loki-mode 9.73.0 → 9.75.0

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 v9.73.0
+# Loki Mode v9.75.0
 
 **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
 
 ---
 
-**v9.73.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v9.75.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-9.73.0
+9.75.0

package/autonomy/council-v2.sh

@@ -20,15 +20,43 @@
 #   - swarm/calibration.py (reviewer calibration)
 #
 # Inline Python (D7): the council runs with the cwd inside the agent's repo, so
-# every python3 here runs -E and drops '' and '.' from sys.path before any other
-# import (the swarm dir is inserted after that); a committed json.py or
-# sitecustomize.py cannot read the votes. Pinned by
-# tests/test-council-no-fabricated-verdict.sh (D7 section).
+# every python here drops '' and '.' from sys.path before any other import (the
+# swarm dir is inserted after that); a committed json.py or sitecustomize.py
+# cannot read the votes. Pinned by tests/test-council-no-fabricated-verdict.sh
+# (D7 section). S-201: -E alone still loads user site-packages, whose .pth
+# "import" lines run before that scrub and can forge json.load, so every site
+# runs the _loki_snapshot_py_tool interpreter -I -S. Pinned by
+# tests/test-council-v2-no-user-site-pth.sh.
 #
 #===============================================================================
 
 COUNCIL_V2_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 
+# Guarded copy for callers that source this file on its own (tests). In
+# production completion-council.sh (or run.sh) has already defined it. run.sh
+# is the source of truth: keep this copy byte-identical to it.
+declare -F _loki_snapshot_py_tool >/dev/null 2>&1 || \
+_loki_snapshot_py_tool() {
+    local c
+    for c in /usr/bin/python3 /bin/python3; do
+        [ -x "$c" ] && [ ! -d "$c" ] && "$c" -I -S -c '' >/dev/null 2>&1 && { printf '%s\n' "$c"; return 0; }
+    done
+    local dir
+    local IFS=:
+    for dir in $PATH; do
+        case "$dir" in
+            /*) ;;
+            *) continue ;;
+        esac
+        if [ -x "$dir/python3" ] && [ ! -d "$dir/python3" ] \
+           && "$dir/python3" -I -S -c '' >/dev/null 2>&1; then
+            printf '%s\n' "$dir/python3"
+            return 0
+        fi
+    done
+    return 1
+}
+
 #===============================================================================
 # council_v2_vote() -- Main entry point for v2 voting
 #
@@ -48,6 +76,9 @@ council_v2_vote() {
     local iteration="${4:-0}"
 
     local council_size="${COUNCIL_SIZE:-3}"
+    # S-201: none resolvable -> every read below takes its fail-safe fallback.
+    local _c2_py
+    _c2_py="$(_loki_snapshot_py_tool)" || _c2_py=""
 
     log_header "COMPLETION COUNCIL v2 - Iteration $iteration"
     log_info "Convening ${council_size}-member blind review council..."
@@ -103,7 +134,7 @@ council_v2_vote() {
             votes_json="$votes_json$vote_content"
 
             local verdict
-            verdict=$(echo "$vote_content" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "UNKNOWN")
+            verdict=$(echo "$vote_content" | "$_c2_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import 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
@@ -141,14 +172,21 @@ council_v2_vote() {
 
     # Step 4: Sycophancy detection
     local sycophancy_score
-    sycophancy_score=$(python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+    sycophancy_score=$("$_c2_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import sys
 sys.path.insert(0, '${COUNCIL_V2_DIR}/../swarm')
 from sycophancy import detect_sycophancy
 import json
 votes = json.loads(sys.argv[1])
 print('{:.3f}'.format(detect_sycophancy(votes)))
-" "$votes_json" 2>/dev/null || echo "0.000")
+" "$votes_json" 2>/dev/null) || sycophancy_score=""
+    # S-201: a crashed or missing detector did not measure 0.000. Record the
+    # score as unmeasured (null keeps summary.json valid JSON) so Step 5
+    # challenges instead of reading "independent".
+    case "$sycophancy_score" in
+        [0-9]*.[0-9][0-9][0-9]) ;;
+        *) sycophancy_score="null" ;;
+    esac
 
     log_info "Sycophancy score: $sycophancy_score"
 
