loki-mode 9.23.1 → 9.25.2

package/README.md

@@ -15,8 +15,6 @@ _The free, source-available autonomous coding agent by [Autonomi](https://www.au
 
 [Website](https://www.autonomi.dev/) | [Documentation](wiki/Home.md) | [Installation](docs/INSTALLATION.md) | [Changelog](CHANGELOG.md)
 
-**Current release: v9.16.0**
-
 </div>
 
 ---
@@ -41,6 +39,21 @@ Upgrade with `loki self-update`. Long form: [Installation Guide](docs/INSTALLATI
 
 </details>
 
+<details>
+<summary>Claude Code plugin (adds /loki-grill, /loki-spec-status, /loki-verify)</summary>
+
+```bash
+claude plugin marketplace add asklokesh/loki-mode
+claude plugin install loki-mode@loki-mode
+```
+
+Adds three slash commands and the Loki MCP server (memory, task queue, code
+search, build management) to Claude Code. It calls the CLI rather than bundling
+it, so install `loki-mode` above first. Verify with `claude plugin list`:
+a healthy install reports `Status: enabled`.
+
+</details>
+
 ## Use it
 
 ```bash
@@ -466,6 +479,10 @@ OpenRouter serves **only** the OpenAI-shaped `/v1/chat/completions`; it has no
 Anthropic `/v1/messages` endpoint. Pointing `ANTHROPIC_BASE_URL` at it does not
 work, which earlier versions of this README incorrectly suggested.
 
+[OrcaRouter](https://www.orcarouter.ai) is an OpenAI-compatible gateway that
+also serves the Anthropic Messages API, so unlike OpenRouter the same key works
+through Route 2 below as well as Route 1 here.
+
 *Route 2 -- Anthropic-protocol gateways.* `ANTHROPIC_BASE_URL` routes Claude
 Code itself, so the endpoint must speak the Anthropic Messages API. LiteLLM,
 Bedrock proxies, and self-hosted gateways can:
@@ -481,6 +498,14 @@ export ANTHROPIC_BASE_URL=https://your-gateway.internal/v1
 export ANTHROPIC_API_KEY=...
 export LOKI_MODEL_OVERRIDE=<whatever your gateway calls the model>
 loki start prd.md
+
+# OrcaRouter (one key for both routes; model ids are namespaced by provider).
+# The bare host is deliberate here: the Anthropic SDK appends /v1/messages
+# itself, so adding /v1 would double it.
+export ANTHROPIC_BASE_URL=https://api.orcarouter.ai
+export ANTHROPIC_API_KEY=sk-orca-...
+export LOKI_MODEL_OVERRIDE=<namespaced id, e.g. anthropic/claude-sonnet-5>
+loki start prd.md
 ```
 
 **Set both variables.** `LOKI_MODEL_OVERRIDE` is what makes the alt-provider

package/SKILL.md

@@ -3,7 +3,7 @@ name: loki-mode
 description: Autonomous spec-driven build system with a built-in trust layer. It does not call work done until it is verified (RARV-C closure loop, 8 quality gates, completion council, verified-completion evidence gate). Triggers on "Loki Mode". Takes a spec (PRD, GitHub issue, OpenAPI doc, etc.) to deployed product with minimal human intervention. Provider-agnostic. Requires --dangerously-skip-permissions flag.
 ---
 
-# Loki Mode v9.23.1
+# Loki Mode v9.25.2
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -470,4 +470,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v9.23.1 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v9.25.2 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-9.23.1
+9.25.2

