loki-mode 8.0.3 → 8.1.0

package/README.md

@@ -239,18 +239,31 @@ This needs Bun on your PATH (the SDK loop runs on the Bun runtime). `loki doctor
 
 **With a coding-agent CLI.** The classic path, and still the default: Loki drives a separate CLI (Claude Code is the recommended one) plus a couple of common tools on your PATH.
 
-**With a different model or provider.** Loki is not tied to Anthropic. Anything
-that speaks the Anthropic Messages API works: OpenRouter, Ollama, LiteLLM, vLLM,
-or your own gateway. Point `ANTHROPIC_BASE_URL` at the endpoint and name the
-model with `LOKI_MODEL_OVERRIDE`:
+**With a different model or provider.** Loki is not tied to Anthropic, but
+*how* you reach another model depends on which API the endpoint speaks. There
+are two routes, and picking the wrong one fails confusingly.
+
+*Route 1 -- OpenAI-shaped endpoints (OpenRouter, and most hosted open models).*
+Use a provider that speaks that API natively. `aider` and `cline` both do, and
+Loki now defaults them to open-weight models rather than Claude:
 
 ```bash
-# OpenRouter (hundreds of models, one key)
-export ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1
-export ANTHROPIC_API_KEY=sk-or-...
-export LOKI_MODEL_OVERRIDE=<model-id-from-openrouter.ai/models>
-loki start prd.md
+loki provider set aider
+export OPENROUTER_API_KEY=sk-or-...
+loki start prd.md                          # defaults to deepseek-v3.2
+
+export LOKI_AIDER_MODEL=openrouter/z-ai/glm-4.6   # or pick your own
+```
 
+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.
+
+*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:
+
+```bash
 # Ollama, fully local (no API key, no per-token cost)
 export ANTHROPIC_BASE_URL=http://localhost:11434/v1
 export LOKI_MODEL_OVERRIDE=<model you have pulled, e.g. the output of `ollama list`>

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 v8.0.3
+# Loki Mode v8.1.0
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -469,4 +469,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v8.0.3 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v8.1.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-8.0.3
+8.1.0

package/autonomy/council-v2.sh

@@ -82,6 +82,9 @@ council_v2_vote() {
     local first=true
     local approve_count=0
     local reject_count=0
+    # Reviewers we could not obtain a verdict from. Never folded into
+    # reject_count -- see the TRUST comment in the tally below.
+    local inconclusive_count=0
     for vote_file in "${vote_files[@]}"; do
         if [ -f "$vote_file" ]; then
             local vote_content
@@ -95,16 +98,40 @@ council_v2_vote() {
 
             local verdict
             verdict=$(echo "$vote_content" | python3 -c "import sys,json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "UNKNOWN")
+            # TRUST: a reviewer that could not be REACHED did not vote REJECT.
+            #
+            # Every provider arm below used to substitute a literal
+            # {"verdict":"REJECT","reasoning":"review execution failed"} on any
+            # miss -- CLI absent, timeout, transient error, unparseable output --
+            # and this tally counted "anything not APPROVE" as a rejection. The
+            # engine therefore fabricated a reviewer vote the model never gave,
+            # and the run was BLOCKED by it.
+            #
+            # That is the same defect class as a receipt attesting to a diff stat
+            # it did not measure: the artifact claims a fact nobody established.
+            # On a weak model that formats poorly it is worse still -- low format
+            # compliance becomes a permanent BLOCK that reads to the user as
+            # "Loki is broken" rather than "your model could not answer".
+            #
+            # INCONCLUSIVE is counted separately and never as a rejection.
             if [ "$verdict" = "APPROVE" ]; then
                 ((approve_count++))
-            else
+            elif [ "$verdict" = "REJECT" ]; then
                 ((reject_count++))
+            else
+                ((inconclusive_count++))
             fi
         fi
     done
     votes_json="$votes_json]"
 
-    log_info "Blind review results: $approve_count APPROVE / $reject_count REJECT"
+    if [ "$inconclusive_count" -gt 0 ]; then
+        log_warn "Blind review results: $approve_count APPROVE / $reject_count REJECT / $inconclusive_count INCONCLUSIVE"
+        log_warn "  $inconclusive_count reviewer(s) produced no verdict (CLI missing, timeout, or unparseable output)."
+        log_warn "  These are NOT rejections. The council decides on the verdicts it actually obtained."
+    else
+        log_info "Blind review results: $approve_count APPROVE / $reject_count REJECT"
+    fi
 
     # Step 4: Sycophancy detection
     local sycophancy_score
@@ -137,8 +164,13 @@ print('{:.3f}'.format(detect_sycophancy(votes)))
 
             if [ -f "$da_vote" ]; then
                 local da_verdict
-                da_verdict=$(cat "$da_vote" | python3 -c "import sys,json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "REJECT")
-                if [ "$da_verdict" = "REJECT" ]; then
+                # Same trust rule as the main tally: an unparseable devil's
+                # advocate did not vote REJECT. Defaulting to REJECT here would
+                # silently overturn a unanimous APPROVE on a transient failure.