@@ -156,10 +194,17 @@ print('{:.3f}'.format(detect_sycophancy(votes)))
     if [ "$approve_count" -eq "$council_size" ]; then
         local threshold="${LOKI_COUNCIL_SYCOPHANCY_THRESHOLD:-0.6}"
         local should_challenge
-        should_challenge=$(python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; print('yes' if float('$sycophancy_score') >= float('$threshold') else 'no')" 2>/dev/null || echo "no")
+        # S-201: an unmeasured score, or a compare that cannot run, challenges.
+        # Only a measured score below the threshold skips the devil's advocate.
+        if [ "$sycophancy_score" = "null" ]; then
+            should_challenge="yes"
+        else
+            should_challenge=$("$_c2_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; print('no' if float(sys.argv[1]) < float(sys.argv[2]) else 'yes')" "$sycophancy_score" "$threshold" 2>/dev/null) || should_challenge="yes"
+            [ "$should_challenge" = "no" ] || should_challenge="yes"
+        fi
 
         if [ "$should_challenge" = "yes" ]; then
-            log_warn "Sycophancy score $sycophancy_score >= $threshold -- adding devil's advocate"
+            log_warn "Sycophancy score $sycophancy_score (threshold $threshold, null = unmeasured) -- adding devil's advocate"
             # Run devil's advocate with fresh evidence (no visibility of other votes)
             local da_dir
             da_dir=$(mktemp -d)
@@ -173,7 +218,7 @@ print('{:.3f}'.format(detect_sycophancy(votes)))
                 # 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 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "INCONCLUSIVE")
+                da_verdict=$(cat "$da_vote" | "$_c2_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json; print(json.load(sys.stdin).get('verdict','').upper())" 2>/dev/null || echo "INCONCLUSIVE")
                 if [ "$da_verdict" = "INCONCLUSIVE" ]; then
                     log_warn "Devil's advocate produced no verdict -- unanimous approval left UNCHANGED (not overturned)"
                 elif [ "$da_verdict" = "REJECT" ]; then
@@ -207,7 +252,7 @@ print('{:.3f}'.format(detect_sycophancy(votes)))
         final_decision="reject"
     fi
 
-    python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+    "$_c2_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import sys
 sys.path.insert(0, '${COUNCIL_V2_DIR}/../swarm')
 from calibration import CalibrationTracker
@@ -271,6 +316,8 @@ council_v2_run_reviewer() {
     local role="$1"
     local review_dir="$2"
     local output_file="$3"
+    local _c2_py
+    _c2_py="$(_loki_snapshot_py_tool)" || _c2_py=""
 
     # Build review prompt based on role
     local prompt
@@ -394,7 +441,7 @@ Respond ONLY with a valid JSON object. No markdown fencing."
                             --json-schema "$_c2_schema_content" --output-format json 2>/dev/null)" || _c2_json=""
                         if [ -n "$_c2_json" ]; then
                             # Pull the schema-validated object out of the envelope.
-                            result="$(printf '%s' "$_c2_json" | python3 -E -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]; import json
+                            result="$(printf '%s' "$_c2_json" | "$_c2_py" -I -S -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]; import json
 try:
     e=json.load(sys.stdin)
     p=e.get("structured_output")
@@ -496,7 +543,7 @@ except Exception:
     # or neither, stays INCONCLUSIVE -- never guessed.
     if [ -z "$extracted" ] && [ -n "${result:-}" ]; then
         local _recovered
-        _recovered="$(printf '%s' "$result" | _LOKI_RAW="$result" python3 -E -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]
+        _recovered="$(printf '%s' "$result" | _LOKI_RAW="$result" "$_c2_py" -I -S -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]
 import os, re, sys, json
 raw = os.environ.get("_LOKI_RAW", "")
 # Standalone words only: "APPROVE" not "approved-by", and not inside a URL.

