Skip to content

CHANGELOG.md

Latest commit

 

History

History
245 lines (200 loc) · 10.8 KB

CHANGELOG.md

File metadata and controls

245 lines (200 loc) · 10.8 KB

Changelog

All notable changes to the Hydra framework will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.3.1] - 2026-05-09

Added

  • Statusline savings indicator — green ↓$X.XX after the cost field, showing real USD saved vs an all-Opus baseline. Hidden when savings <$0.01 or session has no delegations yet.
  • /hydra:stats retail-price effect — strikethrough on hypothetical all-Opus cost, bold green on actual cost. Visually reinforces savings.
  • Shared hooks/hydra-token-math.js helper — single source of truth for token parsing and savings math. Statusline and /hydra:stats both call it so numbers never disagree.
  • Terminal capability detection — strikethrough auto-enabled on iTerm2, kitty, Alacritty, WezTerm, Windows Terminal, VS Code, Apple_Terminal. Auto-disabled on Git Bash (MSYSTEM) and Cygwin where rendering is inconsistent. User override: HYDRA_STRIKETHROUGH=0|1.
  • 15-second in-memory cache on token math so statusline refreshes don't re-parse the full session JSONL on every tick.

Fixed

  • Statusline hides the savings field at session start when no delegations have happened yet (avoids $0.00 ↓$0.00).

Notes

  • Strikethrough is purely additive — terminals without support get a clean Baseline / Actual cost labeled fallback.
  • Statusline savings match /hydra:stats exactly (shared codepath).

[2.3.0] - 2026-05-06

Added

  • /hydra:stats — Real token tracking command. Reads Claude Code session JSONL directly to show actual token usage by model, delegation rate, actual cost, hypothetical all-Opus cost, and real savings in USD. No AI estimation. Cross-platform Node.js implementation works on Windows, macOS, and Linux. Respects CLAUDE_CONFIG_DIR env override.
  • Internal Compression Protocol — two-tier compression strategy:
    • Tier 1 (Orchestrator): Light compression of Opus responses. Drops filler, pleasantries, and hedging while keeping natural prose. Average user won't notice the difference.
    • Tier 2 (Sub-Agents): Aggressive compression of subagent output. Structured terse output (tables, key:value, JSON). Sub-agent output goes to Opus, not the user, so this can be aggressive without UX impact.
  • Auto-Clarity rule — orchestrator drops compression for security warnings, destructive operations, multi-step instructions, and onboarding/learning contexts. Safety-critical output stays full prose.

Changed

  • Slash commands increased from 10 to 11.
  • Every subagent now has an "Output Format — Compressed (MANDATORY)" section in its system prompt with role-specific format hints.
  • SKILL.md includes a new Response Compression Protocol — Orchestrator section near the top of the framework spec.

Improved

  • Subagent output ~60-70% smaller, allowing Opus's context to last longer in extended sessions.
  • Orchestrator responses ~30-40% shorter without losing readability.
  • Real token tracking lets users verify Hydra's savings claims with their own session data — no marketing fluff, just the numbers from Anthropic's API responses.

2.2.0 - 2026-04-11

Added

  • /hydra:preflight — Two-phase environment and compatibility validation command for new projects. Solves the "15-hour debugging session caused by a 2-minute pre-flight check that never happened" problem.
    • Phase 1 — hydra-preflight (Haiku 4.5): Detects runtime versions, runs GPU/CUDA probe scripts (torch.cuda.is_available(), etc.), inventories installed packages, diffs .env.example against .env, verifies build tools, and checks service connectivity. Returns a structured PREFLIGHT_INVENTORY JSON.
    • Phase 2 — hydra-analyst (Sonnet 4.6): Cross-references the inventory against known compatibility matrices (PyTorch/CUDA, React/Next, Python/TF, Node/native addons). Produces three-state verdicts: ✅ COMPATIBLE, ⚠️ KNOWN RISK, ❌ CONFIRMED BREAK. Probe output is treated as ground truth over matrix knowledge.
    • Auto-suggests pinned-version fixes for ❌ verdicts. Flags unknown combinations as UNVERIFIED rather than silently passing.
  • hydra-preflight agent (10th head) — Haiku 4.5; tools: Read, Bash, Glob.
  • Preflight rate metric added to the Measuring Impact section in SKILL.md (target: 100% of new project sessions).

Changed

  • Agent count increased from 9 to 10. "The Nine Heads" → "The Ten Heads" throughout SKILL.md and README.md.
  • Slash command count increased from 9 to 10.
  • npm-package/files/SKILL.md resynced with top-level SKILL.md — this closes a pre-existing v2.1.1 drift where the "Blocking vs Non-Blocking Dispatch" → "Sequential vs Parallel Dispatch" rename had not been propagated to the npm-published copy.

2.1.1 — Bug Fix

Fixed

  • Ghost message bug (#2): Background agent dispatch (fire-and-forget) caused an empty user turn to appear in Claude Code after agent completion, triggering Claude to respond to nothing. All agents are now dispatched in parallel foreground mode — Opus waits for every dispatched agent before presenting results. Same speed benefit (agents still run simultaneously), no ghost messages.