+                da_verdict=$(cat "$da_vote" | python3 -c "import sys,json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "INCONCLUSIVE")
+                if [ "$da_verdict" = "INCONCLUSIVE" ]; then
+                    log_warn "Devil's advocate produced no verdict -- unanimous approval left UNCHANGED (not overturned)"
+                elif [ "$da_verdict" = "REJECT" ]; then
                     log_warn "Devil's advocate REJECTED unanimous approval"
                     approve_count=$((approve_count - 1))
                     reject_count=$((reject_count + 1))
@@ -374,42 +406,42 @@ except Exception:
                 fi
                 # Fall through to the plain text call (+ sed-carve) on any miss.
                 if [ -z "$result" ]; then
-                    result=$(echo "$full_prompt" | CAVEMAN_DEFAULT_MODE=off claude "${_c2_argv[@]}" -p 2>/dev/null || echo '{"verdict":"REJECT","reasoning":"review execution failed","issues":[]}')
+                    result=$(echo "$full_prompt" | CAVEMAN_DEFAULT_MODE=off claude "${_c2_argv[@]}" -p 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
                 fi
             else
-                result='{"verdict":"REJECT","reasoning":"reviewer CLI unavailable","issues":[]}'
+                result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
             fi
             ;;
         codex)
             if command -v codex &>/dev/null; then
-                result=$(codex exec -q "$full_prompt" 2>/dev/null || echo '{"verdict":"REJECT","reasoning":"review execution failed","issues":[]}')
+                result=$(codex exec -q "$full_prompt" 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
             else
-                result='{"verdict":"REJECT","reasoning":"reviewer CLI unavailable","issues":[]}'
+                result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
             fi
             ;;
         gemini)
             if command -v gemini &>/dev/null; then
-                result=$(echo "$full_prompt" | gemini 2>/dev/null || echo '{"verdict":"REJECT","reasoning":"review execution failed","issues":[]}')
+                result=$(echo "$full_prompt" | gemini 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
             else
-                result='{"verdict":"REJECT","reasoning":"reviewer CLI unavailable","issues":[]}'
+                result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
             fi
             ;;
         cline)
             if command -v cline &>/dev/null; then
-                result=$(cline -y "$full_prompt" 2>/dev/null || echo '{"verdict":"REJECT","reasoning":"review execution failed","issues":[]}')
+                result=$(cline -y "$full_prompt" 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
             else
-                result='{"verdict":"REJECT","reasoning":"reviewer CLI unavailable","issues":[]}'
+                result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
             fi
             ;;
         aider)
             if command -v aider &>/dev/null; then
-                result=$(aider --message "$full_prompt" --yes-always --no-auto-commits --no-git 2>/dev/null || echo '{"verdict":"REJECT","reasoning":"review execution failed","issues":[]}')
+                result=$(aider --message "$full_prompt" --yes-always --no-auto-commits --no-git 2>/dev/null || echo '{"verdict":"INCONCLUSIVE","reasoning":"review execution failed (no verdict obtained; NOT a rejection)","issues":[]}')
             else
-                result='{"verdict":"REJECT","reasoning":"reviewer CLI unavailable","issues":[]}'
+                result='{"verdict":"INCONCLUSIVE","reasoning":"reviewer CLI unavailable (no verdict obtained; NOT a rejection)","issues":[]}'
             fi
             ;;
         *)
-            result='{"verdict":"REJECT","reasoning":"review not supported for this provider","issues":[]}'
+            result='{"verdict":"INCONCLUSIVE","reasoning":"review not supported for this provider (no verdict obtained; NOT a rejection)","issues":[]}'
             ;;
     esac
 
@@ -421,7 +453,11 @@ except Exception:
         extracted=$(echo "$result" | sed 's/^```json//;s/^```//' | sed -n '/^{/,/^}/p' | head -50)
     fi
     if [ -z "$extracted" ]; then
-        extracted='{"verdict":"REJECT","reasoning":"failed to parse review output","issues":[]}'
+        # The weak-model case, and the reason this matters most: a model whose
+        # prose could not be carved into JSON did not vote REJECT. Recording one
+        # here would turn "your model formats poorly" into "your code was
+        # rejected" -- a block the user cannot act on and did not earn.
+        extracted='{"verdict":"INCONCLUSIVE","reasoning":"failed to parse review output (no verdict obtained; NOT a rejection)","issues":[]}'
     fi
 
     echo "$extracted" > "$output_file"

package/autonomy/lib/prd-enrich.sh

@@ -224,6 +224,11 @@ loki_prd_enrich() {
     # improve tasks deterministically (content-derived user_story) so even
     # offline users get informative tasks, then return without a model call.
     if ! _loki_prd_enrich_provider_ok; then
+        # HONESTY: say so. This used to degrade to the deterministic path with no
+        # log line at all, so a user on a non-Claude provider believed their PRD
+        # had been model-enriched when it had not. A capability the run did not
+        # get must never be indistinguishable from one it did.
+        log_warn "PRD enrichment: model-assisted pass SKIPPED (provider='${LOKI_PROVIDER:-claude}'), using the deterministic pass only."
         _loki_prd_enrich_deterministic "$pending_path"
         return 0
     fi

package/autonomy/lib/proof-generator.py

@@ -429,6 +429,19 @@ def _collect_termination(loki_dir, session_exit_code=None):
         "exit_code": None,
         "outcome": "",
         "run_status": "",
+        # Which gate stopped the run, if one did. The receipt previously named
+        # only a bare outcome ("intervention"), so a user reading the artifact
+        # could not tell WHICH gate blocked them or how close it came to its
+        # threshold -- the single most actionable fact about a blocked run. The
+        # engine already writes it to .loki/signals/GATE_ESCALATION.json and
+        # already surfaces it in COMPLETION.txt and PAUSED.md; the signed
+        # receipt was the one surface that stayed silent.
+        #
+        # These are deterministic FACTS read from a file the engine wrote, not
+        # an AI assessment, so they belong in the facts block.
+        "blocking_gate": "",
+        "blocking_gate_failures": None,
+        "blocking_gate_threshold": None,
     }
     state_paths = [os.path.join(loki_dir, "autonomy-state.json")]
     sessions_dir = os.path.join(loki_dir, "sessions")
@@ -455,6 +468,20 @@ def _collect_termination(loki_dir, session_exit_code=None):
             out["status"] = outcome
             if outcome not in ("complete", "completed", "success"):
                 out["reason"] = outcome
+    # Gate escalation: name the gate that stopped the run. Best-effort and
+    # non-fatal -- a missing or corrupt signal simply leaves the fields empty,
+    # which reads as "no gate escalation recorded", never as a false claim.
