct and conservative, but it means the release path is planned-never-executed until request tracking is declared: noted on RAM-45.

Closes RAM-48. Provider relief was the biggest memory win breadram had and it could not happen: the adapters could describe an operation and the coordinator could execute one, but there was no transport to send it and no executor to turn an `AdapterOperation` into a write. ## What this adds - **`adapter::HttpTransport`**, the real bounded loopback transport: one connection per request, connect/read/write timeouts, `Connection: close`, content-length and chunked parsing, a response size cap, credentials never rendered, and every resolved address tried in turn. `Transport` carries probes and `OperationTransport` carries writes, so a read-only path cannot be handed a write request. 13 tests. - **`adapter::AdapterExecutor`**, an `Executor` for the provider-backed actions: it revalidates the capability and the workload generation, refuses a release that needs proven idleness when activity is unknown, active, or too recent, checks every URL against the configured endpoints before anything is sent, applies, confirms with a read-only probe, and reports the memory the provider holds before and after so the effect is measured. 17 tests. - **`coordinator::ChainedExecutor`**, so host controls and provider actions run in one pass, dispatched by what each executor claims, with the first real measurement winning on either memory axis. 4 tests. - **Adapter observation**, so a provider is read before it is reasoned about: `WorkloadAdapter::observe` with `ProbeOutcome`, `AdapterRegistry::observe`, `adapter::refresh_adapters`, and `inventory_is_current`. The daemon refreshes every configured endpoint once per tick, records a failed read as a failure rather than leaving the previous success in place, and an empty current reading is a measured zero rather than a missing reading. - **A journal entry now carries the consequence**: when an executor returns a note it is recorded with the entry, so the cold start a model release costs is in the journal, not only in the pass that reported it. - **Wire-level tests** (`crates/breadram-core/tests/provider_execution.rs`) against an in-process fake Ollama service on a real socket: an idle release planned, performed, confirmed, and journalled with a measured effect; a backend with a live request never written to; an unreachable provider reported and unable to release; a provider that ignores the unload journalled as unconfirmed; monitor mode writing nothing. No provider is installed on this host, so the end-to-end release is proven over a real socket against the fake, and the daemon's own read path is covered by a unit test against a closed port. ## Verification `cargo fmt --all --check` clean, `cargo clippy --workspace --all-targets -- -D warnings` clean, `cargo test --workspace` 376 passed, 0 failed (up from 349). ## Deviations, stated rather than dropped - `unix:` endpoints are not supported, and say so: the hook protocol is a process protocol rather than HTTP, and needs a framed message format and a permission check that does not exist yet (RAM-15). - The requested host delta on a provider action is the plan's estimate from the workload's memory, while the provider knows its own figure. Both are now recorded in the entry rather than one being silently preferred. - With the default `TrackingCompleteness::Unknown` a resident model reports `Active`, so nothing is planned against it. Correct and conservative, but it means the release path is planned-never-executed until request tracking is declared: noted on RAM-45.

Breadway added 1 commit

2026-10-11 16:17:56 +08:00

feat: execute provider actions through a real transport and executor a25f352e44

Provider relief was the biggest memory win breadram had and it could not
happen: the adapters could describe an operation and the coordinator could
execute one, but there was no transport to send it and no executor to turn
an AdapterOperation into a write.

Add the real bounded loopback transport, an adapter executor on the same
Executor boundary as the host controls, and a chain so both run in one pass.
The daemon now reads every configured endpoint before policy runs and feeds
the answers back into the adapters, so a release is decided on a current
reading rather than a stale one and a failed read is never left as an old
success. A journal entry for a release now carries the cold start it costs.

Breadway merged commit 774fe49e5f into main

2026-10-11 16:18:01 +08:00

Breadway deleted branch feature/ram-48-provider-execution

2026-10-11 16:18:01 +08:00

Breadway referenced this pull request from a commit

2026-10-11 16:18:02 +08:00

Merge pull request 'Execute provider actions through a real transport and executor' (#17) from feature/ram-48-provider-execution into main

Breadway referenced this pull request

2026-10-11 16:18:20 +08:00

Execute provider actions through a real transport and executor #16