@hilbras/omninode 2.0.0-alpha.1 → 2.0.0-alpha.11

package/README.md

@@ -18,7 +18,9 @@ existing intelligence; it does not try to become another model.
 
 ## Status
 
-**v2.0.0-alpha.1 — v2 Phase 1: Architecture & API Stabilization** (the v2
+**v2.0.0-alpha.11 — v2 Phase 11: Planner v2**
+(the reliability & interoperability line: architecture → execution →
+protocol → adapters → pipelines → providers → OmniHilbras → memory → reports) (the v2
 roadmap is [docs/ROADMAP_V2.md](docs/ROADMAP_V2.md); the v1 plan is complete:
 the
 full loop from §28 runs in one command —
@@ -27,14 +29,75 @@ planner → plan → execution → result` — with audit logging, error recover
 production hardening (see [docs/DEVELOPMENT_PLAN.md](docs/DEVELOPMENT_PLAN.md)
 for the plan and [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the design).
 
-### What changed in v2 Phase 1
-
-- All local stores now build on one shared `JsonFileStore` (atomic writes,
-  serialized read-modify-write, schema-versioned files, corruption tolerance)
-  with a `ProjectStores` facade.
-- Dependency direction corrected (adapters no longer import engine modules).
-- API surface stabilized and guarded: [API_STABILITY.md](docs/API_STABILITY.md),
-  [API.md](docs/API.md), [DEPRECATIONS.md](docs/DEPRECATIONS.md).
+### What's new in v2 (so far)
+
+- **Phase 11 — Planner v2**: the planner now receives the full §15 input —
+  task, project context, agent reports, **aggregated findings (consensus and
+  conflicts included)**, memory and **constraints** (settable per pipeline
+  step). Plans carry an explicit goal, every generated plan is **schema
+  validated**, invalid output produces a **structured error** or falls back to
+  heuristic planning, and a guard test proves the planner module depends on no
+  specific provider.
+- **Phase 10 — Aggregation Improvements**: when agents disagree,
+  OmniNode now says so instead of flattening the answer — **conflict
+  detection** preserves every position (severity spread, divergent
+  recommended fixes), multi-agent **agreements** are recorded as
+  consensus, confidence is preserved per source, and the combined
+  report carries provenance-rich metadata. Still fully deterministic.
+- **Phase 9 — Report System v2**: reports are validated machine-readable
+  objects — `validateReport()` runs on everything entering the system and
+  malformed reports are rejected with diagnostics instead of poisoning
+  aggregation. Findings gain description/confidence/source/recommendation,
+  evidence is typed (`file`, `line`, `url`, `command`, `observation`,
+  `artifact`), and reports carry artifacts and run provenance.
+  Details: [REPORTS.md](docs/REPORTS.md).
+- **Phase 8 — Remembera / Memory v2**: the memory contract is formalized
+  (`retrieve` / `store` / `search` / `metadata` with v1 `query` / `write`
+  aliases kept working), memory gains structured **categories** (decisions,
+  problems, conventions, tasks…), retrieval is **task-oriented and budgeted**
+  (never "load all memory"), and `memory.required: true` makes backend
+  failures fatal instead of tolerated. Details: [MEMORY.md](docs/MEMORY.md).
+- **Phase 7 — OmniHilbras Integration v2**: the dedicated adapter now
+  captures gateway-level metadata (version, tier, region) and attaches it
+  to discovered models, honors provider-level `timeout_ms`, and supports
+  streaming completions (`stream()` over SSE) — while a guard test proves
+  the core references OmniHilbras *only* through the provider layer, and a
+  full pipeline runs with no OmniHilbras configured at all.
+- **Phase 6 — Provider Infrastructure v2**: standardized provider contract
+  (`providerId`, secret-free `authentication`, declared `capabilities`,
+  `connect()` / `getModel()` discovery), richer model metadata (context
+  window, input/output modalities, tool/vision/structured-output/streaming
+  support), and **unified error normalization** — every provider failure now
+  carries a stable `kind` (auth, rate limit, invalid request, model not
+  found, timeout, network, server, unknown) plus `retryable` and
+  `retryAfterMs`. Details: [PROVIDERS.md](docs/PROVIDERS.md).
+- **Phase 5 — Pipeline Lifecycle**: first-class pipeline cancellation
+  (`omninode pipeline cancel <run-id>`: stop scheduling → cancel running
+  tasks → terminate agent processes → persist a `cancelled` run), `pipeline
+  validate` for pre-run checks (deps, cycles, agents, model references),
+  explicit run/step **attempts** in the run record, the
+  `IPipelineExecutor` contract, and recovery awareness — runs left `running`
+  without a `finishedAt` are flagged as interrupted in `pipeline runs`.
+- **Phase 4 — Agent Adapter Hardening**: process-tree cleanup (agents run in
+  their own process group; timeout, cancel and Ctrl-C reap the whole tree, so
+  no orphaned children), explicit environment policies
+  (`inherit` / `allowlist` / `denylist` / `explicit`), working-directory
+  validation against project boundaries, per-stream output limits that kill
+  runaway agents, binary-safe output decoding, human-readable exit/signal
+  reporting, and a native `IAgentAdapter` registry so new integration styles
+  plug in without touching core.
+- **Phase 3 — Agent Protocol v2**: versioned protocol
+  (`omninode-agent-protocol/2`), fully correlated envelopes, the complete
+  message-type set, HELLO handshake, QUESTION/RESPONSE round-trips, and
+  malformed input that can never crash OmniNode. Spec:
+  [PROTOCOL.md](docs/PROTOCOL.md).
+- **Phase 2 — Execution Reliability**: a task killed after dispatch is
+  `unknown`, **not** `failed` — retries never duplicate side effects blindly;
+  bounded retries with backoff, execution identity and history.
+- **Phase 1 — Architecture & API Stabilization**: one shared `JsonFileStore`
+  behind every local store, corrected dependency direction, and a guarded
+  public API ([API_STABILITY.md](docs/API_STABILITY.md), [API.md](docs/API.md),
+  [DEPRECATIONS.md](docs/DEPRECATIONS.md)).
 
 ### What's in v1.0
 
@@ -296,7 +359,7 @@ required for the core engine.
 - [Architecture](docs/ARCHITECTURE.md) — module map, workflow, local state
 - [Security model](docs/SECURITY.md) — secrets, process execution, env isolation, audit log
 - [API stability policy](docs/API_STABILITY.md) and [public API inventory](docs/API.md)
-- [Deprecations](docs/DEPRECATIONS.md) · [Changelog](CHANGELOG.md) · [v2 roadmap](docs/ROADMAP_V2.md)
+- [Agent Protocol v2 spec](docs/PROTOCOL.md) · [Providers](docs/PROVIDERS.md) · [Memory](docs/MEMORY.md) · [Reports](docs/REPORTS.md) · [Deprecations](docs/DEPRECATIONS.md) · [Changelog](CHANGELOG.md) · [v2 roadmap](docs/ROADMAP_V2.md)
 - [Example project](examples/demo/) — a runnable end-to-end demo (`./demo.sh`)
 
 Prefer a container? `docker build -t omninode .` then