blun-king-cli 9.1.447 → 9.1.449

package/bin/cognitive-action-checkpoint.cjs

@@ -13,7 +13,7 @@ const TRIGGER_KINDS = new Set([
   'immediate', 'external_event', 'time', 'dependency', 'user_decision',
 ]);
 const MODEL_KEYS = new Set([
-  'revision', 'phase', 'evidenceBasis', 'epistemicState', 'lastVerified', 'nextAction', 'expectedEvidence', 'nextTrigger', 'problemFrame', 'updatedAt',
+  'revision', 'phase', 'evidenceBasis', 'epistemicState', 'lastVerified', 'nextAction', 'expectedEvidence', 'verificationProof', 'nextTrigger', 'problemFrame', 'updatedAt',
 ]);
 const RUNTIME_KEYS = new Set([...MODEL_KEYS, 'evidenceReceipt']);
 const PROBLEM_FRAME_KEYS = new Set([
@@ -22,11 +22,28 @@ const PROBLEM_FRAME_KEYS = new Set([
 ]);
 const NEXT_TRIGGER_KEYS = new Set(['kind', 'condition', 'dueAt']);
 const EVIDENCE_INPUT_KEYS = new Set([
+  'turnId', 'toolCallId', 'toolName', 'decision', 'outcome', 'durationMs', 'toolArgs',
+]);
+const REQUIRED_EVIDENCE_INPUT_KEYS = new Set([
   'turnId', 'toolCallId', 'toolName', 'decision', 'outcome', 'durationMs',
 ]);
 const EVIDENCE_DIGEST_RE = /^[a-f0-9]{16}$/u;
 const DECISION_BASIS_RE = /^[a-f0-9]{16}$/u;
 const ISO_TIMESTAMP_RE = /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}(?:\.\d{1,9})?(?:Z|[+-]\d{2}(?::?\d{2})?)$/u;
+const ACTION_ONLY_TOOL_NAMES = new Set([
+  'CreateGoal', 'CronCreate', 'CronDelete', 'DubVideo', 'Edit', 'EnterPlanMode',
+  'ExitPlanMode', 'GenerateImage', 'GenerateSpeech', 'GenerateVideo', 'LipSyncMedia',
+  'MistakeRecord', 'SetGoalBudget', 'TaskStop', 'TaskUpdate', 'UpdateGoal', 'Write',
+]);
+const VERIFICATION_TOOL_NAMES = new Set([
+  'codebasesearch', 'cronlist', 'fetchurl', 'getgoal', 'getmedia', 'glob', 'grep',
+  'read', 'readmediafile', 'taskoutput', 'test', 'understandimage', 'understandvideo',
+  'websearch',
+]);
+const COMMAND_TOOL_NAMES = new Set(['bash', 'command', 'exec_command', 'shell']);
+const MUTATING_COMMAND = /(?:^|(?:&&|\|\||;|\s))(?:(?:cp|mv|rm|mkdir|touch|tee|install|scp|sftp|ssh)\b|(?:sed\s+-i\b)|(?:git\s+(?:add|commit|push|checkout|switch|reset|clean|merge|rebase)\b)|(?:Set-Content|Add-Content|Copy-Item|Move-Item|Remove-Item|New-Item|Start-Process)\b)/iu;
+const VERIFICATION_COMMAND = /(?:^|(?:&&|\|\||;|\s))(?:(?:node\s+--(?:test|check))|(?:node\s+[^\s;&|]*(?:check|test|verify|validate|lint|gate|probe)[^\s;&|]*\b)|(?:(?:npm|pnpm|yarn)\s+(?:test|(?:run\s+)?(?:test|lint|check|typecheck|build)))|(?:python(?:3)?\s+-m\s+pytest)|(?:pytest)|(?:go\s+test)|(?:cargo\s+test)|(?:dotnet\s+test)|(?:npx\s+)?(?:tsc|eslint|biome\s+check)|(?:git\s+(?:diff(?:\s+--check)?|fsck|status|rev-parse|show))|(?:sha(?:1|256|512)sum|shasum|cmp|diff|wc|rg|grep|cat|ls|stat)\b|(?:certutil\s+-hashfile)|(?:Get-FileHash|Get-Content|Get-Item|Test-Path|Compare-Object|Measure-Object)\b)/iu;
+const MAX_SUCCESSFUL_TOOL_DIGESTS = 32;
 
 function bounded(value, field, max = 512) {
   const text = String(value ?? '')
@@ -148,6 +165,86 @@ function normalizedTurnId(value) {
   return turnId;
 }
 
