loki-mode 7.31.0 → 7.32.1

package/autonomy/mcp-launch.sh

@@ -113,15 +113,22 @@ _ml_python() {
 # `--check-sdk` probe, which runs the exact loader the server uses and exits 0
 # only when FastMCP loaded.
 #
-# Critical: we set PYTHONPATH to the install root and DO NOT cd into it, so the
-# probe exercises the SAME module resolution as the real launch (which preserves
-# the user's cwd). The redirect of stdin from /dev/null is insurance: if the
-# pip SDK's own `mcp.server` were ever reached, its stub starts a stdio receive
-# loop; the EOF makes it exit instead of hanging.
+# Critical: we probe with the SAME FILE-EXEC form the launch uses
+# (`"$root/mcp/server.py"`, NOT `-m mcp.server`), with PYTHONPATH set to the
+# install root and WITHOUT cd-ing into it, so the probe exercises byte-identical
+# module resolution to the real launch (which preserves the user's cwd). This
+# matters because `-m mcp.server` puts the user's cwd at sys.path[0] AHEAD of
+# PYTHONPATH=$root, so a cwd that happens to contain a regular `mcp/` python
+# package would shadow Loki's server during the probe (false SDK-missing) while
+# the file-exec launch -- immune to cwd shadowing -- would succeed. Probing by
+# file path keeps probe and launch resolving the IDENTICAL module. The redirect
+# of stdin from /dev/null is insurance: if the pip SDK's own `mcp.server` were
+# ever reached, its stub starts a stdio receive loop; the EOF makes it exit
+# instead of hanging.
 _ml_sdk_importable() {
     local py="$1" root="$2"
     PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" \
-        "$py" -m mcp.server --check-sdk </dev/null >/dev/null 2>&1
+        "$py" "$root/mcp/server.py" --check-sdk </dev/null >/dev/null 2>&1
 }
 
 # _ml_print_manual <root> <venv>: print the honest manual install commands.
@@ -135,7 +142,7 @@ _ml_print_manual() {
     printf 'Install the MCP server dependencies manually:\n' >&2
     printf "  python3 -m venv '%s'\n" "$venv" >&2
     printf "  '%s/bin/pip' install -r '%s/mcp/requirements.txt'\n" "$venv" "$root" >&2
-    printf "  PYTHONPATH='%s' '%s/bin/python' -m mcp.server\n" "$root" "$venv" >&2
+    printf "  PYTHONPATH='%s' '%s/bin/python' '%s/mcp/server.py'\n" "$root" "$venv" "$root" >&2
 }
 
 _ml_help() {
@@ -189,7 +196,8 @@ EOF
 # mcp_launch_main: dispatcher invoked by cmd_mcp() (autonomy/loki) or directly.
 mcp_launch_main() {
     # Split argv into launcher-owned flags (consumed here) and server argv
-    # (forwarded verbatim to `python -m mcp.server`). The server's argparse only
+    # (forwarded verbatim to the file-exec launch `python "$root/mcp/server.py"`).
+    # The server's argparse only
     # accepts --transport/--port/--check-sdk; forwarding a launcher flag like
     # --yes would make it abort with exit 2, so launcher flags MUST be stripped.
     # A bare `--` ends launcher parsing: everything after it is forwarded as-is
@@ -255,20 +263,23 @@ mcp_launch_main() {
     fi
 