package/autonomy/loki

@@ -7657,9 +7657,17 @@ cmd_web() {
             cmd_web_redirect_to_dashboard "$@"
             ;;
         stop)
+            # `loki web start` redirects to the dashboard (above), so stopping
+            # Purple Lab here left the user unable to stop what they had just
+            # started: start hit 57374, stop hit 57375. Stop BOTH -- the
+            # dashboard because that is what start launched, and Purple Lab
+            # because a long-lived pre-deprecation process may still be up.
+            cmd_dashboard_stop
             cmd_web_stop
             ;;
         status)
+            # Same reasoning as stop: report on what start actually launched.
+            cmd_dashboard_status
             cmd_web_status
             ;;
         logs)
@@ -7704,6 +7712,7 @@ cmd_web_redirect_to_dashboard() {
     # --port is shared and passes straight through.
     local orig_args=("$@")
     local dash_args=()
+    local _web_dropped_prd=""
     while [[ $# -gt 0 ]]; do
         case "$1" in
             --no-open)
@@ -7711,10 +7720,17 @@ cmd_web_redirect_to_dashboard() {
                 shift
                 ;;
             --prd)
-                # Purple-Lab-only prefill; not supported by the dashboard.
+                # Purple-Lab-only prefill; the dashboard cannot accept it.
+                # Dropping it is correct, doing so SILENTLY is not: the user
+                # named a spec file and nothing would tell them it was ignored,
+                # so they would sit in front of an empty dashboard believing
+                # their PRD had been loaded. Say so, and name the command that
+                # actually takes a PRD.
+                _web_dropped_prd="${2:-}"
                 shift 2 2>/dev/null || shift
                 ;;
             --prd=*)
+                _web_dropped_prd="${1#--prd=}"
                 shift
                 ;;
             *)
@@ -7728,6 +7744,11 @@ cmd_web_redirect_to_dashboard() {
     if ! _deprecated_alias_should_suppress ${orig_args[@]+"${orig_args[@]}"}; then
         echo "Launching the dashboard (http://localhost:${DASHBOARD_DEFAULT_PORT}) instead of the deprecated Purple Lab." >&2
     fi
+    if [ -n "$_web_dropped_prd" ]; then
+        echo "Note: --prd '${_web_dropped_prd}' was NOT loaded. The dashboard does not take a PRD;" >&2
+        echo "      it watches runs that are already going. To build from that spec:" >&2
+        echo "        loki start '${_web_dropped_prd}'" >&2
+    fi
     cmd_dashboard_start ${dash_args[@]+"${dash_args[@]}"}
 }
 
@@ -31217,7 +31238,15 @@ _docs_invoke_provider() {
             # structurally identical captured `claude -p` resolution subcall).
             # env-prefix form because of the optional $t_prefix timeout wrapper.
             # No-op when caveman is absent.
-            result=$($t_prefix env CAVEMAN_DEFAULT_MODE=off claude -p "$prompt" 2>/dev/null) || exit_code=$?
+            # Model pin (v9.25.0). This site makes EIGHT sequential provider
+            # calls per doc-gen run, and without --model every one of them runs
+            # on whatever the account default is -- Opus for many users, at 5x
+            # Sonnet's input price and 5x its output price. Documentation is
+            # summarization over a context this code already assembled; it is
+            # not the workload that needs the top tier. Pinned to the same
+            # default the review subcall uses (run.sh:13762), and overridable
+            # for anyone who wants a different trade.
+            result=$($t_prefix env CAVEMAN_DEFAULT_MODE=off claude -p --model "${LOKI_DOCS_MODEL:-claude-sonnet-5}" "$prompt" 2>/dev/null) || exit_code=$?
             ;;
         codex)
             result=$($t_prefix codex exec --sandbox workspace-write "$prompt" 2>/dev/null) || exit_code=$?
@@ -31347,47 +31376,56 @@ _docs_generate() {
         echo -e "  ${CYAN}Generating${NC} $doc_file ..."
 
