Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
125 lines (91 loc) · 7.04 KB

RELEASE_NOTES.md

File metadata and controls

125 lines (91 loc) · 7.04 KB
Raw

Shannon v0.1.1 — Release notes

Date: 2026-05-28 Status: Release candidate (structural; brownfield-planning patch applied). User-action steps: see RELEASE_CHECKLIST.md.

What is this release

Shannon v0.1.0 is the foundational scaffold of the framework — the smallest interesting surface that delivers the 5 pillars on disk:

  1. Embedded sub-agent skills (Architecture C)
  2. Orchestration (single-message multi-Task dispatch)
  3. Iron Rule validation (no mocks, no stubs, no *.test.* files)
  4. Meta-judge consensus (rubric YAML before any judge)
  5. Self-instrumented (scripts/doctor.py mechanical contract check)

What's in the box

Layer Count Files
Skills 32 skills/<name>/SKILL.md + references/*.md
Agents 9 agents/<name>/manifest.yml + AGENT.md + _built/skills/
Commands 20 commands/*.md
Hooks 7 hooks/*.js + hooks.json + _lib.js
Build scripts 3 build/embed-skills.py, build/verify-build.py, build/README.md
Doctor 1 scripts/doctor.py
Harness 5 scripts/harness/sdk_runner.py, tmux_runner.py, lib.py, run-all.sh, benchmarks/*.yml (×5)
Docs 8 README.md + docs/*.md (INSTALL, QUICK_START, SKILLS_CATALOG, ARCHITECTURE, FUNCTIONAL_VALIDATION_GUIDE, CONTRIBUTING, DEV_HARNESS)

Headline metrics

  • 34 embedding relationships across the 9 agents (23 unique skills embedded)
  • 0 cascade conflicts — build uses straight copy + AGENT.md inlining
  • 0 phantom references across 19 commands × 31 skills × 9 agents × 7 hooks
  • 0 doctor mismatches (8/8 checks PASS)
  • 5/5 pillar benchmark coverage in harness
  • 0 broken internal doc links
  • 0 forbidden Iron Rule patterns (*.test.*, mock_*, etc.) in scripts/harness/

Phase-by-phase build evidence

Phase Output Verdict
0 PRD approval docs/PRD-V1.md, DECISIONS.md, BRIEF.md, ROADMAP.md PASS
1 Skill curation (67→31) skills/ (31 dirs); cut-skills.md; KEEP_PROPOSAL.md USER APPROVED PASS
2 Agent embedding (Architecture C) agents/ (9 manifests + AGENT.md + _built/); build/embed-skills.py; embedding-map.md APPROVED PASS
3 Command consolidation (27→19) commands/ (19); cut-commands.md; 0 phantom refs PASS
4 Hook curation (16→7) + contract hooks/ (7); hooks.json; 8 skills with required_hooks; scripts/doctor.py PASS
5 Validation harnesses scripts/harness/ (SDK + Tmux + 5 benchmarks); docs/DEV_HARNESS.md PASS
6 Documentation (8 docs) README.md + docs/*.md; 0 broken links PASS
7 Release (this file) + RELEASE_CHECKLIST.md + FINAL_SUMMARY.md STRUCTURAL PASS

Each phase's full validation evidence is in .planning/phases/<phase>/PHASE-*-VALIDATION.md.

v0.1.1 brownfield-planning patch

After v0.1.0 structural PASS, the user surfaced a capability gap: brownfield projects need a deep code survey + skill-inventory pass before planning. This patch closes it:

  • New command /shannon:scope — read-only brownfield reconnaissance. Runs codebase-analysis (5 parallel scientists) + skill-inventory (runtime MCP-tool discovery) + observability-report (recent session context) in parallel via Task: team-builder; synthesizes a scope-report.md for downstream planning.
  • New skill skill-inventory — calls mcp__skills__list_skills + mcp__plugins__list_plugins at runtime; maps installed capabilities to the task at hand. Refuses to fabricate recommendations.
  • /shannon:plan --mode brownfield — auto-runs /shannon:scope first; plan cites specific files + specific skills + specific proving-commands.
  • /shannon:cook --brownfield — same flow at the cook level; recommended for any non-trivial change on existing code.
  • planner agent — now embeds codebase-analysis + skill-inventory (5 → 7 embedded skills); brownfield-survey is in the agent's spawn context.
  • 37 stale /shannon:forge references swept — context-aware substitution across 14 source files; 0 forge mentions remain.

See .planning/patches/01-brownfield-planning.md for the patch detail.

What's intentionally NOT in v0.1.x

These were considered and deferred:

  • 3-oracle quorum review — the pre-execution + post-execution review from legacy forge. Approximated by Task: critic in v0.1.x; true 3-oracle quorum deferred to v1.x.
  • Agents coordinator, oracle, researcher, red-teamer, dispatch-judge — absorbed into existing v1 agents (e.g., critic covers oracle + red-teamer; team-builder covers coordinator + researcher).
  • Github-based marketplace publishing — v0.1.0 supports local-directory marketplace install. v0.2+ may add GitHub marketplace flow once the API surface stabilizes.
  • MCP-required workflows — Context7 and sequential-thinking are recommended but not required. Other MCPs are deferred.
  • Live runtime validation in CI — Phase 5 harness covers dry-run validation in sandboxes; live runs require user-action with Anthropic API access.

Known limitations

  1. Local marketplace only. Publishing to a GitHub-hosted marketplace is a v0.2+ enhancement.
  2. Live harness requires user-action. The sandbox where v0.1.0 was built has no API access, so the live-mode harness runs are deferred to user execution per RELEASE_CHECKLIST.md.
  3. The SKILLS_CATALOG.md is hand-regenerated. The Phase 6 script regenerates it from disk; in v1.x this could be promoted to a scripts/sync-catalog.py that runs on every doctor invocation.
  4. No CI integration. v0.1.0 ships with the harness; CI wiring (GitHub Actions calling bash scripts/harness/run-all.sh --live) is a v0.2+ task.

How to install

# Local install:
/plugin marketplace add /path/to/shannon-framework
/plugin install shannon@shannon-framework

# Restart Claude Code
/shannon:enforce on
/shannon:doctor       # expect 0 mismatches

See docs/INSTALL.md for details.

How to verify the release

cd v1/shannon
python3 build/embed-skills.py        # exit 0
python3 build/verify-build.py        # mismatches: 0
python3 scripts/doctor.py            # mismatches: 0
bash scripts/harness/run-all.sh --dry-run   # both runners exit 0

All four must exit 0 with 0 mismatches. The release-candidate evidence is at .planning/phases/07-release/evidence/.

Next: v0.2 (planned)

  • Live harness execution as a CI step
  • GitHub-hosted marketplace publishing
  • Promote SKILLS_CATALOG.md generation to scripts/sync-catalog.py
  • Re-evaluate forge command based on user feedback
  • Re-evaluate coordinator, oracle, researcher, red-teamer as standalone agents based on usage signals

Credits

Architecture C is the user's central design insight. The 5 pillars + 7-phase ROADMAP are grounded in v5/v6/v7 prototypes shipped in this repo + Anthropic's ecosystem (anthropic-skills, claude-code-guide, sadd, oh-my-claudecode). The aggressive-merge pattern across all phases was authorized by the user with the clarifying directive: "You're allowed to merge what you're allowed to merge together."

License

MIT.