     # 3. If the venv already has the SDK, use it directly. The server is launched
-    #    with PYTHONPATH=$root (NOT by cd-ing) so the user's cwd is preserved for
+    #    by FILE PATH ($root/mcp/server.py) rather than `-m mcp.server`, with
+    #    PYTHONPATH=$root (NOT by cd-ing) so the user's cwd is preserved for
     #    .loki resolution; see _ml_sdk_importable for why.
-    #    Known narrow residual: if the user's cwd itself contains a Python
-    #    package literally named mcp/ with a server submodule, python -m puts
-    #    the cwd ahead of PYTHONPATH and that package wins. Essentially never
-    #    true for real projects; documented rather than fought.
+    #    Running the file directly avoids the runpy RuntimeWarning that `-m`
+    #    emits (the local mcp/ package is imported during SDK-namespace setup
+    #    before runpy executes mcp.server). server.py uses only absolute imports
+    #    (e.g. `from mcp._sdk_loader import ...`), which resolve via PYTHONPATH=$root
+    #    under file execution. File-exec also removes the old narrow cwd-shadowing
+    #    residual: an explicit path can never be shadowed by a cwd `mcp/` package.
     if [ -x "$venv_py" ] && _ml_sdk_importable "$venv_py" "$root"; then
-        exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$venv_py" -m mcp.server "$@"
+        exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$venv_py" "$root/mcp/server.py" "$@"
     fi
 
     # 4. If the BASE python already has the SDK (e.g. user pip-installed it),
     #    use it -- no venv needed.
     if _ml_sdk_importable "$base_py" "$root"; then
-        exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$base_py" -m mcp.server "$@"
+        exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$base_py" "$root/mcp/server.py" "$@"
     fi
 
     # 5. SDK missing. Decide whether we may bootstrap.
@@ -374,7 +385,7 @@ mcp_launch_main() {
         return 2
     fi
     printf "%sMCP dependencies ready. Launching server over stdio ...%s\n" "$_ML_BOLD" "$_ML_NC" >&2
-    exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$venv_py" -m mcp.server "$@"
+    exec env PYTHONPATH="$root${PYTHONPATH:+:$PYTHONPATH}" "$venv_py" "$root/mcp/server.py" "$@"
 }
 
 # Executed directly (tests, manual): run the dispatcher.

package/autonomy/run.sh

@@ -12328,13 +12328,30 @@ except Exception as exc:
             # helpers (which expect tier names) resolve correctly. Unknown
             # model strings are passed through as-is; provider loaders fall
             # back to a sane default.
-            case "${LOKI_SESSION_MODEL:-sonnet}" in
+            #
+            # Normalize case + surrounding whitespace BEFORE the match so
+            # 'OPUS' and ' opus ' resolve identically to 'opus'. We do NOT use
+            # loki_normalize_model_alias here: that helper is the narrow
+            # OVERRIDE-file allowlist (haiku|sonnet|opus|fable) and would strip
+            # the documented tier-name pins (planning|development|fast) to
+            # empty, collapsing them onto the default tier. The session pin
+            # legitimately accepts tier names (skills/model-selection.md), and
+            # the estimator + dashboard mirror this exact tier route, so the
+            # canonical session-pin rule is trim+lowercase WITHOUT the alias
+            # allowlist. Interior whitespace is preserved (so 'fab le' stays a
+            # junk value that falls through the '*' default arm), matching the
+            # estimator/dashboard ports.
+            local _session_pin="${LOKI_SESSION_MODEL:-sonnet}"
+            _session_pin="${_session_pin#"${_session_pin%%[![:space:]]*}"}"
+            _session_pin="${_session_pin%"${_session_pin##*[![:space:]]}"}"
+            _session_pin="$(printf '%s' "$_session_pin" | tr '[:upper:]' '[:lower:]')"
+            case "$_session_pin" in
                 opus)   CURRENT_TIER="planning" ;;
                 sonnet) CURRENT_TIER="development" ;;
                 haiku)  CURRENT_TIER="fast" ;;
                 fable)  CURRENT_TIER="fable" ;;
-                planning|development|fast) CURRENT_TIER="${LOKI_SESSION_MODEL}" ;;
-                *)      CURRENT_TIER="${LOKI_SESSION_MODEL}" ;;
+                planning|development|fast) CURRENT_TIER="$_session_pin" ;;
+                *)      CURRENT_TIER="$_session_pin" ;;
             esac
         fi
         # Architect opt-in (LOKI_FABLE_ARCHITECT=1): route ONLY the first

