loki-mode 9.16.0 → 9.17.2

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 v9.16.0
+# Loki Mode v9.17.2
 
 **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
 
 ---
 
-**v9.16.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v9.17.2 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-9.16.0
+9.17.2

package/autonomy/issue-providers.sh

@@ -380,26 +380,6 @@ repo = data.get('repo', '')
 
 labels_str = ', '.join(labels) if labels else ''
 
-_ac_text = (title + ' ' + body).lower()
-_ac_rules = [
-    (['save', 'persist', 'store', 'databas', 'crud'],
-     'Data the change writes survives a restart (a real store, not in-memory state).'),
-    (['auth', 'login', 'sign in', 'session', 'permission'],
-     'The auth path is exercised end to end including the denied case (401/403), not only the happy path.'),
-    (['api', 'endpoint', 'rest', 'graphql', 'route'],
-     'Each affected endpoint returns the documented status codes and is callable without a browser.'),
-    (['payment', 'stripe', 'billing', 'invoice', 'subscription'],
-     'The payment path runs against provider test mode; no mocked charge stands in for the integration.'),
-    (['bug', 'fix', 'regression', 'broken', 'crash', 'error'],
-     'A test reproduces the reported failure and FAILS before the fix, then passes after it.'),
-    (['perf', 'slow', 'latency', 'timeout', 'memory leak'],
-     'The improvement is measured before and after, and the numbers appear in the change.'),
-    (['security', 'vulnerab', 'injection', 'xss', 'csrf'],
-     'A test demonstrates the vulnerable behavior is refused after the change.'),
-]
-_hits = [c for kws, c in _ac_rules if any(k in _ac_text for k in kws)]
-derived_ac = ('\n'.join('- ' + h for h in _hits) + '\n') if _hits else ''
-
 prd = f'''# PRD: {title}
 
 **Source:** {provider.replace('_', ' ').title()} Issue [{number}]({url})
@@ -428,7 +408,6 @@ Based on the issue description, implement the following:
 2. Ensure backward compatibility (unless explicitly breaking changes are requested)
 3. Add appropriate tests for new functionality
 4. Update documentation as needed
-{derived_ac}
 
 ---
 

package/autonomy/lib/verdict.py

@@ -91,25 +91,9 @@ def _authorship_line(loki_dir, run_id=None):
 
 
 def _grounding_line(loki_dir):
-    # The completion route now persists the per-claim result here
-    # (run.sh:_loki_check_claim_grounding, written at the non-destructive claim
-    # peek). Absent file means the run never reached a completion claim, which
-    # is UNKNOWN -- not a pass, and not a failure.
-    d = _read_json(os.path.join(loki_dir, "state", "claim-grounding.json"))
-    if not isinstance(d, dict) or d.get("status") != "measured":
-        return UNKNOWN, "no completion claim was checked against the diff"
-    named = d.get("paths_named") or []
-    if not named:
-        # Reported by name rather than scored: a claim naming no path is
-        # UNGROUNDABLE, which is a different fact from a claim that checked out.
-        return UNKNOWN, "the completion claim named no file path, so it cannot be grounded"
-    ungrounded = d.get("ungrounded") or []
-    if d.get("has_ungrounded_claim"):
-        return "finding", (
-            f"the completion claim names {len(ungrounded)} path(s) absent from the diff: "
-            + ", ".join(str(p) for p in ungrounded[:3])
-        )
-    return "measured", f"all {len(named)} path(s) named in the completion claim are in the diff"
+    # Grounding is computed per claim at completion time, not stored, so this
+    # reports availability rather than inventing a stale result.
+    return UNKNOWN, "run `loki verify` to check the completion claim against the diff"
 
 
 def _model_line(loki_dir):
@@ -177,12 +161,7 @@ def render_markdown(v):
     out.append("| Signal | Status | Detail |")
     out.append("|---|---|---|")
     for r in v["signals"]:
-        # Three states, not two. Collapsing everything non-measured to UNKNOWN
-        # would render a real finding ("the claim names a file absent from the
-        # diff") identically to "we could not check" -- the false equivalence
-        # this whole module exists to refuse. Anything unrecognised still
-        # degrades to UNKNOWN, so an unmeasured signal can never read as a pass.
-        status = r["status"] if r["status"] in ("measured", "finding") else UNKNOWN
+        status = "measured" if r["status"] == "measured" else "UNKNOWN"
         out.append(f"| {r['signal']} | {status} | {r['detail']} |")
     out.append("")
     out.append("Every line is derived from a file in `.loki/` that you can read "