+    gate = _read_json(
+        os.path.join(loki_dir, "signals", "GATE_ESCALATION.json"), default=None
+    )
+    if isinstance(gate, dict):
+        gate_name = str(gate.get("gate") or "").strip()
+        if gate_name:
+            out["blocking_gate"] = gate_name
+            count = gate.get("count")
+            thr = gate.get("threshold")
+            out["blocking_gate_failures"] = _to_int(count, None)
+            out["blocking_gate_threshold"] = _to_int(thr, None)
     if session_exit_code is not None:
         out["exit_code"] = session_exit_code
         if session_exit_code != 0 and not out["reason"]:

package/autonomy/lib/proof-template.html

@@ -640,6 +640,23 @@
                 ? " (" + esc(tOutcome) + ")" : "") +
              "</div>";
     }
+    // Name the GATE that stopped the run, with how close it came to its
+    // threshold. "intervention" tells a reader that something blocked them;
+    // "code_review (failed 3 times, threshold 3)" tells them what to fix. The
+    // engine already surfaced this in COMPLETION.txt and PAUSED.md -- the
+    // signed receipt was the only surface that stayed silent about it.
+    // Display-only, and silent when no gate escalation was recorded.
+    var tGate = String(g(p, "facts.execution.blocking_gate", "") || "").trim();
+    if (tGate) {
+      var gFail = g(p, "facts.execution.blocking_gate_failures", null);
+      var gThr = g(p, "facts.execution.blocking_gate_threshold", null);
+      var detail = "";
+      if (typeof gFail === "number" && typeof gThr === "number") {
+        detail = " (failed " + esc(String(gFail)) + " times, threshold " +
+                 esc(String(gThr)) + ")";
+      }
+      term += '<div class="h-sub">Blocked by: ' + esc(tGate) + detail + "</div>";
+    }
     el.className = "honesty " + v.cls;
     el.innerHTML =
       '<span class="h-mark" aria-hidden="true">' + v.mark + "</span>" +

package/autonomy/loki

@@ -954,10 +954,20 @@ show_help() {
     echo "  # Providers + model routing"
     echo "  loki provider list            # Show 4 providers (claude/codex/cline/aider)"
     echo "  loki provider set codex       # Switch active provider"
-    echo "  # OpenRouter / Ollama routing (Phase I v7.5.25+):"
-    echo "  export ANTHROPIC_BASE_URL=https://openrouter.ai/api/v1 \\\\"
-    echo "         LOKI_MODEL_OVERRIDE=anthropic/claude-sonnet-4.5"
-    echo "  loki start ./prd.md           # Routes to OpenRouter via Claude Code"
+    echo "  # Cheap / open models (DeepSeek, GLM, MiniMax, Kimi) via OpenRouter:"
+    echo "  loki provider set aider       # aider + cline reach any OpenRouter model"
+    echo "  export OPENROUTER_API_KEY=sk-or-..."
+    echo "  loki start ./prd.md           # defaults to deepseek-v3.2 (open weights)"
+    echo "  # Pick a specific model:"
+    echo "  export LOKI_AIDER_MODEL=openrouter/z-ai/glm-4.6"
+    echo ""
+    echo "  # Anthropic-PROTOCOL gateways (LiteLLM, Bedrock proxies, self-hosted):"
+    echo "  export ANTHROPIC_BASE_URL=https://your-gateway/v1 \\\\"
+    echo "         LOKI_MODEL_OVERRIDE=<model-id-your-gateway-serves>"
+    echo "  # NOTE: this routes Claude Code itself, so the endpoint must speak the"
+    echo "  #       Anthropic /v1/messages API. OpenRouter does NOT -- it serves"
+    echo "  #       only the OpenAI-shaped /v1/chat/completions. Use the aider or"
+    echo "  #       cline route above for OpenRouter."
     echo ""
     echo "  # Cross-project context (Phase F v7.5.23+)"
     echo "  # Drop .loki/app.json with {schema_version:1,app_id:myapp,members:[ui,api]}"
@@ -3871,7 +3881,10 @@ cmd_why() {
     fi
 
     if [ "$as_json" = "1" ]; then
+        local _why_json_head_sha
+        _why_json_head_sha="$(git rev-parse HEAD 2>/dev/null || echo "")"
         _LOKI_WHY_STATE="$state_file" _LOKI_WHY_COMPLETION="$completion_file" \
+        _LOKI_WHY_HEAD_SHA="$_why_json_head_sha" \
         _LOKI_WHY_LAST_ERROR="$loki_dir/state/LAST_ERROR.json" python3 - <<'WHYJSON'
 import json, os
 def load(p):
@@ -3881,7 +3894,26 @@ def load(p):
 state = load(os.environ.get("_LOKI_WHY_STATE", ""))
 comp = load(os.environ.get("_LOKI_WHY_COMPLETION", ""))
 last_error = load(os.environ.get("_LOKI_WHY_LAST_ERROR", ""))
-print(json.dumps({"state": state, "completion": comp, "last_error": last_error}, indent=2))
+
+# The human report labels a stale completion record "(from previous completed
+# run)". The JSON consumer got the same record with NO such marker, so a script
+# reading `completion.branch` or `completion.pr_url` after a crashed run would
+# attribute a PREVIOUS run's branch and PR to the current one -- silently, and
+# with no field to check. Emit the identical determination as explicit fields.
+# Kept byte-identical in logic to the human path below; the two must not drift.
