The agent harness performance optimization system. Skills, instincts, memory, security, and research-first development for Claude Code, Codex, Opencode, Cursor and beyond.
fix(context): honor directive order and singular rejections
Yann Roberto2d2712b89912
diff --git a/scripts/lib/context-selection.js b/scripts/lib/context-selection.jsindex dba849cc7..0a1a8745e 100644--- a/scripts/lib/context-selection.js+++ b/scripts/lib/context-selection.js@@ -29,7 +29,8 @@ const FALLBACK_MARGIN = 1.1;// v4: an explicit empty proposal (decline) is honored; the tier-2 fallback no// longer overrides declines at the launch/selection call sites.// v5: indirect name citations cannot bypass review through exact or scored admission.-const ROUTING_POLICY_VERSION = 5;+// v6: ordered directives and singular named references share the same admission guard.+const ROUTING_POLICY_VERSION = 6;const TASK_KEYS = new Set(['sessionId', 'taskId', 'revision', 'phase', 'query', 'explicitIds', 'proposedIds', 'noWorkflow']);const DIRECTIVE_VERB = /\b(use|apply|invoke|run|follow|load)\s+(the\s+)?/i;@@ -37,37 +38,56 @@ const normalizedQueryName = text => text.replace(/n['\u2019]t\b/gi, ' not').replace(/([a-z0-9])([A-Z])/g, '$1 $2').replace(/_/g, ' ').toLowerCase().replace(/[^a-z0-9]+/g, ' ').trim();-// Classify each named skill locally: a rejected/quoted X must not suppress an-// affirmative Y in another clause. Ambiguous references stay available for the-// bounded proposal, but cannot admit themselves via either retrieval threshold.-function citationFor(candidate, query) {- if (!candidate.exact) return 'none';- const aliases = [...new Set([candidate.exactAlias, candidate.id.slice('skill:'.length)]- .filter(Boolean).map(normalizedQueryName))];- const quotedParts = query.match(/"[^"]*"|\u201c[^\u201d]*\u201d|(?:^|[\s(:])'[^']*'|\u2018[^\u2019]*\u2019|\x60[^\x60]*\x60/g) || [];- let directive = false;- let indirect = false;- for (const clause of query.match(/[^.!?;\n]+[.!?;\n]*/g) || []) {+/** Return canonical/native aliases plus conservative singular forms. Singular+ * multiword aliases only guard indirect references; they never create exact admission. */+function citationAliases(candidate, nativeName) {+ const names = [...new Set([candidate.exactAlias, candidate.id.slice('skill:'.length), nativeName]+ .filter(Boolean).map(normalizedQueryName).filter(Boolean))];+ return names.flatMap(name => [{ name, exact: true },+ ...(/\s\w+[^s]s$/.test(name) ? [{ name: name.slice(0, -1), exact: false }] : [])]);+}++/** Classify a retrieval candidate's references in query order. Genuine later+ * directives/rejections replace earlier instructions; questions, quotations and+ * reported speech do not withdraw a directive. Returns directive, indirect or none.+ * @param {object} candidate Ranked skill candidate with its canonical ID and exact alias.+ * @param {string} query Task text to classify without loading skill contents.+ * @param {string} nativeName Registry name, including aliases absent from exact retrieval.+ * @returns {'directive'|'indirect'|'none'} Admission evidence for this skill only.+ */+function citationFor(candidate, query, nativeName) {+ const aliases = citationAliases(candidate, nativeName);+ let citation = 'none';+ // Mask quoted instructions before splitting clauses so punctuation inside a+ // quotation cannot look like a separate instruction. Quoted names remain usable.+ const unquoted = query.replace(/"[^"]*"|\u201c[^\u201d]*\u201d|(?:^|[\s(:])'[^']*'|\u2018[^\u2019]*\u2019|\x60[^\x60]*\x60/g, part => {+ const text = normalizedQueryName(part);+ if (!DIRECTIVE_VERB.test(text)) return part;+ if (aliases.some(({ name }) => new RegExp('\\b' + name + '\\b').test(text))) citation = 'indirect';+ return part.replace(/[^.!?;\n]/g, ' ');+ });+ for (const clause of unquoted.match(/[^.!?;\n]+[.!?;\n]*/g) || []) {const text = normalizedQueryName(clause);- for (const name of aliases) {- const escaped = name.replace(/[.*+?^${}()|[\]\\]/g, '\\$&');- const directivePattern = DIRECTIVE_VERB.source + '(skill\\s*:?\\s*)?' + escaped + '\\b';- if (quotedParts.some(part => new RegExp(directivePattern, 'i').test(normalizedQueryName(part)))) indirect = true;- for (const mention of text.matchAll(new RegExp('\\b' + escaped + '\\b', 'gi'))) {- const prefix = text.slice(0, mention.index);- const polite = /^(?:(?:please )?(?:can|could|would|will) you (?:please )?|do )(?:use|apply|invoke|run|follow|load)\b/.test(text);- const question = !polite && (clause.includes('?')- || /^(can|could|would|should|may|might|do|does|did|is|are|why|how|what|when|where|which)\b/.test(text));- const quoted = /^\s*["'\u201c\u2018\x60]/.test(clause);- if (question || quoted || /\b(do not|never|no|not|avoid|says|said|reads|told|mentions?|quoted?|document)\b/.test(prefix)) {- indirect = true;- } else if (new RegExp(DIRECTIVE_VERB.source + '(skill\\s*:?\\s*)?$', 'i').test(prefix)) {- directive = true;- }+ const mentions = aliases.flatMap(({ name, exact }) => [...text.matchAll(new RegExp('\\b' + name + '\\b', 'g'))]+ .map(mention => ({ index: mention.index, exact }))).sort((a, b) => a.index - b.index);+ const polite = /^(?:(?:please )?(?:can|could|would|will) you (?:please )?|do )(?:use|apply|invoke|run|follow|load)\b/.test(text);+ const question = !polite && (clause.includes('?')+ || (!/^do not\b/.test(text)+ && /^(can|could|would|should|may|might|do|does|did|is|are|why|how|what|when|where|which)\b/.test(text)));+ for (const mention of mentions) {+ const prefix = text.slice(0, mention.index);+ if (question || /^\s*["'\u201c\u2018\x60]/.test(clause)+ || /\b(say|says|said|reads|told|mentions?|quoted?|document)\b/.test(prefix)) {+ if (citation === 'none') citation = 'indirect';+ } else if (/\b(do not|never|no|not|avoid)\b/.test(prefix)) {+ citation = 'indirect';+ } else if (mention.exact && candidate.exact+ && new RegExp(DIRECTIVE_VERB.source + '(skill\\s*:?\\s*)?$', 'i').test(prefix)) {+ citation = 'directive';}}}- return indirect ? 'indirect' : directive ? 'directive' : 'none';+ return citation;}function validateTask(task) {@@ -218,7 +238,7 @@ function resolveTaskContext({ repoRoot = DEFAULT_REPO_ROOT, task, profileId = 'l// questions, negations, reported speech, multiple cited names — never admit// implicitly. Everything else keeps the bounded-proposal path so the// primary agent decides ambiguous cases during work it was already doing.- const citations = new Map(candidates.map(candidate => [candidate.id, citationFor(candidate, task.query || '')]));+ const citations = new Map(candidates.map(candidate => [candidate.id, citationFor(candidate, task.query || '', byId.get(candidate.id).name)]));const exactAnchors = candidates.filter(candidate => citations.get(candidate.id) === 'directive');let autoSelection = null;if (!task.noWorkflow && selectionMode === 'auto' && !reused && !explicitIds.length && !proposedIds.length && candidates.length) {diff --git a/tests/lib/context-profile-launch.test.js b/tests/lib/context-profile-launch.test.jsindex b87c72836..2edc088a1 100644--- a/tests/lib/context-profile-launch.test.js+++ b/tests/lib/context-profile-launch.test.js@@ -157,7 +157,7 @@ test('spawn failures and timeout signals remain unsuccessful without a native ex}));for (const query of ["Don't use the feature skill.", 'Don\u2019t use the feature skill.',- 'Can I use the feature skill?', 'The README says use the feature skill.']) {+ 'Can I use the feature skill?', 'The README says use the feature skill.', 'The docs say use feature.']) {test('indirect citation uses the proposal path and honors decline: ' + query, () => withFixture(repoRoot => {const phases = [];const result = launchTaskContext({ repoRoot, task: { ...input, query, explicitIds: [] },@@ -173,3 +173,33 @@ for (const query of ["Don't use the feature skill.", 'Don\u2019t use the featureassert.equal(result.selection.reason, 'agent-declined-selection');}));}++test('a final directive after a question reaches the task with its context', () => withFixture(repoRoot => {+ const phases = [];+ const result = launchTaskContext({ repoRoot,+ task: { ...input, query: 'Should we use feature? Use feature.', explicitIds: [] },+ execute(command, args, options) {+ phases.push(options.phase);+ assert.match(options.input, /# feature/);+ return { status: 0, stdout: 'ok' };+ } });+ assert.deepEqual(phases, ['task']);+ assert.equal(result.routingCalls, 0);+ assert.deepEqual(result.selection.loadedIds, ['skill:feature']);+}));++test('a singular rejected name cannot inject scored context at launch', () => {+ const phases = [];+ const query = "Don't use the database migration skill. Review a PostgreSQL migration that adds an indexed nullable column without downtime.";+ const result = launchTaskContext({ task: { ...input, query, explicitIds: [] },+ execute(command, args, options) {+ phases.push(options.phase);+ if (options.phase === 'selection') return { status: 0, stdout: '{"selectedIds":[]}' };+ assert.doesNotMatch(options.input, /# Database Migration Patterns/);+ return { status: 0, stdout: 'ok' };+ } });+ assert.deepEqual(phases, ['selection', 'task']);+ assert.equal(result.routingCalls, 1);+ assert.deepEqual(result.selection.loadedIds, []);+ assert.equal(result.selection.fallback, null);+});diff --git a/tests/lib/context-selection.test.js b/tests/lib/context-selection.test.jsindex f0b94ecf2..682615c81 100644--- a/tests/lib/context-selection.test.js+++ b/tests/lib/context-selection.test.js@@ -319,6 +319,9 @@ for (const query of ['Should we use the feature skill?','Why use the feature skill?','The README says to use the feature skill.',+ 'The docs say use feature.',+ 'The docs say to use feature.',+ 'The docs said use feature.','"Use the feature skill" is an example.',]) {test('indirect name citation needs an agent decision: ' + query, () => withFixture(repoRoot => {@@ -372,7 +375,7 @@ test('routing policy changes invalidate receipts created by the old citation polconst { compileContextProfile } = require('../../scripts/lib/context-profiles');const plan = compileContextProfile({ repoRoot, selectionMode: 'auto' });const bindingDigest = digestObject({ sessionId: 'session-1', taskId: 'task-1', revision: 1,- phase: 'implement', planDigest: plan.planDigest, routingPolicyVersion: 4,+ phase: 'implement', planDigest: plan.planDigest, routingPolicyVersion: 5,triggersDigest: digestObject({}), queryDigest: digestObject(query) });const { receiptDigest: _receiptDigest, ...receipt } = first.receipt;const oldReceipt = { ...receipt, bindingDigest, explicitIds: [],@@ -382,3 +385,62 @@ test('routing policy changes invalidate receipts created by the old citation polassert.equal(result.reused, false);assert.deepEqual(result.loadedIds, []);}));++for (const query of [+ 'Should we use feature? Use feature.',+ 'Use feature. Should we use feature?',+ 'The docs say use feature. Use feature.',+ 'Use feature. The docs say use feature.',+ 'The phrase "Use feature." is an example. Use feature.',+ 'The docs say "Use feature." Use feature.',+ 'The docs say "Should we use feature?" Use feature.',+ `Use feature. The docs say "don’t use feature."`,+ 'Use feature. The phrase "Use feature." is an example.',+ "Don't use feature. Use feature.",+ 'The docs say never use feature. Use feature.',+ 'Use feature. The docs say never use feature.',+ 'Use feature to inspect the essay.',+]) {+ test('a genuine directive survives surrounding discussion: ' + query, () => withFixture(repoRoot => {+ const result = resolve(repoRoot, { query }, { load: true });+ assert.deepEqual(result.loadedIds, ['skill:feature']);+ assert.equal(result.reason, 'auto-selection');+ assert.equal(result.receipt.autoSelection.exact, true);+ }));+}++for (const query of [+ "Use feature. Don't use feature.",+ 'Use feature. Never use feature.',+ "Use feature. Don't use feature. Should we use feature?",+]) {+ test('a later rejection withdraws an earlier directive: ' + query, () => withFixture(repoRoot => {+ const result = resolve(repoRoot, { query }, { load: true });+ assert.deepEqual(result.loadedIds, []);+ assert.equal(result.fallback, null);+ }));+}++for (const prefix of [+ "Don't use the database migration skill.",+ 'Don\u2019t use the database migration skill.',+ 'Never use database migration.',+ 'Should we use the database migration skill?',+ 'The docs say use the database migration skill.',+]) {+ test('singular named references cannot bypass citation review: ' + prefix, () => {+ const query = prefix + ' Review a PostgreSQL migration that adds an indexed nullable column without downtime.';+ const result = resolveTaskContext({ task: task({ query }), load: true });+ const candidate = result.candidates.find(value => value.id === 'skill:database-migrations');+ assert.ok(candidate && candidate.bm25 >= 20, 'Exercise the BM25 auto-admission threshold');+ assert.equal(candidate.exact, false, 'Exercise the non-exact candidate path');+ assert.deepEqual(result.loadedIds, []);+ assert.equal(result.fallback, null);+ });+}++test('singular forms do not create exact directive admission', () => {+ const result = resolveTaskContext({ task: task({ query: 'Use the database migration skill.' }), load: true });+ assert.ok(result.candidates.some(candidate => candidate.id === 'skill:database-migrations'));+ assert.ok(!result.receipt.autoSelection || !result.receipt.autoSelection.exact);+});