loki-mode 7.104.5 → 7.106.0

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 v7.104.5
+# Loki Mode v7.106.0
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -408,4 +408,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v7.104.5 | [Autonomi](https://www.autonomi.dev/) flagship product | ~260 lines core**
+**v7.106.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~260 lines core**

package/VERSION

@@ -1 +1 @@
-7.104.5
+7.106.0

package/autonomy/completion-council.sh

@@ -1517,6 +1517,112 @@ HELDOUT_EOF
     return 0
 }
 
+# v7.106.0: reverse-classical test provenance.
+# FrontierCode insight: a NEW test only proves the change if it FAILS on the
+# pre-change base and PASSES on HEAD. A test that passes on BOTH is tautological
+# (or was written to pass) and proves nothing -- it must NOT count as affirmative
+# "tests green" evidence. This helper runs the run's NEW/CHANGED test files
+# against the run-start base in an ISOLATED git worktree (NEVER touching the live
+# tree -- the self-delete/self-mutate hazard) and classifies the provenance.
+#
+# INVARIANT (safety): this only ever DOWNGRADES unearned affirmative -> inconclusive.
+# It NEVER grants affirmative, and NEVER turns inconclusive into a block. Any
+# failure to establish provenance (no base, no git, no test files, scratch
+# materialization error, runner error) returns "unknown" and the caller keeps the
+# existing behavior -- fail OPEN toward the current gate, never a false block.
+#
+# Echoes exactly one token:
+#   tautological-> the FULL suite (incl. the new/changed tests) PASSES on base, so
+#                  the new tests added no failing signal on the old code = they
+#                  prove nothing = downgrade to inconclusive.
+#   unknown     -> cannot determine, so caller no-ops (keeps existing behavior).
+#                  Covers: no base/git/tests, materialization/runner error, AND a
+#                  base-suite FAILURE (rc != 0) -- because a whole-suite failure
+#                  cannot be attributed to the new test specifically, we do NOT
+#                  claim "confirmed provenance" from it (that would be a false
+#                  signal). So there is deliberately NO affirmative-granting token:
+#                  this helper can only ever DOWNGRADE, never confirm.
+# Opt out with LOKI_TEST_PROVENANCE=0 (returns unknown immediately).
+_loki_test_provenance() {
+    [ "${LOKI_TEST_PROVENANCE:-1}" = "0" ] && { echo "unknown"; return 0; }
+    local base_sha="$1" test_cmd="$2"
+    local target="${TARGET_DIR:-.}"
+    [ -n "$base_sha" ] || { echo "unknown"; return 0; }
+    command -v git >/dev/null 2>&1 || { echo "unknown"; return 0; }
+    ( cd "$target" 2>/dev/null && git rev-parse --is-inside-work-tree >/dev/null 2>&1 ) || { echo "unknown"; return 0; }
+    ( cd "$target" 2>/dev/null && git cat-file -e "${base_sha}^{commit}" 2>/dev/null ) || { echo "unknown"; return 0; }
+
+    # Identify NEW/CHANGED test files in the diff vs base (committed + working tree).
+    # Heuristic test-path match across common ecosystems; conservative (only files
+    # that look like tests). If none, provenance is not applicable -> unknown.
+    local test_files
+    test_files=$( cd "$target" 2>/dev/null && {
+        git diff --name-only "$base_sha" -- 2>/dev/null
+        git ls-files --others --exclude-standard 2>/dev/null
+    } | LC_ALL=C sort -u | grep -iE '(^|/)(test|tests|spec|__tests__)/|(\.|_|-)(test|spec)\.[a-z0-9]+$|_test\.[a-z0-9]+$|test_[^/]*\.py$' || true )
+    [ -n "$test_files" ] || { echo "unknown"; return 0; }
+
+    # Materialize base in an ISOLATED worktree; never checkout in the live tree.
+    local scratch; scratch="$(mktemp -d "${TMPDIR:-/tmp}/loki-prov-XXXXXX" 2>/dev/null)" || { echo "unknown"; return 0; }
+    local wt="$scratch/base"
+    if ! ( cd "$target" && git worktree add --detach --quiet "$wt" "$base_sha" 2>/dev/null ); then
+        rm -rf "$scratch" 2>/dev/null; echo "unknown"; return 0
+    fi
+
+    # Inject the current NEW/CHANGED test files into the base worktree so we run
+    # the AGENT'S tests against the OLD code. Copy from the live tree (HEAD/working
+    # version of each test file). Missing source (deleted test) is skipped.
+    local f
+    while IFS= read -r f; do
+        [ -n "$f" ] || continue
+        [ -f "$target/$f" ] || continue
+        mkdir -p "$wt/$(dirname "$f")" 2>/dev/null || true
+        cp -f "$target/$f" "$wt/$f" 2>/dev/null || true
+    done <<< "$test_files"
+
+    # Run the recorded test command in the base worktree, bounded. If the runner
+    # cannot run at all (missing deps on base, command error distinct from a test
+    # failure), we cannot conclude tautological -> unknown (fail open).
+    #
+    # LIVENESS SAFETY: the base run MUST be bounded, else a base suite that hangs
+    # (e.g. waits on a service/dep the new code introduced) would stall the
+    # completion-time evidence gate. If no timeout binary is available (stock
+    # macOS with no coreutils), we do NOT run the base suite unbounded -- we skip
+    # and return unknown (a safe no-op that keeps the existing affirmative-or-
+    # council behavior). This trades a provenance check we cannot run safely for
+    # guaranteed liveness; it can never cause false-green or false-block.
+    local rc timeout_bin=""
+    command -v gtimeout >/dev/null 2>&1 && timeout_bin="gtimeout"
+    [ -z "$timeout_bin" ] && command -v timeout >/dev/null 2>&1 && timeout_bin="timeout"
+    if [ -z "$timeout_bin" ]; then
+        ( cd "$target" && git worktree remove --force "$wt" 2>/dev/null ); rm -rf "$scratch" 2>/dev/null
+        echo "unknown"; return 0
+    fi
+    ( cd "$wt" && eval "$timeout_bin 300 $test_cmd" >/dev/null 2>&1 )
+    rc=$?
+
+    ( cd "$target" && git worktree remove --force "$wt" 2>/dev/null ); rm -rf "$scratch" 2>/dev/null
+
+    # Interpretation (deliberately conservative -- the runner runs the WHOLE test
+    # command, not only the injected tests, so a base failure CANNOT be attributed
+    # to the new test specifically: the base suite may fail for unrelated reasons,
+    # or because the feature's non-test code is absent):
+    #   rc == 0  -> the FULL suite (incl. the injected new tests) passes on the
+    #               OLD code. The new tests therefore added NO failing signal on
+    #               base = tautological. This is the ONLY safe, actionable
+    #               conclusion, and it only ever DOWNGRADES affirmative.
+    #   rc != 0  -> UNKNOWN. We will not claim "confirmed provenance" from a
+    #               whole-suite failure we cannot attribute to the new test (that
+    #               would be a meaningless/false signal). Caller no-ops (keeps the
+    #               existing affirmative-or-council behavior). Timeout (124) is
+    #               likewise unknown.
+    # This keeps the feature strictly a safe downgrade: it can turn an unearned
+    # green into inconclusive, and can never fabricate a "provenance-confirmed"
+    # affirmative.
+    if [ "$rc" = "0" ]; then echo "tautological"; else echo "unknown"; fi
+    return 0
+}
+
 #===============================================================================
 # Council Evidence Hard Gate (v7.19.1) - "verified completion"
 #===============================================================================
@@ -1687,6 +1793,56 @@ else:
         fi
     fi
 
+    # --- v7.106.0: reverse-classical test provenance ---------------------------
+    # A green test suite is only affirmative evidence if the run's NEW/CHANGED
+    # tests actually FAIL on the pre-change base (they exercise the change). Tests
+    # that pass on both base and HEAD are tautological and prove nothing. When the
+    # tests are an affirmative candidate (they ran and passed, and are not already
+    # inconclusive), verify provenance against the base in an ISOLATED worktree.
+    # SAFETY: this can only DOWNGRADE affirmative -> inconclusive; "confirmed" and
+    # "unknown" both leave the existing state untouched, and inconclusive is
+    # pass-through (never a block). No-op on greenfield (unknown: no base/tests)
+    # and when LOKI_TEST_PROVENANCE=0.
+    if [ "$test_fails" != "true" ] && [ "$test_inconclusive" != "true" ] \
+       && [ "$test_pass" = "true" ] && [ "$test_runner" != "none" ] \
+       && [ -n "$base_sha" ]; then
+        local _prov_cmd="${LOKI_TEST_COMMAND:-}"
+        # Fall back to a sensible per-runner default when no explicit command was set.
+        # For pytest, prefer python3 (many machines have only python3, not python);
+        # a missing interpreter just makes the base run error -> unknown -> safe no-op.
+        if [ -z "$_prov_cmd" ]; then
+            case "$test_runner" in
+                pytest)
+                    if command -v python3 >/dev/null 2>&1; then _prov_cmd="python3 -m pytest -q"
+                    else _prov_cmd="python -m pytest -q"; fi ;;
+                node-test) _prov_cmd="node --test" ;;
+                *) _prov_cmd="npm test" ;;
+            esac
+        fi
+        local _prov
+        _prov="$(_loki_test_provenance "$base_sha" "$_prov_cmd")"
+        if [ "$_prov" = "tautological" ]; then
+            # The new/changed tests pass on the OLD code too -> they prove nothing
+            # about the change. DOWNGRADE the test signal from affirmative to
+            # inconclusive (pass-through; the diff axis + council still decide, so
+            # a legitimate refactor/greenfield-in-existing-repo with a non-empty
+            # diff still completes -- inconclusive never blocks).
+            test_inconclusive="true"
+            test_inconclusive_reason="test_provenance_unconfirmed"
+            log_info "[Evidence] New/changed tests pass on the run-start base too (tautological); test signal downgraded to inconclusive, not affirmative"
+            mkdir -p "${TARGET_DIR:-.}/.loki/state" 2>/dev/null || true
+            printf '{"status":"tautological","reason":"new/changed tests pass on the run-start base; not affirmative provenance","base_sha":"%s","runner":"%s"}' \
+                "$base_sha" "$test_runner" > "${TARGET_DIR:-.}/.loki/state/test-provenance.json" 2>/dev/null || true
+            if command -v emit_event_json >/dev/null 2>&1; then
+                emit_event_json "test_provenance" "status" "tautological" "runner" "$test_runner" 2>/dev/null || true
+            fi
+        fi
+        # _prov == "unknown" -> no-op: greenfield (no base/tests), or a base-suite
+        # failure we cannot attribute to the new test. We keep the existing
+        # affirmative-or-council behavior; we never fabricate a "confirmed" from an
+        # unattributable base failure.
+    fi
+
     # --- v7.28.0: inconclusive-baseline lifecycle -------------------------------
     # When the gate cannot establish a diff baseline (no git repo, or no run-start
     # SHA) it does NOT block (would break non-git projects), but completion is no
@@ -3226,10 +3382,36 @@ council_should_stop() {
         circuit_triggered=true
     fi
 
