loki-mode 9.8.0 → 9.12.0

package/README.md

@@ -15,7 +15,7 @@ _The free, source-available autonomous coding agent by [Autonomi](https://www.au
 
 [Website](https://www.autonomi.dev/) | [Documentation](wiki/Home.md) | [Installation](docs/INSTALLATION.md) | [Changelog](CHANGELOG.md) | [Purple Lab -- deprecated v7.44.0](#purple-lab)
 
-**Current release: v8.5.2**
+**Current release: v9.8.1**
 
 </div>
 
@@ -194,7 +194,7 @@ public key can verify offline. See [docs/SIGNED-RECEIPTS.md](docs/SIGNED-RECEIPT
 - **Legacy system healing** -- `loki modernize heal` archaeology/stabilize/isolate/modernize/validate phases (v6.67.0, see `skills/healing.md`)
 - **MCP server** -- 34 tools (including ChromaDB code search) plus 3 resources and 2 prompts (`mcp/server.py`, with magic tools registered from `mcp/magic_tools.py` and the managed-memory tool from `mcp/managed_tools.py`). Of the 34, 33 are always available; `loki_memory_redact` is registered but only succeeds when `LOKI_MANAGED_AGENTS=true` and `LOKI_MANAGED_MEMORY=true`. Launch with `loki mcp` (bootstraps the Python MCP SDK on first run).
 - **Full-stack output** -- Source code, tests, Docker Compose stacks (multi-service with healthchecks), CI/CD pipelines, audit logs
-- **Provider-agnostic** -- runs on Claude, Codex, Cline, or Aider with automatic failover (`loki-ts/src/runner/providers.ts`); no vendor lock-in. Gemini CLI deprecated v7.5.18.
+- **Provider-agnostic** -- runs on Claude, Cline, Codex, Aider, or opencode with automatic failover (`loki-ts/src/runner/providers.ts`); no vendor lock-in. When `LOKI_PROVIDER` is unset, Loki auto-detects in that order; an explicit choice always wins and is never silently substituted. Gemini was removed as a provider in v7.5.18.
 - **Source-available (BUSL-1.1)** -- Free for personal, internal, and academic use.
 
 ---
@@ -410,7 +410,7 @@ loki doctor                                    # check your setup before the fir
 
 Required:
 
-- An agent provider CLI: [Claude Code](https://docs.claude.com/en/docs/claude-code) (`claude`, Tier 1, recommended and E2E-verified - the provider Loki Mode is built for). Codex, Cline, and Aider are supported as experimental providers (wiring in place; not yet E2E-verified by us). Loki cannot run a build without one of these installed and authenticated.
+- An agent provider CLI: [Claude Code](https://docs.claude.com/en/docs/claude-code) (`claude`, Tier 1, recommended and E2E-verified - the provider Loki Mode is built for). Cline, Codex, Aider, and opencode are supported as experimental providers (wiring in place; not yet E2E-verified by us). Loki cannot run a build without one of these installed and authenticated.
 - Python 3.10+ (`python3`) for the dashboard, memory system, and orchestration helpers.
 - Git 2.x (`git`) for checkpoints and worktrees.
 - `curl` for installation and network calls.
@@ -656,7 +656,7 @@ The historical feature set (platform pages, Monaco IDE workspace, AI chat panel)
 | Feature | Loki Mode | bolt.new | Replit | Lovable |
 |---------|:---------:|:--------:|:------:|:-------:|
 | Self-hosted / your keys | Yes | No | No | No |
-| Multi-provider failover (4 providers) | Yes | No | No | No |
+| Multi-provider failover (5 providers) | Yes | No | No | No |
 | 8 quality gates | Yes | No | No | No |
 | Blind code review | Yes | No | No | No |
 | Enterprise auth (OIDC token + scoped RBAC) | Yes | No | Yes | No |
@@ -665,22 +665,23 @@ The historical feature set (platform pages, Monaco IDE workspace, AI chat panel)
 | Source-available (BUSL-1.1) | Yes | No | No | No |
 | Free tier | Source-available | Yes | Yes | Yes |
 
-Loki Mode is the only platform that is fully self-hosted, source-available (BUSL-1.1), and includes automated quality verification. Your code, your keys, your infrastructure.
+Among the four tools in this table, Loki Mode is the one that is fully self-hosted, source-available (BUSL-1.1), and includes automated quality verification. Your code, your keys, your infrastructure. We have not surveyed every tool on the market, so read this as a comparison against the named three, not a claim about the whole category.
 
