mtech-ai-local-runner 0.1.92 → 0.1.94

package/README.md

@@ -1,3 +1,46 @@
+## Unreleased: bounded report drafting and offline preflight
+
+The final writing guidance targets 1,200 UTF-16 code units in `report`, below
+the unchanged 2,000-unit acceptance limit. Findings and uncertainty belong in
+the short report; detailed citations belong in `evidence`. No automatic
+truncation, content repair, model retry, new tool or launch permission is added.
+Prompt guidance alone cannot guarantee model compliance: an oversized final
+response still fails closed. These changes are source-only, not part of the
+installed/reviewed 0.1.93 package.
+
+From the `runners/local-agent` source checkout, before another live validation,
+replay saved **final response text**, not a transcript or stream-JSON wrapper,
+through the production completion validator (the script is a repository-only
+rehearsal utility, not an installed Runner command):
+
+```sh
+npm run build
+node scripts/check-completion-report.mjs /absolute/path/to/final-response.txt
+```
+
+This offline script reads a bounded UTF-8 file, emits only counts and fixed
+diagnostics, and never rewrites the input, calls a model, accesses Gateway or
+starts work. Exit 0 means valid transport (including a truthful `blocked`
+report), 1 means rejected transport, and 2 means usage/read/encoding/input-size
+failure. A reported `completed` status does not verify the investigation,
+citations or Work Item completion. The exact 2026-11-05 failure is preserved as
+an untrusted technical test fixture; its claims are not accepted as true.
+
+No package publication, Gateway registration, Runner replacement or new live
+execution has been performed for this source change. Those require separate
+release review, exact hashes and execution approval.
+
+## 0.1.93: strict investigation JSON completion guidance
+
+Claude uses native-stream progress guidance and receives scoped tool/correction
+context before the final JSON completion contract. A draft for owner review
+still returns the structured report to the Runner without answering the original
+request. Markdown remains rejected; no repair call or automatic retry is added.
+Source verification does not prove live model compliance. Existing attempts,
+contracts, Loop limits and security policies are unchanged. Installation needs
+freshly reviewed package/runtime hashes; Gateway deployment alone does not
+replace an installed Runner.
+
 ## Unreleased: Gateway-managed verification budget (P3E)
 
 D6 owner revision `fbc30fbe-cac9-45d1-9eff-bccec6ee6a54` selects a separate

package/dist/src/adapters/claude-code/prompt-compiler.js

@@ -27,5 +27,18 @@ export function compilePrompt(taskPackage, options) {
     const { toolInventory = "local-tools-v1", ...brief } = options;
     localToolNames(toolInventory);
     const correction = renderBundleCorrectionBrief(taskPackage);
-    return `${compileBrief(taskPackage, brief)}\n\n${repositoryScope(taskPackage)}\n\n${LOCAL_TOOL_GUIDANCE}\n\n${claudePathRules(toolInventory)}${correction}`;
+    // Claude progress is observed from its native stream, not model-authored lifecycle JSON.
+    // Keep the final report contract last, including after untrusted correction findings.
+    return compileBrief(taskPackage, { ...brief, progressReporting: "native_messages" }, [
+        repositoryScope(taskPackage),
+        LOCAL_TOOL_GUIDANCE,
+        `${claudePathRules(toolInventory)}${correction}`,
+        "## Investigation result and transport boundary\n" +
+            "Put concise findings in report and source citations in evidence, not a standalone Markdown report. " +
+            "A draft for owner review still requires this final JSON transport report; returning it to the Runner " +
+            "does not submit an answer to the original request or authorize external publication. " +
+            "Distinguish source inspection from tests actually executed; a successful read is not a passed test. " +
+            "Before finalizing, check the required keys, report/evidence size limits and accurate pending_work. " +
+            "If unfinished, return one blocked report; never invent evidence or claim completion to satisfy the format.",
+    ]);
 }

package/dist/src/adapters/completion-report-preflight.js

@@ -0,0 +1,20 @@
+import { completionEnvelope, completionEvent, COMPLETION_REPORT_LIMITS } from './completion-report.js';
+/** Offline diagnostics only: never rewrites output, creates evidence, or authorizes execution. */
+export function completionReportPreflight(raw) {
+    const options = { earlierCodeBlocks: true };
+    const envelope = completionEnvelope(raw, options);
+    const event = completionEvent(raw, undefined, options);
+    const rejection = event.payload?.completion_rejection;
+    const valid = event.eventType === 'completed' || event.payload?.reason === 'executor_blocked';
+    return {
+        valid_transport: valid,
+        // This is transport status only, not independent review or Work Item completion.
+        completed: event.eventType === 'completed',
+        status: valid ? (event.eventType === 'completed' ? 'completed' : 'blocked') : 'rejected',
+        rejection_code: rejection?.code ?? null,
+        report_code_units: typeof envelope?.report === 'string' ? envelope.report.length : null,
+        report_limit: COMPLETION_REPORT_LIMITS.reportCodeUnits,
+        evidence_count: Array.isArray(envelope?.evidence) ? envelope.evidence.length : null,
+        pending_work_count: Array.isArray(envelope?.pending_work) ? envelope.pending_work.length : null,
+    };
+}