+function successfulToolDigest(toolName) {
+  const name = bounded(toolName, 'verificationProof toolName', 128);
+  return crypto.createHash('sha256').update(`tool:${name}`).digest('hex').slice(0, 16);
+}
+
+function successfulToolCallDigest(turnId, toolCallId, toolName) {
+  const turn = normalizedTurnId(turnId);
+  const callId = bounded(toolCallId, 'verificationProof toolCallId', 256);
+  const name = bounded(toolName, 'verificationProof toolName', 128);
+  return crypto.createHash('sha256')
+    .update(`turn:${turn}\0call:${callId}\0tool:${name}`)
+    .digest('hex')
+    .slice(0, 16);
+}
+
+function isActionOnlyTool(toolName) {
+  return ACTION_ONLY_TOOL_NAMES.has(String(toolName ?? '').trim());
+}
+
+function verificationCommand(toolArgs) {
+  if (!toolArgs || typeof toolArgs !== 'object' || Array.isArray(toolArgs)) return '';
+  const command = toolArgs.command ?? toolArgs.cmd;
+  return typeof command === 'string' ? command.trim().replace(/\s+/gu, ' ') : '';
+}
+
+function isVerificationToolCall(toolName, toolArgs) {
+  const normalizedName = String(toolName ?? '').trim().toLowerCase();
+  if (VERIFICATION_TOOL_NAMES.has(normalizedName)) return true;
+  if (!COMMAND_TOOL_NAMES.has(normalizedName)) return false;
+  const command = verificationCommand(toolArgs);
+  return command.length > 0
+    && !MUTATING_COMMAND.test(command)
+    && VERIFICATION_COMMAND.test(command);
+}
+
+function normalizeSuccessfulToolDigests(value) {
+  if (value === undefined) return Object.freeze([]);
+  if (!Array.isArray(value) || value.length > MAX_SUCCESSFUL_TOOL_DIGESTS
+      || value.some((item) => typeof item !== 'string' || !EVIDENCE_DIGEST_RE.test(item))
+      || new Set(value).size !== value.length) {
+    throw new TypeError('successful tool digests are invalid');
+  }
+  return Object.freeze([...value]);
+}
+
+function normalizeVerificationProof(input, evidenceReceipt, options = {}) {
+  const keys = input && typeof input === 'object' && !Array.isArray(input)
+    ? Object.keys(input)
+    : [];
+  const legacy = options.allowLegacy === true
+    && keys.length === 2
+    && Object.hasOwn(input, 'toolName')
+    && Object.hasOwn(input, 'claim');
+  const current = keys.length === 3
+    && Object.hasOwn(input, 'toolCallId')
+    && Object.hasOwn(input, 'toolName')
+    && Object.hasOwn(input, 'claim');
+  if (!legacy && !current) {
+    throw new TypeError('verificationProof fields are invalid');
+  }
+  const toolName = bounded(input.toolName, 'verificationProof toolName', 128);
+  const claim = bounded(input.claim, 'verificationProof claim');
+  if (isActionOnlyTool(toolName)) {
+    throw new TypeError('action-only tool cannot serve as verification proof');
+  }
+  const receipt = normalizeActionEvidenceReceipt(evidenceReceipt);
+  if (legacy) {
+    if (!receipt.successfulToolDigests.includes(successfulToolDigest(toolName))) {
+      throw new TypeError('verificationProof must name a successful current-turn tool');
+    }
+    return Object.freeze({ toolName, claim });
+  }
+  const toolCallId = bounded(input.toolCallId, 'verificationProof toolCallId', 256);
+  const callDigest = successfulToolCallDigest(receipt.turnId, toolCallId, toolName);
+  if (!receipt.successfulToolCallDigests.includes(callDigest)) {
+    throw new TypeError('verificationProof must match a successful current-turn verification call');
+  }
+  return Object.freeze({ toolCallId, toolName, claim });
+}
+
 function emptyActionEvidenceReceipt(turnId) {
   const normalized = normalizedTurnId(turnId);
   return Object.freeze({
@@ -155,14 +252,18 @@ function emptyActionEvidenceReceipt(turnId) {
     completedTools: 0,
     successfulTools: 0,
     failedTools: 0,
+    successfulToolDigests: Object.freeze([]),
+    successfulToolCallDigests: Object.freeze([]),
     digest: crypto.createHash('sha256').update(`turn:${normalized}:empty`).digest('hex').slice(0, 16),
   });
 }
 