-    # Only run council at check intervals OR if circuit breaker triggered
+    # Only run council at check intervals OR if circuit breaker triggered OR the
+    # agent has EXPLICITLY claimed completion this iteration (v7.105.0 convergence
+    # fix). Rationale, measured from real build telemetry (docs/SPEED-DIAGNOSIS):
+    # a build could be verifiably done at iteration 1 (reviews non-blocking, tests
+    # green, checklist complete) yet the council was not even allowed to evaluate
+    # until the next 5-iteration boundary -- so the loop ground out "next
+    # improvement" iterations the user never asked for (14 iterations for a job
+    # done in ~1). An explicit completion CLAIM is the strongest possible signal
+    # that it is worth asking the council NOW. This changes only WHEN the council
+    # evaluates, never WHETHER it must approve: MIN_ITERATIONS below, the hard
+    # checklist gate, the evidence gate, the aggregate vote, and the devil's
+    # advocate all still run unchanged. A premature or unearned claim is still
+    # rejected (returns 1, loop continues); it is never a forced green.
+    #
+    # The claim signal is the structured one (loki_complete_task MCP tool ->
+    # .loki/signals/COMPLETION_REQUESTED, or the runner exporting
+    # LOKI_COMPLETION_CLAIMED=1 for this iteration), NOT the fuzzy log-grep the
+    # circuit breaker uses -- so it fires on a real, intentional claim only.
+    local completion_claimed=false
+    if [ "${LOKI_COMPLETION_CLAIMED:-0}" = "1" ] \
+       || [ -f "${TARGET_DIR:-.}/.loki/signals/COMPLETION_REQUESTED" ]; then
+        completion_claimed=true
+    fi
+
     local should_check=false
     if [ "$circuit_triggered" = "true" ]; then
         should_check=true
+    elif [ "$completion_claimed" = "true" ]; then
+        should_check=true
+        log_info "[Council] Completion claimed this iteration -- evaluating now (not deferring to the ${COUNCIL_CHECK_INTERVAL}-iteration interval)"
     elif [ $((ITERATION_COUNT % COUNCIL_CHECK_INTERVAL)) -eq 0 ]; then
         should_check=true
     fi

package/autonomy/run.sh

@@ -17530,7 +17530,30 @@ else:
             if type ensure_completion_test_evidence &>/dev/null; then
                 ensure_completion_test_evidence || true
             fi
-            if type council_should_stop &>/dev/null && council_should_stop; then
+            # v7.105.0 convergence: if the agent EXPLICITLY claimed completion this
+            # iteration (structured loki_complete_task / COMPLETION_REQUESTED
+            # signal), tell the council to evaluate NOW instead of deferring to the
+            # 5-iteration check interval. The council still must approve (MIN_
+            # ITERATIONS + hard checklist + evidence gate + vote + devil's advocate
+            # all unchanged); this only stops an already-verified build from
+            # grinding needless "next improvement" iterations. Scoped to this call.
+            #
+            # CRITICAL (council-review finding): we must PEEK at the claim signal
+            # NON-DESTRUCTIVELY. check_task_completion_signal() CONSUMES the signal
+            # (rm -f on read), and the DEFAULT completion-promise route below
+            # (v6.82.0, the live gate-based acceptance path) also consumes it -- so
+            # calling the consuming detector here would drop the claim before that
+            # route sees it, re-introducing the v7.28 "claim drop" bug. Instead we
+            # only test the signal FILES' existence (the same two paths
+            # check_task_completion_signal reads), leaving consumption to the
+            # existing single owner. The council side likewise peeks
+            # ($TARGET_DIR/.loki/signals/COMPLETION_REQUESTED) without consuming.
+            local _loki_completion_claimed=0
+            if [ -f "${TARGET_DIR:-.}/.loki/signals/TASK_COMPLETION_CLAIMED" ] \
+               || [ -f "${TARGET_DIR:-.}/.loki/signals/COMPLETION_REQUESTED" ]; then
+                _loki_completion_claimed=1
+            fi
+            if type council_should_stop &>/dev/null && LOKI_COMPLETION_CLAIMED="$_loki_completion_claimed" council_should_stop; then
                 # bash-F1: council_should_stop returns 0 from a genuine approval
                 # AND from two force-stop safety valves (stagnation flood /
                 # repeated done-signals). A force-stop is NOT a verified-complete

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "7.104.5"
+__version__ = "7.106.0"
 
 # Expose the control app for easy import
 try:

package/docs/FOUNDER-STATUS-AND-DECISIONS-2026-07-01.md

@@ -0,0 +1,106 @@
+# Founder Status + Decision List (2026-07-01)
+
+The consolidated "where everything stands + the decisions only you can make" report.
+Written after: 5 patch releases shipped, top-100 backlog enumerated + triaged, ecosystem
+currency assessed, no-HITL product principle recorded.
+
+## 1. Shipped this session (all live on npm/Docker/GH, council-gated)
+
+| Version | What | Verified |
+|---|---|---|
+| v7.104.2 | Native/Keychain Claude login recognized; `loki start` no longer false-blocks a logged-in user; `loki doctor` confirms login (bash+Bun parity) | reproduced + 8-case test + live Docker |
+| v7.104.3 | Task-list accuracy: honest done column, no empty cards / dups / fake-green (failed iters honestly labeled) | live vs real anonima, browser screenshots, council 3/3 |
+| v7.104.4 | `/api/tasks` never 500s on malformed dashboard-state.json | RED/GREEN-proven test |
+| v7.104.5 | Memory panel correct + self-consistent on JSON-backed projects (completes PR #178, @iizotov co-authored, PR closed with thank-you) | live vs anonima, parity tests |
+| (earlier) v7.104.2 auth + v7.104.3 task-list | (same as above; two-patch train) | |
+
+Also: closed issue #177 (spam advisory). PR #178 merged-in-substance + contributor thanked.
+
+## 2. Backlog reality (honest)
+
+The top-100 was enumerated (116 raw -> 100 ranked) and then TRIAGED against current source.
+Key finding: **the autonomous critical/high queue is essentially ALREADY DONE.**
+
+| Rank | Item | Triage verdict |
+|---|---|---|
+| 1 | run.sh self-delete EXIT trap | already fixed (guarded to /tmp/loki-run-*; verified safe) |
+| 2 | saas cross-tenant leak + stuck-Building | CLOSED (commits a7667d0/44f7c29/8b05b36; 187/187 tests) |
+| 9 | anti-fake-green + gate parity | shipped (verify 282 tests; coherentStatus guard) |
+| 10 | VS-7 worker cutover | CLOSED (commit e6b96ed; BFF 181 + worker 102 tests) |
+| 17 | base_unreachable suppresses empty_diff | already wired (completion-council.sh:1616) |
+| 22 | 3 parallel reviewers | already implemented (run.sh:10461-10954) |
+
+The autonomous well is mostly dry BECAUSE the real remaining leverage is founder-gated
+(the trust-as-product / enterprise surface). That is the honest headline: the next 10x is
+not another patch, it is the decisions below.
+
+## 3. Ecosystem currency (what loki CONSUMES from the Claude ecosystem)
+
+CLI invocation contract (claude 2.1.198): STABLE, no drift. Auth: fixed this session.
+Model catalog: current (opus-4-8, sonnet-5, haiku-4-5, fable-5). Real deltas are small and
+tracked in docs/ECOSYSTEM-CURRENCY-PLAN.json (EC-1..EC-4). Adjacent Anthropic PRODUCTS
+(Managed Agents, Claude Tag, loop-engineering, agent identity) assessed as peer/inspiration
+categories, NOT dependencies -- declined with reason (the CTO triage move).
+
+The one real net-new build from this: **EC-1, the external-tool contract adapter + drift
+test** -- the keychain incident this session proves it would catch a user-facing break
+before a user does. Building autonomously now.
+
+## 4. DECISIONS ONLY YOU CAN MAKE (founder-gated, prepped to one-approval-away)
+
+These are the real blockers to "easiest to integrate with anyone" + the trust-as-product moat.
+Per the no-HITL principle, these are the LEGITIMATE exceptions: irreversible / strategic /
+brand / legal / spend decisions a human must own. Each is ready for a one-word go.
+
+### A. Autonomi Verify productization (the trust-as-product core)
+1. **License decision (rank 5, S, BLOCKS everything commercial).** BUSL-1.1 vs open for
+   @autonomi/verify. No hosted/commercial launch until decided. DECISION: which license?
+2. **npm registry + package name (rank 4, M).** @autonomi/verify is currently unpublishable
+   (private:true, no exports/types/build, .ts bin). And a naming collision exists
+   (loki-vaas vs loki-verify bin vs `loki verify` subcommand, rank 13). DECISION: publish to
+   public npm as `@autonomi/verify`? Confirm the brand/bin name to resolve the collision.
+   -> Once decided, I can make it genuinely npm-installable + fix the collision autonomously.
+3. **Hosted /verify deploy (rank 6, M + rank 36/42 sandbox).** The signed neutral REST
+   endpoint is the category-defining asset but not deployed (Bun.serve-only, ephemeral key).
+   DECISION: deploy target (Fly/Render/AWS/your infra) + signing-key storage (secret manager)?
+   -> I can stage the Node entry + Dockerfile + serve script + 12-factor key guard to
+   one-command-deploy; I will NOT deploy or spend without your go.
+4. **wire verifyCompletion() pipeline (rank 3, XL).** The SDK/MCP verdict path is stubbed;
+   the whole value prop is unreachable until gate->integrity->council->cost->sign is wired.
+   Depends on (2)/(3) decisions. Large; I can start once the license + registry are set.
+
+### B. Enterprise scale (needed for any enterprise sale)
+5. **RBAC / SSO / API-key auth / metering (rank 15, XL)** + **on-prem persistence backend
+   (rank 16, L, Postgres/SQLite)** + **sandbox backend (rank 14, L, gVisor/Firecracker/Kata)**.
+   DECISION: which enterprise deploy shape do we target first (hosted multi-tenant vs on-prem)?
+   That choice sequences 14/15/16/36/42. -> I can build the PersistenceBackend interface +
+   RBAC scaffolding once you pick the shape.
+6. **GitHub App verification integration (rank 46, L, greenfield).** DECISION: build now or later?
+
+### C. Benchmark spend
+7. **SWE-bench Pro 119 resume (issue #174, paused 35/119).** Resume harness exists. DECISION:
+   spend the compute to finish? (This is a $ + time decision, hence gated.)
+
+### D. Outreach (rank/issue #172, low-risk but external-posting)
+8. MCP registry submission + Glama claim + awesome-list PRs. I can prepare the submissions;
+   the actual external POSTS need your nod (per action rules). DECISION: proceed with posts?
+
+## 5. What I am doing autonomously RIGHT NOW (no decision needed)
+- EC-1 external-contract adapter + drift test (the one real net-new build).
+- Small no-HITL audit: find every place the PRODUCT hard-stops/prompts a human; fix genuine
+  gaps so it decides from model+memory+cross-project context (degrading to honest-inconclusive,
+  never fake-green). Not a speculative refactor -- a triage + targeted fixes.
+- NOT building: dashboard steering (#44) -- it is a human-in-the-loop feature, deprioritized by
+  the no-HITL mandate; architect plan is durable and waits. NOT building "autonomous-everywhere"
+  as a speculative refactor.
+
+## 6. The one-line asks
+Reply with any subset:
+- **License:** BUSL / MIT / Apache / other for @autonomi/verify
+- **npm:** publish `@autonomi/verify` publicly? + confirm bin name (resolve loki-verify collision)
+- **Hosted deploy:** target + key storage (or "not yet")
+- **Enterprise shape:** hosted-multitenant-first vs on-prem-first (sequences the enterprise stories)
+- **SWE-bench:** spend to finish 119? (yes/no)
+- **Outreach:** post the MCP-registry/Glama/awesome-list submissions? (yes/no)
+
+Everything else, I continue autonomously.

package/docs/INSTALLATION.md

@@ -2,7 +2,7 @@
 
 The flagship product of [Autonomi](https://www.autonomi.dev/). Loki Mode is a spec-driven autonomous builder with a built-in trust layer that takes any spec to a deployed product and verifies completion with evidence (quality gates plus a completion council), not just a "done" claim. Complete installation instructions for all platforms and use cases.
 