package/dist/src/adapters/completion-report.js

@@ -22,6 +22,14 @@ export function reportRejectionCounts(value, total) {
 const MAX_EVIDENCE_ITEMS = 20;
 const MAX_EVIDENCE_CHARS = 500;
 const MAX_REPORT_ENVELOPE_CHARS = 16_000;
+/** Construction/rehearsal hints; acceptance always goes through completionEvent. */
+export const COMPLETION_REPORT_LIMITS = Object.freeze({
+    reportCodeUnits: MAX_EVENT_SUMMARY_CHARS,
+    targetReportCodeUnits: Math.min(1200, MAX_EVENT_SUMMARY_CHARS),
+    listItems: MAX_EVIDENCE_ITEMS,
+    itemCodeUnits: MAX_EVIDENCE_CHARS,
+    envelopeCodeUnits: MAX_REPORT_ENVELOPE_CHARS,
+});
 const REPORT_FAILURE_CODES = ["CODEX_COMPLETION_REPORT_MISSING", "CODEX_COMPLETION_REPORT_INVALID", "CODEX_COMPLETION_REPORT_MULTIPLE", "CODEX_COMPLETION_REPORT_NOT_FINAL", "CODEX_TURN_INCOMPLETE"];
 const REPORT_DIAGNOSTIC_CODES = [...REPORT_FAILURE_CODES, "CODEX_COMPLETION_REPORT_BLOCKED"];
 /** Pair counts only; no item IDs, text, stable digests or arbitrary keys. */
@@ -68,6 +76,11 @@ export const COMPLETION_REPORT_NOTICE = "Do the assigned work yourself using the
     "Your final response must be one JSON object, without Markdown fences: " +
     '{"status":"completed","report":"actual findings and results","evidence":["file path and observed finding, or command and actual result"],"pending_work":[]}. ' +
     `The report must be nonempty and at most ${MAX_EVENT_SUMMARY_CHARS} characters. ` +
+    `Aim for at most ${COMPLETION_REPORT_LIMITS.targetReportCodeUnits} UTF-16 code units in the decoded report string; the hard limit remains ${MAX_EVENT_SUMMARY_CHARS}. ` +
+    "Count decoded string length, not tokens, words, UTF-8 bytes or JSON escape characters; an emoji can count as two units. " +
+    "Use short findings for the requested topics, then uncertainty and verification limits. Put detailed source citations in evidence instead of repeating them in report. " +
+    "Do not fill the available limit or duplicate a narrative before the JSON. Before sending your single final response, shorten your own draft to the target and recheck all fields. " +
+    "Never drop uncertainty or pending work to fit the report; use blocked and accurate pending_work when work is incomplete. " +
     `Include 1-${MAX_EVIDENCE_ITEMS} nonempty evidence strings, each at most ${MAX_EVIDENCE_CHARS} characters. ` +
     "Address the assigned objective and each applicable acceptance criterion with observed evidence, not a plan or promise. " +
     "Only attest to executor work you observed; Gateway records transport, completion, and review receipts separately. " +

package/dist/src/adapters/prompt-compiler.js

@@ -120,7 +120,8 @@ function continuationContext(taskPackage) {
         + "Do not replay a completed side effect. Never answer a human-only question or approve on another actor's behalf. "
         + "Required review, completion and publication gates remain separate; report only observed work and evidence, not delivery or operational success inferred from a local result.";
 }
-export function compilePrompt(taskPackage, options) {
+/** Adapter-owned context stays ahead of the final wire-format contract. */
+export function compilePrompt(taskPackage, options, adapterSections = []) {
     const contract = record(taskPackage.contract);
     if (!contract) {
         throw new Error("task package is missing a contract");
@@ -167,6 +168,7 @@ export function compilePrompt(taskPackage, options) {
         options.progressReporting === "native_messages"
             ? `## Native progress and terminal reporting\n${NATIVE_PROGRESS_NOTICE}`
             : `## Event and checkpoint obligations\n${EVENT_SCHEMA_NOTICE}`,
+        ...adapterSections,
         `## Final completion report\n${COMPLETION_REPORT_NOTICE}`,
     ];
     return sections.join("\n\n");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mtech-ai-local-runner",
-  "version": "0.1.92",
+  "version": "0.1.94",
   "private": false,
   "license": "UNLICENSED",
   "bin": {