package/bin/loki

@@ -65,6 +65,21 @@ elif [ "${BUN_FROM_SOURCE:-0}" = "1" ] || [ "${BUN_FROM_SOURCE:-}" = "true" ]; t
     fi
 elif [ -f "$REPO_ROOT/loki-ts/dist/loki.js" ]; then
     BUN_CLI="$REPO_ROOT/loki-ts/dist/loki.js"
+    # Stale-dist freshness guard. dist/loki.js is gitignored (loki-ts/.gitignore)
+    # and rebuilt by package.json's prepack and the release Docker build, so on
+    # released channels (npm/Docker/brew) src/cli.ts is absent and this guard is
+    # a no-op -- dist is always current there. On a DEV machine / worktree the
+    # gitignored dist can predate the current dispatcher: e.g. a new shim->Bun
+    # route (`report kpis`) added in src that the old dist bundle does not know,
+    # which would make the canonical form fail with "Unknown command" while a
+    # deprecated alias the old dist still knows silently works -- the exact
+    # deprecation inversion the report-kpis route exists to prevent. So when src
+    # exists AND is newer than dist (`-nt`, available on bash 3.2), prefer the src
+    # form so dev runs never execute a stale dispatcher. Released channels have no
+    # src, so they keep using dist unchanged.
+    if [ -f "$REPO_ROOT/loki-ts/src/cli.ts" ] && [ "$REPO_ROOT/loki-ts/src/cli.ts" -nt "$REPO_ROOT/loki-ts/dist/loki.js" ]; then
+        BUN_CLI="$REPO_ROOT/loki-ts/src/cli.ts"
+    fi
 elif [ -f "$REPO_ROOT/loki-ts/src/cli.ts" ]; then
     BUN_CLI="$REPO_ROOT/loki-ts/src/cli.ts"
 else
@@ -131,6 +146,43 @@ if [ "${1:-}" = "trust" ]; then
     done
 fi
 
+# CLI consolidation (Phase B): `report kpis` is the canonical form of the
+# Bun-only KPI snapshot. The `report` noun is otherwise bash-owned (every other
+# subcommand -- session/metrics/cost/export/share/dogfood -- routes to bash
+# cmd_report), so it is NOT in the Bun allowlist below. But kpis has no bash
+# implementation (it reuses the canonical cost arithmetic in runner/budget.ts),
+# so the canonical `report kpis` must reach the Bun handler -- otherwise the
+# canonical form would print the honest "requires Bun" message on a Bun machine
+# while the deprecated `kpis` alias actually worked, inverting the deprecation.
+# Route `report kpis` to Bun when `kpis` is the report SUBCOMMAND, i.e. the FIRST
+# non-flag token after `report`. This satisfies `report kpis`, `report kpis
+# --json`, and `report --json kpis` (the flag-anywhere orderings the trust-detail
+# precedent established), while NOT hijacking `kpis` when it appears as a
+# positional VALUE of a different report subcommand. `report export json kpis`
+# (kpis = the export output filename) must keep working exactly as on main
+# (v7.31.0): exit 0, file `kpis` created. So we scan past leading flags, take the
+# first real token, and only route to Bun if it is literally `kpis`. Fire the
+# same cli_command telemetry the Bun case-arm below fires (command=report), so a
+# Bun-routed `report kpis` is not invisible to usage analytics (the v7.8.2 parity
+# the case-arm comment documents). Backgrounded, FD-detached, opt-out honored by
+# loki_telemetry itself.
+if [ "${1:-}" = "report" ]; then
+    _report_first_sub=""
+    for _report_arg in "${@:2}"; do
+        case "$_report_arg" in
+            -*) continue ;;
+            *) _report_first_sub="$_report_arg"; break ;;
+        esac
+    done
+    if [ "$_report_first_sub" = "kpis" ]; then
+        if command -v curl &>/dev/null && [ -f "$REPO_ROOT/autonomy/telemetry.sh" ]; then
+            ( SCRIPT_DIR="$REPO_ROOT/autonomy"; source "$SCRIPT_DIR/telemetry.sh" 2>/dev/null && loki_telemetry "cli_command" "command=${1:-}" 2>/dev/null ) >/dev/null 2>&1 </dev/null &
+            disown 2>/dev/null || true
+        fi
+        exec bun "$BUN_CLI" "$@"
+    fi
+fi
+
 # Commands ported in Phase 2 -- route to Bun. Everything else goes to bash.
 # Two-token routes (provider show/list, memory list/index) match on the first
 # token only; the Bun dispatcher handles subcommand routing internally.

