Skip to content

CHANGELOG.md

Latest commit

 

History

History
248 lines (163 loc) · 19.7 KB

CHANGELOG.md

File metadata and controls

248 lines (163 loc) · 19.7 KB

Changelog

2026-05-06

  • VPS Agent: hex-relay — Telegram control-plane extracted from ln-030 references into standalone agents/hex-relay/ package; Fastify+Zod typed routes, runProcess utility, /health build/runtime metadata, structured god-session error classification with deduped admin alerts
  • Per-user god-sessions — systemd template instances *-god@<telegram_user_id>.service with bubblewrap-based agent-sandbox, per-project tmux sockets, per-user runtime under /var/lib/${PROJECT_NAME}/users/<id>; restart-on-update across all active instances
  • ln-030-vps-bootstrap skill (new) — L3 standalone one-shot idempotent VPS bootstrap for Claude Code + Codex workloads; configurable RELAY_HOOK_PORT/DISPATCH_COMMAND_NAME, REPO_URL/REPO_REF git verification, VPS-local fleet registry under /etc/agent-fleet, register-telegram-commands automation for global + all_private_chats scope
  • Inbound enrichment — local voice→text via ffmpeg + whisper.cpp (env-gated RELAY_VOICE_TRANSCRIPTION), durable queued message model with retry/abandon, relay-bot v6 text-only DB schema with control lane and 24h-throttled task-poll notifications; /tasks provider polling every 15 min with handoff UI
  • MCP packages — hex-line v1.27.1→v1.29.0 (Windows CRLF summary fix, generic-external-repo fallback scenarios, anchor-extraction de-flake), hex-graph v0.16.0→v0.19.0 (audit_workspace bounded preview with limit/clone_member_limit, gitignore-aware file discovery via git ls-files, full-rebuild resetProjectGraph on index_project), hex-ssh v1.6.1→v1.7.0 (per-tool timeout doc rows, regression guards)

2026-04-21

  • hex-line v1.17→v1.27 — PROTOCOL.md Response grammar (action-line + section/entry/detail prefixes), prose-noise cleanup (-37% output bytes), boundary-echo auto-strip preserves structural }/)/]/}); delimiters, post-edit lexical brace-balance advisory, range_checksum on replace_between, new warnings field on edit_file
  • hex-graph v0.13.0→v0.15.0 — text-only MCP protocol (content[0].text grammar; structuredContent/outputSchema retired), expansion pointers carry canonical selectors (symbol_id > workspace_qualified_name > qualified_name > name+file), new #quality section exposing family/support tier/languages, 102/102 semantic tests
  • hex-ssh v1.4.2→v1.6.0 — per-tool timeout field scoping (exec-path tools expose execTimeoutMs, transfer tools expose transferTimeoutMs), per-call timeout overrides for connect/keepalive/exec/transfer
  • Response grammar migration — hex-line graph-enrich output drops emoji prose (⚠ Semantic impact: → #semantic_impact, indented facts → . prefix, ⚠ N clone(s): → !clone_siblings count=N list=...); MCP output families documented (structured vs text-grammar) in MCP_OUTPUT_CONTRACT_GUIDE and MCP_TOOL_DESIGN_GUIDE
  • Docs & CI — five Claude hook events documented, plan-mode enforcement for mutating MCP tools, Monitor pattern docs, AGENTS docs unified with MCP result envelope, CI test suite repaired (hash collision test, 11 missing suites, deleted-runtime cleanup), Windows temp-path hint in coordinator CLI errors

2026-04-14

  • Advisor routing simplified — external advisor routing now uses the active registry and self-review fallback. Cleanup across agent_registry.json, agent_runner.mjs, and the shared agent delegation/review/memory references.

2026-04-10

  • hex-line v1.12→v1.16 — progressive disclosure read (edit_ready/verbosity), line-ending preservation (LF/CRLF/CR), allow_external scope boundary, graph staleness auto-refresh, summary-first grep
  • hex-graph v0.8→v0.12 — summary-first output with available_expansions/provenance, expanded use-case API, SCIP module fix, 92/92 semantic tests
  • Artifact-first coordinators — coordinators advance only from machine-readable artifacts; workers keep persisted runtime state and emit JSON summaries; new runtime contracts across all coordinator families
  • Planning runtime overhaul — new/updated contracts for scope decomposition, epic/story/task planning, dependency and worker runtimes with coordinator summary contract and guard tests
  • Skill updates — ln-200, ln-201, ln-210–ln-310, ln-840 refreshed with runtime alignment

