feat: self-improving skills loop — observe, inspect, amend, evaluate

- Add skill health observation layer (execution logging, success/failure tracking)
- Add skill health inspector (trace recurring failures across runs)
- Add amendify mechanism (propose SKILL.md patches from failure evidence)
- Add evaluation scaffolding (compare amended vs original performance)
- Wire into session-inspect CLI: skills:health, skills:amendify, skills:evaluate
- 1145/1145 tests passing (+3 new)
This commit is contained in:
Affaan Mustafa
2026-03-14 23:21:18 -07:00
parent 2b2777915e
commit c53bba9e02
7 changed files with 753 additions and 6 deletions

View File

@@ -0,0 +1,89 @@
'use strict';
const { buildSkillHealthReport } = require('./health');
const AMENDMENT_SCHEMA_VERSION = 'ecc.skill-amendment-proposal.v1';
function createProposalId(skillId) {
return `amend-${skillId}-${Date.now()}`;
}
function summarizePatchPreview(skillId, health) {
const lines = [
'## Failure-Driven Amendments',
'',
`- Focus skill routing for \`${skillId}\` when tasks match the proven success cases.`,
];
if (health.recurringErrors[0]) {
lines.push(`- Add explicit guardrails for recurring failure: ${health.recurringErrors[0].error}.`);
}
if (health.recurringTasks[0]) {
lines.push(`- Add an example workflow for task pattern: ${health.recurringTasks[0].task}.`);
}
if (health.recurringFeedback[0]) {
lines.push(`- Address repeated user feedback: ${health.recurringFeedback[0].feedback}.`);
}
lines.push('- Add a verification checklist before declaring the skill output complete.');
return lines.join('\n');
}
function proposeSkillAmendment(skillId, records, options = {}) {
const report = buildSkillHealthReport(records, {
...options,
skillId,
minFailureCount: options.minFailureCount || 1
});
const [health] = report.skills;
if (!health || health.failures === 0) {
return {
schemaVersion: AMENDMENT_SCHEMA_VERSION,
skill: {
id: skillId,
path: null
},
status: 'insufficient-evidence',
rationale: ['No failed observations were available for this skill.'],
patch: null
};
}
const preview = summarizePatchPreview(skillId, health);
return {
schemaVersion: AMENDMENT_SCHEMA_VERSION,
proposalId: createProposalId(skillId),
generatedAt: new Date().toISOString(),
status: 'proposed',
skill: {
id: skillId,
path: health.skill.path || null
},
evidence: {
totalRuns: health.totalRuns,
failures: health.failures,
successRate: health.successRate,
recurringErrors: health.recurringErrors,
recurringTasks: health.recurringTasks,
recurringFeedback: health.recurringFeedback
},
rationale: [
'Proposals are generated from repeated failed runs rather than a single anecdotal error.',
'The suggested patch is additive so the original SKILL.md intent remains auditable.'
],
patch: {
format: 'markdown-fragment',
targetPath: health.skill.path || `skills/${skillId}/SKILL.md`,
preview
}
};
}
module.exports = {
AMENDMENT_SCHEMA_VERSION,
proposeSkillAmendment
};

View File

@@ -0,0 +1,59 @@
'use strict';
const EVALUATION_SCHEMA_VERSION = 'ecc.skill-evaluation.v1';
function roundRate(value) {
return Math.round(value * 1000) / 1000;
}
function summarize(records) {
const runs = records.length;
const successes = records.filter(record => record.outcome && record.outcome.success).length;
const failures = runs - successes;
return {
runs,
successes,
failures,
successRate: runs > 0 ? roundRate(successes / runs) : 0
};
}
function buildSkillEvaluationScaffold(skillId, records, options = {}) {
const minimumRunsPerVariant = options.minimumRunsPerVariant || 2;
const amendmentId = options.amendmentId || null;
const filtered = records.filter(record => record.skill && record.skill.id === skillId);
const baseline = filtered.filter(record => !record.run || record.run.variant !== 'amended');
const amended = filtered.filter(record => record.run && record.run.variant === 'amended')
.filter(record => !amendmentId || record.run.amendmentId === amendmentId);
const baselineSummary = summarize(baseline);
const amendedSummary = summarize(amended);
const delta = {
successRate: roundRate(amendedSummary.successRate - baselineSummary.successRate),
failures: amendedSummary.failures - baselineSummary.failures
};
let recommendation = 'insufficient-data';
if (baselineSummary.runs >= minimumRunsPerVariant && amendedSummary.runs >= minimumRunsPerVariant) {
recommendation = delta.successRate > 0 ? 'promote-amendment' : 'keep-baseline';
}
return {
schemaVersion: EVALUATION_SCHEMA_VERSION,
generatedAt: new Date().toISOString(),
skillId,
amendmentId,
gate: {
minimumRunsPerVariant
},
baseline: baselineSummary,
amended: amendedSummary,
delta,
recommendation
};
}
module.exports = {
EVALUATION_SCHEMA_VERSION,
buildSkillEvaluationScaffold
};

View File

@@ -0,0 +1,118 @@
'use strict';
const HEALTH_SCHEMA_VERSION = 'ecc.skill-health.v1';
function roundRate(value) {
return Math.round(value * 1000) / 1000;
}
function rankCounts(values) {
return Array.from(values.entries())
.map(([value, count]) => ({ value, count }))
.sort((left, right) => right.count - left.count || left.value.localeCompare(right.value));
}
function summarizeVariantRuns(records) {
return records.reduce((accumulator, record) => {
const key = record.run && record.run.variant ? record.run.variant : 'baseline';
if (!accumulator[key]) {
accumulator[key] = { runs: 0, successes: 0, failures: 0 };
}
accumulator[key].runs += 1;
if (record.outcome && record.outcome.success) {
accumulator[key].successes += 1;
} else {
accumulator[key].failures += 1;
}
return accumulator;
}, {});
}
function deriveSkillStatus(skillSummary, options = {}) {
const minFailureCount = options.minFailureCount || 2;
if (skillSummary.failures >= minFailureCount) {
return 'failing';
}
if (skillSummary.failures > 0) {
return 'watch';
}
return 'healthy';
}
function buildSkillHealthReport(records, options = {}) {
const filterSkillId = options.skillId || null;
const filtered = filterSkillId
? records.filter(record => record.skill && record.skill.id === filterSkillId)
: records.slice();
const grouped = filtered.reduce((accumulator, record) => {
const skillId = record.skill.id;
if (!accumulator.has(skillId)) {
accumulator.set(skillId, []);
}
accumulator.get(skillId).push(record);
return accumulator;
}, new Map());
const skills = Array.from(grouped.entries())
.map(([skillId, skillRecords]) => {
const successes = skillRecords.filter(record => record.outcome && record.outcome.success).length;
const failures = skillRecords.length - successes;
const recurringErrors = new Map();
const recurringTasks = new Map();
const recurringFeedback = new Map();
skillRecords.forEach(record => {
if (!record.outcome || record.outcome.success) {
return;
}
if (record.outcome.error) {
recurringErrors.set(record.outcome.error, (recurringErrors.get(record.outcome.error) || 0) + 1);
}
if (record.task) {
recurringTasks.set(record.task, (recurringTasks.get(record.task) || 0) + 1);
}
if (record.outcome.feedback) {
recurringFeedback.set(record.outcome.feedback, (recurringFeedback.get(record.outcome.feedback) || 0) + 1);
}
});
const summary = {
skill: {
id: skillId,
path: skillRecords[0].skill.path || null
},
totalRuns: skillRecords.length,
successes,
failures,
successRate: skillRecords.length > 0 ? roundRate(successes / skillRecords.length) : 0,
status: 'healthy',
recurringErrors: rankCounts(recurringErrors).map(entry => ({ error: entry.value, count: entry.count })),
recurringTasks: rankCounts(recurringTasks).map(entry => ({ task: entry.value, count: entry.count })),
recurringFeedback: rankCounts(recurringFeedback).map(entry => ({ feedback: entry.value, count: entry.count })),
variants: summarizeVariantRuns(skillRecords)
};
summary.status = deriveSkillStatus(summary, options);
return summary;
})
.sort((left, right) => right.failures - left.failures || left.skill.id.localeCompare(right.skill.id));
return {
schemaVersion: HEALTH_SCHEMA_VERSION,
generatedAt: new Date().toISOString(),
totalObservations: filtered.length,
skillCount: skills.length,
skills
};
}
module.exports = {
HEALTH_SCHEMA_VERSION,
buildSkillHealthReport
};