package/autonomy/loki

@@ -1044,11 +1044,11 @@ show_help() {
     echo "  import init intent"
     echo "  issue kpis logs magic mcp memory metrics migrate modernize monitor"
     echo "  next notify onboard open optimize otel outcomes own (handoff) pause plan preview"
-    echo "  projects proof (receipt) provider quick quickstart rc readiness remote report reset"
+    echo "  projects proof (receipt) provider quick quickstart rc remote report reset"
     echo "  resume review rollback run sandbox secrets secure self-update sentrux"
     echo "  serve setup-skill share ship spec start state stats status steer stop"
     echo "  syslog telemetry template test tour trigger trust trust-metrics"
-    echo "  ultracode update verdict verify version voice watch watchdog web welcome why"
+    echo "  ultracode update verify version voice watch watchdog web welcome why"
     echo "  wiki worktree (wt)"
     echo ""
     echo "Any command: loki <command> --help"
@@ -11905,14 +11905,7 @@ STALE_DAYS = 90
 p = os.environ['LOKI_CATALOG_PATH']
 try:
     updated = json.load(open(p))['updated']
-    # UTC, matching the --json path below (loki:12337) and the Bun route
-    # (doctor.ts:545). This is the TEXT-mode twin of that computation, and it was
-    # missed when the JSON one was fixed: the parity gate compares BOTH surfaces,
-    # so fixing only --json left doctor text-mode still diverging 5 vs 6 and the
-    # gate still red. Two copies of one calculation is the actual defect here;
-    # they are left as two only because the text path prints and the JSON path
-    # returns a dict.
-    age = (datetime.datetime.now(datetime.timezone.utc).date() - datetime.date.fromisoformat(updated)).days
+    age = (datetime.date.today() - datetime.date.fromisoformat(updated)).days
 except Exception:
     print('warn|Catalog unreadable or missing an ISO \"updated\" date -- cannot determine age')
 else:
@@ -12036,17 +12029,6 @@ else:
             local _blk_key="other"
             case "$_doctor_blockers" in
                 *"No AI provider CLI"*) _blk_key="no_provider" ;;
-                # This doctor DETECTS the logged-out / expired wall above
-                # (:11610, :11615) but had no arm for it, so the single most
-                # common post-install failure was reported as `other` -- the one
-                # bucket that cannot be acted on. not_logged_in is already a
-                # first-class enum value (telemetry.sh:204) and the Bun doctor
-                # already reports it, so without this arm the SAME host answered
-                # `other` on bash and `not_logged_in` on Bun and the two routes'
-                # counts could not be added together. Ordered directly after
-                # no_provider because install-vs-authenticate need opposite fixes
-                # and a missing provider is the earlier wall.
-                *"not logged in"*|*"login has expired"*) _blk_key="not_logged_in" ;;
                 *"Node.js is not installed"*|*"Node.js must be"*) _blk_key="node" ;;
                 *"Python 3 is not installed"*|*"Python 3 must be"*) _blk_key="python3" ;;
                 *"jq is not installed"*) _blk_key="jq" ;;
@@ -12330,18 +12312,7 @@ _cat_path = os.environ['LOKI_CATALOG_PATH']
 try:
     import datetime as _dt
     _cat_updated = json.load(open(_cat_path))['updated']
-    # UTC, not local. date.today() is the HOST's date, while the Bun route parses
-    # both sides at UTC midnight (doctor.ts:545 says so, to keep the day count
-    # from shifting with the host timezone). Anywhere west of UTC the two
-    # disagree for the hours between local midnight and UTC midnight, and
-    # doctor --json is compared BYTE FOR BYTE between routes by bun-parity.
-    #
-    # Measured on this host at 02:45 UTC / 22:45 local: bash reported age_days 5
-    # and Bun reported 6, from the same file at the same instant. It reproduced
-    # across two full gate runs hours apart, so it is not a midnight-rollover
-    # flake. It is a real parity defect that stays invisible while the local date
-    # and the UTC date agree, which is most of the day.
-    _cat_age = (_dt.datetime.now(_dt.timezone.utc).date() - _dt.date.fromisoformat(_cat_updated)).days
+    _cat_age = (_dt.date.today() - _dt.date.fromisoformat(_cat_updated)).days