2026-04-05

  • hex-line v1.12.0 — conservative conflict recovery (retry_edit/retry_edits/retry_plan), hook policy extraction, canonical output contracts for verify/changes
  • hex-graph v0.8.0 — output contract alignment (pruneEmpty, STATUS/ACTION constants, error mapping, DB busy handling), refactored platform and use-case API
  • hex-ssh v1.4.0 — remotePlatform parameter (auto/posix/windows), requirePosixRemotePath validation, transfer improvements
  • GitHub Actions Node.js 24 — upgraded checkout@v6, setup-node@v6, deploy-pages@v5
  • MCP Registry fix — shortened tool descriptions to ≤100 chars for hex-line and hex-graph

2026-04-01

  • hex-line v1.10.0 — tool descriptions refresh, graph-enrich refactor, hook safeExit, edit/read improvements
  • hex-graph v0.6.0 — new flow.mjs dataflow module with find_dataflows source/sink API, store expansion
  • CI fixes — guards.mjs split path fix, hex-ssh sftp test isolation, hex-graph cross-platform workspace discovery, better-sqlite3 native binary rebuild on CI

2026-03-31

  • Advisor model selection cleanup — removed hardcoded external advisor model flags from the agent registry. ln-011 post-install disables Conseca safety checker
  • hex-line search output cap — block-aware 80K char limit prevents CC truncation on large search results; emits OUTPUT_CAPPED diagnostic with narrowing guidance
  • Agent review suggestion schema v2 — added file, line_start, line_end, recommended_action fields to all review schemas and prompt templates; relaxed read-only constraint to allow trivial fixes
  • model: claude-sonnet-4-6 in skills — 32 SKILL.md files now declare explicit model; skills run on Sonnet 4.6 by default instead of inheriting parent model
  • hex-line hook safeExit + path filter — writeSync-based flush before exit prevents output loss; path filter simplified to project-dir check (drops complex settings.json allowlist)

2026-03-30

  • hex-graph workspace-aware API — tool parameters updated with explicit path, canonical selector rules, precision controls (min_confidence), and expanded language coverage (Python, C#, PHP with optional precise overlay)
  • hex-ssh file transfer and interop — new transfer.mjs module, interop test fixtures, and expanded SSH server capabilities
  • State machine guards audit — guards.mjs and store.mjs hardened across 12 runtimes with final_result/resumablePhases enforcement; new guard coverage tests for all runtime families
  • CI workflow — added .github/workflows/test.yml for automated testing
  • review-skills R9 false positive fix — auditor count check no longer triggers when site correctly omits volatile numbers

2026-03-29

  • ln-310 quality-based refinement — Codex iterative loop now exits by quality (MEDIUM/HIGH findings drive continuation), not just iteration count; added Architecture Gate, patience requirements, and risk mitigation criterion
  • Process cleanup enforcement — mandatory --verify-dead {pid} after every Codex call in ln-310, ln-510, ln-813; runtime guard blocks self-check pass without process verification
  • Exit reason enum canonicalized — CONVERGED/CONVERGED_LOW_IMPACT/MAX_ITER/ERROR/SKIPPED unified across workflow, runtime contract, status catalog, and smoke tests
  • Perspective-based refinement — 5 iterations now use different review angles (Generic, Dry-Run, New Dev, Adversarial, Final Sweep) instead of repeating the same criteria; exits on 2 consecutive APPROVED
  • Guard integration tests — new guards.mjs test validates all machine-enforced state transitions (exit_reason, processes_verified_dead)

2026-03-28

  • ln-015-hex-line-uninstaller — standalone skill to remove hex-line hooks, output style, and settings from the system
  • hex-line auto-sync — setup_hooks tool removed; hooks and output style now auto-sync on MCP server startup via content comparison

2026-03-27

  • Marketplace install docs — updated README, plugin docs, and GitHub Pages site to use the current Claude Code marketplace flow (/plugin marketplace add + /plugin install plugin@marketplace) instead of unsupported /plugin add
  • Audit runtime unification — 6XX coordinators and workers migrated to run-scoped artifacts, evaluation runtime CLIs, and JSON summary contract with summaryArtifactPath
  • Runtime status catalog — canonical status sets codified across active runtime contracts (story-gate, optimization, evaluation, environment)
  • Worker independence cleanup — removed 36 coordinator-aware DoD references, fixed 3 false positives in review-skills checks (R12/R13/R16)

2026-03-26

  • State platform runtime — shared coordinator runtime now enforces manifest, state, checkpoints, history, active-pointer, and status schemas across stateful skill families
  • Planning and environment migration — ln-010, ln-220, and ln-300 now run on identifier-scoped runtimes with pause/resume, replay, and standalone worker summaries
  • Coordinator family alignment — ln-310, ln-400, ln-500, ln-810, and ln-1000 now follow the same run-scoped runtime discipline and normalized CLI/status contracts
  • Review and contract hardening — ln-162 and review-skills now catch standalone-worker drift, reverse coupling, and non-run-scoped runtime artifacts
  • Docs and skill compression — planning, environment, and pipeline skills/docs were rewritten into shorter runtime-first contracts with updated public docs and site pages