 ---
 
 <details>
 <summary><strong>Provider matrix -- per-provider status, autonomous flags, parallelism, install (includes deprecated Gemini)</strong></summary>
 
-Loki's autonomy and quality loop are the product; the underlying coding CLI is swappable. Loki runs on any of the providers below so you are never locked to one vendor.
+Loki's autonomy and quality loop are the product; the underlying coding CLI is swappable. Loki runs on any of the providers below so you are never locked to one vendor. With `LOKI_PROVIDER` unset, Loki auto-detects the first installed provider in the order the table lists (claude, cline, codex, aider, opencode); setting it explicitly always wins and is never silently substituted.
 
 | Provider | Status | Autonomous Flag | Parallel Agents | Install |
 |----------|--------|:-:|:-:|---------|
 | **Claude Code** | Active (Tier 1, E2E-verified) | `--dangerously-skip-permissions` | Yes (10+) | `npm i -g @anthropic-ai/claude-code` |
-| **Codex CLI** | Experimental (Tier 3) | `--sandbox workspace-write --skip-git-repo-check` | Sequential | `npm i -g @openai/codex` |
-| **Cline CLI** | Experimental (Tier 2) | `-y` | Sequential | `npm i -g @anthropic-ai/cline` |
+| **Cline CLI** | Experimental (Tier 2) | `-y` | Sequential | `npm install -g cline` |
+| **Codex CLI** | Experimental (Tier 3) | `exec --sandbox workspace-write --skip-git-repo-check` | Sequential | `npm i -g @openai/codex` |
 | **Aider** | Experimental (Tier 3) | `--yes-always` | Sequential | `pip install aider-chat` |
-| **Google Gemini CLI** | DEPRECATED v7.5.18 | -- | -- | Upstream deprecated; runtime removed. `LOKI_PROVIDER=gemini` exits with migration message. |
+| **opencode** | Experimental | `--auto` | Sequential | `npm install -g opencode-ai` |
+| **Google Gemini CLI** | REMOVED v7.5.18 | -- | -- | Upstream deprecated; runtime removed. `LOKI_PROVIDER=gemini` exits with a migration message. |
 
 Status legend: "E2E-verified" means we run real spec-to-code builds on it ourselves. Claude Code is the primary, fully supported provider and the one Loki Mode is built for; it gets full features (subagents, parallelization, MCP, Task tool). "Experimental" means the wiring is in place but we have not produced an end-to-end verified build ourselves; treat as community-tested. Experimental providers run sequentially. Auto-failover switches providers when rate-limited. See [Provider Guide](skills/providers.md).
 
@@ -753,6 +754,7 @@ Loki Mode's accuracy and autonomy behaviors are default-on. Each is an opt-out e
 | `LOKI_CONFIDENCE_SPIKE` | `1` (on) | Forces one EXTRA verification pass when the agent's self-reported confidence spikes, instead of trusting the claim. Strictly additive -- it can never skip a gate. Set `0` to opt out; tune with `LOKI_CONFIDENCE_SPIKE_DELTA` (default `40`) and `LOKI_CONFIDENCE_SPIKE_MIN` (default `90`). |
 | `LOKI_GOAL_SCORING` | `1` (on) | Flags a goal with no measurable success condition and asks for a threshold, metric, or concrete artifact. Advisory only -- never blocks a build or rewrites the goal. Set `0` to opt out. |
 | `LOKI_SMART_RETRY` | `1` (on) | Stops early on a positively-identified permanent failure (bad credentials, unknown model, exhausted quota) rather than burning retries. Unrecognized errors and rate limits still retry as before. Set `0` to retry every failure. |
+| `LOKI_SIMPLE` | `0` (off) | EXPERIMENTAL. Strips the coaching half of the system prompt -- the RARV cycle, SDLC phases and memory habits that a frontier model already does natively. Per-iteration state (which gate failed, self-heal output, checklist status) is never touched, because that is information the model cannot derive. Measured at -78% prompt size, ~1562 tokens per iteration, on both the bash and Bun routes. INERT on degraded providers (Codex, Aider): those take an earlier return path whose prompt is already minimal by design, so the flag has nothing to strip there -- a measured zero, not an untested case. Whether it changes build speed or quality is NOT yet measured, so treat it as an experiment, not a tuning knob: run `benchmarks/run-prompt-ablation.sh` on your own workload before adopting it. |
 
