kc-beta 0.7.5 → 0.8.3

package/README.md

@@ -21,12 +21,59 @@ npm install -g kc-beta
 kc-beta onboard      # configure provider + API key
 cd my-project        # a folder containing rules/ and samples/
 kc-beta              # launch the agent
+
+# v0.8.1+ unattended runs: type the slash command inside the kc-beta TUI
+> /marathon Verify the new regulation against samples/. Iterate twice. If
+  most workflows work with regex, build another version using tier1+tier2
+  worker LLMs more aggressively.
+# Marathon mode chains turns automatically. /marathon off disengages.
+# /marathon status shows the driver state.
 ```
 
 Requires **Node.js 20+**. See [QUICKSTART.md](./QUICKSTART.md) for the full setup walkthrough.
 
 ---
 
+## v0.8 Highlights
+
+- **`/marathon <goal>` slash command** (v0.8.1; replaces the v0.8.0
+  separate-process `kc-marathon` CLI). Activates an inline driver inside
+  the running kc-beta TUI. Goal embedded in the command. `/marathon off`
+  to disengage manually; `/marathon status` to inspect. F5
+  one-phase-per-prompt stays enabled for interactive sessions and
+  bypasses cleanly when marathon is active. Status-bar shows
+  `🏃 MARATHON` only when active — no clutter in normal mode. v0.8.0's
+  separate-process driver was scrapped after E2E #11 found drivers
+  died silently when their parent terminal closed (SIGHUP unhandled).
+- **Skill usage counter** — passive Layer-B measurement of which skills
+  the engine actually ships to the LLM. `skill_byte_send` events go to
+  events.jsonl; the audit script aggregates per-phase × per-skill.
+  Agent-blind by design.
+- **`worker_llm_call` batch mode** — `prompts: [...]` array input with
+  concurrency control (default 5, max 10). Pairs with a canonical
+  `workflows/common/llm_client.py` shim (taught in `skill-to-workflow`)
+  so distilled workflows route through the engine where possible and
+  log to `output/llm_ledger.jsonl` when they don't.
+- **`sandbox_exec` timeout model** — default 120s (was 30s); per-call
+  `timeout_ms` up to 600s for known-slow commands. Configurable via
+  `KC_EXEC_DEFAULT_TIMEOUT_MS` + `KC_EXEC_MAX_TIMEOUT_MS`.
+- **Prescriptive phase-advance hints** — refusal messages now name
+  concrete next-action artifacts (`workflows/<rule_id>/workflow_v1.py`,
+  `output/results/production_qc_results.json`, etc.) instead of
+  descriptive `engineCounts` only.
+- **`check.py` substantiveness audit** — engine detects stub-shaped
+  rule_skills/<id>/check.py files (NOT_APPLICABLE-only returns with no
+  workflow delegation). Surfaced in milestones; opt-in enforcement via
+  `KC_ENFORCE_CHECK_PY_SUBSTANTIVE=1`.
+
+Plus engine-fix carryover from the v0.7.5 audit cycle: H3 calibration
+aggregator schema, milestone-derivation gaps (review_001.json + multi-
+path coverage_report.md), VLM runtime hardcode (workspace .env overlay),
+heap_mb periodic-write fix, stale release detection, taskboard skill
+availability in every phase. Full list: see DEV_LOG v0.8 entry.
+
+---
+
 ## What It Does
 
 KC drives a single coding agent through seven phases:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "kc-beta",
-  "version": "0.7.5",
+  "version": "0.8.3",
   "description": "KC Agent — LLM document verification agent (pure Node.js CLI). Dual-licensed: PolyForm Noncommercial 1.0.0 for personal/noncommercial use; commercial license required for enterprise production. See LICENSE and LICENSE-COMMERCIAL.md.",
   "type": "module",
   "bin": {
@@ -18,7 +18,8 @@
   ],
   "scripts": {
     "build:meme": "node scripts/build-meme.js",
-    "prepublishOnly": "node scripts/build-meme.js"
+    "prepublishOnly": "node scripts/build-meme.js",
+    "test": "node scripts/run-tests.js"
   },
   "homepage": "https://github.com/kitchen-engineer42/kc-cli",
   "repository": {

package/src/agent/context.js

@@ -152,10 +152,26 @@ export class ContextAssembler {
    * @param {string} [opts.projectMemory] - v0.7.0 B3: rules/PATTERNS.md
    *   content. Capped at ~5 KB by the caller. Surfaced for phases the
    *   work-decomposition skill operates in (skill_authoring + skill_testing).
+   * @param {string} [opts.marathonGoal] - v0.8.2 P12-A: the active marathon
+   *   goal text. Pinned at the system-prompt layer (never windowed) for the
+   *   duration of the marathon session. Surfaced only when marathon mode is
+   *   active; absent otherwise. Fixes the v0.8.1 regression where the goal
+   *   user_message got evicted by context_windowed before distillation, so
+   *   agents reverted to default behavior mid-run.
    * @returns {string}
    */
-  build({ agentMd, pipelineState, workspaceState, skillIndex, projectMemory } = {}) {
+  build({ agentMd, pipelineState, workspaceState, skillIndex, projectMemory, marathonGoal } = {}) {
     const parts = [AGENT_IDENTITY];
+    if (marathonGoal) {
+      parts.push(
+        "## Marathon goal (pinned for the duration of this session)\n\n" +
+        marathonGoal.trim() + "\n\n" +
+        "You are running in marathon mode — no manual user check-ins between " +
+        "phases. This goal is your north star; keep returning to it as you " +
+        "advance through the pipeline. If a continuation prompt focuses on " +
+        "phase mechanics, the goal above tells you *why*.",
+      );
+    }
     if (agentMd) parts.push(agentMd);
     if (skillIndex) parts.push(skillIndex);
     if (projectMemory) {