-**Version:** v7.104.5
+**Version:** v7.106.0
 
 ---
 

package/docs/SPEED-DIAGNOSIS-2026-07-01.md

@@ -0,0 +1,101 @@
+# Loki Build Speed Diagnosis (2026-07-01)
+
+Measured from real build telemetry (anonima `.loki/events.jsonl`, stage_complete +
+iteration timeline), NOT from reading run.sh. Method per advisor: measure before optimizing.
+
+## Data hygiene
+anonima's events.jsonl contained **9 interleaved builds/sessions** (9 session_starts) plus
+session-long manual poking. Naive sums ("5 min/iter x 14") are contaminated. Split on
+session_start and analyzed the real ones.
+
+## The pathological build (session 3): the smoking gun
+A simple topic-based chat app. 14 ACT iterations, ~97 min of real work (idle gaps excluded).
+
+| iteration | real work |
+|---|---|
+| 1 | 2.1 min |
+| 6 | 9.9 min |
+| 10 | 15.3 min |
+| 12 | **29.6 min** |
+| (others) | 3-6 min each |
+
+**13 completion claims across 14 iterations.** The FIRST claim already stated:
+"All PRD requirements implemented and tests passing. Evidence: npm test => 19/19 pass...
+PRD checklist (14 items) fully implemented." Yet the loop ran 13 more iterations.
+
+## Root cause: NON-CONVERGENCE, not per-stage slowness
+- **Code review / 3-reviewer council is NOT the bottleneck** (0.1-2.8 min per iteration,
+  mostly <1 min). Cutting the council would harm accuracy for ~no speed gain. Ruled out.
+- **The bottleneck is iteration COUNT + a few pathological iterations.** The app was
+  essentially done early (real passing tests, checklist complete) but the RARV loop's
+  "there is NEVER a finished state -- always find the next improvement" behavior kept it
+  grinding polish iterations. For a user who wants a fullstack app fast, this is THE
+  frustration: loki keeps going after done.
+- iter 10 (15 min) and iter 12 (30 min) did disproportionate work -- candidate over-scoping
+  or re-analysis; needs per-iteration content inspection.
+
+## Buckets (advisor framework)
+1. **WASTE (cut freely):** polish iterations after genuine completion; the PRD-reuse
+   spurious-update re-reconciliation (already diagnosed, design fix pending). These add
+   wall-clock with no trust value.
+2. **VERIFICATION COST (keep -- this is the product):** council, gates, completion vote.
+   Fast already. The moat. Do NOT cut for speed.
+3. **STALE SCAFFOLDING (test per-stage):** machinery built for weaker models may be dead
+   weight on Opus 4.8 / Sonnet 5 (cf. Anthropic's "context resets became dead weight on
+   Opus 4.5"). Hypothesis -- must be tested against the benchmark, not assumed.
+
+## The fix direction (speed AND accuracy, never speed-by-cutting-verification)
+**Convergence:** stop as soon as the completion council genuinely agrees it is done
+(honest verified completion), instead of running "next improvement" iterations. This is a
+SPEED win that INCREASES trust alignment (stop when verified-done = the moat working), not a
+verification cut. The user's mandate is "accuracy AND speed"; convergence delivers both.
+
+## HARD PREREQUISITE (before any council/RARV-C change): the benchmark
+Both the speed directive and the "update council/RARV-C to latest research, show before/after
+accuracy" directive require a **clean reproducible benchmark on a fixed spec** emitting
+wall-clock + iteration count + completion verdict + pass/fail vs known acceptance criteria.
+Without it, changing the verification core is unmeasurable churn on the moat (architecture-
+level fake-green). Build the benchmark FIRST, capture before-numbers, fix the clearest waste
+(convergence), re-run, show real before->after. Then gate any council/RARV-C change on that
+benchmark, one named change at a time, kept only if accuracy goes up and wall-clock does not
+regress.
+
+## MEASURED before/after (2026-07-01, same greet-CLI spec, isolated, council+gates ON)
+
+| metric | before v7.104.1 (no fix) | after v7.105.0 (convergence fix) |
+|---|---|---|
+| wall clock | 28.6 min | 7.4 min |
+| ACT iterations | 13 | 1 |
+| completion claims | 11 | 1 |
+| engine_completed | true | true |
+| greet.js built correctly | yes | yes |
+
+First after-run: 3.9x faster (-74% wall clock), 13->1 iterations, SAME correct output (accuracy
+preserved -- not speed-by-cutting-verification; the council still approved).
+
+HONEST CAVEATS (do not overclaim): (1) n=1 per side; LLM builds are stochastic, so a single
+1-iteration after-run could be a low draw. Running 2-3 more after-runs for a real range before
+presenting a headline multiplier. (2) before=v7.104.1 vs after=v7.105.0 differ by 7 commits;
+convergence is the only one plausibly affecting iteration count, so it is the likely driver, but
+the clean claim is "v7.105.0 vs v7.104.1 on this spec", not "convergence alone". (3) this is one
+trivial spec; larger specs may show a smaller relative gain.
+
+## UPDATE: after-run consistency (2 of 3 repeats in; n>1)
+- after #1: 7.4 min, 1 iteration, correct
+- after #2: 7.0 min, 1 iteration, correct
+Both after-runs converged at 1 ACT iteration / ~7 min / correct output -- consistent, NOT a
+single lucky draw. The convergence fix reliably stops at verified-done. Avg ~7.2 min vs 28.6 min
+before = ~4x faster, 13x fewer iterations, accuracy preserved (council still approves). The
+residual ~7 min is the single build+verify iteration's real inference cost (the honest floor for
+this spec WITH verified-completion). Next speed lever: per-iteration cost on larger specs (the
+fat-iteration investigation), measured via this same benchmark.
+
+## CONFIRMED (n=3 after-runs): the convergence fix delivers 4.0x, reliably
+- before (v7.104.1): 28.6 min, 13 iterations
+- after (v7.105.0): 7.4 / 7.0 / 7.1 min -> avg 7.2 min, EVERY run exactly 1 iteration, all correct
+- Result: 4.0x faster wall-clock, 13 -> 1 iterations, 100% correct across all 3 runs.
+The tight range (7.0-7.4 min, always 1 iteration) confirms this is the reliable behavior, not a
+single lucky draw. Attribution: v7.105.0 vs v7.104.1 on the greet-CLI spec; the convergence fix
+(completion-claim triggers immediate council evaluation) is the sole plausible driver of the
+13->1 iteration collapse. Residual ~7.2 min = one build+verify iteration's real inference cost
+(the honest floor WITH verified-completion). NEXT speed lever: per-iteration cost on larger specs.

package/docs/TOP-100-BACKLOG.md

@@ -0,0 +1,122 @@
+# Loki + Autonomi: Top-100 Backlog (Master Plan)
+
+Generated 2026-07-01 by the top100-backlog-enumeration ultracode workflow (7 agents, 116 raw items -> 100 ranked). North star: easiest to integrate with anyone; ranked by leverage toward zero-friction adoption, modularization, enterprise-grade, and scale.
+
+## Themes
+- Trust and correctness bugs (data-loss, cross-tenant, fake-green safety)
+- Autonomi Verify productization (stub -> shipped: gate/council/cost/sign wiring)
+- Zero-friction adoption (reduce user setup and error surface)
+- Modularization (extract libraries, break monoliths, fix import boundaries)
+- Enterprise-scale and multi-tenant (RBAC/SSO/persistence/sandbox, mostly founder-gated)
+- autonomi-saas evidence and proof pipeline (receipt phases, correlation, worker cutover)
+- CI, test, and distribution hardening (coverage, unreachable tests, publish gates)
+- Runtime-agnostic and Bun migration (bash sunset, write-path port)
+
+## Legend
+- **gate**: autonomous (I execute) | founder-gated (prep to one-approval-away) | community
+- **effort**: S/M/L/XL | **value**: critical/high/medium/low
+
+## Full ranked list
+
+| # | value | eff | gate | area | item |
+|---|---|---|---|---|---|
+| 1 | critical | S | autonomous | engine | run.sh self-delete EXIT trap nukes canonical source on LOKI_RUNNING_FROM_TEMP |
+| 2 | critical | M | autonomous | saas | autonomi-saas proof correlation: cross-tenant leak + stuck-Building fix (per-build run_id) |
+| 3 | critical | XL | founder-gated | verify | Verify: verifyCompletion() stub throws not-implemented (wire gate->integrity->council->cost->sign pipeline) |
+| 4 | critical | M | founder-gated | verify | Verify: npm @autonomi/verify unpublishable (private:true, no exports/types/build, .ts bin, undeclared zod) |
+| 5 | critical | S | founder-gated | saas | Verify: BUSL-1.1 vs open licensing undecided (blocks hosted/commercial launch) |
+| 6 | critical | M | founder-gated | saas | Verify: hosted /verify REST API not deployed (Node entry + Dockerfile + serve script + persistent signing key) |
+| 7 | critical | M | autonomous | modularization | Verify: extract verification-core library (port council_evidence_gate to TS) |
+| 8 | critical | S | autonomous | verify | Verify: 5 deterministic gates parity with completion-council.sh (staged, verify + ship) |
+| 9 | critical | S | autonomous | saas | autonomi-saas anti-fake-green safety: verdict strictly from engine honesty.headline (verify + guard) |
+| 10 | critical | M | autonomous | saas | autonomi-saas VS-7 worker cutover: BFF enqueues, worker sole executor (verify live e2e) |
+| 11 | high | M | founder-gated | verify | Signed neutral records on SDK/MCP path (verifyCompletion returns UNSIGNED_SENTINEL; only REST signs) |
+| 12 | medium | S | autonomous | verify | Verify: fix embed import boundary (review/* drags LLM council into offline core; refresh regresses) |
+| 13 | high | S | founder-gated | verify | Verify: resolve naming collision (loki-vaas vs loki-verify bin vs loki verify subcommand) |
+| 14 | critical | L | founder-gated | engine | Sandbox backend implementation (createSandbox throws; choose gVisor/Firecracker/Kata) |
+| 15 | critical | XL | founder-gated | enterprise | Enterprise RBAC / SSO / API-key auth / metering (hosted + on-prem control plane) |
+| 16 | high | L | founder-gated | enterprise | On-prem persistence backend interface + first prod store (Postgres/SQLite) |
+| 17 | high | S | autonomous | engine | base_unreachable signal in ObservedDiff (suppress empty_diff false-block once sandbox lands) |
+| 18 | high | M | founder-gated | engine | Real Anthropic LLM provider wiring for production verify (council against live model) |
+| 19 | high | M | founder-gated | verify | Verify: LLM council wired into default verdict path (module exists, verifyCompletion stub does not invoke) |
+| 20 | high | S | autonomous | engine | Signed verification records + ed25519 audit trail (staged, verify + ship) |
+| 21 | high | M | autonomous | verify | Golden-output conformance corpus for byte-identical CLI-invariance across implementations |
+| 22 | high | M | autonomous | engine | Code review gate: run 3 parallel reviewers (currently single-reviewer path) |
+| 23 | high | S | founder-gated | saas | HTTP /verify endpoint production wiring (handler complete, README lists as planned) |
+| 24 | high | S | autonomous | saas | MCP verify tools (verify_completion, run_evidence_gate) wiring + distribution (staged) |
+| 25 | high | M | community | test | Memory failure-loop (FAILURE_MEMORY) persistence bug (strict-xfail; recency read incomplete) |
+| 26 | high | XL | autonomous | engine | Bun runtime Phase 2: migrate write-path commands (feat/bun-migration) |
+| 27 | high | M | autonomous | cli | Bash runtime sunset (loki run-shell removal, Phase 6, after clean soak) |
+| 28 | high | M | community | saas | autonomi-saas Evidence Receipt Phase 2: screenshot capture + BFF stream route (in progress) |
+| 29 | high | M | community | saas | autonomi-saas Evidence Receipt Phase 3: test results from engine workspace |
+| 30 | high | L | autonomous | saas | autonomi-saas Cost tab + budget-breaker UX (web + enforcement audit) |
+| 31 | high | S | community | saas | autonomi-saas BFF overlays stored saas_evidence onto engine proof (verify) |
+| 32 | medium | M | founder-gated | saas | E4 no-server run output capture for CLI/library builds (bounded inline + artifact) |
+| 33 | high | M | autonomous | dashboard | Prompt Optimizer: wire real LLM-as-judge (currently heuristic-only) |
+| 34 | high | M | community | engine | Cost-honesty check implementation (normalizeModel + quote==dispatched) |
+| 35 | high | S | community | saas | POST /cost-honesty API handler (blocked on cost core) |
+| 36 | critical | L | founder-gated | saas | Hosted metered runner with per-verification container sandbox |
+| 37 | high | M | community | ci | CI parity-matrix flake root-cause (state cooldown after test-cli-commands.sh) |
+| 38 | high | M | community | ci | Post-release smoke sign-off gate before npm/Docker go public |
+| 39 | high | M | community | test | Document the 3 pre-existing failing shell tests (no public known-failures list) |
+| 40 | medium | S | autonomous | dashboard | PR #178: dashboard memory summary reports 0 on JSON-backed projects |
+| 41 | high | M | community | test | ARM64 runtime verification (emulation-only in CI; needs real Apple Silicon run) |
+| 42 | high | XL | founder-gated | saas | Multi-tenant sandbox backend (gVisor/Firecracker/Kata) selection spike |
+| 43 | critical | L | community | test | Real provider end-to-end tests (cost + secrets + duration make CI-unreachable) |
+| 44 | high | XL | founder-gated | enterprise | Enterprise deploy: object-store sync + queue fleet modes + #691 --config loader |
+| 45 | high | M | founder-gated | enterprise | Compliance scanner (loki-verify) converged as module in trust-layer SKU |
+| 46 | high | L | founder-gated | saas | GitHub App verification integration (webhook wrapper, greenfield) |
+| 47 | high | L | community | test | Long-duration stress tests (leak/queue-churn) unreachable in CI |
+| 48 | medium | L | community | saas | autonomi-saas Evidence Receipt Phase 4: clip + full test-log streaming + build_artifacts table |
+| 49 | high | XL | community | test | SWE-bench Pro 119 resume (paused 35/119) |
+| 50 | medium | M | autonomous | engine | Multi-reviewer completion council option in hosted verify (council:true dropped today) |
+| 51 | medium | M | autonomous | engine | Integrity detectors (mock-only, tautological, test-weakening) wired into evidence block |
+| 52 | medium | S | community | test | Bun test coverage threshold enforcement on branches (70% baseline exists) |
+| 53 | medium | L | autonomous | dashboard | Dashboard mid-run steering / builder control (mid-flight model switch exists, broader control deferred) |
+| 54 | medium | S | community | test | Dashboard static-asset freshness gate in local-ci (like loki-ts/dist) |
+| 55 | medium | L | community | test | Windows/WSL runtime verification (no Windows CI runner) |
+| 56 | low | M | autonomous | cli | KPI reporting (loki report kpis) on bash legacy route (Bun-only today) |
+| 57 | medium | M | autonomous | cli | Voice mode (Issue #85) |
+| 58 | medium | M | community | test | Mutation testing (stryker) workflow re-enable + fix stale 5-provider assertions |
+| 59 | medium | XL | founder-gated | engine | Managed Agents multiagent path (LOKI_EXPERIMENTAL_MANAGED_*, research preview) |
+| 60 | low | M | autonomous | engine | Sentrux iteration-loop auto-gating (manual-only today) |
+| 61 | medium | S | founder-gated | adoption | npm registry publish prep for @autonomi/verify SDK (distribution wiring) |
+| 62 | low | M | founder-gated | adoption | Disclosed default-on PostHog telemetry (Phase L, blocked + deferred) |
+| 63 | medium | S | community | saas | MCP check_cost_honesty tool registration (deferred until cost core lands) |
+| 64 | medium | L | founder-gated | dashboard | Hosted dashboard with shareable verification badges |
+| 65 | medium | M | founder-gated | adoption | MCP registry submission + Glama claim + awesome-list PRs |
+| 66 | medium | M | autonomous | saas | autonomi-saas UI/UX pass: build-watching pane + finished cues from trust verdict |
+| 67 | medium | XL | autonomous | dashboard | Server.py route modularization (~10K-line monolith, no domain separation) |
+| 68 | medium | L | autonomous | dashboard | Large dashboard-ui components modularization (5 files 1K-1.9K lines) |
+| 69 | medium | L | founder-gated | engine | Wave-13 deferred trust fixes + PRD-reuse spurious-update design fix |
+| 70 | medium | M | community | dashboard | Dashboard dark-toggle iframe theme not following SPA (harness/product parity) |
+| 71 | medium | M | community | ci | Sonnet-5 calibration follow-ups: test-verify --hosted + gate missing-tool handling |
+| 72 | medium | S | community | ci | Integrity audit (SBOM/provenance) blocking on publish (currently informational) |
+| 73 | medium | S | community | distribution | Python SDK version pre-check before publish (mismatch risk if publish fails post-tag) |
+| 74 | medium | S | community | ci | Docker buildx early-detection for uncommitted loki-ts/dist (silent COPY failure risk) |
+| 75 | medium | S | community | ci | loki-enterprise workflow enabled in main CI (dispatch-only today) |
+| 76 | medium | S | community | distribution | Brew tap real-install verification on clean macOS host post-release |
+| 77 | low | L | founder-gated | perf | Horizontal scaling / distributed rate-limit + persistence (in-memory today) |
+| 78 | low | S | community | saas | Optional build_artifacts table for enumeration/lifecycle/reaping (deferred) |
+| 79 | medium | S | community | saas | Headless browser (playwright-core) dep + Docker image size cap for screenshot capture |
+| 80 | low | M | community | engine | Wave-8/9 deferred tail: spawn_timeout removal, heredoc guards, MED/LOW findings |
+| 81 | medium | L | community | adoption | v7.19.1 traction: uncertainty-gated escalation + shareable proof + integrate 8 tutorials |
+| 82 | medium | M | autonomous | verify | Benchmarks HumanEval/SWE-bench: publish results (runners + datasets exist) |
+| 83 | low | S | autonomous | cli | loki run removal (deprecated alias for loki start, next major) |
+| 84 | low | S | autonomous | cli | Fable model tier wiring (collapses to Opus until API-available) |
+| 85 | low | S | community | distribution | Homepage/blog release update automation (Slack-notify-only today) |
+| 86 | low | S | community | distribution | Wiki-sync scheduled trigger (manual/dispatch only, skips without API key) |
+| 87 | low | L | autonomous | dashboard | Dashboard React app: implement four stub views (experimental app) |
+| 88 | low | S | community | test | Embeddings edge-case tests run without numpy (skip today) |
+| 89 | low | S | community | test | Hybrid-search test runs without chromadb skip |
+| 90 | low | S | community | test | PID-recycling test platform-portability (ps lstart skip for pid 1) |
+| 91 | low | S | community | test | Cloud integration tests (aider-cloud) run in a credentialed gated job |
+| 92 | low | S | community | test | Sentrux real-binary nightly promoted from best-effort to tracked |
+| 93 | low | S | community | test | Model-catalog probe: surface staleness in tests (issue-create skipped without admin scope) |
+| 94 | low | S | community | test | Phase 6 readiness check re-enable (cron disabled; issues disabled on repo) |
+| 95 | low | S | community | distribution | TS/Bun SDK (@loki-mode/sdk) discoverability from main package/docs |
+| 96 | low | S | community | distribution | Docker Hub description update resilience (PAT scope failures non-blocking) |
+| 97 | low | S | community | distribution | VS Code extension: keep source marked deprecated, not silently removed |
+| 98 | low | S | autonomous | dashboard | Web-app docs: fill video placeholder (coming-soon text) |
+| 99 | low | M | autonomous | cli | C# provider support (roslyn analyzers + dotnet build, deferred) |
+| 100 | low | XL | founder-gated | engine | BMAD-METHOD v6 integration (net-new, needs founder planning) |

