loki-mode 7.31.0 → 7.32.0

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '7.31.0'
+__version__ = '7.32.0'

package/mcp/server.py

@@ -537,6 +537,27 @@ def _build_fastmcp():
 
 mcp = _build_fastmcp()
 
+# Propagate the loki-mode VERSION into serverInfo.version.
+#
+# FastMCP 1.x exposes no `version=` kwarg (see _build_fastmcp above), so the
+# kwarg branch there never fires on the installed SDK. FastMCP DOES forward to
+# an underlying lowlevel `Server` (mcp._mcp_server), whose `version` attribute
+# is what `create_initialization_options()` reads into serverInfo at the
+# initialize handshake -- falling back to importlib.metadata.version("mcp")
+# (the SDK's OWN pip version, e.g. 1.27.x) when it is None. That fallback is
+# why the listing surfaced the SDK version instead of ours. Setting this
+# attribute is the only mechanism the installed SDK exposes to override
+# serverInfo.version; `version=` is a documented public parameter on the
+# lowlevel Server.__init__, so this is the supported field, not an internal
+# hack. Guarded so a future SDK that already set a version (e.g. via the
+# _build_fastmcp kwarg branch) is left untouched.
+try:
+    _inner = getattr(mcp, "_mcp_server", None)
+    if _inner is not None and getattr(_inner, "version", None) in (None, ""):
+        _inner.version = _version
+except Exception:  # pragma: no cover - defensive: never block server startup
+    pass
+
 # ============================================================
 # TOOLS - Functions Claude can call
 # ============================================================

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "loki-mode",
   "mcpName": "io.github.asklokesh/loki-mode",
-  "version": "7.31.0",
+  "version": "7.32.0",
   "description": "Loki Mode by Autonomi. Autonomous spec-to-product system: takes a PRD, GitHub issue, OpenAPI/JSON/YAML, or one-line brief to a deployed app via the RARV-C closure loop with 11 quality gates. Provider-agnostic (Claude Code, OpenAI Codex, Cline, Aider).",
   "keywords": [
     "agent",

package/skills/model-selection.md

@@ -5,7 +5,7 @@
 Since v6.81.0, the main RARV loop uses **one pinned model per session** instead of rotating models per RARV phase. This eliminates per-iteration tier churn and keeps the main loop behavior predictable.
 
 - Controlled by `LOKI_SESSION_MODEL` (default: `sonnet`).
-- Accepted values: `opus`, `sonnet`, `haiku`, or a raw tier name (`planning`, `development`, `fast`). Model names are mapped to abstract tiers (opus -> planning, sonnet -> development, haiku -> fast) so provider helpers resolve correctly.
+- Accepted values: `opus`, `sonnet`, `haiku`, or a raw tier name (`planning`, `development`, `fast`). The value names a **tier**, not a fixed model: it is mapped to an abstract tier (opus -> planning, sonnet -> development, haiku -> fast) which is then resolved to a concrete model through the provider config (`PROVIDER_MODEL_PLANNING/DEVELOPMENT/FAST`). On stock config the default `sonnet` pin = the **development** tier = `PROVIDER_MODEL_DEVELOPMENT` = **opus** (the model the runner actually dispatches). Under `LOKI_ALLOW_HAIKU=true` the development tier lowers to sonnet (and fast to haiku). This tier-to-model resolution is why `loki plan` quotes Opus on the stock default path: the quote names the dispatched model, not the pin alias. A mid-flight `.loki/state/model-override` is different: that alias is fed straight to `--model` (a `sonnet` override dispatches sonnet, no tier route).
 - `get_rarv_tier()` is retained and still used for subagent Task-tool dispatches (planning vs implementation vs fast work), not for the main loop.
 