         if [ "$provider_available" = true ]; then
-            # Build a focused prompt for each doc type
+            # Build a focused prompt for each doc type.
+            #
+            # CONTEXT FIRST, AND THE ORDER IS LOAD-BEARING. All eight doc types
+            # share the same $context (measured at 4,269 tokens on this repo).
+            # With the per-type instruction first, the eight prompts share no
+            # common prefix, so every call re-sends the whole context at full
+            # input price: ~29,883 redundant input tokens per doc-gen run.
+            # Leading with the identical context makes it one cache write plus
+            # seven cache reads at 0.1x. Do not "tidy" the instruction back to
+            # the top -- that silently restores the redundant billing.
             case "$doc_type" in
                 README)
-                    prompt="Based on the following project context, generate a comprehensive README.md file. Include: project title, description, features, installation instructions, usage examples, and contribution guidelines. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a comprehensive README.md file. Include: project title, description, features, installation instructions, usage examples, and contribution guidelines. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 ARCHITECTURE)
-                    prompt="Based on the following project context, generate an ARCHITECTURE.md file. Include: system overview, high-level design, component diagram (as text), data flow, directory structure explanation, key design decisions, and technology choices. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate an ARCHITECTURE.md file. Include: system overview, high-level design, component diagram (as text), data flow, directory structure explanation, key design decisions, and technology choices. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 API)
-                    prompt="Based on the following project context, generate an API.md file. Document all public APIs: REST endpoints, exported functions, classes, and their signatures. Include parameter types, return types, and brief descriptions. If no API is detected, document the main public interfaces. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate an API.md file. Document all public APIs: REST endpoints, exported functions, classes, and their signatures. Include parameter types, return types, and brief descriptions. If no API is detected, document the main public interfaces. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 SETUP)
-                    prompt="Based on the following project context, generate a SETUP.md file. Include: prerequisites, installation steps, environment variables, database setup, running locally, running in Docker (if applicable), and common troubleshooting. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a SETUP.md file. Include: prerequisites, installation steps, environment variables, database setup, running locally, running in Docker (if applicable), and common troubleshooting. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 COMPONENTS)
-                    prompt="Based on the following project context, generate a COMPONENTS.md file. Document each major component/module/directory: its purpose, key files, public interface, and dependencies on other components. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a COMPONENTS.md file. Document each major component/module/directory: its purpose, key files, public interface, and dependencies on other components. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 TESTING)
-                    prompt="Based on the following project context, generate a TESTING.md file. Include: test strategy, test types (unit, integration, e2e), how to run tests, test configuration, coverage goals, and CI integration. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a TESTING.md file. Include: test strategy, test types (unit, integration, e2e), how to run tests, test configuration, coverage goals, and CI integration. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 DECISIONS)
-                    prompt="Based on the following project context, generate a DECISIONS.md file with Architectural Decision Records (ADRs). Infer key decisions from the technology stack, directory structure, and configuration. Each ADR should have: title, status, context, decision, and consequences. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a DECISIONS.md file with Architectural Decision Records (ADRs). Infer key decisions from the technology stack, directory structure, and configuration. Each ADR should have: title, status, context, decision, and consequences. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
                 CLAUDE)
-                    prompt="Based on the following project context, generate a CLAUDE.md file for AI agent context. Include: project overview, key commands (build, test, run, lint), project structure with brief descriptions of each directory, coding conventions, important files, and any gotchas. This file helps AI coding assistants understand the project quickly. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations.
+                    prompt="$context
 
-$context"
+Based on the project context above, generate a CLAUDE.md file for AI agent context. Include: project overview, key commands (build, test, run, lint), project structure with brief descriptions of each directory, coding conventions, important files, and any gotchas. This file helps AI coding assistants understand the project quickly. Use markdown formatting. Do NOT use emojis. Output ONLY the markdown content, no explanations."
                     ;;
             esac
 

package/autonomy/quickstart.sh

@@ -1346,7 +1346,10 @@ cmd_quickstart() {
     printf '\n'
     printf 'Starting your build. Progress streams here in the terminal.\n'
     printf '  PRD saved to: %s\n' "$target"
-    printf "  Tip: run 'loki dashboard' in another terminal to watch in a browser.\n"
+    # `loki dashboard` with no subcommand prints help and exits -- it does not
+    # start. Naming the bare command here sent every first-run user to a help
+    # screen at the exact moment they wanted to watch their build.
+    printf "  Tip: run 'loki dashboard start' in another terminal to watch in a browser.\n"
     printf '\n'
 
     # ----- Compose with cmd_start -------------------------------------------

package/autonomy/run.sh

@@ -4245,6 +4245,27 @@ build_completion_summary() {
         stopped)        outcome_label="Stopped";          notify_title="Run stopped" ;;
         failed)         outcome_label="Failed";           notify_title="Run failed" ;;
         intervention)   outcome_label="Needs input";      notify_title="Input needed" ;;
+        # Every outcome below reached this case and fell through to the `*` arm,
+        # so the user's headline label was the raw enum string --
+        # "council_force_approved", "max_duration" -- at the exact moment they
+        # were deciding whether to trust the build. The guidance block further
+        # down already handles several of these properly; only the label was
+        # missing.
+        force_stopped)  outcome_label="Stopped without approval"
+                        notify_title="Run stopped (not approved)" ;;
+        budget_exceeded) outcome_label="Stopped at spend cap"
+                        notify_title="Run stopped (budget cap)" ;;
+        max_duration)   outcome_label="Time limit reached"
+                        notify_title="Run stopped (time limit)" ;;
+        max_retries_exceeded) outcome_label="Retries exhausted"
+                        notify_title="Run stopped (retries exhausted)" ;;
+        inconclusive_spec_contradiction) outcome_label="Spec contradiction"
+                        notify_title="Run stopped (spec contradiction)" ;;
+        # Force-approval is NOT the same as council approval, and labelling both
+        # "Completed" hid the difference on a product whose whole claim is a
+        # checkable receipt. Name it.
+        council_force_approved) outcome_label="Completed (force-approved)"
+                        notify_title="Run complete (force-approved)" ;;
         *)              outcome_label="$outcome";          notify_title="Run finished" ;;
     esac
 
@@ -12547,8 +12568,31 @@ auto_generate_docs_if_needed() {
             -not -path '*/node_modules/*' -not -path '*/.loki/*' \
             -not -path '*/.git/*' -not -path '*/dist/*' 2>/dev/null | head -40 | wc -l | tr -d ' ')
         _doc_src="${_doc_src:-0}"
-        # <=3 source files cannot need an architecture suite. 90s still allows a
-        # README + USAGE pass, which is all the gate asks of a small project.
+        # <=3 source files cannot need an architecture suite. This used to cap
+        # the timeout at 90s and still run; measurement showed that on a small
+        # project the run reaches the cap and is KILLED (exit 124 below), so the
+        # 90s bought nothing -- the gate then scored on whatever files already
+        # existed, exactly as it does when generation is skipped. On the one
+        # profiled build doc_generation was 90s of a 960s wall clock, 9%,
+        # producing no document (benchmarks/results/gate-profile.json).
+        #
+        # So skip outright rather than pay for a timeout. This is not a quality
+        # trade: the outcome for the gate is identical, and the 90s is returned
+        # to the user. LOKI_DOCS_TIMEOUT is still honored -- the whole block is
+        # inside `if [ -z "${LOKI_DOCS_TIMEOUT:-}" ]`, so anyone who explicitly
+        # asks for doc generation on a tiny project still gets it.
+        # Gate the skip on the SIMPLE tier, not on the file count alone.
+        # A standard/complex project can legitimately have few source files and
+        # still need its full doc suite -- tests/test-doc-scope-generator.sh
+        # exists precisely to assert that "quality at any complexity is
+        # preserved", and a count-only skip broke it. The simple tier already
+        # returns early further up for the same reason, so this only shortens a
+        # doomed run for projects that were never getting the full suite.
+        if [ "$_doc_src" -le 3 ] && [ "${DETECTED_COMPLEXITY:-}" = "simple" ]; then
+            log_info "Auto-documentation: ${_doc_src} source file(s) on the simple tier -- skipping generation (it times out before producing a document; set LOKI_DOCS_TIMEOUT to force it)"
+            return 0
+        fi
+        # Everything else keeps the previous behavior: cap the timeout, still run.
         if [ "$_doc_src" -le 3 ] && [ "$_doc_to" -gt 90 ]; then
             log_info "Auto-documentation: ${_doc_src} source file(s) -- capping generation at 90s (was ${_doc_to}s)"
             _doc_to=90
@@ -16802,7 +16846,32 @@ start_dashboard() {
         return 1
     fi
 