Changed

  • SKILL.md: "Blocking vs Non-Blocking Dispatch" section renamed to "Sequential vs Parallel Dispatch". All fire-and-forget language removed. Parallel dispatch is now the model for independent agents running simultaneously.

2.1.0 - 2026-03-26

Added

  • Codebase Map — persistent JSON dependency map built by hydra-scout
    • File-level import/export tracking with automatic reverse indexing
    • Risk scores per file (low/medium/high/critical based on dependent count)
    • Environment variable index (tracks every env var reference across the codebase)
    • Test coverage mapping (covered/partial/untested per file)
    • Git hash staleness detection — incremental updates, only re-maps changed files
    • Stored at .claude/hydra/codebase-map.json, no external dependencies
  • /hydra:map slash command — view summary, query blast radius, force rebuild
  • Map-aware sentinel scanning — instant blast-radius lookups instead of grep
  • Risk-based sentinel triggering — critical files always get deep analysis, low-risk files get fast-track acceptance
  • Map-aware orchestration — Opus uses risk scores for smarter routing decisions

Changed

  • hydra-scout now builds and maintains the codebase map alongside exploration
  • hydra-sentinel-scan uses map for dependency checks (falls back to grep if no map)
  • hydra-sentinel reads map for blast radius context during deep analysis
  • SKILL.md updated with Map-Aware Orchestration protocol
  • Routing guide updated with map-aware examples
  • Slash commands increased from 8 to 9

Improved

  • Sentinel scanning ~3-5x faster with map (JSON lookup vs grep)
  • Sentinel scanning ~3-5x fewer tokens with map (reads only blast radius files)
  • Scout exploration dramatically faster on repeat sessions (reads map, skips full scan)
  • Risk-proportional verification reduces unnecessary sentinel runs on low-risk changes

2.0.3 - 2026-03-16

Changed

  • Replaced soft delegation guidelines with mandatory hard rules in SKILL.md
  • Routing guide rewritten with agent-based examples and 7 mandatory delegation patterns
  • Added Plan Mode behavior: built-in Explore ok for planning, Hydra agents for execution
  • Added parallel dispatch enforcement for multi-file tasks
  • Added delegation overhead budget (max 2-3 self-handled tasks per session)
  • Added dispatch log accountability showing delegation vs direct ratio

2.0.2 - 2026-03-15

Fixed

  • Enforce sentinel execution with 3-layer trigger system
  • Sentinel pipeline now fires reliably on all code changes

2.0.1 - 2026-03-11

Fixed

  • Skip overwrite prompt in non-interactive mode (--global/--local/--both)

2.0.0 - 2026-03-11

Added

  • hydra-sentinel-scan (Haiku 4.5) — fast integration sweep after code changes
  • hydra-sentinel (Sonnet 4.6) — deep integration analysis when issues found
  • Persistent agent memory (memory: project) across all 9 agents
  • Orchestrator memory via CLAUDE.md Hydra Notes
  • Sentinel Protocol in SKILL.md — two-tier verification pipeline
  • Post-Code-Change Pipeline — code changes held until sentinel + guard complete
  • Fix decision tree: auto-fix trivial, offer medium, report complex
  • Compaction warning in statusline at 70%+ context usage

Changed

  • Agent count increased from 7 to 9
  • All agent frontmatter updated with memory: project
  • SKILL.md substantially rewritten with Sentinel Protocol and Orchestrator Memory
  • README overhauled with Sentinel section, Memory section, updated features, new FAQ

Breaking

  • Code changes now block until sentinel + guard verification completes

1.2.3 - 2026-03-08

Changed

  • CI: Publish to both npm and GitHub Packages on release
  • CI: Gracefully skip publish if version already exists

Added

  • npm lifetime downloads badge in README

1.2.2 - 2026-03-05

Fixed

  • Remove model annotations from agent name fields (was causing display issues)
  • Target npmjs.com registry instead of GitHub Packages

1.2.1 - 2026-02-24

Fixed

  • Move skill to skills/hydra/, update frontmatter description
  • Add hydra-guard & hydra-git agents, dynamic version, fix head count display

Added

  • GitHub Packages publish workflow

Changed

  • npm package README corrected to show skills/hydra/ install tree

1.1.0 - 2026-02-23

Added

  • npx installer package (npx hail-hydra-cc)
  • 7 specialized agents (scout, runner, scribe, guard, git, coder, analyst)
  • Auto-guard system
  • Configuration system (hydra.config.md)
  • "Why I Built This" section in README

Changed

  • Corrected pricing information

1.0.0 - 2026-02-18

Added

  • Initial release
  • SKILL.md orchestrator instructions
  • Reference documents (routing guide, model capabilities)
  • Wave execution, verification reports, handoff protocol
  • 4 orchestrator-level speed optimizations