+live_statuses = {"running", "exited"}
+head_sha = os.environ.get("_LOKI_WHY_HEAD_SHA", "")
+comp_head = comp.get("head_sha", "")
+comp_is_stale = bool(state.get("status") in live_statuses) or bool(
+    head_sha and comp_head and comp_head != head_sha
+)
+print(json.dumps({
+    "state": state,
+    "completion": comp,
+    "last_error": last_error,
+    "completion_is_stale": comp_is_stale,
+    "head_sha": head_sha or None,
+}, indent=2))
 WHYJSON
         return 0
     fi

package/autonomy/run.sh

@@ -22823,6 +22823,8 @@ check_human_intervention() {
                 log_warn "PAUSE file created by budget limit - NOT auto-clearing in perpetual mode"
                 log_warn "Budget limit reached. Remove .loki/signals/BUDGET_EXCEEDED and .loki/PAUSE to continue."
                 notify_intervention_needed "Budget limit reached - execution paused" 2>/dev/null || true
+                # Same-instant snapshot as COMPLETION.txt (see the PAUSE-file site).
+                (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
                 # Genuinely blocking pause: write the durable intervention record
                 # now (state-only; the ping above already fired). This is the
                 # correct site for the durable file because the run actually halts
@@ -22835,6 +22837,7 @@ check_human_intervention() {
                 if [ "$pause_result" -eq 1 ]; then
                     # STOP requested DURING the pause: relabel the durable record
                     # as stopped (state-only; the user typed STOP and is aware).
+                    (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
                     build_completion_summary stopped 2>/dev/null || true
                     return 2
                 fi
@@ -22849,6 +22852,13 @@ check_human_intervention() {
         fi
         log_warn "PAUSE file detected - pausing execution"
         notify_intervention_needed "Execution paused via PAUSE file"
+        # Refresh STATUS.txt from the queue BEFORE writing the durable record, so
+        # the two files snapshot the same instant. The status monitor refreshes
+        # every 2s but is not running by the time we block here, so STATUS.txt
+        # was frozen at its last tick: measured on a real paused run reporting
+        # "Failed: 0" while COMPLETION.txt -- written now, from the same
+        # queue/failed.json -- reported failed=1. Same source, different age.
+        (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
         # Genuinely blocking pause: write the durable intervention record now
         # (state-only; the ping above already fired).
         build_completion_summary intervention 2>/dev/null || true
@@ -22859,6 +22869,7 @@ check_human_intervention() {
         if [ "$pause_result" -eq 1 ]; then
             # STOP was requested during pause: relabel the durable record as
             # stopped (state-only; the user typed STOP and is aware).
+            (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
             build_completion_summary stopped 2>/dev/null || true
             return 2
         fi
@@ -22872,6 +22883,8 @@ check_human_intervention() {
             rm -f "$loki_dir/PAUSE_AT_CHECKPOINT"
             notify_intervention_needed "Execution paused at checkpoint"
             touch "$loki_dir/PAUSE"
+            # Same-instant snapshot as COMPLETION.txt (see the PAUSE-file site).
+            (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
             # Genuinely blocking pause: write the durable intervention record now
             # (state-only; the ping above already fired).
             build_completion_summary intervention 2>/dev/null || true
@@ -22881,6 +22894,7 @@ check_human_intervention() {
             rm -f "$loki_dir/PAUSE"
             if [ "$pause_result" -eq 1 ]; then
                 # STOP requested during pause: relabel as stopped (state-only).
+                (cd "${TARGET_DIR:-.}" && update_status_file) 2>/dev/null || true
                 build_completion_summary stopped 2>/dev/null || true
                 return 2
             fi

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "8.0.3"
+__version__ = "8.1.0"
 
 # Expose the control app for easy import
 try:

package/events/emit.sh

@@ -44,36 +44,63 @@ safe_append_event_jsonl() {
     if command -v flock >/dev/null 2>&1; then
         # flock path: bind FD 9 to the sentinel file (created if absent),
         # take an exclusive lock, append, release on subshell exit.
+        #
+        # `-w 5` is LOAD-BEARING: a bare `flock -x 9` blocks FOREVER. A holder
+        # that is killed mid-write (a reaped CI run, a Ctrl-C'd build) leaves
+        # every subsequent emit hung, and because emit.sh is spawned per CLI
+        # invocation those hangs ACCUMULATE. Measured 2026-07-29: 59 orphaned
+        # emit.sh processes, the oldest alive 21 HOURS, on a machine at 0% idle.
+        # Observability must never outlive the thing it observes.
         (
-            flock -x 9
+            flock -w 5 -x 9 || exit 1
             printf '%s\n' "$line" >> "$events_path"
         ) 9>"$lock_target"
-        return $?
+        local frc=$?
+        if [ "$frc" -ne 0 ]; then
+            # Lock unavailable within the timeout: append unlocked rather than
+            # block. A rare interleaved line is strictly better than a hung CLI.
+            printf '%s\n' "$line" >> "$events_path" 2>/dev/null || true
+        fi
+        return 0
     fi
 
-    # Fallback: mkdir-based mutex. mkdir is atomic on POSIX.
+    # Fallback: mkdir-based mutex. mkdir is atomic on POSIX. macOS ships NO
+    # flock(1), so this is the path real users take -- it must be the robust one.
     local lock_dir="${events_path}.lockdir"
     local attempts=0
-    local max_attempts=500   # ~5s at 10ms sleep
+    local max_attempts=100   # ~1s at 10ms sleep
+    local stale_after=30
     while ! mkdir "$lock_dir" 2>/dev/null; do
+        # Check staleness EVERY iteration, not only after exhausting attempts.