     if _cat_age > CATALOG_STALE_DAYS:
         # Deliberately NOT counted. The block comment above states catalog age is
         # excluded from pass/fail/warn and from 'ok' so a stale catalog can never
@@ -13803,76 +13774,12 @@ set_ttfv_lightweight_profile() {
 # never pollutes the v7.8.1 generated-PRD-reuse signature logic. The brief text
 # is the project intent; the rest is a minimal scaffold the agent fills in.
 # Usage: synthesize_brief_prd <output_file> <brief_text>
-# _brief_acceptance_criteria <brief_text>: acceptance criteria derived from what
-# the user ACTUALLY asked for, not constants.
-#
-# WHY. Every one-liner used to get byte-identical Requirements and Success
-# Criteria: "build a todo app" and "build a Stripe billing dashboard" produced
-# the same acceptance criteria, and the user's own words appeared exactly once,
-# under Overview. So the completion council, the checklist and the evidence gate
-# were all checking generic prose rather than the request. That is the weakest
-# input shape getting the least specific help, which is backwards -- a cheap
-# model's output quality depends more on how precisely the target is stated than
-# on the model.
-#
-# DETERMINISTIC ON PURPOSE. No model call: this runs before a provider is even
-# selected, must work with no API key, and must not add latency or cost to the
-# first thing a new user does. It is keyword-to-obligation mapping, which is
-# honest about being shallow -- it turns stated nouns into checkable lines and
-# claims nothing about intent it cannot see. Anything cleverer belongs in the
-# spec-interrogation grill, which already runs after this and does call a model.
-_brief_acceptance_criteria() {
-    local t
-    t="$(printf '%s' "${1:-}" | tr '[:upper:]' '[:lower:]')"
-    local out=""
-    _bac() { out="${out}- ${1}"$'\n'; }
-
-    # Persistence. The single most common churn report is "I submitted the form
-    # and nothing happened", so a stated store or form becomes an explicit
-    # survives-a-reload obligation rather than an implied one.
-    case "$t" in
-        *save*|*persist*|*store*|*databas*|*crud*|*todo*|*note*|*task*|*record*)
-            _bac "Data the user creates survives a page reload and a server restart (it is written to a real store, not held in memory)." ;;
-    esac
-    case "$t" in
-        *form*|*submit*|*signup*|*"sign up"*|*contact*|*upload*|*checkout*)
-            _bac "Every form actually submits: the happy path writes real data and the user sees a confirmation, and a validation failure shows an inline error." ;;
-    esac
-    case "$t" in
-        *auth*|*login*|*"log in"*|*"sign in"*|*account*|*user*|*password*|*session*)
-            _bac "Authentication works end to end: a real signup, a real login, and a protected route that returns 401 when logged out." ;;
-    esac
-    case "$t" in
-        *api*|*endpoint*|*rest*|*graphql*|*backend*|*server*)
-            _bac "Each endpoint returns real data with correct status codes, and is callable with curl without a browser." ;;
-    esac
-    case "$t" in
-        *payment*|*stripe*|*billing*|*subscription*|*checkout*|*invoice*)
-            _bac "The payment path is wired to the provider's test mode and a test transaction completes; no mocked charge stands in for the integration." ;;
-    esac
-    case "$t" in
-        *search*|*filter*|*sort*)
-            _bac "Search or filtering queries the real dataset and returns different results for different inputs." ;;
-    esac
-    case "$t" in
-        *dashboard*|*chart*|*graph*|*analytic*|*report*|*metric*)
-            _bac "Every figure shown traces to a real query. No hardcoded sample numbers." ;;
-    esac
-    case "$t" in
-        *page*|*landing*|*site*|*website*|*ui*|*app*|*frontend*)
-            _bac "The page renders with real content, no lorem ipsum, and no placeholder image where a real asset belongs." ;;
-    esac
-    printf '%s' "$out"
-}
-
 synthesize_brief_prd() {
     local out_file="$1"
     local brief_text="$2"
     local out_dir
     out_dir="$(dirname "$out_file")"
     mkdir -p "$out_dir" 2>/dev/null || true
-    local _derived
-    _derived="$(_brief_acceptance_criteria "$brief_text")"
     cat > "$out_file" << BRIEFEOF
 # Project Brief
 