 function normalizeActionEvidenceReceipt(input) {
   if (!input || typeof input !== 'object' || Array.isArray(input)) throw new TypeError('evidence receipt must be an object');
-  const keys = new Set(['turnId', 'completedTools', 'successfulTools', 'failedTools', 'digest']);
-  if (!Object.keys(input).every((key) => keys.has(key)) || Object.keys(input).length !== keys.size) {
+  const requiredKeys = new Set(['turnId', 'completedTools', 'successfulTools', 'failedTools', 'digest']);
+  const allowedKeys = new Set([...requiredKeys, 'successfulToolDigests', 'successfulToolCallDigests']);
+  if (!Object.keys(input).every((key) => allowedKeys.has(key))
+      || ![...requiredKeys].every((key) => Object.hasOwn(input, key))) {
     throw new TypeError('evidence receipt fields are invalid');
   }
   const receipt = {
@@ -170,6 +271,8 @@ function normalizeActionEvidenceReceipt(input) {
     completedTools: Number(input.completedTools),
     successfulTools: Number(input.successfulTools),
     failedTools: Number(input.failedTools),
+    successfulToolDigests: normalizeSuccessfulToolDigests(input.successfulToolDigests),
+    successfulToolCallDigests: normalizeSuccessfulToolDigests(input.successfulToolCallDigests),
     digest: String(input.digest ?? ''),
   };
   if (![receipt.completedTools, receipt.successfulTools, receipt.failedTools]
@@ -184,8 +287,8 @@ function normalizeActionEvidenceReceipt(input) {
 function advanceActionEvidenceReceipt(current, input) {
   const prior = normalizeActionEvidenceReceipt(current);
   if (!input || typeof input !== 'object' || Array.isArray(input)
-      || Object.keys(input).length !== EVIDENCE_INPUT_KEYS.size
-      || !Object.keys(input).every((key) => EVIDENCE_INPUT_KEYS.has(key))) {
+      || !Object.keys(input).every((key) => EVIDENCE_INPUT_KEYS.has(key))
+      || ![...REQUIRED_EVIDENCE_INPUT_KEYS].every((key) => Object.hasOwn(input, key))) {
     throw new TypeError('evidence input fields are invalid');
   }
   const turnId = normalizedTurnId(input.turnId);
@@ -200,14 +303,30 @@ function advanceActionEvidenceReceipt(current, input) {
     throw new TypeError('evidence input values are invalid');
   }
   const successful = decision === 'passed' && outcome === 'success';
+  const successfulToolDigests = [...prior.successfulToolDigests];
+  const successfulToolCallDigests = [...prior.successfulToolCallDigests];
+  const toolDigest = successfulToolDigest(toolName);
+  if (successful && !successfulToolDigests.includes(toolDigest)) {
+    successfulToolDigests.push(toolDigest);
+    if (successfulToolDigests.length > MAX_SUCCESSFUL_TOOL_DIGESTS) successfulToolDigests.shift();
+  }
+  const verificationCall = successful && isVerificationToolCall(toolName, input.toolArgs);
+  if (verificationCall) {
+    const callDigest = successfulToolCallDigest(turnId, toolCallId, toolName);
+    if (!successfulToolCallDigests.includes(callDigest)) successfulToolCallDigests.push(callDigest);
+    if (successfulToolCallDigests.length > MAX_SUCCESSFUL_TOOL_DIGESTS) successfulToolCallDigests.shift();
+  }
   const digest = crypto.createHash('sha256').update([
     prior.digest, String(turnId), toolCallId, toolName, decision, outcome, String(durationMs),
+    String(verificationCall),
   ].join('\0')).digest('hex').slice(0, 16);
   return Object.freeze({
     turnId,
     completedTools: prior.completedTools + 1,
     successfulTools: prior.successfulTools + (successful ? 1 : 0),
     failedTools: prior.failedTools + (successful ? 0 : 1),
+    successfulToolDigests: Object.freeze(successfulToolDigests),
+    successfulToolCallDigests: Object.freeze(successfulToolCallDigests),
     digest,
   });
 }
@@ -293,6 +412,14 @@ function normalizeActionCheckpoint(input, options = {}) {
       ? normalizeActionEvidenceReceipt(input.evidenceReceipt)
       : undefined;
   if (evidenceReceipt !== undefined) checkpoint.evidenceReceipt = evidenceReceipt;
+  if (input.verificationProof !== undefined) {
+    if (phase !== 'verify' || evidenceBasis !== 'runtime_tool') {
+      throw new TypeError('verificationProof requires a runtime_tool verify checkpoint');
+    }
+    checkpoint.verificationProof = normalizeVerificationProof(input.verificationProof, evidenceReceipt, {
+      allowLegacy: replay,
+    });
+  }
   return Object.freeze(checkpoint);
 }
 