+        # The old code waited for all 500 attempts before its first staleness
+        # check, so a stale lock cost ~5s AND ~500 forked sleep helpers per
+        # event; under concurrent emits the locks kept going stale and the
+        # processes piled up. Measured before this fix: 10s and ~500 forks for
+        # a SINGLE append against an already-stale lock.
+        local age=0
+        local mtime
+        mtime=$(stat -f%m "$lock_dir" 2>/dev/null || stat -c%Y "$lock_dir" 2>/dev/null || echo "")
+        if [ -n "$mtime" ]; then
+            age=$(( $(date +%s) - mtime ))
+        else
+            # Lock vanished between the failed mkdir and the stat: retry at once.
+            continue
+        fi
+        if [ "$age" -gt "$stale_after" ]; then
+            rmdir "$lock_dir" 2>/dev/null || rm -rf "$lock_dir" 2>/dev/null || true
+            continue
+        fi
         attempts=$((attempts + 1))
         if [ "$attempts" -ge "$max_attempts" ]; then
-            # Stale lock: if the dir is older than 30s, force-remove it.
-            local age
-            age=$(( $(date +%s) - $(stat -f%m "$lock_dir" 2>/dev/null \
-                                    || stat -c%Y "$lock_dir" 2>/dev/null \
-                                    || echo 0) ))
-            if [ "$age" -gt 30 ]; then
-                rmdir "$lock_dir" 2>/dev/null || rm -rf "$lock_dir" 2>/dev/null || true
-                attempts=0
-                continue
-            fi
-            # Give up -- best-effort write so observability never blocks.
+            # Give up fast -- best-effort write so observability never blocks.
             printf '%s\n' "$line" >> "$events_path" 2>/dev/null || true
-            return 1
+            return 0
         fi
-        # Sleep ~10ms (perl avoids `sleep 0.01` portability issues).
-        perl -e 'select(undef,undef,undef,0.01)' 2>/dev/null || sleep 1
+        # Sleep ~10ms WITHOUT forking when the shell supports fractional sleep
+        # (bash's `read -t` needs no external process). perl/sleep are fallbacks.
+        read -r -t 0.01 _unused_ < /dev/null 2>/dev/null \
+            || perl -e 'select(undef,undef,undef,0.01)' 2>/dev/null \
+            || sleep 1
     done
     # Critical section.
     printf '%s\n' "$line" >> "$events_path"

package/loki-ts/dist/loki.js

@@ -1,5 +1,5 @@
 // @bun
-var k_=Object.create;var{getPrototypeOf:x_,defineProperty:oK,getOwnPropertyNames:S_}=Object;var y_=Object.prototype.hasOwnProperty;function b_(Z){return this[Z]}var __,f_,h_=(Z,X,Q)=>{var Y=Z!=null&&typeof Z==="object";if(Y){var J=X?__??=new WeakMap:f_??=new WeakMap,z=J.get(Z);if(z)return z}Q=Z!=null?k_(x_(Z)):{};let K=X||!Z||!Z.__esModule?oK(Q,"default",{value:Z,enumerable:!0}):Q;for(let $ of S_(Z))if(!y_.call(K,$))oK(K,$,{get:b_.bind(Z,$),enumerable:!0});if(Y)J.set(Z,K);return K};var VQ=(Z,X)=>()=>(X||Z((X={exports:{}}).exports,X),X.exports);var v_=(Z)=>Z;function g_(Z,X){this[Z]=v_.bind(null,X)}var c0=(Z,X)=>{for(var Q in X)oK(Z,Q,{get:X[Q],enumerable:!0,configurable:!0,set:g_.bind(X,Q)})};var p=(Z,X)=>()=>(Z&&(X=Z(Z=0)),X);var t0=import.meta.require;var EO={};c0(EO,{lokiDir:()=>j0,homeLokiDir:()=>I4,findRepoRootForVersion:()=>tK,REPO_ROOT:()=>Q8});import{resolve as s7,dirname as rK}from"path";import{fileURLToPath as m_}from"url";import{existsSync as qQ}from"fs";import{homedir as u_}from"os";function p_(){let Z=DO;for(let X=0;X<6;X++){if(qQ(s7(Z,"VERSION"))&&qQ(s7(Z,"autonomy/run.sh")))return Z;let Q=rK(Z);if(Q===Z)break;Z=Q}return s7(DO,"..","..","..")}function tK(Z){let X=Z;for(let Q=0;Q<6;Q++){if(qQ(s7(X,"VERSION"))&&qQ(s7(X,"autonomy/run.sh")))return X;let Y=rK(X);if(Y===X)break;X=Y}return s7(Z,"..","..","..")}function j0(){return process.env.LOKI_DIR??s7(process.cwd(),".loki")}function I4(){return s7(u_(),".loki")}var DO,Q8;var G8=p(()=>{DO=rK(m_(import.meta.url));Q8=p_()});import{readFileSync as d_}from"fs";import{resolve as c_,dirname as l_}from"path";import{fileURLToPath as i_}from"url";function _3(){if(f5!==null)return f5;let Z="8.0.3";if(typeof Z==="string"&&Z.length>0)return f5=Z,f5;try{let X=l_(i_(import.meta.url)),Q=tK(X);f5=d_(c_(Q,"VERSION"),"utf-8").trim()}catch{f5="unknown"}return f5}var f5=null;var GQ=p(()=>{G8()});var kO={};c0(kO,{runOrThrow:()=>Wf,run:()=>D0,readStreamCapped:()=>HQ,commandVersion:()=>qf,commandExists:()=>H9,ShellError:()=>eK,MAX_STDOUT_BYTES:()=>RO});async function HQ(Z,X=RO){let Q=Z.getReader(),Y=new TextDecoder,J="",z=0;try{while(z<X){let{done:K,value:$}=await Q.read();if(K)break;if(!