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "7.31.0"
+__version__ = "7.32.1"
 
 # Expose the control app for easy import
 try:

package/dashboard/server.py

@@ -2233,6 +2233,29 @@ def _normalize_session_model(raw: str | None) -> str:
     return val if val in _SESSION_MODEL_ALLOWLIST else ""
 
 
+# Session-pin allowlist is BROADER than the override-file allowlist above.
+# run.sh's session-pin case (run.sh:12331) accepts the four model aliases AND
+# the three raw tier names (planning|development|fast) -- documented at
+# skills/model-selection.md:8. The OVERRIDE file / POST path keeps the narrow
+# _SESSION_MODEL_ALLOWLIST because that value is fed straight to `claude
+# --model`, where tier names are not valid. The session pin is a tier route, so
+# tier names ARE valid pins.
+_SESSION_PIN_ALLOWLIST = _SESSION_MODEL_ALLOWLIST + ("planning", "development", "fast")
+
+
+def _normalize_session_pin(raw: str | None) -> str:
+    """Normalize a LOKI_SESSION_MODEL pin value (aliases + raw tier names).
+
+    Mirrors run.sh's session-pin case: trim + lowercase, accept the four model
+    aliases and the three tier names. Interior whitespace is preserved (so
+    "fab le" stays junk and falls through to the default tier, exactly like the
+    runner's "*" arm). Use this for the session-pin (no-override) derivation;
+    use _normalize_session_model for the override-file / POST path.
+    """
+    val = (raw or "").strip().lower()
+    return val if val in _SESSION_PIN_ALLOWLIST else ""
+
+
 # Provider-config model resolution mirror.
 #
 # SYNC: This is a byte-faithful python port of the claude provider's tier->model
@@ -2268,6 +2291,17 @@ def _provider_model_development() -> str:
     )
 
 
+def _provider_model_planning() -> str:
+    # claude.sh:65 -> LOKI_CLAUDE_MODEL_PLANNING > LOKI_MODEL_PLANNING > opus.
+    # CLAUDE_DEFAULT_PLANNING is always opus (LOKI_ALLOW_HAIKU lowers only the
+    # development and fast defaults, not planning).
+    return (
+        os.environ.get("LOKI_CLAUDE_MODEL_PLANNING")
+        or os.environ.get("LOKI_MODEL_PLANNING")
+        or "opus"
+    )
+
+
 def _clamp_to_max_tier(alias: str) -> str:
     """Apply the operator LOKI_MAX_TIER ceiling to a model alias.
 
@@ -2297,44 +2331,102 @@ def _clamp_to_max_tier(alias: str) -> str:
     return alias
 
 
+def _resolve_session_pin(alias: str) -> str:
+    """Resolve a session-pin alias the way the runner's NO-OVERRIDE path does.
+
+    The runner does NOT feed a session pin straight to --model. It maps the alias
+    to an abstract TIER (run.sh:12331 -- opus->planning, sonnet->development,
+    haiku->fast, fable->fable) and resolves that tier through
+    resolve_model_for_tier (claude.sh:353), then applies
+    loki_apply_max_tier_clamp(model, REAL_tier). This DIFFERS from
+    _clamp_to_max_tier (the override-path clamp): a 'sonnet' SESSION pin
+    dispatches OPUS (development tier -> PROVIDER_MODEL_DEVELOPMENT=opus on stock
+    config), whereas a 'sonnet' OVERRIDE file dispatches sonnet (fed straight to
+    --model). Use this for the no-override `default`/`effective` derivation so the
+    dashboard reports the model the run actually dispatches on the default path.
+
+    SYNC: byte-faithful with run.sh's session-pin case + claude.sh
+    resolve_model_for_tier + loki_apply_max_tier_clamp, and with the estimator's
+    _resolve_session_pin in autonomy/loki. Locked by the session-pin parity matrix
+    in tests/test-model-override.sh.
+    """
+    pin_tier = {
+        "opus": "planning",
+        "sonnet": "development",
+        "haiku": "fast",
+        "fable": "fable",
+        # Raw tier-name pins (run.sh:12336 passthrough arm) map to their own
+        # tier, NOT through the alias table. pin=fast -> fast tier ->
+        # PROVIDER_MODEL_FAST, matching the runner's dispatch instead of
+        # collapsing onto development.
+        "planning": "planning",
+        "development": "development",
+        "fast": "fast",
+    }.get((alias or "").strip().lower(), "development")
+    if pin_tier == "planning":
+        model = _provider_model_planning()
+    elif pin_tier == "fast":
+        model = _provider_model_fast()
+    elif pin_tier == "fable":
+        model = "fable"
+    else:  # development (and the unknown-alias '*' fallthrough)
+        model = _provider_model_development()
+    max_tier = (os.environ.get("LOKI_MAX_TIER") or "").strip().lower()
+    if not max_tier:
+        return model
+    if max_tier == "haiku":
+        return _provider_model_fast()
+    if max_tier == "sonnet":
+        # claude.sh sonnet-cap downgrades planning/fable tiers (or a fable model)
+        # to PROVIDER_MODEL_DEVELOPMENT; development/fast pass through.
+        if pin_tier in ("planning", "fable") or model == "fable":
+            return _provider_model_development()
+        return model
+    if max_tier == "opus":
+        return "opus" if model == "fable" else model
+    return model
+
+
 @app.get("/api/session/model", dependencies=[Depends(auth.require_scope("read"))])
 async def get_session_model():
     """Report the live run's model override and the effective default.
 
     `override` is the alias currently written to .loki/state/model-override
-    (None when no override is active). `default` is the session model the run
-    falls back to when there is no override (LOKI_SESSION_MODEL or the catalog
-    default). `effective` is the model the next iteration will actually use,
-    after the LOKI_MAX_TIER cost ceiling is applied (so the dashboard never
-    reports a model the run would clamp down).
-
-    The clamp resolves through the SAME provider config the runner uses
-    (LOKI_ALLOW_HAIKU plus the LOKI_CLAUDE_MODEL_FAST/DEVELOPMENT and
-    LOKI_MODEL_FAST/DEVELOPMENT overrides): _clamp_to_max_tier mirrors
-    providers/claude.sh loki_apply_max_tier_clamp byte-for-byte (locked by the
-    resolver parity matrix in tests/test-model-override.sh). So for the OVERRIDE
-    case -- the feature this endpoint exists for -- the reported `effective` model
-    equals the model the runner's mid-flight override path dispatches, given the
-    same environment.
+    (None when no override is active). `default` is the session pin alias the run
+    falls back to when there is no override (LOKI_SESSION_MODEL or "sonnet").
+    `effective` is the model the next iteration will actually DISPATCH, resolved
+    on the SAME route the runner uses for the active case, so the dashboard never
+    reports a model that differs from what the run runs:
+
+      - OVERRIDE active: the runner feeds the alias straight to --model via
+        loki_apply_max_tier_clamp(alias, alias). `effective` = _clamp_to_max_tier
+        (the override-path clamp). A "sonnet" override dispatches sonnet.
+      - NO override (session pin): the runner maps the pin through a tier
+        (opus->planning, sonnet->development, haiku->fast) and resolves the tier
+        through PROVIDER_MODEL_* (then the cost-ceiling clamp). `effective` =
+        _resolve_session_pin. A "sonnet" pin dispatches OPUS (development tier ->
+        PROVIDER_MODEL_DEVELOPMENT=opus on stock config).
+
+    Both routes resolve through the SAME provider config the runner uses
+    (LOKI_ALLOW_HAIKU plus the LOKI_CLAUDE_MODEL_PLANNING/FAST/DEVELOPMENT and
+    LOKI_MODEL_* overrides) and the SAME LOKI_MAX_TIER ceiling, mirroring
+    providers/claude.sh byte-for-byte. The agreement (estimator == dashboard ==
+    runner) on BOTH routes -- including the no-override stock path -- is locked by
+    the cross-route cases and the session-pin parity matrix in
+    tests/test-model-override.sh. (Before task 568 the no-override path applied the
+    override-path clamp to the pin, so a stock "sonnet" pin reported "sonnet" while
+    the run dispatched opus; that gap is now closed.)
 
     KNOWN LIMITATION (cross-process env divergence): the resolution reads
     LOKI_MAX_TIER, LOKI_ALLOW_HAIKU, LOKI_SESSION_MODEL and the model-override env
     vars from the DASHBOARD process's environment, which is usually a different
     process than the live run. So if the run was launched with a different
     environment than the dashboard, the no-override `default`/`effective` may not
-    reflect the run's real pinned tier or ceiling (e.g. a run pinned to opus still
-    reads "sonnet" here). The override case reads the run's own state file, so its
-    alias is always accurate and the clamp is exact whenever the dashboard shares
+    reflect the run's real pinned tier or ceiling (e.g. a run launched with
+    LOKI_SESSION_MODEL=opus while the dashboard's env has no pin still reads the
+    default here). The override case reads the run's own state file, so its alias
+    is always accurate and the resolution is exact whenever the dashboard shares
     the run's environment.
-
-    SCOPE NOTE (no-override default path): when there is no override, `effective`
-    applies the override-path clamp to the session default. The runner's
-    no-override route instead maps a session pin through a tier
-    (resolve_model_for_tier: opus->planning, sonnet->development), which can differ
-    from the override-path clamp in one cell (e.g. an opus pin under sonnet cap +
-    LOKI_ALLOW_HAIKU: the tier route yields sonnet, the override-path clamp yields
-    opus). That session-pin modeling gap is pre-existing and out of scope here;
-    the override case this endpoint serves is exact.
     """
     override = None
     try:
@@ -2343,8 +2435,16 @@ async def get_session_model():
             override = _normalize_session_model(p.read_text()) or None
     except OSError:
         override = None
-    default = _normalize_session_model(os.environ.get("LOKI_SESSION_MODEL")) or "sonnet"
-    effective = _clamp_to_max_tier(override or default)
+    # Session pin accepts tier names too (run.sh:12336), so use the broader
+    # session-pin normalizer here (NOT the narrow override allowlist).
+    default = _normalize_session_pin(os.environ.get("LOKI_SESSION_MODEL")) or "sonnet"
+    # Resolve on the route the runner will actually take: override-path clamp when
+    # an override file is present, session-pin tier route otherwise. This closes
+    # the task-568 stock-path gap (a "sonnet" pin dispatches opus).
+    if override is not None:
+        effective = _clamp_to_max_tier(override)
+    else:
+        effective = _resolve_session_pin(default)
     return {
         "override": override,
         "default": default,

package/docs/INSTALLATION.md

@@ -2,7 +2,7 @@
 
 The flagship product of [Autonomi](https://www.autonomi.dev/). Loki Mode is a spec-driven autonomous builder with a built-in trust layer that takes any spec to a deployed product and verifies completion with evidence (quality gates plus a completion council), not just a "done" claim. Complete installation instructions for all platforms and use cases.
 
-**Version:** v7.31.0
+**Version:** v7.32.1
 
 ---