 This is a subset. See the [wiki](wiki/Home.md) for the full env-var reference and the RARV-C closure knobs (`LOKI_INJECT_FINDINGS`, `LOKI_OVERRIDE_COUNCIL`, `LOKI_AUTO_LEARNINGS`, `LOKI_HANDOFF_MD`).
 
@@ -776,12 +778,15 @@ See [BMAD Integration Validation](docs/architecture/bmad-integration-validation.
 <details>
 <summary><strong>Enterprise Features</strong></summary>
 
-Enterprise features are included but require env var activation. Self-audit: 35/45 capabilities working, 0 broken, 1,314 tests passing.
+Enterprise features are included but require env var activation.
 
 ```bash
-export LOKI_TLS_ENABLED=true
-export LOKI_OIDC_PROVIDER=google
-export LOKI_AUDIT_ENABLED=true
+export LOKI_ENTERPRISE_AUTH=true                 # token auth (dashboard/auth.py)
+export LOKI_OIDC_ISSUER=https://accounts.google.com
+export LOKI_OIDC_CLIENT_ID=your-client-id        # OIDC needs issuer + client id
+export LOKI_ENTERPRISE_AUDIT=true                # force audit logging on
+export LOKI_TLS_CERT=/path/cert.pem              # HTTPS: set BOTH cert and key
+export LOKI_TLS_KEY=/path/key.pem
 loki enterprise status
 ```
 
@@ -823,7 +828,7 @@ See [benchmarks/](benchmarks/) for methodology.
 | **Code Gen** | Full-stack apps from PRDs | Complex domain logic may need human review |
 | **Deploy** | Generates configs, Dockerfiles, CI/CD; `loki deploy` prints the exact deploy command | Does not deploy -- human runs the printed deploy command (Loki never runs a cloud CLI or git push) |
 | **Testing** | 8 automated quality gates | Test quality depends on AI assertions |
-| **Providers** | 4 providers with auto-failover | Non-Claude providers lack parallel agents |
+| **Providers** | 5 providers with auto-failover | Non-Claude providers lack parallel agents |
 | **Dashboard** | Real-time single-machine monitoring | No multi-node clustering |
 
 > **What "autonomous" means:** The system runs RARV cycles without prompting. It does NOT access your cloud accounts, payment systems, or external services unless you provide credentials. Human oversight is expected for deployment, API keys, and critical decisions.

package/SKILL.md

@@ -3,7 +3,7 @@ name: loki-mode
 description: Autonomous spec-driven build system with a built-in trust layer. It does not call work done until it is verified (RARV-C closure loop, 8 quality gates, completion council, verified-completion evidence gate). Triggers on "Loki Mode". Takes a spec (PRD, GitHub issue, OpenAPI doc, etc.) to deployed product with minimal human intervention. Provider-agnostic. Requires --dangerously-skip-permissions flag.
 ---
 
-# Loki Mode v9.8.0
+# Loki Mode v9.11.0
 
 **You are an autonomous agent. You make decisions. You do not ask questions. You do not stop.**
 
@@ -369,6 +369,7 @@ that cannot succeed.
 | `LOKI_CONFIDENCE_SPIKE_MIN` | `90` | Absolute level that counts as a spike on first arrival |
 | `LOKI_GOAL_SCORING=0` | on | Disable the goal-measurability advisory |
 | `LOKI_SMART_RETRY=0` | on | Retry every failure, including non-retryable ones |
+| `LOKI_SIMPLE=1` | off | Strip the coaching half of the system prompt (-78%, ~1562 tokens/iteration). Experimental ablation arm. |
 
 - **Prompt-cache discipline.** The prompt is split into a cache-stable
   `<loki_system>` prefix and a volatile `<dynamic_context>` tail at an explicit
@@ -469,4 +470,4 @@ See `CHANGELOG.md` entries [7.5.7], [7.5.8], [7.5.13] for the per-fix list and r
 
 ---
 
-**v9.8.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**
+**v9.11.0 | [Autonomi](https://www.autonomi.dev/) flagship product | ~410 lines core**

package/VERSION

@@ -1 +1 @@
-9.8.0
+9.12.0

package/autonomy/loki

@@ -19316,7 +19316,40 @@ main() {
                 # instead of re-checking every time a command is added.
                 local _help_target="$1"
                 shift
-                LOKI_HELP_ONLY=1 "$0" "$_help_target" --help "$@"
+                # FORK BOMB GUARD, and this is not hypothetical: it filled the
+                # process table with 3,788 chained `bash autonomy/loki --help
+                # --help` processes on this machine, after which even `pkill`
+                # could not fork.
+                #
+                # `loki help --help` sets _help_target="--help" and delegates as
+                # `"$0" --help --help`. That re-enters THIS case, and "$@"
+                # accumulates one more --help every pass, so the recursion is
+                # unbounded and spawns a process per level.
+                #
+                # LOKI_HELP_ONLY was described above as making delegation "safe
+                # by construction". It does -- against SIDE EFFECTS. It says
+                # nothing about SELF-REFERENCE, which is a different hazard: the
+                # danger is not that a command does something, it is that the
+                # help router asks itself for help forever.
+                #
+                # A help-shaped target has no command help to delegate to, so
+                # print the front page directly instead of recursing.
+                case "$_help_target" in
+                    --help|-h|help|--h|-help)
+                        show_help
+                        ;;
+                    *)
+                        # Second, independent stop: if we are ALREADY inside a
+                        # delegated help call, do not delegate again. Belt and
+                        # braces, because the pattern list above is a blacklist
+                        # and a future spelling would slip past it.
+                        if [ "${LOKI_HELP_ONLY:-}" = "1" ]; then
+                            show_help
+                        else
+                            LOKI_HELP_ONLY=1 "$0" "$_help_target" --help "$@"
+                        fi
+                        ;;
+                esac
             else
                 show_help
             fi
@@ -33430,6 +33463,7 @@ cmd_bench() {
             echo "  vs <task>       Run all configured tools on a task-spec (head-to-head)"
             echo "  list            List available task-specs"
             echo "  verify <file>   Recompute task_hash + check tool versions for a result.json"
+            echo "  oracles         Are the graders trustworthy? Hash-bound, no spend, no provider"
             echo "  report <files>  Build results.json + RESULTS.md from per-tool result-rows"
             echo ""
             echo "Options:"
@@ -33623,6 +33657,10 @@ cmd_proof() {
             echo ""
             echo "Subcommands:"
             echo "  list                 List proof-of-run artifacts in .loki/proofs/"
+            echo "  releases [--json]    Release history from git tags; version_is_ahead"
+            echo "                       reads UNKNOWN when it cannot be computed"
+            echo "  phases [--json]      Measured phase history from real phase_change"
+            echo "                       events; unrecorded boundaries read UNKNOWN"
             echo "  show <id>            Pretty-print .loki/proofs/<id>/proof.json"
             echo "  verify <id>          Re-check a receipt against the repo (tamper + drift);"
             echo "                       exit 0 clean, 1 tamper/drift. Verify it yourself."
@@ -33645,6 +33683,89 @@ cmd_proof() {
             [ "$sub" = "" ] && exit 1
             exit 0
             ;;
+        releases)
+            # CLI PARITY with GET /api/operator/releases. Same reader
+            # (dashboard/api_releases.py), so the two surfaces cannot disagree
+            # about which versions shipped.
+            _rel_json=0
+            [ "${1:-}" = "--json" ] && _rel_json=1
+            python3 - "$_rel_json" "$SKILL_DIR" <<'PYREL'
+import json, os, sys
+sys.dont_write_bytecode = True
+# Resolve the repo from THIS SCRIPT's location, not the cwd. Using getcwd()
+# made `loki proof releases` fail with "No module named dashboard" whenever it
+# was run from any other directory -- and it exited 0 while doing so, which is
+# a failure reported as success.
+# SKILL_DIR is the CLI's own install root, computed at loki:216. sys.argv[0]
+# is "-" inside a heredoc, so deriving the path from it does not work.
+_repo = sys.argv[2] if len(sys.argv) > 2 else ""
+for _cand in (_repo, os.getcwd()):
+    if os.path.isdir(os.path.join(_cand, "dashboard")):
+        sys.path.insert(0, _cand)
+        break
+as_json = sys.argv[1] == "1"
+try:
+    from dashboard import api_releases
+except Exception as exc:
+    sys.stderr.write("release history unavailable: %s\n" % exc)
+    raise SystemExit(2)
+env = api_releases.list_releases(os.getcwd())
+if as_json:
+    print(json.dumps(env, indent=2))
+    raise SystemExit(0)
+rows = env.get("releases") or []
+if not rows:
+    print("no releases readable: %s" % (env.get("reason") or "unknown"))
+    raise SystemExit(3)
+# version_is_ahead is None when it could not be computed. None is NOT False:
+# "I cannot compare" is a different claim from "it is not ahead".
+ahead = env.get("version_is_ahead")
+print("VERSION %s | newest tag %s | ahead: %s" % (
+    env.get("version") or "UNKNOWN",
+    env.get("newest_tag") or "UNKNOWN",
+    "UNKNOWN" if ahead is None else ("yes" if ahead else "no")))
+for r in rows:
+    print("  %-14s %s%s" % (r.get("tag"), r.get("date") or "date UNKNOWN",
+                            "  (current)" if r.get("is_current") else ""))
+PYREL
+            return $?
+            ;;
+        phases)
+            # CLI PARITY with GET /api/operator/phases. Both call the SAME
+            # reader (dashboard/api_phases.py), so the two surfaces cannot
+            # drift into disagreeing about the same run.
+            _phases_json=0
+            [ "${1:-}" = "--json" ] && _phases_json=1
+            python3 - "$loki_dir" "$_phases_json" <<'PYPHASES'
+import json, os, sys
+sys.dont_write_bytecode = True
+sys.path.insert(0, os.getcwd())
+loki_dir, as_json = sys.argv[1], sys.argv[2] == "1"
+try:
+    from dashboard import api_phases
+except Exception as exc:
+    sys.stderr.write("phase history unavailable: %s\n" % exc)
+    raise SystemExit(2)
+env = api_phases.phase_history(loki_dir)
+if as_json:
+    print(json.dumps(env, indent=2))
+    raise SystemExit(0)
+segs = env.get("segments") or []
+if not segs:
+    # An empty result states WHY: "no phases" and "could not read" are
+    # different operator situations and must not print identically.
+    print("no measured phase history: %s" % (env.get("reason") or "unknown"))
+    raise SystemExit(3)
+lead = env.get("leading_phase")
+rows = ([lead] if lead else []) + list(segs)
+for seg in rows:
+    print("  %-12s start=%-14s end=%s" % (
+        seg.get("phase"),
+        seg.get("start") if seg.get("start") is not None else "UNKNOWN",
+        seg.get("end") if seg.get("end") is not None else "UNKNOWN"))
+PYPHASES
+            return $?
+            ;;
         list)
             if [ ! -d "$proofs_dir" ]; then
                 echo -e "${YELLOW}No proofs found.${NC} Run 'loki start' to generate one."

package/autonomy/run.sh

@@ -19874,8 +19874,47 @@ except Exception:
 
     # STATIC PREFIX (cache-stable across iterations).
     # Order is deterministic so the prefix is byte-identical for iter N and N+1.
+    #
+    # LOKI_SIMPLE=1 -- THE ABLATION ARM. Default off; the emitted bytes are
+    # unchanged unless it is explicitly set, so parity fixtures do not move.
+    #
+    # WHY THIS EXISTS. Every instruction below was written to correct a model
+    # that needed correcting. Anthropic deleted ~80% of Claude Code's system
+    # prompt for Opus 5 on the finding that the corrections had become dead
+    # weight -- and that the model measured slightly MORE capable without them.
+    # Their method was ablation: delete, then add back only what a measured
+    # failure demands. Nothing here had ever been measured at all.
+    #
+    # THE DISTINCTION THIS FLAG IS BUILT AROUND, and the reason it strips the
+    # prefix while leaving the tail completely alone:
+    #
+    #   The prefix is COACHING -- how to work. "Use a Reason-Act-Reflect-Verify
+    #   cycle", "execute all SDLC phases", "consult memory". A frontier model
+    #   does these natively; being told costs attention and buys nothing.
+    #
+    #   The tail is STATE -- what happened. Which gate failed, what the
+    #   self-heal found, what the checklist still shows open. That is
+    #   information the model cannot derive from anywhere else, and deleting
+    #   it would be deleting the run's memory, not its lecture.
+    #
+    # So this ablates coaching ONLY. The dynamic tail below is untouched, and
+    # so is every gate, receipt, and verification path: the trust core is not
+    # prompt correction, and it is never an ablation arm.
+    #
+    # prd_anchor stays in both arms -- it names the task, which is the one
+    # thing the model genuinely cannot infer.
+    #
+    # MEASURED, with its provenance: 8090 -> 1776 bytes (-78%) from a LIVE
+    # build_prompt call under the fixture-1 ENVIRONMENT with a gate-failure
+    # file present -- not from the fixture file itself, which is 7909 bytes.
+    # The distinction matters because a number attributed to the wrong source
+    # cannot be reproduced by the next person who tries.
+    #
+    # The strip is bounded: only the block below is gated, so the anchor and
+    # the surrounding tags survive. Prefix size is a CEILING on the saving.
     printf '<loki_system>\n'
     printf '%s\n' "$prd_anchor"
+    if [ "${LOKI_SIMPLE:-0}" != "1" ]; then
     printf '%s\n' "$rarv_instruction"
     printf '%s\n' "$sdlc_instruction"
     printf '%s\n' "$autonomous_suffix"
@@ -19885,6 +19924,7 @@ except Exception:
     printf '%s\n' "$compose_instruction"
     printf '%s\n' "$lsp_grounding_instruction"
     printf '%s\n' "$agents_md_instruction"
+    fi
     # v8 (3c): goal-measurability advisory. Empty (and therefore not emitted at
     # all) for a measurable goal, an absent goal, or perpetual mode. Sits in the
     # static prefix because COMPLETION_PROMISE is fixed for the run, so it stays
@@ -25730,12 +25770,19 @@ except Exception:
         case "$_final_status" in
             council_approved|council_force_approved|deterministic_gates_passed|completion_promise_fulfilled|paused|interrupted|stopped)
                 result=0 ;;
-            # force_stopped belongs HERE too, for the same reason. A council
+            # force_stopped is in the result=20 arm below, NOT here. A council
             # force-stop (stagnation, or a flood of done-signals) means the run
             # gave up WITHOUT verifying the work -- the code already says so in
             # its header, its warning, and its refusal to open a PR. Reporting
             # it as a clean stop made it indistinguishable from success to the
             # only consumer that matters to automation: the exit code.
+            #
+            # This comment previously read "belongs HERE too" while the status
+            # appeared in NEITHER arm, so it fell through to `*)` and returned
+            # the incoming code unchanged -- a force-stop after a nonzero
+            # iteration exited nonzero, and one after a zero exited 1. The
+            # diagnosis was written and never applied; the wording is corrected
+            # here so the comment cannot be read as describing current behavior.
             # budget_exceeded belongs HERE, not with the human-controlled stops.
             # It sat in the result=0 arm on the rationale that "a human will
             # resume", which is true of `paused` (a human pressed pause) and
@@ -25752,7 +25799,7 @@ except Exception:
             # The operator raises the cap (or narrows the spec) and submits a
             # NEW Job -- the same remedy as max_iterations_reached, which is why
             # it shares that code.
-            failed|max_iterations_reached|max_retries_exceeded|budget_exceeded|max_duration_reached|policy_blocked|inconclusive_spec_contradiction)
+            failed|max_iterations_reached|max_retries_exceeded|budget_exceeded|max_duration_reached|policy_blocked|inconclusive_spec_contradiction|force_stopped)
                 result=20 ;;
             *)
                 # Unknown/running/exited terminal: leave $result as-is (nonzero on a

package/dashboard/__init__.py

@@ -7,7 +7,7 @@ Modules:
     control: Session control API (start/stop/pause/resume)
 """
 
-__version__ = "9.8.0"
+__version__ = "9.12.0"
 
 # Expose the control app for easy import
 try: