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

@@ -5,6 +5,10 @@ const fs = require('fs');
const path = require('path');
const { createAdapterRegistry, inspectSessionTarget } = require('./lib/session-adapters/registry');
const { readSkillObservations } = require('./lib/skill-improvement/observations');
const { buildSkillHealthReport } = require('./lib/skill-improvement/health');
const { proposeSkillAmendment } = require('./lib/skill-improvement/amendify');
const { buildSkillEvaluationScaffold } = require('./lib/skill-improvement/evaluate');
function usage() {
console.log([
@@ -18,12 +22,17 @@ function usage() {
' claude:latest Most recent Claude session history entry',
' claude:<id|alias> Specific Claude session or alias',
' <session.tmp> Direct path to a Claude session file',
' skills:health Inspect skill failure/success patterns from observations',
' skills:amendify Propose a SKILL.md patch from failure evidence',
' skills:evaluate Compare baseline vs amended skill outcomes',
'',
'Examples:',
' node scripts/session-inspect.js .claude/plan/workflow.json',
' node scripts/session-inspect.js workflow-visual-proof',
' node scripts/session-inspect.js claude:latest',
' node scripts/session-inspect.js latest --target-type claude-history',
' node scripts/session-inspect.js skills:health',
' node scripts/session-inspect.js skills:amendify --skill api-design',
' node scripts/session-inspect.js claude:a1b2c3d4 --write /tmp/session.json'
].join('\n'));
}
@@ -39,14 +48,57 @@ function parseArgs(argv) {
const targetTypeIndex = args.indexOf('--target-type');
const targetType = targetTypeIndex >= 0 ? args[targetTypeIndex + 1] : null;
const skillIndex = args.indexOf('--skill');
const skillId = skillIndex >= 0 ? args[skillIndex + 1] : null;
const amendmentIndex = args.indexOf('--amendment-id');
const amendmentId = amendmentIndex >= 0 ? args[amendmentIndex + 1] : null;
const observationsIndex = args.indexOf('--observations');
const observationsPath = observationsIndex >= 0 ? args[observationsIndex + 1] : null;
const writeIndex = args.indexOf('--write');
const writePath = writeIndex >= 0 ? args[writeIndex + 1] : null;
return { target, adapterId, targetType, writePath, listAdapters };
return { target, adapterId, targetType, writePath, listAdapters, skillId, amendmentId, observationsPath };
}
function inspectSkillLoopTarget(target, options = {}) {
const observations = readSkillObservations({
cwd: options.cwd,
projectRoot: options.cwd,
observationsPath: options.observationsPath
});
if (target === 'skills:health') {
return buildSkillHealthReport(observations, {
skillId: options.skillId || null
});
}
if (target === 'skills:amendify') {
if (!options.skillId) {
throw new Error('skills:amendify requires --skill <id>');
}
return proposeSkillAmendment(options.skillId, observations);
}
if (target === 'skills:evaluate') {
if (!options.skillId) {
throw new Error('skills:evaluate requires --skill <id>');
}
return buildSkillEvaluationScaffold(options.skillId, observations, {
amendmentId: options.amendmentId || null
});
}
return null;
}
function main() {
const { target, adapterId, targetType, writePath, listAdapters } = parseArgs(process.argv);
const { target, adapterId, targetType, writePath, listAdapters, skillId, amendmentId, observationsPath } = parseArgs(process.argv);
if (listAdapters) {
const registry = createAdapterRegistry();
@@ -59,12 +111,20 @@ function main() {
process.exit(1);
}
const inspectTarget = targetType ? { type: targetType, value: target } : target;
const snapshot = inspectSessionTarget(inspectTarget, {
const skillLoopPayload = inspectSkillLoopTarget(target, {
cwd: process.cwd(),
adapterId
skillId,
amendmentId,
observationsPath
});
const payload = JSON.stringify(snapshot, null, 2);
const payloadObject = skillLoopPayload || inspectSessionTarget(
targetType ? { type: targetType, value: target } : target,
{
cwd: process.cwd(),
adapterId
}
);
const payload = JSON.stringify(payloadObject, null, 2);
if (writePath) {
const absoluteWritePath = path.resolve(writePath);