loki-mode 8.6.0 → 8.6.1

package/SKILL.md

@@ -3,7 +3,7 @@ name: loki-mode
 description: Autonomous spec-driven build system with a built-in trust layer. It does not call work done until it is verified (RARV-C closure loop, 8 quality gates, completion council, verified-completion evidence gate). Triggers on "Loki Mode". Takes a spec (PRD, GitHub issue, OpenAPI doc, etc.) to deployed product with minimal human intervention. Provider-agnostic. Requires --dangerously-skip-permissions flag.
 ---
 
-# Loki Mode v8.6.0
+# Loki Mode v8.6.1
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -469,4 +469,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v8.6.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v8.6.1 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-8.6.0
+8.6.1

package/autonomy/loki

@@ -1017,6 +1017,7 @@ show_help() {
     echo "Cost controls (budget, iteration and wall-clock caps): docs/cost-controls.md"
     echo "Already have a codebase? Start read-only: docs/brownfield.md"
     echo "No egress allowed? What runs disconnected: docs/air-gapped.md"
+    echo "Where this is going: docs/STRATEGY-2026-2028.md"
     echo ""
     echo "Scripting Loki? Exit codes for every command are documented in"
     echo "docs/exit-codes.md. Severity rises with the code, and 'loki start'"

package/autonomy/run.sh

@@ -13321,6 +13321,17 @@ ${dependency_context}"
             "diff_bytes=$_review_diff_bytes" \
             "limit_bytes=$_review_max_bytes" \
             "iteration=${ITERATION_COUNT:-0}" 2>/dev/null || true
+        # Classify this as INFRASTRUCTURE, not a finding. The discriminator is
+        # otherwise set at :14912, which this early return never reaches, so the
+        # variable stayed "" from the reset at :12959 and the consumer at :22126
+        # read a file-size condition as "Critical/High findings" -- escalating to
+        # PAUSE over a diff that was merely too large to send.
+        #
+        # Still fail-closed: the review did NOT pass, and nothing here converts a
+        # skipped review into a green one. It only records WHY it could not run,
+        # which is the difference between "your code is bad" and "we could not
+        # look at it".
+        _LOKI_REVIEW_FAILURE_KIND="infrastructure_inconclusive"
         return 1
     fi
 

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "8.6.0"
+__version__ = "8.6.1"
 
 # Expose the control app for easy import
 try:

package/docs/STRATEGY-2026-2028.md

@@ -0,0 +1,124 @@
+# Two-year adoption strategy
+
+Written 2026-07-31 from measured inputs: registry download data, competitor
+documentation fetched the same day, and benchmark results from this repository.
+Where a claim is inference rather than measurement, it says so.
+
+## The one number that matters
+
+**Floor: ~94 downloads/day. Peak: 1,263.** The 13x swing tracks our own
+release activity -- eight releases landed on 07-30. A curve that rises when we
+publish and falls when we stop is CI and mirrors, not word of mouth.
+
+Organic growth is a **rising floor**. Everything below is judged against that
+single number, measured on days we ship nothing.
+
+Two years from now the question is not "how many releases did we cut." It is
+"what is the floor, and does it rise when we are quiet."
+
+## What we actually sell, stated so it survives a demo
+
+Competitor documentation, fetched 2026-07-31:
+
+- **Lovable** runs a security scan on every publish and admins can block the
+  publish outright.
+- **Claude Code** has a review step that checks findings against actual code
+  behavior.
+- **Replit** says its agent tests its own work.
+
+So **"we verify and they don't" is false**, and a founder demo against Lovable
+would expose it. That framing is retired.
+
+What is true and unoccupied across all seven competitors: **nobody ships a
+persisted, portable, diff-bound artifact.** Theirs live in a dashboard --
+Lovable's is a findings count in a dialog, Claude Code's check run is
+deliberately non-blocking. Ours is a file: bound to a diff by `diff_sha256`,
+recording what was NOT proven as prominently as what was, verifiable by someone
+who never installed us.
+
+That is the sentence. Portable, diff-bound, honest about gaps.
+
+## The second thing we sell, now measured
+
+**The harness carries quality, not the model.** On `hard-2-ledger`, a task
+authored so a bare model fails it:
+
+| arm | result | cost |
+|---|---|---|
+| haiku, harness off | 1/4 passed | $0.86 |
+| haiku, harness on | 1/1 passed | $0.54 |
+
+Same model. Same prompt. The harness is the only variable. A correct
+implementation cost **less** than the failing ones, because a cheap failure is
+not cheap.
+
+Caveat, stated because it will be checked: n is small and trials are still
+accumulating. This demonstrates the mechanism. The rate needs more trials, and
+those are now worth buying -- before this task existed, the baseline passed
+everything and more trials bought precision around a ceiling.
+
+## Where we win, and where we should not fight
+
+**Do not fight on:** hosted preview URLs, visual editing, zero-install browser
+onboarding, managed backend primitives. Those are structural properties of a
+hosted product. Lovable publishes to `[name].lovable.app` free at zero credit
+balance; we cannot and should not try.
+
+**Win on the three axes competitors structurally cannot occupy:**
+
+1. **Air-gapped operation.** Measured with egress severed: `version`, `doctor`,
+   `plan --json`, `proof verify`, `heal --assess` all return real results. No
+   competitor can do this -- "Devin's brain always resides within Cognition's
+   Cloud." For defence, government, and regulated banking this is winnable on
+   this axis alone. One required egress (model inference), disclosed.
+
+2. **In-place brownfield.** Lovable **cannot import an existing repository at
+   all**. Replit and Cursor import into *their* environment. For a private
+   monorepo with internal dependencies, that is frequently not permitted. We
+   run where the code already lives.
+
+3. **Cost per correct result.** Haiku-plus-harness beat opus-baseline in the
+   aggregate at roughly an eighth the cost. If that holds under more trials, it
+   is a procurement argument, not a benchmark curiosity.
+
+## The two-year sequence
+
+**Year 1, first half -- earn the floor.** Every item judged by whether a first
+run reaches a result. Ship `first_run_blocked` (done, v8.6.0), read what it
+says, fix the top blocker, repeat. The floor is the scoreboard.
+
+Concretely already done and testable: 43 of 112 commands were unreachable from
+`loki help` including `loki proof`; a first-run dead end on provider-less hosts;
+`loki proof md` so the receipt travels into a PR or Slack.
+
+**Year 1, second half -- make the receipt the artefact people forward.** A
+receipt in a PR is a person showing a colleague. That is the only word-of-mouth
+mechanic available to a CLI, and it costs no infrastructure.
+
+**Year 2 -- enterprise pull, not push.** Air-gapped + in-place brownfield +
+signed provenance is a procurement story no competitor can match today. It sells
+to the buyer who cannot use the others at all, and those buyers talk to each
+other.
+
+## What would falsify this
+
+Stated so it is checkable rather than reassuring:
+
+- **The floor does not rise** over the next quarter despite first-run fixes ->
+  the bottleneck is not discoverability, and this plan is wrong.
+- **`first_run_blocked` shows trials dying on something we did not predict** ->
+  follow the data, not this document.
+- **The harness lift does not survive more trials** -> the cost argument
+  collapses and the differentiator narrows to portability alone.
+- **A competitor ships a portable signed receipt** -> the wedge is gone and we
+  compete on cost and air-gap only.
+
+## What is deliberately not here
+
+No 80-item backlog. The items that exist are the ones with a measured
+mechanism. Padding this list to look comprehensive would be the fabrication
+this project has already paid for once.
+
+Release cadence is explicitly **not** a growth lever: eight releases in one day
+produced a 1,263 spike and a 94 floor. If the floor is the goal, cadence is
+noise.

package/docs/adoption-baseline-2026-07-31.md

@@ -0,0 +1,84 @@
+# Adoption baseline, 2026-07-31
+
+The first numbers in this project taken from the registry rather than from
+intuition. Recorded so the next measurement has something to compare against.
+
+## What we actually have
+
+**4,456 npm downloads in the last 7 days.** Daily:
+
+| Day | Downloads |
+|---|---|
+| 07-24 | 863 |
+| 07-25 | 1,160 |
+| 07-26 | 196 |
+| 07-27 | 94 |
+| 07-28 | 386 |
+| 07-29 | 494 |
+| 07-30 | 1,263 |
+
+Two things follow, and the second matters more than the first.
+
+**1. We have users.** Roughly 4.5k downloads a week is not a project nobody
+has heard of. Every strategy discussion in this repo has proceeded as though
+adoption were hypothetical. It is not.
+
+**2. The shape is release-driven, not organic.** The 13x swing between 07-27
+(94) and 07-30 (1,263) tracks publishing activity -- eight releases landed on
+07-30 alone. A curve that rises when we publish and falls when we stop is
+mirrors and CI, not word of mouth. Organic growth would show a floor that
+rises over time; this shows a floor near 94.
+
+**Do not read 4,456 as 4,456 humans.** npm counts mirrors, CI, and Docker
+layer pulls. The honest statement is that the ceiling is real and the floor is
+what needs to move.
+
+## Why the floor is the metric
+
+The founder's goal is word-of-mouth growth over two years. The number that
+measures it is the **trough**, not the peak: how many installs happen on a day
+we publish nothing. Today that is ~94.
+
+Peaks are bought with releases. Floors are earned by people telling other
+people. Every adoption item should be judged against whether it moves the
+floor.
+
+## What we still cannot see, and what changed today
+
+Until v8.6.0 shipped this morning, we could see that a first run was ATTEMPTED
+and nothing about whether it succeeded. `first_run_blocked` (v8.6.0) now names
+the class of dependency that stops a first run -- enum-clamped, once per
+install, strict opt-in.
+
+That data does not exist yet: the release is hours old and the telemetry is
+off by default behind a second opt-in. It will accumulate slowly and from a
+minority of users, which is the correct trade for not exfiltrating anyone's
+environment.
+
+So the sequence is: floor today ~94/day -> ship things that plausibly move it
+-> watch the floor, not the peak.
+
+## The three things measured this session that plausibly move it
+
+Ranked by how directly they affect someone's first ten minutes:
+
+1. **43 of 112 commands were unreachable from `loki help`**, including
+   `loki proof` -- the Evidence Receipt, the thing the product argues on. Fixed
+   and gated. A user who cannot find the differentiator does not repeat it to
+   anyone.
+2. **A first-run dead end on hosts with no provider CLI.** One route named the
+   blockers and pointed at `loki tour` (no provider, no key, no spend); the
+   other said "some required prerequisites are missing" and stopped. That is
+   the exact moment an evaluator decides whether to continue.
+3. **`loki proof md`** puts the receipt in a form a person can paste into a PR
+   or a Slack message. Competitors' verification output lives in their
+   dashboard; a file is the only artifact that travels.
+
+None of these is proven to move the floor. They are the candidates with a
+plausible mechanism, and the floor is now being watched.
+
+## Reproduce
+
+```sh
+curl -s "https://api.npmjs.org/downloads/range/last-week/loki-mode"
+```