-    sleep 2
+    # Wait for the dashboard to come up, but only as long as it actually takes.
+    # This was a flat `sleep 2` on the critical path of every build, before the
+    # first iteration, spent entirely on a process that is typically serving in
+    # a fraction of that. Poll the endpoint the reuse path above already trusts
+    # (/api/status), so this returns the moment the server really serves rather
+    # than on a fixed guess.
+    #
+    # The floor is deliberate and NOT an optimization target. A process that
+    # starts and then dies at t=1.5s would pass an early `kill -0` where the old
+    # flat sleep would have caught it, so polling alone would trade a real
+    # liveness check for a second of wall clock. We keep polling until the
+    # endpoint answers AND require the process to still be alive at the end,
+    # which is strictly stronger than the old single check at t=2s.
+    _dash_ready=0
+    for _ in $(seq 1 40); do
+        kill -0 "$DASHBOARD_PID" 2>/dev/null || break
+        if curl -fsS -m 1 "http://127.0.0.1:${DASHBOARD_PORT}/api/status" >/dev/null 2>&1; then
+            _dash_ready=1
+            break
+        fi
+        sleep 0.05
+    done
+    # A server that never answered still gets the original grace period: some
+    # environments have no curl, and the endpoint is not the only thing that
+    # makes a dashboard useful. Falling back keeps behavior identical there.
+    [ "$_dash_ready" = "1" ] || sleep 2
 
     if kill -0 "$DASHBOARD_PID" 2>/dev/null; then
         DASHBOARD_LAST_ALIVE=$(date +%s)
@@ -26073,6 +26142,18 @@ main() {
         exit 1
     fi
 
+    # BOOT WINDOW (v9.24.0). Everything from here to setup_agent_branch is
+    # pre-loop setup: prerequisites, provider detection, complexity detection,
+    # dashboard start, branch setup. None of it was timed, and none of the nine
+    # existing emit_stage_complete sites lies outside the iteration body -- so
+    # on the one profiled build, stage_total_s summed to 723s against a 960s
+    # wall clock and 237s (25%) was attributed by GUESS, not measurement
+    # (benchmarks/results/gate-profile.json). The guess named "rsync of the
+    # engine copy", which measurement later falsified: the only rsync in the
+    # repo is in the benchmark harness, costs 1.09s, and runs before the timer
+    # starts. Bracket the window instead of arguing about it.
+    _boot_t0=$(date +%s 2>/dev/null)
+
     # Check prerequisites (unless skipped)
     if [ "$SKIP_PREREQS" != "true" ]; then
         if ! check_prerequisites; then
@@ -26288,6 +26369,11 @@ main() {
     # Setup agent branch protection (isolates agent changes to a feature branch)
     setup_agent_branch
 
+    # Close the boot window. Purely additive: emit_stage_complete appends one
+    # event line and swallows every error, so it cannot alter a gate verdict or
+    # control flow (see its contract at run.sh:2523-2531).
+    emit_stage_complete "boot" "pass" "$_boot_t0"
+
     # Log session start for audit
     audit_log "SESSION_START" "prd=$PRD_PATH,dashboard=$ENABLE_DASHBOARD,staged_autonomy=$STAGED_AUTONOMY,parallel=$PARALLEL_MODE"
     audit_agent_action "session_start" "Session started" "prd=$PRD_PATH,provider=${PROVIDER_NAME:-claude}"
@@ -26384,6 +26470,11 @@ main() {
         # a stuck "Planning" state.
         _advance_current_phase "BUILDING"
         run_autonomous "$PRD_PATH" || result=$?
+        # TEARDOWN WINDOW (v9.24.0): the other half of the unmeasured 237s.
+        # Everything after the loop -- pre-edit snapshot, commit, handoff and
+        # learnings writers, proof generation, metrics aggregation -- runs here
+        # and was never timed. Opened immediately so nothing below is missed.
+        _teardown_t0=$(date +%s 2>/dev/null)
         # PRE-EDIT SNAPSHOT: freeze the agent's raw diff HERE, the first
         # instruction after the loop returns, because everything below this line
         # can change the tree -- commit_session_changes commits the work (after
@@ -26578,6 +26669,14 @@ except Exception:
         generate_proof_of_run "$result" || true
     fi
 
+    # Close the teardown window here rather than after cleanup: everything below
+    # is process reaping and file removal, while everything above is the work a
+    # user waits on (commit, PR, summary, proof). Emitting before cleanup also
+    # guarantees the event is written even if a later reap kills this shell.
+    if [ -n "${_teardown_t0:-}" ]; then
+        emit_stage_complete "teardown" "pass" "$_teardown_t0"
+    fi
+
     # Cleanup
     if type app_runner_cleanup &>/dev/null; then
         app_runner_cleanup

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "9.23.1"
+__version__ = "9.25.2"
 
 # Expose the control app for easy import
 try:

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:** v9.23.1
+**Version:** v9.25.2
 
 ---