@@ -309,6 +436,12 @@ function projectActionCheckpoint(checkpoint) {
   ];
   lines.push(`Next action: ${value.nextAction}`);
   lines.push(`Expected evidence: ${value.expectedEvidence}`);
+  if (value.verificationProof !== undefined) {
+    const call = value.verificationProof.toolCallId === undefined
+      ? value.verificationProof.toolName
+      : `${value.verificationProof.toolName} call ${value.verificationProof.toolCallId}`;
+    lines.push(`Verification proof: ${call} - ${value.verificationProof.claim}`);
+  }
   if (value.nextTrigger !== undefined) {
     lines.push(`Next trigger: ${value.nextTrigger.kind.replaceAll('_', ' ')} - ${value.nextTrigger.condition}`);
     if (value.nextTrigger.dueAt !== undefined) lines.push(`Due at: ${value.nextTrigger.dueAt}`);
@@ -344,5 +477,8 @@ module.exports = {
   assertActionCheckpointRevision,
   emptyActionEvidenceReceipt,
   normalizeActionCheckpoint,
+  normalizeVerificationProof,
   projectActionCheckpoint,
+  successfulToolCallDigest,
+  successfulToolDigest,
 };

package/bin/goal-completion-evidence-policy.cjs

@@ -1,5 +1,9 @@
 'use strict';
 
+const {
+  normalizeVerificationProof,
+} = require('./cognitive-action-checkpoint.cjs');
+
 function hasCompletionCriterion(goal) {
   return typeof goal?.completionCriterion === 'string'
     && goal.completionCriterion.trim().length > 0;
@@ -10,6 +14,24 @@ function successfulRuntimeEvidence(checkpoint) {
   return Number.isSafeInteger(successfulTools) && successfulTools > 0;
 }
 
+function verificationProofGaps(checkpoint) {
+  if (!checkpoint?.verificationProof) {
+    return ['Bind the completion claim to a successful verification tool from the checkpoint turn.'];
+  }
+  try {
+    normalizeVerificationProof(checkpoint.verificationProof, checkpoint.evidenceReceipt, { allowLegacy: true });
+    return [];
+  } catch (error) {
+    if (/action-only tool/u.test(String(error?.message ?? ''))) {
+      return ['The completion proof names an action-only tool, not a verification tool.'];
+    }
+    if (/successful current-turn (?:tool|verification call)/u.test(String(error?.message ?? ''))) {
+      return ['The completion proof does not match a successful verification call from the checkpoint turn.'];
+    }
+    return ['The completion verification proof is malformed.'];
+  }
+}
+
 function evaluateGoalCompletionEvidence(goal) {
   if (!hasCompletionCriterion(goal)) {
     return {
@@ -40,9 +62,16 @@ function evaluateGoalCompletionEvidence(goal) {
   if (checkpoint.epistemicState !== 'verified') {
     gaps.push('The completion proof must be classified as verified.');
   }
-  if (!successfulRuntimeEvidence(checkpoint)) {
+  const hasRuntimeEvidence = successfulRuntimeEvidence(checkpoint);
+  if (!hasRuntimeEvidence) {
     gaps.push('The runtime evidence receipt must contain at least one successful tool result.');
   }
+  if (checkpoint.phase === 'verify'
+      && checkpoint.evidenceBasis === 'runtime_tool'
+      && checkpoint.epistemicState === 'verified'
+      && hasRuntimeEvidence) {
+    gaps.push(...verificationProofGaps(checkpoint));
+  }
 
   return {
     required: true,

package/blun.mjs

@@ -260348,6 +260348,11 @@ function createActionCheckpointInputSchema(problemFrameSchema, requireProblemFra
 		lastVerified: string().min(1).max(512),
 		nextAction: string().min(1).max(512),
 		expectedEvidence: string().min(1).max(512),
+		verificationProof: object({
+			toolCallId: string().min(1).max(256),
+			toolName: string().min(1).max(128),
+			claim: string().min(1).max(512)
+		}).strict().optional(),
 		nextTrigger: object({
 			kind: _enum(["immediate", "external_event", "time", "dependency", "user_decision"]),
 			condition: string().min(1).max(512),
@@ -262733,6 +262738,7 @@ var init_turn = __esmMin((() => {
 				this.toolCallDupType.set(event.toolCallId, dupType === "cross_step" ? "cross_step" : "normal");
 				this.toolCallStartedAt.set(event.toolCallId, {
 					name: event.name,
+					args: event.args,
 					startedAt: Date.now()
 				});
 				this.agent.feedRootMissionContract("start", {
@@ -262763,7 +262769,8 @@ var init_turn = __esmMin((() => {
 					toolCallId: event.toolCallId,
 					toolName: started.name,
 					outcome,
-					durationMs: Date.now() - started.startedAt
+					durationMs: Date.now() - started.startedAt,
+					toolArgs: started.args
 				});
 				this.agent.telemetry.track("tool_call", properties);
 				this.agent.feedRootMissionContract("result", {
@@ -262868,6 +262875,7 @@ var init_outcome_prompts = __esmMin((() => {}));
 var update_goal_default;
 var init_update_goal$1 = __esmMin((() => {
 	update_goal_default = "Update the current autonomous goal. Set `status` only for a lifecycle change. After a coherent work slice, save `actionCheckpoint` with a monotone revision, the last verified result, exact next action, expected evidence, exact `nextTrigger`, and an explicit evidence basis. Persist the exact `nextTrigger` that releases `nextAction`: use `immediate` outside the `wait` phase; while waiting, name the external event, time, dependency, or user decision instead of pretending work can continue. A `time` trigger must include the exact ISO timestamp in `dueAt`; no other trigger kind may include `dueAt`. Use `runtime_tool` only when a successful tool in this turn measured the result; use `user_statement` for a direct user assertion, `external_report` for a report not independently measured here, and `carried_forward` only when the last verified text is unchanged. Classify knowledge as `verified`, `credible_unverified`, `hypothesis`, `uncertain_memory`, `stale`, or `unknown`; never present a weaker state as verified, and preserve the state on carry-forward. Start at revision 1 and increment the currently projected revision by exactly one; stale writers fail closed. This is durable progress state, not permission, and should change only when the facts change. A checkpoint-only call keeps the goal active.\n\n- `active` — resume a paused or blocked goal when the user explicitly asks you to work on that goal.\n- `complete` — the objective is fully satisfied, all files are written, all tests pass, and any stated validation has passed. When the goal has a completion criterion, first save a `verify` checkpoint with `runtime_tool`, `verified`, and a successful runtime evidence receipt.\n- `blocked` — a genuine external condition or required user decision prevents progress.\n- `paused` — set the goal aside for now.\n\nDo not mark complete after a plan or partial result. If useful work remains, checkpoint it and continue. Do not ask for permission merely to execute an already authorized checkpoint; ask only at a real rights boundary or missing user decision.\n";
+	update_goal_default += "\nBefore completing a goal with a criterion, bind the verified claim to the exact successful current-turn verification call in `verificationProof`, including its `toolCallId`. A write, edit, copy, deploy, or other action is not proof that the changed behavior works, even when it shares a mixed-use tool such as `Bash` with tests. Run a separate observation, test, or counterexample probe and cite that exact successful call plus the claim it supports.\n";
 	update_goal_default += "\nFor a non-trivial or unfamiliar problem, preserve `problemFrame` with the success criterion, missing knowledge, bounded candidate actions, selected action and reason, support choice, risk, and reversibility. The selected action must match one candidate. Bind each selected action to the projected durable facts or assumptions it relies on by copying their explicit refs into `decisionBasis`. A stale or unknown decision basis requires replanning before execution. Problem framing is descriptive state and never grants permission.\n";
 }));
 //#endregion

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.447",
+  "version": "9.1.449",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {