v1.32.2

v1.32.2

Fixes

• Hub node secret cache TTL: Added TTL-based expiration to the hub node secret cache, preventing stale 403 errors when secrets rotate (#258)
• completeDistillation state variable: Fixed unnecessary variable rename (state2 -> state) in completeDistillation

Improvements

• Error logging in silent catches: Added console.warn to previously silent catch {} blocks in a2aProtocol.js (transport/heartbeat paths) and assetStore.js (file I/O operations), improving debuggability without changing behavior
• var -> const/let migration: Migrated all var declarations to const/let in skillDistiller.js for modern JavaScript standards

Files changed

• src/gep/a2aProtocol.js -- error logging in 4 catch blocks
• src/gep/assetStore.js -- error logging in 8 catch blocks
• src/gep/skillDistiller.js -- full var -> const/let migration
• src/gep/learningSignals.js -- minor fix

v1.32.0

What's New in v1.32.0

Improvements

• feat: structured learning signals -- expand raw signals into categorized tags (problem:X, area:X, action:X, risk:X) for better gene selection and evolution feedback
• feat: adaptive strategy policy -- detect repair/failure streaks and auto-adjust blast radius and innovation pressure
• feat: gene learning history -- track success/failure outcomes per gene with anti-pattern recording for smarter selection
• feat: failure-driven candidate extraction -- group recurring failed capsules into improvement candidates
• feat: auto-distillation -- synthesize new genes from successful patterns without requiring LLM interaction
• feat: soft/hard failure classification in solidify -- distinguish retryable validation failures from destructive constraint breaches

Bug Fixes

• fix: tool frequency threshold raised from 2 to 3 to reduce noise in candidate extraction

Other

• chore: comprehensive test coverage for all new learning closure features (6 new test files)

Contributors

• @firstdraft-work (PR #260) -- Strengthened the evolver learning closure with structured signals, adaptive policy, and auto-distillation

v1.31.0

What's New

Skill Store Fetch Command

• New evolver fetch --skill <id> CLI command to download skills from the Skill Store
• Supports -s <id>, --skill=<id>, --skill <id>, and positional argument
• Auto-registers with Hub if no node_secret exists
• Downloads SKILL.md + bundled files to ./skills/<skill_id>/ (customizable via --out=<dir>)
• The npx @evomap/evolver fetch --skill <id> command shown on skill detail pages now works

Bug Fixes

• writeStateForSolidify: ensure the correct directory (evolutionDir) exists before writing state file, preventing ENOENT on scoped or first runs
• rejectPendingRun: use atomic tmp+rename write to prevent state file corruption on crash
• EVOLVE_BRIDGE: respect pre-set environment variable in loop mode instead of unconditionally overwriting to false
• selectedBy: remove duplicate declaration in solidify try-block
• .env loading: use getRepoRoot() in index.js to match evolve.js path resolution
• WORKSPACE_ROOT: use getWorkspaceRoot() from paths.js for consistent path resolution

Internal

• Export getHubUrl from a2aProtocol.js
• Remove stale .env rewrite rule from public.manifest.json
• Sanitize skill_id in output path to prevent path traversal

v1.30.2

What's New in v1.30.2

Bug Fixes

• fix: remove destructive git rollback from loop auto-reject (#257 by @firstdraft-work)
  • rejectPendingRun no longer runs git checkout -- . and git clean -fd, which could delete untracked work files during loop-mode auto-reject
  • State-only rejection is now used, preserving all working directory contents

Other

• test: add loop-mode auto-reject test (test/loopMode.test.js)
• refactor: export main, readJsonSafe, rejectPendingRun, isPendingSolidify from index.js for testability

v1.29.9

What's New in v1.29.9

Skill Distillation Quality Overhaul

Distilled skills now meet marketplace-quality standards. Previously, distilled skills often contained timestamps in IDs, signals, and summaries, had minimal strategy steps, and generated bare-bones SKILL.md content.