package/docs/research-2026-07/GAP-ANALYSIS-BACKLOG.md

@@ -0,0 +1,232 @@
+# Loki + Autonomi: Competitive Gap-Analysis Backlog (2026-07-01)
+
+From studying ToolHive, brood-box, FrontierCode, SWE-1.6/Model-UX, Devin verification/async/productivity, and Anthropic's Code Modernization Playbook, gap-analyzed against loki/autonomi's real current source. Accuracy is the moat; zero adoption friction is the top need. Research-first: each item has acceptance criteria + a validation plan. Full data: gap-analysis-backlog.json.
+
+## Themes
+- Accuracy is the moat: the highest-leverage items make loki's core 'tests green' and 'checklist passed' signals actually mean something (test-provenance, source-grounded checklists, annotate-before-act, golden-master equivalence, runtime-boot proof). These are what verified-completion / trust-as-product cashes out to.
+- Adoption friction: an honest, calibrated engineering-hours/ROI estimate on the trust rail (proof.json) is the enterprise-sales lever; deterministic per-repo setup recipes reduce the top user need (reliably reaching a runnable state).
+- Anti-overthinking is a prompt-orchestration problem, not a training problem: loki can stop telling the model 'never finished' and can lower the no-claim council floor; these complement the in-flight completion-claim convergence fix.
+- Hosting (server-side sandbox backend, durable session, multi-tenant isolation, egress firewall) is real but scoped to autonomi-saas and already founder-gated; it does not outrank the accuracy moat despite source 'critical' labels.
+- Measurement instruments (FrontierCode-style mergeability benchmark, estimator calibration set) are prerequisites, not vanity: they are the validation_plan for the accuracy and ROI items and must exist before any headline number is claimed.
+- Honesty over count: five source candidates are adjacent product categories or unconsumable model-lab techniques and are dropped, not padded; the ROI chain and golden-master chain are merged to their load-bearing primitives.
+
+## Ranked backlog
+
+| # | value | eff | product | axis | item |
+|---|---|---|---|---|---|
+| 1 | critical | L | loki-mode | verification | Reverse-classical test-provenance gate (change-mode: issue/verify/healing) |
+| 2 | critical | L | loki-mode | accuracy | Ground the acceptance checklist in source reality, not spec prose alone |
+| 3 | high | L | loki-mode | accuracy | Annotate-before-act: pre-committed expected-outcome ledger, diffed against reality at verify time |
+| 4 | critical | L | loki-mode | accuracy | Golden-master parallel-run equivalence harness + per-boundary characterization (heal/migrate) |
+| 5 | high | L | loki-mode | verification | Runtime/boot smoke gate in loki verify (shipped evidence includes 'it actually runs') |
+| 6 | high | L | loki-mode | adoption-friction | LLM-calibrated engineering-hours estimator emitted into proof.json per build (+ calibration gate) |
+| 7 | high | M | loki-mode | adoption-friction | Persist a deterministic per-repo setup/run recipe (reusable setup skill) consumed by verify |
+| 8 | high | M | loki-mode | speed | Make rarv_instruction mode-aware + add 'stop-when-verified, gates are authority' directive |
+| 9 | high | M | loki-mode | accuracy | Add a mergeability rubric dimension to run_code_review (weighted, tech-lead framing) |
+| 10 | medium | M | loki-mode | accuracy | Code-scope / locality gate in autonomy/verify.sh (advisory-first) |
+| 11 | medium | M | loki-mode | accuracy | Wire grill findings into the verification/completion gate |
+| 12 | high | L | loki-mode | accuracy | FrontierCode-style scored mergeability benchmark harness for loki change-mode output |
+| 13 | medium | M | loki-mode | adoption-friction | Modernization readiness + target-selection triage (loki heal --assess, pre-flight) |
+| 14 | medium | L | loki-mode | verification | Route loki heal modernize/validate through the RARV-C loop + completion council |
+| 15 | medium | M | loki-mode | speed | Lower the no-claim council convergence floor so genuinely-finished no-promise runs stop sooner |
+| 16 | low | S | loki-mode | speed | Instruct the inner dev agent to batch independent tool calls in a single message |
+| 17 | medium | L | autonomi-saas | adoption-friction | autonomi-saas: hours-saved / ROI receipt overlay consuming loki-mode proof.json effort_estimate |
+| 18 | low | S | loki-mode | hosting | Add optional bearer-token auth + explicit loopback bind to loki mcp --transport http |
+| 19 | medium | XL | autonomi-saas | hosting | autonomi-saas: hosted sandbox backend abstraction (gVisor-first) behind a SandboxBackend interface |
+| 20 | medium | L | autonomi-saas | hosting | autonomi-saas: externalize the session as a durable append-only log decoupled from the sandbox |
+
+## Assessed NOT applicable (CTO triage - declined with reason)
+
+- **ToolHive-style per-MCP-server container isolation / curated registry / secrets vault (source: ToolHive)**: Presupposes running UNTRUSTED third-party MCP servers. Loki ships only first-party servers and uses --strict-mcp-config (providers/claude.sh:209) to actively EXCLUDE arbitrary third-party servers. Building a registry/containment layer adds adoption friction and code loki does not need for its threat model (it runs code it wrote). Adjacent product category. Source itself marked applicable=false (effort XL).
+- **ToolHive as inspiration for autonomi hosting IF/when it hosts third-party MCP servers (source: ToolHive)**: Cannot be verified from this tree (autonomi is a separate private repo) and is gated on autonomi actually hosting untrusted third-party MCP servers, which is not a current requirement. Recorded as autonomi-hosting inspiration only, not a loki-mode backlog item. Source marked applicable=false. (Note: the gateway-auth primitive that IS directly consumable is captured as rank 18 for loki's own mcp http command.)
+- **Integrate brood-box or Anthropic Managed Agents as a runtime dependency (source: brood-box / Managed Agents)**: brood-box is a local-developer agent sandbox CLI (adjacent product category loki's own three-tier sandbox already covers for local use); Managed Agents is a hosted Anthropic PLATFORM loki competes with, not an API loki consumes. The repo already made this triage (docs/ECOSYSTEM-CURRENCY-PLAN.json:77). The correct move is borrowing the architectural PATTERNS (decoupling, disposable untrusted sandbox, diff-then-flush) into autonomi's own founder-gated backend (ranks 19-20), not bolting on either tool. Source marked applicable=false.
+- **mutagent-style adaptive test grading (source: FrontierCode)**: An internal benchmark-authoring tool for a third-party eval grader: LLMs surgically adapt a hidden-canonical-answer test env so open-ended solutions grade deterministically. Loki does not grade external agents' solutions against hidden canonical answers, so there is no consumer inside loki/autonomi. Adjacent product category. Source marked applicable=false (effort XL).
+- **Training-time length penalty (source: SWE-1.6 Model UX)**: Cognition's headline fix was a length penalty during model POST-TRAINING. Loki is an orchestration layer over a hosted model (claude -p) and cannot retrain or add a training-time penalty. Adjacent (model-lab technique). The consumable analogs are delivered structurally via ranks 8, 15, 16 (stop telling the model never to stop; lower the no-claim floor; batch tool calls). Source marked applicable=false (effort XL).
+
+## Detail (acceptance criteria + validation plan per item)
+
+### #1 Reverse-classical test-provenance gate (change-mode: issue/verify/healing) (loki-mode, critical/L)
+Before counting an agent's NEW tests as affirmative 'green' evidence in council_evidence_gate, run them against the pre-change base (VERIFY_MERGE_BASE / _LOKI_RUN_START_SHA) and require they FAIL on base and PASS on HEAD. Tests that pass on both base and HEAD are tautological and marked inconclusive, not affirmative. No-op on greenfield (no base) and when LOKI_TEST_PROVENANCE=0. This is the single highest-value accuracy-moat item: it makes loki's strongest verification signal (tests-green) actually mean something and directly attacks the 'test written to pass, proving nothing' failure mode.
+Acceptance criteria:
+- Given a change-mode run with a resolvable base, test files added/modified in the diff are identified
+- New/changed tests are run against base (checkout or reverse-patch); which FAIL on base is recorded
+- In council_evidence_gate (completion-council.sh:1621-1688) a completion claim whose only test evidence is provenance-UNconfirmed (passes on base) is marked inconclusive, never affirmative VERIFIED
+- A trust event + .loki record lists each new test with base-fail/head-pass status
+- Byte-identical no-op on greenfield and when LOKI_TEST_PROVENANCE=0
+Validation: Fixture harness with two change-mode tasks: (A) a genuine fix with a test that fails on base + passes on HEAD, (B) a tautological test that passes on both. Assert: A is counted affirmative, B is downgraded to inconclusive and surfaced as a finding. Prove no-op via byte-diff of build_prompt/gate output on a greenfield run before/after. Cross-validate against the FrontierCode-style mergeability benchmark (rank 12): provenance-confirmed tasks should score higher on the reverse-classical dimension.
+
+### #2 Ground the acceptance checklist in source reality, not spec prose alone (loki-mode, critical/L)
+Extend checklist_oracle_triangulate beyond datastore to triangulate the checklist against the actual codebase: discovered HTTP routes/handlers (does each spec'd endpoint map to a real handler?), CLI commands, exported public API symbols (via the LSP proxy lsp_workspace_symbols / lsp_check_exists), and >=1 high-value domain invariant (passwords hashed not plaintext; money not float). Auto-generate deterministic checks (http_check, lsp_check_exists) instead of relying on the same LLM that wrote the code to author acceptance checks from the same prose. Attacks the single-oracle weakness the code itself flags as deferred (prd-checklist.sh).
+Acceptance criteria:
+- For a web project, every endpoint named in the spec maps to a real route/handler in source; unmatched -> High finding (mirrors existing oracle-datastore-conflict shape)
+- Spec'd public API symbols verified to exist via the LSP proxy, not LLM-authored grep_codebase checks
+- At least one domain-invariant check implemented as a deterministic gate (e.g. password fields hashed)
+- Greenfield/absent (nothing wired yet) does NOT fire a finding, matching existing non-conflict guards
+Validation: Fixture repos: (A) spec names endpoint /orders that exists -> no finding; (B) spec names /orders but no handler -> High finding; (C) plaintext password store -> invariant finding. Measure false-positive rate on 3 clean real repos (must be zero). Prove accuracy lift on the FrontierCode-style benchmark (rank 12): source-grounded checklists should catch wrong-but-internally-consistent implementations that prose-only checklists pass.
+
+### #3 Annotate-before-act: pre-committed expected-outcome ledger, diffed against reality at verify time (loki-mode, high/L)
+Before running verification for a change/checklist-item, the runtime writes a tamper-evident ledger of EXPECTED observable outcomes (e.g. 'GET /health -> 200 {status:ok}', 'test X fails before fix, passes after', 'endpoint Y should now exist'). At verify time, actual results are compared to the pre-committed prediction; any expectation silently dropped or contradicted becomes a finding. Reuse the proof-verify canonical-hash + drift pattern so the ledger is tamper-evident and edits-after-the-fact are detectable. Forces a falsifiable prediction, cutting self-serving post-hoc rationalization.
+Acceptance criteria:
+- .loki/expectations/<iter>.json written BEFORE the verify/checklist run, each entry {id, statement, check_ref, expected}, canonicalized and hashed like proof-generator
+- loki verify (or the completion gate) reads the ledger and emits a finding for any expectation with no matching executed check, or whose actual contradicts expected
+- An unevaluable expectation maps to inconclusive->CONCERNS, never VERIFIED (reuse verify.sh Entanglement-2)
+- The ledger hash is embedded in evidence.json/proof.json so an edited-after-the-fact expectation is detectable
+Validation: Test: seed a ledger with 3 expectations, run verify, assert (1) a met expectation passes, (2) a contradicted one is a finding, (3) a dropped/unexecuted one is a finding, (4) editing the ledger after write breaks the embedded hash. Before/after: same build with ledger disabled vs enabled must surface >=1 contradiction that post-hoc-only grading missed on a deliberately-mismatched fixture.
+
+### #4 Golden-master parallel-run equivalence harness + per-boundary characterization (heal/migrate) (loki-mode, critical/L)