package/autonomy/lib/done-recognition.sh

@@ -38,9 +38,42 @@
 # No emojis. No em dashes. bash 3.2 safe. Honors `set -uo pipefail`.
 #
 # Inline Python (D7): the gate runs with the cwd inside the agent's repo, so
-# every python3 here runs -E and drops '' and '.' from sys.path before any other
-# import; a committed json.py or sitecustomize.py cannot turn the verdict done.
-# Pinned by tests/test-reuse-done-recognition.sh case (d7).
+# every python3 here drops '' and '.' from sys.path before any other import; a
+# committed json.py or sitecustomize.py cannot turn the verdict done. Pinned by
+# tests/test-reuse-done-recognition.sh case (d7).
+#
+# S-200 (BACKLOG 54): -E alone still loads the user site-packages, whose .pth
+# "import" lines run before that scrub and can forge json.load (a forged verdict
+# read fast-stopped an incomplete project as done). Every call site resolves
+# its interpreter through _loki_snapshot_py_tool and runs it -I -S; when none
+# resolves, nothing reads as done or met (the gate builds, the writers write
+# nothing). See that function's comment in run.sh for the rationale and accepted
+# gaps. run.sh sources this file AFTER it defines _loki_snapshot_py_tool, so in
+# production run.sh's definition is kept; the copy below only serves callers
+# that source this file on its own (tests). run.sh is the source of truth: keep
+# this copy byte-identical to it. Pinned by
+# tests/test-done-recognition-no-user-site-pth.sh.
+declare -F _loki_snapshot_py_tool >/dev/null 2>&1 || \
+_loki_snapshot_py_tool() {
+    local c
+    for c in /usr/bin/python3 /bin/python3; do
+        [ -x "$c" ] && [ ! -d "$c" ] && "$c" -I -S -c '' >/dev/null 2>&1 && { printf '%s\n' "$c"; return 0; }
+    done
+    local dir
+    local IFS=:
+    for dir in $PATH; do
+        case "$dir" in
+            /*) ;;
+            *) continue ;;
+        esac
+        if [ -x "$dir/python3" ] && [ ! -d "$dir/python3" ] \
+           && "$dir/python3" -I -S -c '' >/dev/null 2>&1; then
+            printf '%s\n' "$dir/python3"
+            return 0
+        fi
+    done
+    return 1
+}
 
 # Bound the single model call so a huge PRD or test log cannot run away.
 : "${LOKI_DONE_RECOG_TIMEOUT:=180}"           # seconds for the single model call
@@ -133,8 +166,9 @@ _loki_done_recog_invoke() {
                     claude --dangerously-skip-permissions -p "$prompt" \
                     --json-schema "$_dr_schema_content" --output-format json 2>/dev/null) || _dr_json=""
             fi
-            if [ -n "$_dr_json" ]; then
-                _dr_obj=$(printf '%s' "$_dr_json" | python3 -E -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]; import json
+            local _dr_py
+            if [ -n "$_dr_json" ] && _dr_py="$(_loki_snapshot_py_tool)"; then
+                _dr_obj=$(printf '%s' "$_dr_json" | "$_dr_py" -I -S -c 'import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]; import json
 try:
     e=json.load(sys.stdin)
     p=e.get("structured_output")
@@ -206,7 +240,9 @@ _loki_done_recog_prd_sha() {
         sha256sum "$prd_file" 2>/dev/null | awk '{print $1}'
     else
         # Last-resort: a python hash, still deterministic over the same bytes.
-        LOKI_DR_PRD="$prd_file" python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+        local _dr_py
+        _dr_py="$(_loki_snapshot_py_tool)" || { printf ''; return 0; }
+        LOKI_DR_PRD="$prd_file" "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import hashlib, os
 p = os.environ.get('LOKI_DR_PRD','')
 try:
@@ -278,6 +314,13 @@ reuse_done_recognition_gate() {
     fi
     local _test_results="$loki_dir/quality/test-results.json"
 