$)continue;if(z+=$.byteLength,z>X){let W=$.byteLength-(z-X);J+=Y.decode($.subarray(0,W),{stream:!0});break}J+=Y.decode($,{stream:!0})}J+=Y.decode()}finally{try{await Q.cancel()}catch{}Q.releaseLock()}return J}async function D0(Z,X={}){let Q=Bun.spawn({cmd:[...Z],stdout:"pipe",stderr:"pipe",env:X.env?{...process.env,...X.env}:process.env,cwd:X.cwd}),Y,J;if(X.timeoutMs&&X.timeoutMs>0)Y=setTimeout(()=>{try{Q.kill("SIGTERM")}catch{}J=setTimeout(()=>{try{Q.kill("SIGKILL")}catch{}},2000)},X.timeoutMs);try{let[z,K,$]=await Promise.all([HQ(Q.stdout),new Response(Q.stderr).text(),Q.exited]);return{stdout:z,stderr:K,exitCode:$}}finally{if(Y)clearTimeout(Y);if(J)clearTimeout(J)}}async function Wf(Z,X={}){let Q=await D0(Z,X);if(Q.exitCode!==0)throw new eK(`command failed (${Q.exitCode}): ${Z.join(" ")}`,Q.exitCode,Q.stdout,Q.stderr);return Q}async function H9(Z){let X=Vf(Z),Q=await D0(["sh","-c",`command -v ${X}`],{timeoutMs:5000});if(Q.exitCode===0)return Q.stdout.trim()||null;return null}function Vf(Z){if(!/^[A-Za-z0-9._/-]+$/.test(Z))throw Error(`refused to shell-escape suspect token: ${Z}`);return Z}async function qf(Z,X="--version"){if(!await H9(Z))return null;let Y=await D0([Z,X],{timeoutMs:5000});if(Y.exitCode!==0)return null;return((Y.stdout||Y.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var RO=16777216,eK;var k9=p(()=>{eK=class eK extends Error{message;exitCode;stdout;stderr;constructor(Z,X,Q,Y){super(Z);this.message=Z;this.exitCode=X;this.stdout=Q;this.stderr=Y;this.name="ShellError"}}});function n7(Z){return Gf?"":Z}var Gf,M0,k8,l0,gW0,a0,H8,X9,g;var S6=p(()=>{Gf=(process.env.NO_COLOR??"").length>0;M0=n7("\x1B[0;31m"),k8=n7("\x1B[0;32m"),l0=n7("\x1B[1;33m"),gW0=n7("\x1B[0;34m"),a0=n7("\x1B[0;36m"),H8=n7("\x1B[1m"),X9=n7("\x1B[2m"),g=n7("\x1B[0m")});import{existsSync as wf}from"fs";async function F7(){if(R4!==void 0)return R4;let Z="/opt/homebrew/bin/python3.12";if(wf(Z))return R4=Z,Z;let X=await H9("python3.12");if(X)return R4=X,X;let Q=await H9("python3");return R4=Q,Q}async function D7(Z,X={}){let Q=await F7();if(!Q)return{stdout:"",stderr:"python3 not found",exitCode:127};return D0([Q,"-c",Z],X)}var R4;var o7=p(()=>{k9()});var oO={};c0(oO,{runStatus:()=>of});import{existsSync as Q9,readFileSync as h3,readdirSync as pO,statSync as dO}from"fs";import{resolve as h8,basename as mf}from"path";import{homedir as uf}from"os";function cO(Z){let X=Math.trunc(Z);if(X>=1e6)return`${(Math.trunc(X/1e6*10)/10).toFixed(1)}M`;if(X>=1000)return`${(Math.trunc(X/1000*10)/10).toFixed(1)}K`;return String(X)}function lO(Z,X,Q){if(X===0)return null;let Y=Math.trunc(Z*100/X),J=Math.trunc(Z*BQ/X);if(J>BQ)J=BQ;let z=BQ-J,K=k8;if(Y>=80)K=M0;else if(Y>=50)K=l0;let $="=".repeat(Math.max(0,J))+" ".repeat(Math.max(0,z)),W=cO(Z),V=cO(X);return`  ${H8}${Q}${g} ${K}[${$}]${g} ${Y}% (${W} / ${V})`}async function df(){if(await H9("jq"))return!0;return process.stdout.write(`${M0}Error: jq is required but not installed.${g}
+var k_=Object.create;var{getPrototypeOf:x_,defineProperty:oK,getOwnPropertyNames:S_}=Object;var y_=Object.prototype.hasOwnProperty;function b_(Z){return this[Z]}var __,f_,h_=(Z,X,Q)=>{var Y=Z!=null&&typeof Z==="object";if(Y){var J=X?__??=new WeakMap:f_??=new WeakMap,z=J.get(Z);if(z)return z}Q=Z!=null?k_(x_(Z)):{};let K=X||!Z||!Z.__esModule?oK(Q,"default",{value:Z,enumerable:!0}):Q;for(let $ of S_(Z))if(!y_.call(K,$))oK(K,$,{get:b_.bind(Z,$),enumerable:!0});if(Y)J.set(Z,K);return K};var VQ=(Z,X)=>()=>(X||Z((X={exports:{}}).exports,X),X.exports);var v_=(Z)=>Z;function g_(Z,X){this[Z]=v_.bind(null,X)}var c0=(Z,X)=>{for(var Q in X)oK(Z,Q,{get:X[Q],enumerable:!0,configurable:!0,set:g_.bind(X,Q)})};var p=(Z,X)=>()=>(Z&&(X=Z(Z=0)),X);var t0=import.meta.require;var EO={};c0(EO,{lokiDir:()=>j0,homeLokiDir:()=>I4,findRepoRootForVersion:()=>tK,REPO_ROOT:()=>Q8});import{resolve as s7,dirname as rK}from"path";import{fileURLToPath as m_}from"url";import{existsSync as qQ}from"fs";import{homedir as u_}from"os";function p_(){let Z=DO;for(let X=0;X<6;X++){if(qQ(s7(Z,"VERSION"))&&qQ(s7(Z,"autonomy/run.sh")))return Z;let Q=rK(Z);if(Q===Z)break;Z=Q}return s7(DO,"..","..","..")}function tK(Z){let X=Z;for(let Q=0;Q<6;Q++){if(qQ(s7(X,"VERSION"))&&qQ(s7(X,"autonomy/run.sh")))return X;let Y=rK(X);if(Y===X)break;X=Y}return s7(Z,"..","..","..")}function j0(){return process.env.LOKI_DIR??s7(process.cwd(),".loki")}function I4(){return s7(u_(),".loki")}var DO,Q8;var G8=p(()=>{DO=rK(m_(import.meta.url));Q8=p_()});import{readFileSync as d_}from"fs";import{resolve as c_,dirname as l_}from"path";import{fileURLToPath as i_}from"url";function _3(){if(f5!==null)return f5;let Z="8.1.0";if(typeof Z==="string"&&Z.length>0)return f5=Z,f5;try{let X=l_(i_(import.meta.url)),Q=tK(X);f5=d_(c_(Q,"VERSION"),"utf-8").trim()}catch{f5="unknown"}return f5}var f5=null;var GQ=p(()=>{G8()});var kO={};c0(kO,{runOrThrow:()=>Wf,run:()=>D0,readStreamCapped:()=>HQ,commandVersion:()=>qf,commandExists:()=>H9,ShellError:()=>eK,MAX_STDOUT_BYTES:()=>RO});async function HQ(Z,X=RO){let Q=Z.getReader(),Y=new TextDecoder,J="",z=0;try{while(z<X){let{done:K,value:$}=await Q.read();if(K)break;if(!$)continue;if(z+=$.byteLength,z>X){let W=$.byteLength-(z-X);J+=Y.decode($.subarray(0,W),{stream:!0});break}J+=Y.decode($,{stream:!0})}J+=Y.decode()}finally{try{await Q.cancel()}catch{}Q.releaseLock()}return J}async function D0(Z,X={}){let Q=Bun.spawn({cmd:[...Z],stdout:"pipe",stderr:"pipe",env:X.env?{...process.env,...X.env}:process.env,cwd:X.cwd}),Y,J;if(X.timeoutMs&&X.timeoutMs>0)Y=setTimeout(()=>{try{Q.kill("SIGTERM")}catch{}J=setTimeout(()=>{try{Q.kill("SIGKILL")}catch{}},2000)},X.timeoutMs);try{let[z,K,$]=await Promise.all([HQ(Q.stdout),new Response(Q.stderr).text(),Q.exited]);return{stdout:z,stderr:K,exitCode:$}}finally{if(Y)clearTimeout(Y);if(J)clearTimeout(J)}}async function Wf(Z,X={}){let Q=await D0(Z,X);if(Q.exitCode!==0)throw new eK(`command failed (${Q.exitCode}): ${Z.join(" ")}`,Q.exitCode,Q.stdout,Q.stderr);return Q}async function H9(Z){let X=Vf(Z),Q=await D0(["sh","-c",`command -v ${X}`],{timeoutMs:5000});if(Q.exitCode===0)return Q.stdout.trim()||null;return null}function Vf(Z){if(!/^[A-Za-z0-9._/-]+$/.test(Z))throw Error(`refused to shell-escape suspect token: ${Z}`);return Z}async function qf(Z,X="--version"){if(!await H9(Z))return null;let Y=await D0([Z,X],{timeoutMs:5000});if(Y.exitCode!==0)return null;return((Y.stdout||Y.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var RO=16777216,eK;var k9=p(()=>{eK=class eK extends Error{message;exitCode;stdout;stderr;constructor(Z,X,Q,Y){super(Z);this.message=Z;this.exitCode=X;this.stdout=Q;this.stderr=Y;this.name="ShellError"}}});function n7(Z){return Gf?"":Z}var Gf,M0,k8,l0,gW0,a0,H8,X9,g;var S6=p(()=>{Gf=(process.env.NO_COLOR??"").length>0;M0=n7("\x1B[0;31m"),k8=n7("\x1B[0;32m"),l0=n7("\x1B[1;33m"),gW0=n7("\x1B[0;34m"),a0=n7("\x1B[0;36m"),H8=n7("\x1B[1m"),X9=n7("\x1B[2m"),g=n7("\x1B[0m")});import{existsSync as wf}from"fs";async function F7(){if(R4!==void 0)return R4;let Z="/opt/homebrew/bin/python3.12";if(wf(Z))return R4=Z,Z;let X=await H9("python3.12");if(X)return R4=X,X;let Q=await H9("python3");return R4=Q,Q}async function D7(Z,X={}){let Q=await F7();if(!Q)return{stdout:"",stderr:"python3 not found",exitCode:127};return D0([Q,"-c",Z],X)}var R4;var o7=p(()=>{k9()});var oO={};c0(oO,{runStatus:()=>of});import{existsSync as Q9,readFileSync as h3,readdirSync as pO,statSync as dO}from"fs";import{resolve as h8,basename as mf}from"path";import{homedir as uf}from"os";function cO(Z){let X=Math.trunc(Z);if(X>=1e6)return`${(Math.trunc(X/1e6*10)/10).toFixed(1)}M`;if(X>=1000)return`${(Math.trunc(X/1000*10)/10).toFixed(1)}K`;return String(X)}function lO(Z,X,Q){if(X===0)return null;let Y=Math.trunc(Z*100/X),J=Math.trunc(Z*BQ/X);if(J>BQ)J=BQ;let z=BQ-J,K=k8;if(Y>=80)K=M0;else if(Y>=50)K=l0;let $="=".repeat(Math.max(0,J))+" ".repeat(Math.max(0,z)),W=cO(Z),V=cO(X);return`  ${H8}${Q}${g} ${K}[${$}]${g} ${Y}% (${W} / ${V})`}async function df(){if(await H9("jq"))return!0;return process.stdout.write(`${M0}Error: jq is required but not installed.${g}
 `),process.stdout.write(`Install with:
 `),process.stdout.write(`  brew install jq    (macOS)
 `),process.stdout.write(`  apt install jq     (Debian/Ubuntu)
@@ -1203,4 +1203,4 @@ Set LOKI_LEGACY_BASH=1 to force the bash CLI for every command.