View File

@@ -0,0 +1,108 @@
'use strict';
const fs = require('fs');
const path = require('path');
const os = require('os');
const OBSERVATION_SCHEMA_VERSION = 'ecc.skill-observation.v1';
function resolveProjectRoot(options = {}) {
return path.resolve(options.projectRoot || options.cwd || process.cwd());
}
function getSkillTelemetryRoot(options = {}) {
return path.join(resolveProjectRoot(options), '.claude', 'ecc', 'skills');
}
function getSkillObservationsPath(options = {}) {
return path.join(getSkillTelemetryRoot(options), 'observations.jsonl');
}
function ensureString(value, label) {
if (typeof value !== 'string' || value.trim().length === 0) {
throw new Error(`${label} must be a non-empty string`);
}
return value.trim();
}
function createObservationId() {
return `obs-${Date.now()}-${process.pid}-${Math.random().toString(16).slice(2, 8)}`;
}
function createSkillObservation(input) {
const task = ensureString(input.task, 'task');
const skillId = ensureString(input.skill && input.skill.id, 'skill.id');
const skillPath = typeof input.skill.path === 'string' && input.skill.path.trim().length > 0
? input.skill.path.trim()
: null;
const success = Boolean(input.success);
const error = input.error == null ? null : String(input.error);
const feedback = input.feedback == null ? null : String(input.feedback);
const variant = typeof input.variant === 'string' && input.variant.trim().length > 0
? input.variant.trim()
: 'baseline';
return {
schemaVersion: OBSERVATION_SCHEMA_VERSION,
observationId: typeof input.observationId === 'string' && input.observationId.length > 0
? input.observationId
: createObservationId(),
timestamp: typeof input.timestamp === 'string' && input.timestamp.length > 0
? input.timestamp
: new Date().toISOString(),
task,
skill: {
id: skillId,
path: skillPath
},
outcome: {
success,
status: success ? 'success' : 'failure',
error,
feedback
},
run: {
variant,
amendmentId: input.amendmentId || null,
sessionId: input.sessionId || null,
source: input.source || 'manual'
}
};
}
function appendSkillObservation(observation, options = {}) {
const outputPath = getSkillObservationsPath(options);
fs.mkdirSync(path.dirname(outputPath), { recursive: true });
fs.appendFileSync(outputPath, `${JSON.stringify(observation)}${os.EOL}`, 'utf8');
return outputPath;
}
function readSkillObservations(options = {}) {
const observationPath = path.resolve(options.observationsPath || getSkillObservationsPath(options));
if (!fs.existsSync(observationPath)) {
return [];
}
return fs.readFileSync(observationPath, 'utf8')
.split(/\r?\n/)
.filter(Boolean)
.map(line => {
try {
return JSON.parse(line);
} catch {
return null;
}
})
.filter(record => record && record.schemaVersion === OBSERVATION_SCHEMA_VERSION);
}
module.exports = {
OBSERVATION_SCHEMA_VERSION,
appendSkillObservation,
createSkillObservation,
getSkillObservationsPath,
getSkillTelemetryRoot,
readSkillObservations,
resolveProjectRoot
};