+    # --- Resolve the -I -S interpreter (S-200); none -> inconclusive -> build.
+    local _dr_py
+    if ! _dr_py="$(_loki_snapshot_py_tool)"; then
+        log_info "Done-recognition: no isolated python3 interpreter resolved; proceeding to build."
+        return 1
+    fi
+
     # --- Build the model prompt payload (python: bounded, defensive) ---------
     local prompt
     prompt=$(LOKI_DR_PRD="$prd_path" \
@@ -287,7 +330,7 @@ reuse_done_recognition_gate() {
              LOKI_DR_CHECKLIST="$loki_dir/checklist/checklist.json" \
              LOKI_DR_MAX_PRD="${LOKI_DONE_RECOG_MAX_PRD_CHARS}" \
              LOKI_DR_MAX_TEST="${LOKI_DONE_RECOG_MAX_TEST_CHARS}" \
-             python3 -E << 'DR_PROMPT_EOF'
+             "$_dr_py" -I -S << 'DR_PROMPT_EOF'
 import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]
 import json, os, sys
 
@@ -389,7 +432,7 @@ DR_PROMPT_EOF
              LOKI_DR_TESTS="$_test_results" \
              LOKI_DR_ACTION="$action" \
              LOKI_DR_PRD="$prd_path" \
-             python3 -E << 'DR_PARSE_EOF'
+             "$_dr_py" -I -S << 'DR_PARSE_EOF'
 import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]
 import json, os, re, sys
 
@@ -700,11 +743,11 @@ DR_PARSE_EOF
     fi
 
     local verdict
-    verdict=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('verdict',''))" 2>/dev/null)
+    verdict=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('verdict',''))" 2>/dev/null)
 
     case "$verdict" in
         done)
-            _loki_done_recog_finish "$prd_path" "$parsed"
+            _loki_done_recog_finish "$prd_path" "$parsed" || return 1
             return 0
             ;;
         incomplete)
@@ -713,7 +756,7 @@ DR_PARSE_EOF
             ;;
         *)
             local _reason
-            _reason=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('reason','') or 'unverifiable')" 2>/dev/null)
+            _reason=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('reason','') or 'unverifiable')" 2>/dev/null)
             log_info "Done-recognition: could not confirm the existing code already satisfies the reused spec (${_reason}). Proceeding to build to be safe."
             return 1
             ;;
@@ -729,6 +772,10 @@ _loki_done_recog_finish() {
     local parsed="$2"
     local loki_dir="${TARGET_DIR:-.}/.loki"
 
+    # S-200: no -I -S interpreter -> write no done record (the gate then builds).
+    local _dr_py
+    _dr_py="$(_loki_snapshot_py_tool)" || return 1
+
     # The gate runs EARLY in run_autonomous, before the run normally mints these
     # run-scoped ids/baselines (run.sh sets them just after this call site). The
     # council-parity finalizers below (and what they transitively call) expect
@@ -749,10 +796,10 @@ _loki_done_recog_finish() {
     mkdir -p "$loki_dir/state" 2>/dev/null || true
 
     local summary met total axis
-    summary=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('summary','') or 'Project already satisfies its spec.')" 2>/dev/null)
-    met=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('met_count',0))" 2>/dev/null)
-    total=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('total_count',0))" 2>/dev/null)
-    axis=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('tests_axis','unknown'))" 2>/dev/null)
+    summary=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('summary','') or 'Project already satisfies its spec.')" 2>/dev/null)
+    met=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('met_count',0))" 2>/dev/null)
+    total=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('total_count',0))" 2>/dev/null)
+    axis=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('tests_axis','unknown'))" 2>/dev/null)
     [ -n "$met" ] || met=0
     [ -n "$total" ] || total=0
     [ -n "$axis" ] || axis="unknown"
@@ -792,7 +839,7 @@ _loki_done_recog_finish() {
         echo ""
         echo "## Per-requirement evidence"
         echo ""
-        printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+        printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import json, sys
 try:
     d = json.load(sys.stdin)
@@ -811,7 +858,7 @@ for t in d.get('satisfied', []):
     LOKI_DR_MET="$met" \
     LOKI_DR_TOTAL="$total" \
     LOKI_DR_TS="$ts" \
-    python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+    "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import json, os, tempfile
 out = os.environ['LOKI_DR_OUT']
 def i(v):
@@ -861,6 +908,7 @@ except Exception:
         log_info "Verified ${met}/${total} requirements met by code inspection (no passing test run was available to confirm). ${summary}"
     fi
     log_info "To rebuild from scratch run 'loki start --fresh-prd'; to extend it, edit the spec or pass a new/changed PRD."
+    return 0
 }
 