 `),2}case"start":{let{runStart:Y}=await Promise.resolve().then(() => (I_(),E_));return Y(Q)}default:return process.stderr.write(`Unknown command: ${X}
 `),process.stderr.write(P_),2}}mO();process.on("SIGINT",()=>process.exit(130));process.on("SIGTERM",()=>process.exit(143));var TW0=await MW0(Bun.argv.slice(2));process.exit(TW0);
 
-//# debugId=3C2284D252E647F964756E2164756E21
+//# debugId=59A2C45B0211620364756E2164756E21

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '8.0.3'
+__version__ = '8.1.0'

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "loki-mode",
   "mcpName": "io.github.asklokesh/loki-mode",
-  "version": "8.0.3",
+  "version": "8.1.0",
   "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 8 quality gates. Provider-agnostic (Claude Code, OpenAI Codex, Cline, Aider).",
   "keywords": [
     "agent",

package/plugins/loki-mode/.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "$schema": "https://json.schemastore.org/claude-code-plugin-manifest.json",
   "name": "loki-mode",
   "displayName": "Loki Mode",
-  "version": "8.0.3",
+  "version": "8.1.0",
   "description": "Autonomous spec-to-product build system with a built-in trust layer (RARV-C closure loop, 8 quality gates, completion council). Ships Loki's spec-hardening, drift-detection, and deterministic PR verification commands plus the Loki MCP server.",
   "author": {
     "name": "Autonomi",

package/providers/model_catalog.json

@@ -1,7 +1,7 @@
 {
   "_comment": "Canonical model catalog. Update this single file when a provider ships a new model. Providers/web-app/docs read from here.",
   "schema_version": 1,
-  "updated": "2026-06-30",
+  "updated": "2026-07-29",
   "providers": {
     "claude": {
       "latest_planning": "claude-opus-4-8",
@@ -52,31 +52,78 @@
       "latest_development": "gpt-5.3-codex",
       "latest_fast": "gpt-5.3-codex",
       "models": [
-        { "id": "gpt-5.3-codex", "tier": "planning" },
-        { "id": "o3", "tier": "planning" },
-        { "id": "o4-mini", "tier": "fast" }
+        {
+          "id": "gpt-5.3-codex",
+          "tier": "planning"
+        },
+        {
+          "id": "o3",
+          "tier": "planning"
+        },
+        {
+          "id": "o4-mini",
+          "tier": "fast"
+        }
       ],
       "notes": "Codex uses a single model with effort level (xhigh/high/low) for tier differentiation"
     },
     "cline": {
-      "latest_planning": "claude-opus-4-8",
-      "latest_development": "claude-sonnet-5",
-      "latest_fast": "claude-sonnet-5",
+      "latest_planning": "openrouter/deepseek/deepseek-v3.2",
+      "latest_development": "openrouter/deepseek/deepseek-v3.2",
+      "latest_fast": "openrouter/deepseek/deepseek-chat",
       "models": [
-        { "id": "claude-opus-4-8", "tier": "planning" },
-        { "id": "claude-sonnet-5", "tier": "development" },
-        { "id": "gpt-4.1", "tier": "development" }
+        {
+          "id": "openrouter/deepseek/deepseek-v3.2",
+          "tier": "planning",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/deepseek/deepseek-v3.2",
+          "tier": "development",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/deepseek/deepseek-chat",
+          "tier": "fast",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/z-ai/glm-4.6",
+          "tier": "development",
+          "open_weights": true
+        }
       ]
     },
     "aider": {
-      "latest_planning": "claude-opus-4-8",
-      "latest_development": "claude-sonnet-5",
-      "latest_fast": "claude-sonnet-5",
+      "latest_planning": "openrouter/deepseek/deepseek-v3.2",
+      "latest_development": "openrouter/deepseek/deepseek-v3.2",
+      "latest_fast": "openrouter/deepseek/deepseek-chat",
       "models": [
-        { "id": "claude-opus-4-8", "tier": "planning" },
-        { "id": "claude-sonnet-5", "tier": "development" },
-        { "id": "gpt-4.1", "tier": "development" },
-        { "id": "ollama_chat/deepseek-coder", "tier": "fast" }
+        {
+          "id": "openrouter/deepseek/deepseek-v3.2",
+          "tier": "planning",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/deepseek/deepseek-v3.2",
+          "tier": "development",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/deepseek/deepseek-chat",
+          "tier": "fast",
+          "open_weights": true
+        },
+        {
+          "id": "openrouter/minimax/minimax-m2.1",
+          "tier": "development",
+          "open_weights": true
+        },
+        {
+          "id": "ollama_chat/deepseek-coder",
+          "tier": "fast",
+          "open_weights": true
+        }
       ]
     }
   }