+MERGED (harness + boundary-characterization). During archaeology/guardrail, capture outputs at natural system boundaries (CLI stdout/exit, HTTP responses, DB state deltas, file/queue outputs) into .loki/healing/behavioral-baseline/ as golden masters. During modernize/validate, re-run the SAME boundary probes against modernized code and diff. Replace hook_post_healing_modify's current whole-suite pass/fail (migration-hooks.sh:580) with per-boundary comparison so verification is implementation-agnostic and cannot be satisfied by overfit unit tests. Any divergence blocks the phase gate unless documented as intentional. Turns the playbook's 'side-by-side identical outputs' from a prompt suggestion into machine-verified evidence. The behavioral-baseline/ dir is currently mkdir'd (loki:14323) but never populated or compared.
+Acceptance criteria:
+- cmd_heal archaeology (and migrate guardrail) writes >=1 golden-master artifact per detected boundary into .loki/healing/behavioral-baseline/
+- hook_healing_phase_gate for modernize->validate fails closed unless boundary outputs match baseline OR a documented-intentional-change record exists
+- hook_post_healing_modify compares boundary outputs to baseline, not only whole-suite exit code; a change keeping unit tests green but altering a CLI/API boundary is BLOCKED
+- diff report enumerates every changed boundary (old vs new) in loki heal --report and structured evidence JSON
+- Degrades honestly ('no boundaries detected') rather than false-green
+Validation: Fixture: a Python2->Python3 (or C->Java) sample repo. Capture baseline, apply (1) a behavior-preserving refactor -> gate passes; (2) a transform that keeps unit tests green but changes a CLI output -> gate BLOCKS. Assert baseline artifacts exist post-archaeology (grep count > 0, currently 0). Matches references/legacy-healing-patterns.md:156-181 prescribed system-boundary approach.
+
+### #5 Runtime/boot smoke gate in loki verify (shipped evidence includes 'it actually runs') (loki-mode, high/L)
+Promote the build-loop's playwright/http smoke into loki verify as a deterministic runtime gate: detect the start command (or use the persisted setup recipe from rank 9), boot the app with a timeout, hit a health/root path, capture status + a screenshot artifact, tear down, record in evidence.json as reproducible. Closes the biggest trust gap vs Devin (compiles + green-tests != works); the screenshot becomes a proof artifact. Runtime-not-reachable -> inconclusive->CONCERNS, never VERIFIED. Fail-open: without a detectable app, output is byte-identical to today.
+Acceptance criteria:
+- verify_gate_runtime added to verify_main gate list, skipped (not failed) only when no start command/setup recipe is detectable
+- A successful boot records HTTP status + artifact path in evidence.json with reproducible=true
+- Boot failure -> Critical/High finding; boot-not-attempted-but-app-detected -> inconclusive->CONCERNS
+- No regression: without a detectable app, output byte-identical to today (mirrors --hosted fail-open pattern)
+Validation: Fixture web app: (A) healthy boot -> evidence.json has status 200 + screenshot path + reproducible=true; (B) broken start command -> Critical finding, verdict not VERIFIED; (C) library repo with no app -> gate skipped, evidence.json byte-identical to pre-change baseline (diff must be empty).
+
+### #6 LLM-calibrated engineering-hours estimator emitted into proof.json per build (+ calibration gate) (loki-mode, high/L)
+MERGED (estimator + calibration). Replace the iterations x 15min heuristic (loki:23478) with an LLM estimator that reads the actual work (diff stat, files changed, tests written, task/PRD scope, iteration transcript) and predicts equivalent human-engineering-hours plus a low/high band. Persist as an effort_estimate object in per-build proof.json so it rides the trust/evidence rail. Fall back to the heuristic when no LLM is available, labeled 'heuristic (uncalibrated)' vs 'estimated' so the number is never silently fabricated. Ship a validation harness (ground-truth engineer-hour set) that computes an error metric (r_log or MAE-in-log-space); estimates may only be labeled 'calibrated' once the metric clears a threshold. This is the load-bearing primitive both the dollar ROI surface and any guarantee depend on, and it directly serves the 'accuracy is the moat, no fabrication' rule.
+Acceptance criteria:
+- proof.json gains effort_estimate: {hours, low, high, method:'llm'|'heuristic', model, inputs_hash}
+- loki metrics TIME SAVED reads per-build estimates when present, falls back to iterations x 15min only when absent, labeling the method
+- Estimator input is real work (diff stat + files + tests + scope), NOT iteration count alone; unit test feeds two builds of different scope and asserts different hours
+- A validation dataset of builds with engineer-hour ground truth exists in benchmarks/; estimator error metric computed and reported; threshold gates 'calibrated' vs 'uncalibrated' labeling
+- No LLM -> heuristic path clearly labeled 'heuristic (uncalibrated)', never presented as validated
+Validation: Unit test: two builds of clearly different scope (2-iter schema tweak vs full-stack feature) must yield materially different hours (heuristic gives both 30min; assert estimator does not). Calibration: score LLM estimates against the ground-truth set, report r_log/MAE; assert copy stays 'uncalibrated' until threshold cleared. This IS the validation gate for the entire ROI chain.
+
+### #7 Persist a deterministic per-repo setup/run recipe (reusable setup skill) consumed by verify (loki-mode, high/M)
+On first successful build/boot, save .loki/setup-recipe.json capturing deterministic steps to install, seed, auth, and start the app (commands + env var NAMES + health path). Replay on subsequent verify/e2e runs instead of re-detecting heuristically. This is Devin's 'saved login/setup skill'; it reduces the top adoption need (reliably reaching a runnable state) and enables the runtime gate (rank 5) for repos needing auth/seed. Distinct from cmd_setup_skill (which installs the Loki skill, not a repo recipe). Secrets never persisted (only env var NAMES; reuse proof_redact.py).
+Acceptance criteria:
+- .loki/setup-recipe.json written after a verified boot with {install, seed, env_keys(no secret values), start, health_path}
+- verify_gate_runtime prefers the recipe over heuristic detection when present
+- Secrets never persisted into the recipe (only env var NAMES); reuse proof_redact.py patterns
+- Recipe replayed idempotently and its use recorded in evidence.json
+Validation: Fixture repo needing seed+auth: first run writes recipe; second run reaches runnable state via recipe (assert no heuristic re-detection ran, e.g. log marker). Assert recipe contains env var NAMES but zero secret VALUES (grep for known secret pattern must return nothing). Measure: time-to-runnable second run < first run.
+
+### #8 Make rarv_instruction mode-aware + add 'stop-when-verified, gates are authority' directive (loki-mode, high/M)
+MERGED (mode-aware rarv + stop directive). build_prompt currently emits the same 'There is NEVER a finished state - always find the next improvement' rarv_instruction (run.sh:13770) in every mode, including PRD/checkpoint runs just auto-switched to finite scope (a direct self-contradiction). Gate it on AUTONOMY_MODE/PERPETUAL_MODE: for finite runs replace with 'When the PRD requirements are implemented and completion gates pass, claim done via loki_complete_task and STOP; do not add unrequested improvements. Verify once -- the completion gates (tests, checklist, evidence) are the authority on done; do not re-verify redundantly.' Keep never-finished text only for genuine perpetual mode. Delete the unverifiable '2-3x quality improvement' clause (biases toward redundant self-verification). This is the orchestration analog of Cognition's length penalty and the prompt root of the 14-iter convergence bug. COMPLEMENTARY to the in-flight 'council evaluates on completion claim' fix (that changes when the council checks; this stops the prompt telling the model never to stop) -- not a duplicate.
+Acceptance criteria:
+- In a PRD run (auto-switched to checkpoint) the emitted build_prompt no longer contains 'There is NEVER a finished' or 'always find the next improvement'
+- In a perpetual/no-promise run the never-finished text still appears (behavior preserved)
+- The '2-3x quality improvement' clause removed from all modes; finite-mode prompt contains the concise 'stop when verified-done, gates are the authority' directive
+- Deterministic completion gates (council_evidence_gate/checklist/heldout/assumption) unchanged and still block on failure
+- loki-ts/src/runner/build_prompt.ts parity string updated (parity-locked)
+Validation: String assertions on emitted build_prompt for each mode (finite lacks 'NEVER finished', perpetual retains it; '2-3x' absent everywhere). Behavioral: a spec satisfiable in ~1 iteration claims completion within a small bounded iteration count on a benchmark spec, with all completion gates still passing (verification not weakened -- run completion-council tests, assert no coverage regression). Confirm dual-route (bash + Bun) parity.
+
+### #9 Add a mergeability rubric dimension to run_code_review (weighted, tech-lead framing) (loki-mode, high/M)
+Extend (do not replace) the existing severity mechanism with a 'maintainer-mergeability' reviewer whose rubric covers scope creep, dead/duplicated code, and conformance to the surrounding code's conventions -- and have the aggregator produce a weighted quality SCORE alongside the binary block verdict (FrontierCode-style: 0 on any blocker, else weighted non-blocker sum). Reuse has_blocking (run.sh:10770) for the blocker layer; add the weighted score as a reported quality metric, not a new hard gate initially. Upgrades loki's review from 'CI + architecture' toward 'tech lead' on the axes it currently misses (the specialist pool is security/test/perf/dependency/architecture; none carries a 'would a maintainer merge this' mandate).
+Acceptance criteria:
+- A mergeability-focused reviewer added to the pool with a rubric prompt covering scope, dead code, and convention-conformance
+- Aggregator emits a numeric quality score (0 if any blocker, else weighted non-blocker sum) in aggregate.json
+- Existing Critical/High = block, Medium/Low = non-blocking behavior preserved unchanged
+- Score reported/surfaced but not a new hard gate without explicit opt-in flag
+- No emojis/em-dashes in prompt text; dual-route (bash + Bun) parity maintained
+Validation: Fixture PRs: (A) tight 3-file fix -> high score, no scope blocker; (B) 40-file sprawling diff touching unrelated code -> scope blocker, score 0. Assert existing block behavior unchanged on a known-Critical fixture. Prove the score is discriminating on the FrontierCode-style benchmark (rank 12): mergeable vs unmergeable tasks should separate on the weighted score.
+
+### #10 Code-scope / locality gate in autonomy/verify.sh (advisory-first) (loki-mode, medium/M)
+Turn the already-computed VERIFY_DIFF_FILES/INS/DEL (verify.sh:154-164) into an enforced gate: configurable caps on total files changed and net line growth, plus an optional LLM semantic-locality check that flagged edits stay within the files/functions the task requires. Emit a 'scope' record on violation. Advisory-with-warning by default (record, don't hard-block) to avoid adoption friction; LOKI_SCOPE_ENFORCE=1 makes it blocking. Over-broad diffs are the #1 human merge-rejection reason FrontierCode targets, and loki currently measures the numbers but never uses them as a gate. N/A to greenfield full builds.
+Acceptance criteria:
+- verify.sh emits a structured scope record: files_changed, net_lines, verdict (ok|warn|block) with thresholds used
+- Thresholds configurable via env (LOKI_SCOPE_MAX_FILES, LOKI_SCOPE_MAX_NET_LINES) with sane defaults
+- Advisory by default (surfaces in verify output + SARIF), blocking only under LOKI_SCOPE_ENFORCE=1
+- Optional semantic-locality LLM check gated behind its own flag (off by default, zero added latency/cost)
+- No behavior change on greenfield loki start
+Validation: Fixture change-mode diffs: (A) 3 files within scope -> verdict ok; (B) 40 files -> verdict warn (default) and block under LOKI_SCOPE_ENFORCE=1. Assert greenfield loki start output byte-identical before/after. Assert the scope record appears in evidence.json + SARIF.
+
+### #11 Wire grill findings into the verification/completion gate (loki-mode, medium/M)
+Feed the pre-build Devil's-Advocate grill's surfaced ambiguities/missing-acceptance-criteria into checklist generation and the completion gate: unresolved High grill findings become required checks or block completion until acknowledged (reuse the existing assumption-ledger gate in completion-council.sh). grill.sh currently writes a standalone report never consumed by verify/council. Makes the Devil's-Advocate posture actually harden what gets verified rather than being advisory-only.
+Acceptance criteria:
+- grill High findings converted into checklist items or assumption-ledger entries
+- Unresolved High grill findings block completion via an existing gate path (not a new bespoke one)
+- Opt-out knob preserves current standalone behavior
+- No double-firing when grill was not run (gate inert, like the held-out gate NONE branch)
+Validation: Fixture spec with a known ambiguity: run grill -> High finding -> assert it becomes a required check and blocks completion until acknowledged. Assert gate is inert (no finding) when grill was not run. Confirm reuse of the assumption-ledger path (no new bespoke gate) via code inspection + existing completion-council test suite passing.
+
+### #12 FrontierCode-style scored mergeability benchmark harness for loki change-mode output (loki-mode, high/L)
+Not a runtime feature -- the measurement instrument. Assemble a small suite of change-mode tasks with maintainer rubrics (blocker + weighted non-blocker) and reverse-classical tests, run loki verify / issue-mode against them, and produce a per-task mergeability score. This is what lets loki claim an accuracy/mergeability number and is the validation_plan backbone for ranks 1, 2, 9, 10. Aligns with pending task #49 (reproducible speed+accuracy benchmark) and #51 (council/RARV-C research update, gated on benchmark).
+Acceptance criteria:
+- A handful of tasks with base repo + rubric (blocker/non-blocker) + reverse-classical tests
+- Harness runs loki change-mode and produces a per-task mergeability score
+- Reproducible, documented, wired into the existing benchmarks/ conventions
+- Reports a headline mergeability % comparable across loki versions
+Validation: Self-validating: run harness on current loki, record baseline mergeability %. Re-run after ranks 1/2/9 land; assert the % moves (this IS the before/after proof for the accuracy items). Reproducibility check: two runs on the same loki version produce the same score within tolerance.
+
+### #13 Modernization readiness + target-selection triage (loki heal --assess, pre-flight) (loki-mode, medium/M)
+Add loki heal --assess (or a readiness step auto-run before archaeology) that scans a codebase read-only and outputs: estimated LOC/language mix, technical-debt signals, maintenance-risk indicators, placement on the playbook's 4-level maturity model, and a RANKED list of low-risk high-visibility targets to start with (playbook weeks 1-2). Reduces the top adoption-friction question 'where do I start' to one command, reusing loki's existing complexity-detection machinery. Human-readable + structured JSON for the dashboard.
+Acceptance criteria:
+- Single command emits a readiness report with maturity-level placement + ranked candidate targets + rationale
+- Ranking prefers isolated/low-blast-radius modules (matches playbook weeks 1-2)
+- Runs read-only (no modification), zero required flags
+- Output is both human-readable and structured JSON for the dashboard
+Validation: Fixture legacy repo with a mix of isolated and highly-coupled modules: assert the ranking places the isolated module above the coupled one, assert zero filesystem modifications (git status clean after run), assert valid JSON output parses.
+
+### #14 Route loki heal modernize/validate through the RARV-C loop + completion council (loki-mode, medium/L)
+loki heal is single-shot per phase (claude -p, loki:14465-14502), bypassing the RARV-C loop, completion council, and 8 quality gates that make loki start/migrate trustworthy. For behavior-changing phases (stabilize/isolate/modernize/validate), execute via run_autonomous with LOKI_INJECT_FINDINGS/LOKI_OVERRIDE_COUNCIL/LOKI_AUTO_LEARNINGS auto-enabled and the legacy-healing-auditor in the review pool. Archaeology-only can stay single-shot. Gives modernization builds the same iterative verification and override-council trust loki start already provides.
+Acceptance criteria:
+- cmd_heal for phases stabilize|isolate|modernize|validate invokes the RARV loop (run.sh path), not a bare claude -p call
+- RARV-C closure flags default-on within heal (documented in skills/healing.md, verified by grep of cmd_heal)
+- legacy-healing-auditor fires and can BLOCK on behavioral change without characterization update
+- A deliberately-broken transform is caught + reverted by the loop, proven on a fixture
+Validation: Fixture: run heal modernize with a deliberately-broken transform; assert the RARV loop catches + reverts it (vs current single-shot which would ship it). Grep cmd_heal to confirm run_autonomous path taken for the four phases. Assert legacy-healing-auditor appears in the review pool during heal. Pairs with rank 4 golden-master for the behavioral-change block proof.
+
+### #15 Lower the no-claim council convergence floor so genuinely-finished no-promise runs stop sooner (loki-mode, medium/M)
+For the convergence-detection (NO explicit completion claim) path only, reduce the effective floor: allow council_should_stop to evaluate as early as MIN_ITERATIONS=1 when tests+checklist evidence already indicate done, or check every iteration once evidence is green instead of every 5th (currently CHECK_INTERVAL=5, MIN_ITERATIONS=3). The explicit-claim path already bypasses this (run.sh:17418), so this only helps analysis-mode / no-promise runs -- COMPLEMENTARY to the in-flight 'council evaluates on completion claim' fix (which is the explicit-claim path), not a duplicate. Keep stagnation and evidence gates intact -- changes WHEN the stop check runs, not WHETHER work is verified.
+Acceptance criteria:
+- A no-promise run that is genuinely complete can stop before iteration 5 when evidence gates pass
+- Stagnation and evidence gates (completion-council.sh) still block premature/unverified stops
+- Explicit-claim completion path behavior unchanged
+- Gated behind existing LOKI_COUNCIL_CHECK_INTERVAL/MIN_ITERATIONS so operators can restore old behavior
+Validation: Fixture no-promise/analysis run that is complete by iteration 2: assert it can stop before iteration 5 with evidence gates green. Negative test: an unverified/stagnating run still does NOT stop early (evidence gate blocks). Assert explicit-claim path timing unchanged via existing completion-council tests.
+
+### #16 Instruct the inner dev agent to batch independent tool calls in a single message (loki-mode, low/S)
+Add a short build_prompt directive (all modes): 'When issuing independent read-only operations (reads, greps, file lookups, LSP checks) that do not depend on each other, issue them in a single message so they run in parallel; do not serialize independent lookups.' The direct, cheap analog of Cognition's parallel-tool-call fix at the exact layer loki controls (the prompt to claude -p). Zero adoption friction (internal prompt string), no verification impact. Parity mirror in loki-ts.
+Acceptance criteria:
+- build_prompt emits a parallel/batch tool-call directive in every mode
+- Parity mirror added to loki-ts/src/runner/build_prompt.ts
+- Directive is <= 2 sentences and adds no new env knob
+- A representative run shows the inner agent issuing at least one multi-tool-call message for independent lookups (spot-checked in agent.log / stream-json)
+Validation: String assertion: directive present in all build_prompt mode variants and in loki-ts parity string. Behavioral spot-check: on a representative run, grep agent.log/stream-json for at least one message issuing multiple independent tool calls. Confirm no new env knob added.
+
+### #17 autonomi-saas: hours-saved / ROI receipt overlay consuming loki-mode proof.json effort_estimate (autonomi-saas, medium/L)
+In autonomi-saas, overlay the per-build effort_estimate from loki's proof.json (rank 6) onto the Evidence Receipt and aggregate into an account-level ROI dashboard ('this month: ~N engineer-hours, ~$M at a configurable blended rate'). Aggregation matches the level where the estimator is honest (Devin's aggregation-honest framing). $ requires an operator-set rate; unset shows hours only, never a fabricated dollar figure. The natural home for a Devin-style productivity guarantee (SaaS carries billing/tenancy); the guarantee is founder-gated and blocked until the estimator is calibrated (rank 6). No cross-tenant leakage (respects the per-build two-read split used for #17).
+Acceptance criteria:
+- BFF reads proof.json effort_estimate per build and stores it against the tenant/build
+- Account view aggregates hours (+ $ if rate set) across builds, labeled with the aggregation-honest caveat
+- Guarantee logic (if pursued) founder-gated and blocked until estimator has a validation number
+- No cross-tenant leakage in the aggregate (respects the per-build two-read split for #17)
+Validation: Two-tenant fixture: assert tenant A's aggregate excludes tenant B's builds (cross-tenant leak test, reuse #17 harness). Assert $ absent when no rate configured. Assert guarantee copy is blocked/hidden while estimator method='heuristic' or label='uncalibrated'. Depends on rank 6 shipping first.
+
+### #18 Add optional bearer-token auth + explicit loopback bind to loki mcp --transport http (loki-mode, low/S)
+The one place loki's surface maps directly onto ToolHive's gateway. mcp/server.py:2639 calls mcp.run(transport='http', port=args.port) with no explicit host and no auth. Pass an explicit host (default 127.0.0.1) and support an optional LOKI_MCP_AUTH_TOKEN that the server requires on every request. Low-cost hardening of an already-shipping command; loopback default unchanged (zero adoption friction, token opt-in). Latent-not-exposed today (FastMCP loopback default) but the explicit bind + opt-in token close it honestly.
+Acceptance criteria:
+- loki mcp --transport http binds 127.0.0.1 explicitly (verified via ss/lsof, not 0.0.0.0)
+- When LOKI_MCP_AUTH_TOKEN is set, requests without a matching bearer are rejected 401; when unset, behavior is unchanged
+- Docs note the token env var; default local stdio path untouched
+Validation: Start loki mcp --transport http, verify bind address via lsof/ss is 127.0.0.1 not 0.0.0.0. With LOKI_MCP_AUTH_TOKEN set: request without bearer -> 401, with matching bearer -> 200. With it unset: behavior byte-identical to today. Confirm stdio path unaffected.
+
+### #19 autonomi-saas: hosted sandbox backend abstraction (gVisor-first) behind a SandboxBackend interface (autonomi-saas, medium/XL)
+FOUNDER-GATED. Define one SandboxBackend interface (create/exec/collect-diff/flush/destroy) and implement gVisor as the first backend for per-build/per-verification isolation, mirroring Managed Agents' disposable untrusted-sandbox. Concrete unblock for founder-gated backlog ranks 14/36/42. Borrow brood-box's diff-then-flush and Managed Agents' decoupling as PATTERNS, not dependencies; libkrun-microVM is a strong second backend to evaluate in the selection spike. Sandbox treated as untrusted: no tenant secrets/signing keys reachable from inside. This is real but scoped to autonomi and already triaged/founder-gated in the repo docs; it does NOT outrank the accuracy moat despite the source 'critical' label.
+Acceptance criteria:
+- A SandboxBackend interface exists with create/exec/collect-diff/flush/destroy and at least one working backend (gVisor)
+- Each hosted build/verification runs in its own disposable sandbox destroyed on completion
+- Sandbox is treated as untrusted: no tenant secrets or signing keys reachable from inside
+- Selection spike documents gVisor vs Firecracker vs Kata vs libkrun with a chosen default + rationale (satisfies rank 42)
+- Pre-flush diff review gates change persistence; per-domain egress allowlist (default provider API + git host, deny-by-default option); secret paths (.env*, keys) non-overridably excluded from any flushed change set
+Validation: Isolation test: attempt to read a mounted tenant secret from inside the sandbox -> must fail. Disposability: assert sandbox is destroyed post-build (no lingering container). Egress: request to a non-allowlisted domain blocked, provider API + git host allowed. Diff-then-flush: an unapproved change is not persisted to the tenant workspace. Founder-gated: do not start until the founder unblocks; kept below the accuracy moat.
+
+### #20 autonomi-saas: externalize the session as a durable append-only log decoupled from the sandbox (autonomi-saas, medium/L)
+FOUNDER-GATED. Move build state off local .loki/ files into a durable session store (the persistence backend at backlog rank 16) so the harness/sandbox can crash or scale horizontally without losing progress -- Managed Agents' session/harness/sandbox decoupling. Enables resume-across-workers and starting reasoning before the sandbox is fully provisioned. NOT applicable to loki-mode local (its .loki/ file state is correct for single-user). Scoped to autonomi, founder-gated; below the accuracy moat.
+Acceptance criteria:
+- A PersistenceBackend interface with a first prod store (Postgres or SQLite) exists (satisfies backlog rank 16)
+- A hosted build survives a worker/sandbox restart and resumes from the durable session
+- The harness reads session state independently of any specific sandbox instance
+Validation: Crash-resume test: kill the worker/sandbox mid-build, restart, assert the build resumes from the durable session with no lost progress. Horizontal test: two workers read the same session store without corruption. Assert loki-mode local .loki/ path is untouched (scoped to autonomi only).
+

package/docs/research-2026-07/OBSERVABILITY-TOOL-SPEC.md

@@ -0,0 +1,96 @@
+# Founder Observability Tool - Spec
+
+## Goal
+ONE local, Docker-hosted dashboard for the founder to navigate the whole program's history at a
+glance: releases, before/after metrics, what worked, what did NOT work, mistakes made + fixes,
+and current status - across loki-mode + autonomi-verify + autonomi-saas. Simplest possible form,
+in Autonomi's design language. Runs in Docker locally; opens in the browser when done.
+
+## Absolute rules (never violate)
+- **HONEST DATA ONLY.** Every number/claim must come from a REAL source file (below). NEVER
+  fabricate a metric, a release, or an outcome. Accuracy is the moat; a fabricated founder
+  dashboard destroys it. If a datum is unknown, show "not measured" - never invent it.
+- Include the UGLY: "what didn't work" and "mistakes" are REQUIRED sections and must show real
+  ones (e.g. the reverted PRD-reuse denylist fix; the convergence double-consume bug two Opus
+  reviewers caught before ship; the benchmark harness dying at turn boundaries). Honesty about
+  misses is the point of the tool.
+- No emojis, no em-dashes/en-dashes anywhere (project rule).
+
+## Data sources (read these; do not invent)
+- Releases: `CHANGELOG.md` (parse `## [x.y.z] - date` + the entry body).
+- Before/after metrics: `benchmarks/results/*.json` (speed-before/after-*.json = wall_clock_min,
+  act_iterations, completion_claims, engine_completed, acceptance, per_iteration_work_s).
+- Diagnoses / decisions / plans: `docs/SPEED-DIAGNOSIS-2026-07-01.md`,
+  `docs/FOUNDER-STATUS-AND-DECISIONS-2026-07-01.md`, `docs/TOP-100-BACKLOG.md`,
+  `docs/research-2026-07/GAP-ANALYSIS-BACKLOG.md`.
+- What worked / didn't / mistakes: parse the above docs' honest sections + git log on origin/main
+  (`git log --oneline` for shipped commits; look for "revert" and "fix(...council-review finding)"
+  as mistake->fix signals).
+- Founder-gated decisions (open): section 4/6 of FOUNDER-STATUS-AND-DECISIONS.
+
+## REAL adoption + efficiency metrics (founder wants graphs, real numbers, NO fakes)
+Fetch these live at build/refresh time. If any endpoint fails, show "unavailable" - NEVER fake.
+Verified working 2026-07-01 (sample values shown so you know the shape; re-fetch for current):
+- **GitHub stars/forks/issues**: `gh api repos/asklokesh/loki-mode` (.stargazers_count=998,
+  .forks_count=193, .open_issues_count, .created_at=2025-12-26). 
+- **Per-release download counts + dates**: `gh api repos/asklokesh/loki-mode/releases`
+  (per tag: .tag_name, .published_at, sum of .assets[].download_count). Real per-release adoption.
+- **npm downloads (last-month total + 30-day DAILY series for a line graph)**:
+  `curl -s https://api.npmjs.org/downloads/range/last-month/loki-mode` -> .downloads[] is
+  [{day, downloads}] (30 points; last month total ~23,798; peak 2,598 on 2026-06-17). GRAPH THIS.
+  Also per-version: `https://api.npmjs.org/versions/loki-mode/last-week` (downloads by version).
+- **Docker Hub pulls**: `curl -s https://hub.docker.com/v2/repositories/asklokesh/loki-mode/`
+  -> .pull_count (=56,232), .last_updated. (Cumulative total only; no time-series from Hub.)
+- **Homebrew custom-tap installs**: NOT publicly available for a custom tap (only homebrew/core
+  taps have analytics). Show "not available (custom tap has no public install API)" - do NOT fake.
+- **Efficiency per release**: from `benchmarks/results/speed-*.json` (wall_clock_min,
+  act_iterations) + git-diff churn per release from CHANGELOG. Where a release has no benchmark,
+  show "not benchmarked" honestly. As more before/after runs accumulate, chart wall-clock and
+  iteration-count per version to show the efficiency trend.
+
+GRAPHS REQUIRED (real data only): npm daily-downloads line (30d), per-release GH-download bars,
+cumulative Docker pulls (single stat + timestamp), stars (single stat), and a
+speed/efficiency-per-release chart (wall-clock + iterations) that grows as benchmarks accumulate.
+Use a tiny dependency-light chart approach (inline SVG or a single vendored lib) so it stays
+Dockerizable and offline-safe. A daily/refresh fetch keeps the numbers current.
+
+## Design language (Autonomi)
+Match the shipped Autonomi look. Reference files: `artifacts/whitepaper/Autonomi-Whitepaper.html`
+and `dashboard/static/trust.html`. Palette: dark bg (#0F0B1A / #0d1117), Autonomi purple accent
+(#553DE9 / #7B6BF0), verified teal (#2ED8B6 / #1AAF95), dim ink (#B8B0C8 / #8A857C), warning amber
+(#C4922E), reject red (#C04848). Semantic status colors: verified=teal, reported/pending=amber,
+mistake/reverted=red, roadmap=purple. Clean, calm, founder-readable; big numbers, short labels,
+before->after arrows. Reuse the whitepaper's CSS-variable pattern (--bg, --accent, --verified,
+--ink, --panel, --line) so it feels native.
+
+## Sections (the dashboard)
+1. **At a glance**: current live version (VERSION), # releases this cycle, latest before/after
+   speed number (with an honest "n=1, pending repeat" caveat if only one after-run exists).
+2. **Release timeline**: each version, date, one-line what+why, and its outcome (live/verified).
+3. **Before/After metrics**: the speed benchmark(s) as before->after cards (wall-clock, iterations,
+   completion-claims); mark clearly what engine version each ran on and that attribution is
+   "vX vs vY" not single-cause. If only a before exists, show before + "after pending".
+4. **What worked**: shipped, council-approved, verified fixes (with the evidence: tests, live check).
+5. **What did NOT work / reverted**: honest list (PRD-reuse denylist reverted as unsound; any
+   dropped approach) with the reason.
+6. **Mistakes -> Fixes**: bugs I introduced that review/CI caught and I fixed (convergence
+   double-consume; server.py payload edge; benchmark cross-spec contamination avoided). Each:
+   what went wrong -> how caught -> how fixed. This is the founder's audit trail.
+7. **Open founder decisions**: the 6 one-line asks from FOUNDER-STATUS-AND-DECISIONS.
+8. **Backlog / roadmap**: top ranked items from the gap-analysis + top-100 (accuracy-first).
+
+## Delivery
+- Static-first is fine (a single index.html + a small data.json generated from the sources by a
+  build script), OR a tiny server. Whatever is simplest and robust.
+- Dockerized: a `Dockerfile` + a one-command run (e.g. `docker build -t loki-observability . &&
+  docker run -p <port>:<port> loki-observability`) serving the dashboard. Pick a free port
+  (NOT 57374/57375 which loki uses); e.g. 58080.
+- A build/refresh script that regenerates data.json from the real sources so it stays current.
+- Place under `artifacts/observability/` (or `tools/observability/`).
+- OPEN it in the browser when up (macOS `open http://127.0.0.1:<port>`).
+
+## Validation (before calling done)
+- Every displayed number traces to a source file (spot-check 3).
+- The "mistakes" and "didn't work" sections are NON-EMPTY and real.
+- `docker run` serves the page; the page renders in the Autonomi look; no console errors.
+- No emojis / dashes. Screenshot the running page as proof.

package/loki-ts/dist/loki.js

@@ -1,5 +1,5 @@
 // @bun
-var X8=Object.defineProperty;var K8=($)=>$;function q8($,Q){this[$]=K8.bind(null,Q)}var b=($,Q)=>{for(var Z in Q)X8($,Z,{get:Q[Z],enumerable:!0,configurable:!0,set:q8.bind(Q,Z)})};var L=($,Q)=>()=>($&&(Q=$($=0)),Q);var J$=import.meta.require;var v1={};b(v1,{lokiDir:()=>P,homeLokiDir:()=>T$,findRepoRootForVersion:()=>e$,REPO_ROOT:()=>h});import{resolve as r,dirname as i$}from"path";import{fileURLToPath as J8}from"url";import{existsSync as x$}from"fs";import{homedir as V8}from"os";function W8(){let $=y1;for(let Q=0;Q<6;Q++){if(x$(r($,"VERSION"))&&x$(r($,"autonomy/run.sh")))return $;let Z=i$($);if(Z===$)break;$=Z}return r(y1,"..","..","..")}function e$($){let Q=$;for(let Z=0;Z<6;Z++){if(x$(r(Q,"VERSION"))&&x$(r(Q,"autonomy/run.sh")))return Q;let z=i$(Q);if(z===Q)break;Q=z}return r($,"..","..","..")}function P(){return process.env.LOKI_DIR??r(process.cwd(),".loki")}function T$(){return r(V8(),".loki")}var y1,h;var C=L(()=>{y1=i$(J8(import.meta.url));h=W8()});import{readFileSync as U8}from"fs";import{resolve as H8,dirname as G8}from"path";import{fileURLToPath as B8}from"url";function N$(){if(Q$!==null)return Q$;let $="7.104.5";if(typeof $==="string"&&$.length>0)return Q$=$,Q$;try{let Q=G8(B8(import.meta.url)),Z=e$(Q);Q$=U8(H8(Z,"VERSION"),"utf-8").trim()}catch{Q$="unknown"}return Q$}var Q$=null;var $1=L(()=>{C()});var c1={};b(c1,{runOrThrow:()=>x8,run:()=>F,readStreamCapped:()=>u1,commandVersion:()=>S8,commandExists:()=>f,ShellError:()=>Q1,MAX_STDOUT_BYTES:()=>f1});async function u1($,Q=f1){let Z=$.getReader(),z=new TextDecoder,X="",q=0;try{while(q<Q){let{done:K,value:W}=await Z.read();if(K)break;if(!W)continue;if(q+=W.byteLength,q>Q){let V=W.byteLength-(q-Q);X+=z.decode(W.subarray(0,V),{stream:!0});break}X+=z.decode(W,{stream:!0})}X+=z.decode()}finally{try{await Z.cancel()}catch{}Z.releaseLock()}return X}async function F($,Q={}){let Z=Bun.spawn({cmd:[...$],stdout:"pipe",stderr:"pipe",env:Q.env?{...process.env,...Q.env}:process.env,cwd:Q.cwd}),z,X;if(Q.timeoutMs&&Q.timeoutMs>0)z=setTimeout(()=>{try{Z.kill("SIGTERM")}catch{}X=setTimeout(()=>{try{Z.kill("SIGKILL")}catch{}},2000)},Q.timeoutMs);try{let[q,K,W]=await Promise.all([u1(Z.stdout),new Response(Z.stderr).text(),Z.exited]);return{stdout:q,stderr:K,exitCode:W}}finally{if(z)clearTimeout(z);if(X)clearTimeout(X)}}async function x8($,Q={}){let Z=await F($,Q);if(Z.exitCode!==0)throw new Q1(`command failed (${Z.exitCode}): ${$.join(" ")}`,Z.exitCode,Z.stdout,Z.stderr);return Z}async function f($){let Q=N8($),Z=await F(["sh","-c",`command -v ${Q}`],{timeoutMs:5000});if(Z.exitCode===0)return Z.stdout.trim()||null;return null}function N8($){if(!/^[A-Za-z0-9._/-]+$/.test($))throw Error(`refused to shell-escape suspect token: ${$}`);return $}async function S8($,Q="--version"){if(!await f($))return null;let z=await F([$,Q],{timeoutMs:5000});if(z.exitCode!==0)return null;return((z.stdout||z.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var f1=16777216,Q1;var d=L(()=>{Q1=class Q1 extends Error{message;exitCode;stdout;stderr;constructor($,Q,Z,z){super($);this.message=$;this.exitCode=Q;this.stdout=Z;this.stderr=z;this.name="ShellError"}}});function t($){return D8?"":$}var D8,T,S,_,nZ,I,k,y,J;var c=L(()=>{D8=(process.env.NO_COLOR??"").length>0;T=t("\x1B[0;31m"),S=t("\x1B[0;32m"),_=t("\x1B[1;33m"),nZ=t("\x1B[0;34m"),I=t("\x1B[0;36m"),k=t("\x1B[1m"),y=t("\x1B[2m"),J=t("\x1B[0m")});import{existsSync as c8}from"fs";async function Z$(){if(A$!==void 0)return A$;let $="/opt/homebrew/bin/python3.12";if(c8($))return A$=$,$;let Q=await f("python3.12");if(Q)return A$=Q,Q;let Z=await f("python3");return A$=Z,Z}async function z$($,Q={}){let Z=await Z$();if(!Z)return{stdout:"",stderr:"python3 not found",exitCode:127};return F([Z,"-c",$],Q)}var A$;var V$=L(()=>{d()});var J0={};b(J0,{runStatus:()=>U3});import{existsSync as v,readFileSync as U$,readdirSync as e1,statSync as $0}from"fs";import{resolve as D,basename as Q3}from"path";import{homedir as Z3}from"os";function Q0($){let Q=Math.trunc($);if(Q>=1e6)return`${(Math.trunc(Q/1e6*10)/10).toFixed(1)}M`;if(Q>=1000)return`${(Math.trunc(Q/1000*10)/10).toFixed(1)}K`;return String(Q)}function Z0($,Q,Z){if(Q===0)return null;let z=Math.trunc($*100/Q),X=Math.trunc($*D$/Q);if(X>D$)X=D$;let q=D$-X,K=S;if(z>=80)K=T;else if(z>=50)K=_;let W="=".repeat(Math.max(0,X))+" ".repeat(Math.max(0,q)),V=Q0($),U=Q0(Q);return`  ${k}${Z}${J} ${K}[${W}]${J} ${z}% (${V} / ${U})`}async function X3(){if(await f("jq"))return!0;return process.stdout.write(`${T}Error: jq is required but not installed.${J}
+var X8=Object.defineProperty;var K8=($)=>$;function q8($,Q){this[$]=K8.bind(null,Q)}var b=($,Q)=>{for(var Z in Q)X8($,Z,{get:Q[Z],enumerable:!0,configurable:!0,set:q8.bind(Q,Z)})};var L=($,Q)=>()=>($&&(Q=$($=0)),Q);var J$=import.meta.require;var v1={};b(v1,{lokiDir:()=>P,homeLokiDir:()=>T$,findRepoRootForVersion:()=>e$,REPO_ROOT:()=>h});import{resolve as r,dirname as i$}from"path";import{fileURLToPath as J8}from"url";import{existsSync as x$}from"fs";import{homedir as V8}from"os";function W8(){let $=y1;for(let Q=0;Q<6;Q++){if(x$(r($,"VERSION"))&&x$(r($,"autonomy/run.sh")))return $;let Z=i$($);if(Z===$)break;$=Z}return r(y1,"..","..","..")}function e$($){let Q=$;for(let Z=0;Z<6;Z++){if(x$(r(Q,"VERSION"))&&x$(r(Q,"autonomy/run.sh")))return Q;let z=i$(Q);if(z===Q)break;Q=z}return r($,"..","..","..")}function P(){return process.env.LOKI_DIR??r(process.cwd(),".loki")}function T$(){return r(V8(),".loki")}var y1,h;var C=L(()=>{y1=i$(J8(import.meta.url));h=W8()});import{readFileSync as U8}from"fs";import{resolve as H8,dirname as G8}from"path";import{fileURLToPath as B8}from"url";function N$(){if(Q$!==null)return Q$;let $="7.106.0";if(typeof $==="string"&&$.length>0)return Q$=$,Q$;try{let Q=G8(B8(import.meta.url)),Z=e$(Q);Q$=U8(H8(Z,"VERSION"),"utf-8").trim()}catch{Q$="unknown"}return Q$}var Q$=null;var $1=L(()=>{C()});var c1={};b(c1,{runOrThrow:()=>x8,run:()=>F,readStreamCapped:()=>u1,commandVersion:()=>S8,commandExists:()=>f,ShellError:()=>Q1,MAX_STDOUT_BYTES:()=>f1});async function u1($,Q=f1){let Z=$.getReader(),z=new TextDecoder,X="",q=0;try{while(q<Q){let{done:K,value:W}=await Z.read();if(K)break;if(!W)continue;if(q+=W.byteLength,q>Q){let V=W.byteLength-(q-Q);X+=z.decode(W.subarray(0,V),{stream:!0});break}X+=z.decode(W,{stream:!0})}X+=z.decode()}finally{try{await Z.cancel()}catch{}Z.releaseLock()}return X}async function F($,Q={}){let Z=Bun.spawn({cmd:[...$],stdout:"pipe",stderr:"pipe",env:Q.env?{...process.env,...Q.env}:process.env,cwd:Q.cwd}),z,X;if(Q.timeoutMs&&Q.timeoutMs>0)z=setTimeout(()=>{try{Z.kill("SIGTERM")}catch{}X=setTimeout(()=>{try{Z.kill("SIGKILL")}catch{}},2000)},Q.timeoutMs);try{let[q,K,W]=await Promise.all([u1(Z.stdout),new Response(Z.stderr).text(),Z.exited]);return{stdout:q,stderr:K,exitCode:W}}finally{if(z)clearTimeout(z);if(X)clearTimeout(X)}}async function x8($,Q={}){let Z=await F($,Q);if(Z.exitCode!==0)throw new Q1(`command failed (${Z.exitCode}): ${$.join(" ")}`,Z.exitCode,Z.stdout,Z.stderr);return Z}async function f($){let Q=N8($),Z=await F(["sh","-c",`command -v ${Q}`],{timeoutMs:5000});if(Z.exitCode===0)return Z.stdout.trim()||null;return null}function N8($){if(!/^[A-Za-z0-9._/-]+$/.test($))throw Error(`refused to shell-escape suspect token: ${$}`);return $}async function S8($,Q="--version"){if(!await f($))return null;let z=await F([$,Q],{timeoutMs:5000});if(z.exitCode!==0)return null;return((z.stdout||z.stderr).split(/\r?\n/)[0]?.trim()??"")||null}var f1=16777216,Q1;var d=L(()=>{Q1=class Q1 extends Error{message;exitCode;stdout;stderr;constructor($,Q,Z,z){super($);this.message=$;this.exitCode=Q;this.stdout=Z;this.stderr=z;this.name="ShellError"}}});function t($){return D8?"":$}var D8,T,S,_,nZ,I,k,y,J;var c=L(()=>{D8=(process.env.NO_COLOR??"").length>0;T=t("\x1B[0;31m"),S=t("\x1B[0;32m"),_=t("\x1B[1;33m"),nZ=t("\x1B[0;34m"),I=t("\x1B[0;36m"),k=t("\x1B[1m"),y=t("\x1B[2m"),J=t("\x1B[0m")});import{existsSync as c8}from"fs";async function Z$(){if(A$!==void 0)return A$;let $="/opt/homebrew/bin/python3.12";if(c8($))return A$=$,$;let Q=await f("python3.12");if(Q)return A$=Q,Q;let Z=await f("python3");return A$=Z,Z}async function z$($,Q={}){let Z=await Z$();if(!Z)return{stdout:"",stderr:"python3 not found",exitCode:127};return F([Z,"-c",$],Q)}var A$;var V$=L(()=>{d()});var J0={};b(J0,{runStatus:()=>U3});import{existsSync as v,readFileSync as U$,readdirSync as e1,statSync as $0}from"fs";import{resolve as D,basename as Q3}from"path";import{homedir as Z3}from"os";function Q0($){let Q=Math.trunc($);if(Q>=1e6)return`${(Math.trunc(Q/1e6*10)/10).toFixed(1)}M`;if(Q>=1000)return`${(Math.trunc(Q/1000*10)/10).toFixed(1)}K`;return String(Q)}function Z0($,Q,Z){if(Q===0)return null;let z=Math.trunc($*100/Q),X=Math.trunc($*D$/Q);if(X>D$)X=D$;let q=D$-X,K=S;if(z>=80)K=T;else if(z>=50)K=_;let W="=".repeat(Math.max(0,X))+" ".repeat(Math.max(0,q)),V=Q0($),U=Q0(Q);return`  ${k}${Z}${J} ${K}[${W}]${J} ${z}% (${V} / ${U})`}async function X3(){if(await f("jq"))return!0;return process.stdout.write(`${T}Error: jq is required but not installed.${J}
 `),process.stdout.write(`Install with:
 `),process.stdout.write(`  brew install jq    (macOS)
 `),process.stdout.write(`  apt install jq     (Debian/Ubuntu)
@@ -805,4 +805,4 @@ Set LOKI_LEGACY_BASH=1 to force the bash CLI for every command.
 `),2}default:return process.stderr.write(`Unknown command: ${Q}
 `),process.stderr.write(z8),2}}i1();process.on("SIGINT",()=>process.exit(130));process.on("SIGTERM",()=>process.exit(143));var EZ=await RZ(Bun.argv.slice(2));process.exit(EZ);
 
-//# debugId=CD7A0B6FD83A195A64756E2164756E21
+//# debugId=FD93D5FC3483E3A564756E2164756E21

package/mcp/__init__.py

@@ -57,4 +57,4 @@ try:
 except ImportError:
     __all__ = ['mcp']
 
-__version__ = '7.104.5'
+__version__ = '7.106.0'

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "loki-mode",
   "mcpName": "io.github.asklokesh/loki-mode",
-  "version": "7.104.5",
+  "version": "7.106.0",
   "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": "7.104.5",
+  "version": "7.106.0",
   "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",