Distiller prompt:

• Complete rewrite of the LLM prompt with explicit rules for signals_match (no timestamps/tool names), summary (marketplace-quality), strategy (5-10 concrete steps), and preconditions
• Each rule section includes Good/Bad examples
• Added QUALITY BAR section that frames output as a marketplace listing

Signal sanitization:

• New sanitizeSignalsMatch() strips trailing timestamps, rejects pure-numeric signals, tool-name-only signals (cursor, vscode, bypass, distill, etc.), and signals with 8+ digit sequences
• Signals are now sanitized BEFORE ID derivation (fixes bug where deriveDescriptiveId used dirty data)

Gene validation:

• Strip ALL embedded timestamps from gene IDs (not just trailing)
• Require minimum 3 strategy steps
• Auto-set schema_version: "1.6.0"
• Summary timestamps stripped before ID derivation

SKILL.md generation:

• Title Case display names
• Added "When to Use" and "Metadata" sections
• Strategy steps render with bold leading verb
• Constraints rendered with inline code formatting
• No more duplicate sections
• Gene object is shallow-copied to avoid caller mutation

Bug Fixes

• fix: deriveDescriptiveId no longer reads dirty signals_match
• fix: duplicate Preconditions / When to Use sections removed
• fix: extractStepVerb handles backtick-prefixed steps gracefully
• fix: publishSkillToHub no longer mutates caller's gene object

v1.29.8

What's New in v1.29.8

Bug Fixes

• fix: getHubNodeSecret() now reads A2A_NODE_SECRET environment variable (resolves #245). Deployments that inject the secret solely via env var no longer fail Hub Search with 401. Priority order: A2A_NODE_SECRET > in-memory cache > persisted file > A2A_HUB_TOKEN (compat fallback)
• fix: sanitize newlines in gene name/description before YAML output

Improvements

• feat: enforce descriptive naming rules for distilled skills
• refactor: add warn logging to empty catch blocks in solidify.js (#237)
• docs(security): declare env vars, network endpoints, shell commands and file access in SKILL.md metadata
• chore: add workspace/ and runtime-generated GEP state files to gitignore

v1.29.4

What's New in v1.29.4

Heartbeat Environment Fingerprint

• Evolver now sends env_fingerprint in the first heartbeat after startup, so the Hub can detect evolver version without requiring a fresh hello
• Only marks fingerprint as sent when capture succeeds and includes evolver_version, retries on subsequent heartbeats if the first attempt fails

Bug Fixes

• fix: prevent false "no evolver detected" warning for agents that upgraded evolver but had not re-sent hello
• fix: add logging to critical-path empty catch blocks in evolve.js
• feat: add ESL-1.0 license notice to generated SKILL.md footer
• fix: prevent data loss from parent .git and OOM on large candidates files

v1.29.3

Bug Fixes

• [Critical] Fix parent .git data loss (#198): getRepoRoot() now checks evolver's own directory first. Parent .git repositories are ignored by default -- set EVOLVER_USE_PARENT_GIT=true to opt in. This prevents git reset --hard from running in the wrong scope.

• Fix OOM on large candidates.jsonl (#210): readRecentCandidates() and readRecentExternalCandidates() now use a tail-read strategy for files >1MB instead of loading the entire file into memory.

v1.29.2

What's New in v1.29.2

Skill Store Integration

• Distilled Genes are automatically published to the EvoMap Hub Skill Store as SKILL.md files (Claude/Anthropic Agent Skill format)
• Gene-to-SKILL.md conversion maps signals_match, strategy, constraints, validation to structured markdown
• Auto-update on re-distillation, set SKILL_AUTO_PUBLISH=0 to disable

Other Changes (since v1.27.5)

• feat: add npm publishing as @evomap/evolver
• feat: add commitment tracking for task deadlines
• feat: merge community PRs #217, #218, #226, #232
• fix: recompute asset_id after adding model_name
• fix: resolve evolver_version detection for npm deployments
• fix: commitment deadline edge case and heartbeat meta regression

v1.28.2

fix: recompute asset_id after adding model_name in buildPublishBundle

When EVOLVER_MODEL_NAME was set, model_name was added to gene/capsule objects after asset_id had already been computed. Hub-side verifyAssetId would recompute the hash including model_name, causing gene_asset_id_verification_failed.

Update: git pull && npm install