2026-03-25

  • ln-010 npx cache probe — hex package version detection via npx cache scan instead of npm outdated -g; setup_hooks called unconditionally in Phase 3c verification
  • ln-014 auto-fix — Phase 5b auto-fixes missing Compact Instructions and MCP Tool Preferences sections in instruction files
  • ln-012 probe cleanup — removed global npm/npm ls fallback probes; hex packages are npx-only

2026-03-24

  • skills-catalog rename — skills/ → skills-catalog/ to fix plugin skill duplication (128×6 entries in autocomplete)
  • ln-840-benchmark-compare — new skill in optimization-suite: A/B benchmark (built-in vs hex-line), renamed from ln-015
  • bulk_replace caps — format param (compact/full), per-file diff cap (50L), payload cap (30K chars)

2026-03-23

  • ln-010 assess-dispatch-verify — redesigned from invoke-all to smart dispatch: probes environment once, builds decision matrix, skips workers with nothing to do
  • ln-150 removed — presentation-creator skill deleted; all references cleaned from pipeline, marketplace, site, docs
  • ln-162 Check 18 — new automated check verifies every SKILL.md has **Type:** line; prevents silent bypass of Check 9/17
  • run_checks.sh hardened — Check 5 scoped to skills-catalog/ paths only; Check 9/17 exclude Workers from coordinator-only requirements
  • ln-014 auditor → manager — renamed to ln-014-agent-instructions-manager; creates missing CLAUDE.md and AGENTS.md

2026-03-22

  • GitHub Pages MCP section — hex-line, hex-ssh, hex-graph MCP servers showcased on site index with dedicated detail pages (site/mcp/)
  • Plugin pages enhanced — all 6 plugin detail pages now link each skill row to its SKILL.md source on GitHub
  • GitHub Pages best practices — new standards doc for site development guidelines
  • hex-graph layered rewrite — canonical symbol identities, layered semantic edges, and the current use-case graph analysis surface replaced the old flat impact/context API
  • Agent review simplified — debate protocol (challenge/follow-up rounds) replaced with AGREE/REJECT verification + iterative Codex refinement loop (max 5 iterations)
  • hex-ssh security hardening — 4 new modules: command-policy, edit-validation, host-verify, shell-escape

2026-03-21

  • ln-012 MCP configurator — 3 critical phases added: hooks+outputStyle install (Phase 4b), allowed-tools REPLACE strategy with mcp__* preservation (Phase 4d), MCP Tool Preferences auto-write to CLAUDE.md/AGENTS.md (Phase 4e)
  • 01X consistency audit — ln-010 delegation table and rules aligned with ln-012 sanctioned write paths; ln-013 hooks mentioned in description, duplicate tool mapping removed

2026-03-20

  • hex MCP family — 3 npm MCP servers: hex-line (hash-verified file editing, 10 tools), hex-ssh (remote file ops over SSH, 6 tools), hex-graph (code knowledge graph, 7 tools); rebranded from sharpline
  • hex MCP v2 — output style system, hash-hint fallback, anchor-based editing, benchmark v3 (91-98% savings on multi-step workflows); hex-line v1.1.0, hex-ssh v1.1.0, hex-graph v0.2.0 published to npm
  • Setup Environment plugin — 7th plugin extracted from agile-workflow: ln-010 coordinator + 4 workers (agent installer, MCP configurator, marketplace/config aligner, instructions auditor)
  • Research skills consolidated — ln-001 and ln-002 merged into ln-310 and ln-220 via shared layer; research methodology and docs creation extracted to shared references
  • Best practice guides — MCP Tool Design, Hook Design, Prompt Caching; hooks redesigned with dangerous command blocker

2026-03-19

  • hex MCP family — 3 bundled MCP servers: hex-line (hash-verified file editing, 10 tools), hex-ssh (remote file ops over SSH, 6 tools), hex-graph (code knowledge graph with tree-sitter AST, 7 tools); FNV-1a hashing, security boundaries; npm publishable
  • Agent runner overhaul — Windows spawn fix (whichSync PATHEXT), heartbeat removed (log-based monitoring), registry 4→2 agents with focus_hint, --approval-mode yolo ⚠️ BREAKING
  • Python → Node.js ESM — all runtime scripts (.py) replaced with .mjs: agent_runner, 3 hooks, analyze_test_logs; Python dependency eliminated ⚠️ BREAKING
  • ln-1000 redesign — pipeline orchestration uses sequential Skill() calls; quality gate (ln-500) and test planning (ln-520) can no longer be skipped
  • GitHub Actions — npm auto-publish for hex-line-mcp on tag hex-line-v*