@@ -13888,7 +13795,6 @@ $brief_text
 ## Success Criteria
 - A user can run the result and observe the core behavior described above.
 - No errors on a clean start; the happy path works end to end.
-${_derived}
 
 ## Constraints
 - This is a fast first pass (zero-config first run). Keep scope tight.
@@ -13990,10 +13896,6 @@ cmd_quick() {
     find "$LOKI_DIR" -maxdepth 1 -name 'quick-prd-*.md' -mtime +1 -delete 2>/dev/null || true
     find "$LOKI_DIR" -maxdepth 1 -name 'brief-prd-*.md' -mtime +1 -delete 2>/dev/null || true
     local quick_prd="$LOKI_DIR/quick-prd-$$.md"
-    # Same spec-derived criteria as the brief path: a quick task was getting the
-    # identical generic Success Criteria regardless of what was asked.
-    local _derived
-    _derived="$(_brief_acceptance_criteria "$task_desc")"
     cat > "$quick_prd" << QPRDEOF
 # Quick Task
 
@@ -14010,7 +13912,6 @@ $task_desc
 - Task is completed as described
 - No errors or regressions introduced
 - Code follows project conventions
-${_derived}
 
 ## Constraints
 - This is a quick single-task execution
@@ -19475,15 +19376,6 @@ main() {
             # happened to it afterwards -- reverted, reworked, or survived.
             cmd_outcomes "$@"
             ;;
-        verdict)
-            # All five trust signals in one block. The individual measurements
-            # existed and were unreachable; this is the surface a reviewer reads.
-            cmd_verdict "$@"
-            ;;
-        readiness)
-            # Can an agent verify its own work here? Measured, not LLM-scored.
-            cmd_readiness "$@"
-            ;;
         secure)
             # Secure-by-default gate surface: inspect findings + manage waivers.
             cmd_secure "$@"
@@ -30959,84 +30851,6 @@ cmd_outcomes() {
     LOKI_DIR="${LOKI_DIR:-.loki}" python3 "$lib" "$@"
 }
 
-# loki verdict: the five measured trust signals, in one block a reviewer reads
-# in ten seconds.
-#
-# WHY A COMMAND AND NOT JUST A LIBRARY. We measure five things nobody else does
-# -- did the work survive, does the spec still match intent, was this the agent
-# or a human rescue, does the completion claim name real work, which model
-# decided -- and each was correct, tested, and unreachable. A moat nobody can
-# see is not a moat. This is the surface.
-#
-# There is deliberately NO composite score. Averaging a revert count, a hash
-# comparison, a diff hash, a path match and a model id yields a number whose
-# movement nobody can explain, which is what competitors already ship. UNKNOWN
-# is PRINTED, never suppressed: a reviewer must tell "we checked and it is fine"
-# from "we could not check". See autonomy/lib/verdict.py.
-cmd_verdict() {
-    local lib="${_LOKI_SCRIPT_DIR}/lib/verdict.py"
-    if [ ! -f "$lib" ]; then
-        echo "verdict renderer is not installed at $lib" >&2
-        return 2
-    fi
-    case "${1:-}" in
-        --help|-h|help)
-            echo -e "${BOLD}loki verdict${NC} - the five measured trust signals, in one readable block"
-            echo ""
-            echo "Usage: loki verdict [--json]"
-            echo ""
-            echo "Prints one line each for outcome, intent, authorship, grounding and"
-            echo "model: what survived, whether the spec still matches intent, whether"
-            echo "this was the agent or a human rescue, whether the completion claim"
-            echo "named real work, and which model decided."
-            echo ""
-            echo "Every line is either a measured fact or an explicit UNKNOWN. There is"
-            echo "no composite score: a number nobody can explain is not evidence."
-            echo "Read-only: it never writes to the repo it analyses."
-            return 0
-            ;;
-    esac
-    LOKI_DIR="${LOKI_DIR:-.loki}" python3 "$lib" "$@"
-}
-
-# loki readiness: can an autonomous agent verify its own work in THIS repo?
-#
-# WHY THIS IS NOT A COPY OF FACTORY AI'S AGENT READINESS MODEL. Theirs is
-# LLM-scored -- their report objects record modelUsed and reasoningEffort, so
-# the number is a model's opinion and two runs can disagree about the same
-# commit. Every criterion here is a file that exists or does not, a command
-# present or absent: same commit, same answer, every machine, no key, no spend.
-#
-# No percentage and no letter grade. A composite invites ranking, ranking
-# invites gaming, and the individual signals are the actionable part -- "there
-# is no test command" tells you what to do, "readiness 62%" does not. Criteria
-# that cannot be determined report UNKNOWN by name rather than counting as
-# failures. See autonomy/lib/agent_readiness.py.
-cmd_readiness() {
-    local lib="${_LOKI_SCRIPT_DIR}/lib/agent_readiness.py"
-    if [ ! -f "$lib" ]; then
-        echo "readiness assessor is not installed at $lib" >&2
-        return 2
-    fi
-    case "${1:-}" in
-        --help|-h|help)
-            echo -e "${BOLD}loki readiness${NC} - can an agent verify its own work in this repo?"
-            echo ""
-            echo "Usage: loki readiness [path] [--json]"
-            echo ""
-            echo "Measures whether this repo gives an agent a way to check itself: a"
-            echo "test command, a build, CI config, typed sources, a lockfile. Not"
-            echo "general code quality -- the narrower question every agent depends on."
-            echo ""
-            echo "Deterministic: no model, no key, no spend. Same commit, same answer."
-            echo "Criteria that cannot be determined report UNKNOWN rather than failing."
-            echo "Read-only: it never writes to the repo it analyses."
-            return 0
-            ;;
-    esac
-    python3 "$lib" "$@"
-}
-
 cmd_wiki() {
     local subcmd="${1:-}"
     shift 2>/dev/null || true