 **Rollback (legacy RARV tier rotation in the main loop):**
@@ -32,7 +32,7 @@ Claude Fable 5 (alias `fable`, full id `claude-fable-5`) is Anthropic's most cap
 
 ### Mid-flight model switching
 
-You can change the model a **live run** uses, from the dashboard or by writing a state file. The switch applies at the **next iteration boundary** (each iteration spawns a fresh `claude -p`, which fixes the model per invocation, so it never changes mid-invocation). The override applies to the **current run only**: the runner clears a leftover override at the start of a fresh run, so a switch never silently carries into future runs. The override is also clamped by `LOKI_MAX_TIER`: if the operator set a cost ceiling, an override above it (e.g. `fable` under `LOKI_MAX_TIER=sonnet`) is clamped down with one honest log line, and the dashboard reports the clamped effective model. (The clamp is enforced on the bash runner, which is the live `start` route; the experimental Bun runner does not yet port it.)
+You can change the model a **live run** uses, from the dashboard or by writing a state file. The switch applies at the **next iteration boundary** (each iteration spawns a fresh `claude -p`, which fixes the model per invocation, so it never changes mid-invocation). The override applies to the **current run only**: the runner clears a leftover override at the start of a fresh run, so a switch never silently carries into future runs. The override is also clamped by `LOKI_MAX_TIER`: if the operator set a cost ceiling, an override above it is clamped down with one honest log line, and the dashboard reports the clamped effective model. The clamp is scoped, not blanket: on the override path a `sonnet` ceiling downgrades only `fable` (to `PROVIDER_MODEL_DEVELOPMENT`, opus by default), while a plain `opus` override stays opus; a `haiku` ceiling pins to `PROVIDER_MODEL_FAST`; an `opus` ceiling caps only `fable` back to opus. (The clamp is enforced on the bash runner, which is the live `start` route; the experimental Bun runner does not yet port it.)
 
 - Dashboard: the Model selector in the session-control panel. The Fable option shows its 2x-Opus cost; an inline notice discloses the iteration-boundary timing. It calls `POST /api/session/model`.
 - File / CLI: write an allowlisted alias (`haiku`, `sonnet`, `opus`, `fable`) to `.loki/state/model-override`. Empty or absent file reverts to the tier mapping. Invalid content is ignored (the runtime warns once). The value is allowlist-validated because it is fed straight into `claude --model`.
@@ -58,7 +58,11 @@ For diffs that move money or mutate infrastructure (payments, billing, spend aut
 
 ### Cost estimate honesty
 
-`loki plan` quotes a Fable run honestly and only for levers the runner actually honors. Set `LOKI_SESSION_MODEL=fable`, or have a pending `.loki/state/model-override` of `fable`, and the estimate uses Fable's $10/$50 pricing with a 2x-Opus note. The quote, the dashboard `effective` model, and the actual `claude --model` argument always agree: the same lever that changes the quote changes the run. `LOKI_MAX_TIER` is applied to the quote too, so the estimate never quotes a model above the operator's cost ceiling. (`LOKI_MODEL` is not a session lever and does not affect the run; use `LOKI_SESSION_MODEL`.)
+`loki plan` quotes the model the runner actually dispatches, on **every** path, for the levers the runner honors. This includes the stock no-lever default: the `sonnet` session pin resolves through the development tier to opus, so the quote names Opus (not Sonnet) on the default path. Set `LOKI_SESSION_MODEL=fable`, or have a pending `.loki/state/model-override` of `fable`, and the estimate uses Fable's $10/$50 pricing with a 2x-Opus note. The quote, the dashboard `effective` model, and the actual `claude --model` argument always agree on both routes: the session-pin tier route (no override) and the override-path clamp (override present) are each mirrored exactly in the estimator and the dashboard, so the same lever that changes the quote changes the run. `LOKI_MAX_TIER` is applied to the quote too, so the estimate never quotes a model above the operator's cost ceiling. (`LOKI_MODEL` is not a session lever and does not affect the run; use `LOKI_SESSION_MODEL`.)
+
+**Where the dispatched model is disclosed (precisely):**
+- Override path (a `.loki/state/model-override` alias): the runner logs the override and the clamp line (`model override: <model>`, plus the one honest clamp-down log line when `LOKI_MAX_TIER` reduces it); the dashboard `GET /api/session/model` `effective` field reports the clamped model.
+- Session pin / architect path (no override, the default): the runner logs the RARV effective-model line each iteration (`RARV Phase: <phase> -> Tier: <tier> (<model>)`), and `loki plan` discloses the provenance (`pinned via LOKI_SESSION_MODEL=<pin> -> <tier> tier -> <model>`) so the quoted model is legible against the pin alias. The `LOKI_FABLE_ARCHITECT=1` first-iteration Fable opt-in is disclosed as an extra architecture iteration in the quote.
 
 ---