write_plan tool, review-fix pipeline, and tool-based submission (#562)AskUserQuestionCard option button text wrapping (#567)web-research skill — reusable research primitive that turns a sharp research question into a written plan, parallel-subagent findings, and a cited synthesis report (TL;DR / Findings / Gaps & Limitations / Sources) on disk. Dual-mode: directly invocable by users and programmatically invocable by companion skills (prfaq-beagle, brainstorm-beagle, strategy-interview) via a documented contract (#96)artifact-analysis skill — sibling primitive to web-research that scans local documents and project knowledge (auto-discovering .beagle/concepts/, .planning/, docs/, and root README/brief files by default) via parallel subagents and produces a path-cited synthesis report with fixed sections (Documents Found / Key Insights / User / Market Context / Technical Context / Ideas & Decisions / Raw Detail Worth Preserving / Gaps & Limitations). Dual-mode with the same companion contract used by prfaq-beagle, brainstorm-beagle, and strategy-interview (#96)prfaq-beagle skill — hardcore Working Backwards PRFAQ coach that runs a 5-stage gauntlet (Ignition → Press Release → Customer FAQ → Internal FAQ → Verdict) to filter weak product, internal-tool, or OSS concepts before they consume brainstorm-beagle cycles. Detects concept type in Ignition and calibrates later stages; invokes web-research and artifact-analysis serially for grounding with graceful degradation on web-tools-unavailable and resume-by-default on prior-run-present. Binary pass/fail verdict — on pass, produces a concept brief at .beagle/concepts/<slug>/brief.md that brainstorm-beagle auto-ingests; on fail, produces targeted feedback naming exactly what stage to re-enter and what would need to be true (#96)brainstorm-beagle — relocate spec output from docs/specs/YYYY-MM-DD-<topic>.md to .beagle/concepts/<slug>/spec.md; auto-detect brief.md at startup and ingest it to skip most discovery; when invoking web-research or artifact-analysis mid-session, land their outputs under the shared .beagle/concepts/<slug>/research/ and .beagle/concepts/<slug>/analysis/ folders so the whole concept-forging audit trail lives in one place (#96)cli: Add --deep mode for multi-stack code review with inline PR comments (#45)
A 5-stage pipeline (exploration → TTT intent → TTT alternatives → per-stack fan-out → cross-stack merge) with an optional fix gate that auto-detects the stacks touched by the diff, fans out per-stack reviews in parallel via the matching Beagle skills, merges findings with dedup, and posts the result as a single atomic inline GitHub PR review. Handles mixed-stack PRs (e.g. Python + React) that existing single-stack modes can't review cleanly. Falls back to generic review when a per-stack skill is unavailable.
cli: Add --start-at {ttt,per-stack,merge,fix} for stage-granular resume of --deep runs (#45)
.daydream/deep/ artifacts are preserved across runs so an interrupted pipeline can resume from a later stage without re-running earlier work. Each resume target enforces an artifact precondition and fails with an actionable error naming the missing file.
pr-review: Post inline GitHub PR comments from --ttt and --deep (#45)
Anchor-greps each finding to a real head-SHA line, classifies against diff hunks, and posts a single atomic review via the GitHub API. Cross-stack and off-hunk findings fold into the review body with severity (high/medium/low) and confidence (HIGH/MEDIUM/LOW) breakdowns. y/n gated; non-fatal on failure; payload preserved for retry.
phases: phase_parse_feedback accepts a keyword-only input_path: Path | None parameter (#45)
Default None preserves the existing cwd/REVIEW_OUTPUT_FILE behavior for all existing callers. Explicit paths let the per-stack deep-review fan-out parse multiple review files in parallel without colliding.
runner: Derive skill availability from the Claude Code plugin registry at runtime (#45)
Reads $CLAUDE_CONFIG_DIR/plugins/installed_plugins.json to check whether a beagle-<stack> plugin is installed. When absent, deep mode routes that stack to the generic fallback review instead of letting the call silently fail with a swallowed MissingSkillError.
Multi-agent research platform for parallel workflows, fine-tuning, and reinforcement learning with real-time observability.
Rust · Elixir · gRPC · PostgreSQL
Benchmark suite for evaluating multi-agent coding systems across multi-step, multi-file workflows.
Python · Docker · PostgreSQL
Local-first iOS agent that ingests real-time biometrics, calendar, and location data to plan and adjust fitness, nutrition, and productivity goals.
Metal · SwiftUI · SwiftData · HealthKit