 # incomplete path: write the satisfied-requirements manifest so
@@ -871,6 +919,10 @@ _loki_done_recog_write_manifest() {
     local parsed="$2"
     local loki_dir="${TARGET_DIR:-.}/.loki"
 
+    # S-200: no -I -S interpreter -> record nothing as met (full build).
+    local _dr_py
+    _dr_py="$(_loki_snapshot_py_tool)" || return 1
+
     mkdir -p "$loki_dir/state" 2>/dev/null || true
 
     local prd_sha
@@ -882,7 +934,7 @@ _loki_done_recog_write_manifest() {
     LOKI_DR_PARSED="$parsed" \
     LOKI_DR_SHA="$prd_sha" \
     LOKI_DR_TS="$ts" \
-    python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
+    "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]
 import json, os, sys, tempfile
 out = os.environ['LOKI_DR_OUT']
 try:
@@ -923,7 +975,7 @@ except Exception:
     if [ -f "$loki_dir/queue/pending.json" ]; then
         # Drop prior PRD-sourced tasks so the incremental pass is the source of
         # truth; non-PRD tasks (if any) are preserved.
-        LOKI_DR_PENDING="$loki_dir/queue/pending.json" python3 -E - <<'RESET_EOF' 2>/dev/null || true
+        LOKI_DR_PENDING="$loki_dir/queue/pending.json" "$_dr_py" -I -S - <<'RESET_EOF' 2>/dev/null || true
 import sys; sys.path[:] = [p for p in sys.path if p not in ("", ".")]
 import json, os, tempfile
 p = os.environ.get("LOKI_DR_PENDING", ".loki/queue/pending.json")
@@ -948,8 +1000,8 @@ RESET_EOF
     fi
 
     local met total
-    met=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('met_count',0))" 2>/dev/null)
-    total=$(printf '%s' "$parsed" | python3 -E -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('total_count',0))" 2>/dev/null)
+    met=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('met_count',0))" 2>/dev/null)
+    total=$(printf '%s' "$parsed" | "$_dr_py" -I -S -c "import sys; sys.path[:] = [p for p in sys.path if p not in ('', '.')]; import json;print(json.load(sys.stdin).get('total_count',0))" 2>/dev/null)
     local unmet=$(( ${total:-0} - ${met:-0} ))
     log_info "Done-recognition: ${met:-0} of ${total:-0} requirements already satisfied; building only the ${unmet} unmet. Pass --fresh-prd to rebuild from scratch."
 }

package/autonomy/lib/proof-generator.py

@@ -911,6 +911,7 @@ def _collect_tests(loki_dir):
         "status": "not_run",
         "duration_sec": None,
     }
+    empty = dict(out)
     if not isinstance(raw, dict):
         return out
     out["runner"] = str(raw.get("runner") or "")
@@ -944,6 +945,12 @@ def _collect_tests(loki_dir):
             out["status"] = "inconclusive"
     else:
         out["status"] = "not_run"
+    # BACKLOG 98: a test-results.json left by an earlier iteration is not this
+    # run's evidence, so it cannot lift the headline (the gates' rule). An
+    # earlier red is never discarded: a stale failure still reads failed.
+    if out["status"] != "failed" and not _test_results_fresh(
+            os.path.join(loki_dir, "quality")):
+        return empty
     return out
 
 

package/autonomy/loki

@@ -28444,16 +28444,23 @@ if budget_limit is not None:
 