2026-03-18

  • Agent process tree kill — agent_runner kills entire process tree (not just immediate child) on both timeout and normal completion; --verify-dead CLI flag for safety net checks
  • Python advanced tools mandatory — import-linter, deptry, vulture, pip-audit promoted from optional to required in linter configurator; new config templates added
  • Multi-stack verification matrix — quality setup coordinator now has tool matrix across TypeScript/Python/.NET for all verification checks
  • Epistemic protocol — new shared/references/epistemic_protocol.md for source attribution and anti-hallucination across all research skills; integrated into research_tool_fallback, phase2_research_audit, solution_validation
  • Description triggers + agent timeout 30min — all 125 descriptions rewritten with "Use when..." triggers; ln-162 reviewer gains M6 + CHECK 14; all agent hard timeouts raised to 30 min; 7XX bootstrap skills get Meta-Analysis sections

2026-03-16

  • Codex Windows performance — agent_runner auto-detects Windows and prepends prompt hint directing Codex to prefer built-in file read over PowerShell shell commands (5-15s overhead per call)

2026-03-15

  • Multi-cycle optimization — performance pipeline now iterates (profile → research → validate → execute → repeat) until target met or plateau detected; each cycle discovers new bottlenecks as dominant ones are fixed (Amdahl's law)
  • Cross-service performance profiling — optimization pipeline traces bottlenecks across microservices (monorepo, git submodules, docker-compose); profiles inside accessible services instead of treating them as black boxes
  • Community Engagement plugin — new plugin with skills for automated GitHub community management: triage issues/PRs, compose announcements, launch RFC debates, respond to threads
  • Token efficiency: output normalization — new shared reference normalizes, deduplicates, and groups CLI output before presenting to agent; reduces noise in test runners, build auditors, profilers, log analyzers
  • Skill reviewer automated script — ln-162 Phase 2 checks now run via executable run_checks.sh instead of manual template assembly

2026-03-14

  • Agent sandbox fix — plan files from outside project workspace now materialized for agent access when advisor CLIs require project-local inputs

2026-03-13

  • Test log analysis — new skill classifies errors from Docker/file/Loki logs into 4 categories; only Real Bugs block quality verdict
  • Documentation skill extraction — scan project docs, extract procedural content into reusable .claude/commands with quality review

2026-03-11

  • Pipeline orchestrator hardening — Plan Gate now enforced at all 5 stages, worker prompts consolidated

2026-03-08

  • Plugin marketplace — split into 5 focused plugins installable individually: agile-workflow, documentation-pipeline, codebase-audit-suite, project-bootstrap, optimization-suite
  • Optimization Suite — new plugin for full-stack performance optimization (profile → research → execute), dependency upgrades (npm/NuGet/pip), code modernization (OSS replacement, bundle optimization)
  • Destructive operation safety — all skills now classify destructive actions by severity with human-in-the-loop gates

2026-03-07

  • Two-layer detection — audit skills now use grep pre-filter + AI context analysis instead of pure AI scanning (faster, fewer false positives)

2026-03-06

  • Documentation fact-checker — new skill extracts verifiable claims from .md files (paths, versions, configs) and cross-checks against codebase

2026-02-13

  • Pipeline Orchestrator — one command drives a Story through the full lifecycle: task planning → validation → implementation → quality gate → merge to develop

2026-02-12

  • Multi-round agent debate — advisor sessions now persist across challenge rounds, preserving full reasoning context during disagreements

2026-02-11

  • Multi-model code review — parallel advisor analysis with Critical Verification: the host independently validates each suggestion and debates controversial findings (max 2 rounds)
  • Risk Analysis in validation — 6 risk categories with Impact × Probability scoring before Story approval
  • Persistence performance audit — new skills for query efficiency, transaction correctness, blocking I/O, resource lifecycle analysis

2026-01-10

  • Project Bootstrap — 32 new skills for scaffolding production-ready projects or transforming existing ones to Clean Architecture. Supports React, .NET, Python with Docker, CI/CD, security scanning, and quality tooling setup

2025-12-23

  • RICE prioritization — new skill scores Stories by Reach, Impact, Confidence, Effort with automated market research

2025-12-21

  • Validation overhaul — universal pattern detection (OAuth, REST, ML pipelines) with fast path for trivial CRUD stories

2025-11-21

  • 100x token reduction — coordinators now load Story/Task metadata only (~50 tokens vs ~5,000), delegating full reads to workers

2025-11-14

  • Orchestrator-Worker architecture — 3-level hierarchy (L1 orchestrators → L2 coordinators → L3 workers) with Progressive Disclosure for 24-40% documentation reduction

2025-11-10

  • v1.0.0 — Agile workflow automation end-to-end: scope decomposition, task execution, quality gates, Linear integration, Risk-Based Testing