 # Prefer the recorded cumulative spend; fall back to the current run only when
 # the state file carries no figure. Never silently substitute 0.0 for "unknown".
-if isinstance(budget_used_recorded, (int, float)):
+# A recorded ZERO is not a reading: run.sh check_budget_limit writes
+# "budget_used": 0 for a run it never measured. Same rule as the dashboard
+# (server.py budget readers): only a positive recorded figure counts, and a
+# measured zero comes from current_cost, which collect_efficiency leaves None
+# unless a record passes record_is_measured.
+if isinstance(budget_used_recorded, (int, float)) and budget_used_recorded > 0:
     budget_used = budget_used_recorded
 elif isinstance(current_cost, (int, float)):
     budget_used = current_cost
 else:
-    budget_used = 0.0
+    budget_used = None
 status = "none"
 percent_used = None
 remaining = None
-if budget_limit is not None and budget_limit > 0:
+if budget_limit is not None and budget_limit > 0 and budget_used is None:
+    status = "unknown"
+elif budget_limit is not None and budget_limit > 0:
     percent_used = round(budget_used / budget_limit * 100, 2)
     remaining = max(0.0, budget_limit - budget_used)
     if budget_used >= budget_limit:
@@ -28487,7 +28494,7 @@ if show_json:
         "by_model": {k: round(v, 6) for k, v in by_model.items()},
         "budget": {
             "limit": budget_limit,
-            "used": round(budget_used, 6),
+            "used": round(budget_used, 6) if budget_used is not None else None,
             "remaining": round(remaining, 6) if remaining is not None else None,
             "percent_used": percent_used,
             "status": status,
@@ -28538,13 +28545,17 @@ if budget_limit is None:
     print(DIM + "  When set, Loki warns at 80% and hard-stops at 100%." + NC)
 else:
     col = GREEN
-    if status == "warn":
+    if status in ("warn", "unknown"):
         col = YELLOW
     elif status == "exceeded":
         col = RED
     print("  Cap:       " + _fmt_usd(budget_limit))
-    print("  Used:      " + _fmt_usd(budget_used) + " (" + col + str(percent_used) + "%" + NC + ")")
-    print("  Remaining: " + _fmt_usd(remaining))
+    if budget_used is None:
+        print("  Used:      not recorded (no measured iteration cost)")
+        print("  Remaining: not recorded")
+    else:
+        print("  Used:      " + _fmt_usd(budget_used) + " (" + col + str(percent_used) + "%" + NC + ")")
+        print("  Remaining: " + _fmt_usd(remaining))
     print("  Status:    " + col + BOLD + status.upper() + NC)
     if status == "warn":
         print(YELLOW + "  Warning: at or above 80% of cap. Run continues; hard-stop at 100%." + NC)

package/autonomy/run.sh

@@ -13875,7 +13875,20 @@ sys.stdout.write(t.strip())
                    echo "$LOKI_MONOREPO_TEST_CMD" | grep -qE '[;|`$]|&&|\|\||>>|<<'; then
                     log_error "LOKI_MONOREPO_TEST_CMD rejected (only [A-Za-z0-9_./= -] allowed): $LOKI_MONOREPO_TEST_CMD"
                     test_runner="monorepo-custom-rejected"
-                    details="monorepo-custom: rejected by whitelist (gate skipped, inconclusive)"
+                    # BACKLOG 62: nothing ran, so record the no-runner shape
+                    # (see the runner=="none" return below) instead of falling
+                    # through with test_passed=true, which wrote pass:true and
+                    # touched unit-tests.pass. The runner label stays distinct:
+                    # the council reads a non-boolean pass on a named runner as
+                    # NO_PASS (inconclusive pass-through). The command itself is
+                    # not echoed into the JSON (it failed the whitelist).
+                    rm -f "$quality_dir/unit-tests.pass" 2>/dev/null || true
+                    cat > "$quality_dir/test-results.json" << TREOF
+{"timestamp":"$(date -u +%Y-%m-%dT%H:%M:%SZ)","runner":"monorepo-custom-rejected","pass":"inconclusive","summary":"LOKI_MONOREPO_TEST_CMD rejected by whitelist (not run)","command":null,"exit_code":null,"status":"not_run","passed_count":null,"failed_count":null,"verification_gap":"test_command_rejected"}
+TREOF
+                    printf '%s\n' "${ITERATION_COUNT:-0}" > "$quality_dir/.test-results.iter" 2>/dev/null || true
+                    _LOKI_TEST_SUITE_STATUS=not_run
+                    return 0
                 else
                     test_runner="monorepo-custom"
                     local output

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "9.73.0"
+__version__ = "9.75.0"
 
 # Expose the control app for easy import
 try:

package/dashboard/server.py

@@ -7852,6 +7852,8 @@ def _compute_cost_snapshot() -> dict:
     budget_remaining = None
     # Did ANY record carry an observed value? Not "was a file present".
     cost_recorded = False
+    # Tokens were measured (tracker fallback) but no USD figure was recorded.
+    cost_unknown = False
 
     # Read efficiency files (one JSON file per iteration/task).
     # Use the iteration-*.json pattern so this reader sees the same
@@ -7929,25 +7931,32 @@ def _compute_cost_snapshot() -> dict:
                 total_input = totals.get("total_input", 0)
                 total_output = totals.get("total_output", 0)
                 if total_input > 0 or total_output > 0:
-                    # Real observed tokens from the context tracker: this IS a
-                    # measurement, even if the recorded USD total happens to
-                    # be 0.
+                    # Real observed tokens from the context tracker: the TOKENS
+                    # are a measurement. The USD figure is only a measurement
+                    # when the tracker recorded a number; a missing one is
+                    # unknown (null), while a recorded 0.0 stays 0.0.
                     cost_recorded = True
-                    estimated_cost = totals.get("total_cost_usd", 0.0)
+                    usd = totals.get("total_cost_usd")
+                    if isinstance(usd, (int, float)) and not isinstance(usd, bool):
+                        estimated_cost = usd
+                    else:
+                        cost_unknown = True
+                    # No provider recorded -> the model is unknown, not sonnet.
+                    model = str(ctx.get("provider") or "unknown").lower()
                     # Rebuild by_model and by_phase from per_iteration data
                     for it in ctx.get("per_iteration", []):
                         inp = it.get("input_tokens", 0)
                         out = it.get("output_tokens", 0)
-                        cost = it.get("cost_usd", 0)
-                        model = ctx.get("provider", "sonnet").lower()
+                        cost = it.get("cost_usd")
+                        cost_is_num = isinstance(cost, (int, float)) and not isinstance(cost, bool)
                         if model not in by_model:
-                            by_model[model] = {"input_tokens": 0, "output_tokens": 0, "cost_usd": 0.0}
+                            by_model[model] = {"input_tokens": 0, "output_tokens": 0, "cost_usd": 0.0,
+                                               "measured": False, "tokens_measured": True}
                         by_model[model]["input_tokens"] += inp
                         by_model[model]["output_tokens"] += out
-                        by_model[model]["cost_usd"] += cost
-                        # Observed tokens from the tracker (cost_recorded above).
-                        by_model[model]["measured"] = True
-                        by_model[model]["tokens_measured"] = True
+                        if cost_is_num:
+                            by_model[model]["cost_usd"] += cost
+                            by_model[model]["measured"] = True
             except (json.JSONDecodeError, KeyError, TypeError, AttributeError):
                 pass
 
@@ -7964,7 +7973,7 @@ def _compute_cost_snapshot() -> dict:
             # Spend against the cap only when something was measured: with
             # nothing recorded, "used" is unknown, not $0.00, and "remaining"
             # is unknown, not the whole cap.
-            if budget_limit is not None and cost_recorded:
+            if budget_limit is not None and cost_recorded and not cost_unknown:
                 budget_used = estimated_cost
                 budget_remaining = max(0.0, budget_limit - budget_used)
         except (json.JSONDecodeError, KeyError):
@@ -7988,7 +7997,7 @@ def _compute_cost_snapshot() -> dict:
             total_input + total_output + total_cache_read + total_cache_creation
         ) if cost_recorded else None,
         "cache_hit_ratio": round(total_cache_read / _read_in, 4) if _read_in > 0 else None,
-        "estimated_cost_usd": round(estimated_cost, 6) if cost_recorded else None,
+        "estimated_cost_usd": round(estimated_cost, 6) if cost_recorded and not cost_unknown else None,
         "cost_recorded": cost_recorded,
         "by_phase": {k: {
             "input_tokens": v["input_tokens"] if v.get("tokens_measured") else None,