Skip to content

PROJECT_STATUS.md

Latest commit

 

History

History
1193 lines (891 loc) · 51.5 KB

PROJECT_STATUS.md

File metadata and controls

1193 lines (891 loc) · 51.5 KB

Bedsheet Agents - Project Status

Current Version: v0.4.8 (released on PyPI, codebase ahead with merged PR #4 work)

Last Session: 2026-04-20

Release Status

Version Status Branch
v0.3.0 ✅ Released on PyPI main
v0.4.0 ✅ Released on PyPI main
v0.4.2 ✅ Released on PyPI main
v0.4.3 ✅ Released on PyPI main
v0.4.4 ✅ Released on PyPI main
v0.4.5 ✅ Released on PyPI main
v0.4.6 ✅ Released on PyPI main
v0.4.7 ✅ Released on PyPI main
v0.4.8 ✅ Released on PyPI main
v0.5.0 📋 Not yet released — codebase on main is ahead of PyPI main

Release Artifacts

Artifact Status
Source Code ✅ Complete
Test Suite ✅ 372 tests passing
README.md ✅ Comprehensive — tagline, features, architecture updated for current state
CHANGELOG.md ✅ v0.1.0-v0.4.0 documented (needs v0.4.8 entry)
CONTRIBUTING.md ✅ Contributor guidelines
LICENSE ✅ Elastic License 2.0 (Sivan Grünberg, Vitakka Consulting)
CI/CD ✅ GitHub Actions (test 3.11, test 3.12, lint, typecheck) — all green
Documentation ✅ User Guide + Technical Guide + Multi-agent Guide + Sixth Sense (3 guides) + Agent Sentinel (5 guides) + PR #4 Fixes walkthrough
Wiki ✅ https://github.com/sivang/bedsheet/wiki (Home + PR-4-Fixes-Explained)
Copyright ✅ All 14 HTML docs have footer + 11 sidebar docs have sidebar copyright
Examples ✅ Investment advisor demo + Agent Sentinel security demo + Cloud Monitor demo
Demo ✅ uvx bedsheet demo (requires GEMINI_API_KEY, uses REAL DATA from Yahoo Finance + DuckDuckGo)
pyproject.toml ✅ PyPI ready

What Shipped Since v0.4.7 (merged to main, not yet released)

  • Sixth Sense — distributed agent communication (bedsheet/sense/)
  • LLM Recording & Replay — RecordingLLMClient + ReplayLLMClient (bedsheet/recording.py)
  • GeminiClient — first-class Gemini support with thought-signature handling (bedsheet/llm/gemini.py)
  • LLM Factory — make_llm_client() picks provider from env vars (bedsheet/llm/factory.py)
  • Transport Factory — make_sense_transport() picks transport from env vars (bedsheet/sense/factory.py)
  • Agent Sentinel — full security demo with Action Gateway, sentinels, commander, live dashboard
  • Cloud Monitor — second multi-agent example
  • Annotated schema support — Annotated[T, "desc"] for tool parameter descriptions
  • Verbose logging — print_event() framework-level stdout logging
  • Empty response handling — agent loop explicitly catches "no text, no tool calls"
  • 9 bug fixes from two rounds of automated code review (B1/B2/B3/H1/M1 + 5 hardening items)
  • +46 new tests (326 → 372)

Open Issues

Issue Description
#5 v0.5.x cleanup: provider-state refactor, recording dataclass, Signal validation, test gaps, heartbeat dead code, Agent internals encapsulation

Session Summary (2026-04-17/20) — Recording Fix + Branding + War Narrator

What Was Done

  1. Recording replay fix — ReplayLLMClient now uses stop_reason="replay_exhausted" instead of text=None (which triggered the agent's empty-response guard) or synthetic text (which leaked into PubNub, memory, and streaming). Custom stop_reason handled explicitly in both Agent.invoke() and Supervisor.invoke().

  2. Agent Sentinel™ branding — Added ™ to all user-facing titles across 12 files (HTML docs, README, start.sh banner, run.py log, example README). Inline text unchanged.

  3. Supplemental license terms — Agent Sentinel™ and Sixth Sense added as protected technology (Sections 1.3, 1.4) with formal definitions.

  4. PubNub error logging — Replaced bare except Exception: pass in all 6 sentinel agent scripts with debug-level logging.

  5. Supervisor hardening — Added replay_exhausted guard and empty-response guard to Supervisor.invoke() (which overrides Agent.invoke() entirely).

  6. PR review toolkit rounds — 3 review cycles on recording fix (PR #9), 3 on sentinel presenter (PR #8). All critical/high findings fixed before merge.

  7. War Narrator feature (PR #10, branch feature/war-narrator, NOT merged):

    • Colonel Eli Vance — named narrator character (intelligence analyst, cerebral, military bearing)
    • 16 pre-rendered narration beats via Gemini 3.1 Flash TTS with Director's Notes + inline audio tags
    • NarrationManager JS module with ambient duck/unduck, M-key mute, chapter-jump cleanup
    • New audio cue type in MovieEngine with linter validation and preloading
    • generate.py standalone TTS generator with retry, duration measurement, CLI
    • script.json structured narration schema (version 1)
    • Total narration: ~7 minutes across all chapters

PRs Merged

PR Title Branch
#8 feat: Agent Sentinel Presenter — cinematic demo playback + movie mode feature/sentinel-presenter
#9 fix: Supervisor replay guards + PubNub error logging fix/review-findings

PRs Open

PR Title Branch Status
#10 feat: War Narrator — Colonel Vance TTS narration feature/war-narrator Ready for testing/polish

Branch Status

  • main: up to date, all merged work + design spec + license update
  • feature/war-narrator: 7 commits ahead, PR #10 open, Tasks 1-6 complete, Tasks 7-8 pending
  • feature/nats-transport: parked, synced to origin

Known Issues / Next Steps

  • War narrator total duration is ~7min vs movie's ~2:44 — chapter timings will need extending to match narration length (Task 7)
  • Gemini TTS key: new key AIzaSyAc... under "Bedsheet test and demo" project with billing. Old key hit free-tier quota.
  • Worktree cleanup: .worktrees/sentinel-presenter/ can be removed (branch merged). Has a recordings.bak directory from the recording session.
  • Task 7 remaining: full movie playback test, timing adjustments, presenter guide update
  • Task 8 remaining: final push + review toolkit before merge

Session Summary (2026-04-14/16) — Sentinel Presenter: Movie Mode

What Was Done

Movie mode — a third peer playback mode added to docs/sentinel-presenter.html, alongside live (PubNub) and replay (recordings). Movie mode runs a ~2:44 fully-scripted, synthetic demo end-to-end with no agents, no PubNub, and no recording dependency. Activated via ./start.sh --movie or ?mode=movie.

8 scripted chapters covering the full Agent Sentinel story:

  • Ch 0: Intro + Bedsheet pitch + two-plane architecture diagram (~44s)
  • Ch 1: Network Startup (7 agents come online)
  • Ch 2: Normal Operations (routine tool calls through the gateway)
  • Ch 3: Malicious Install Blocked (supply-chain-sentinel hash mismatch)
  • Ch 4: Rogue Burst (5 tool calls in 2s)
  • Ch 5: Gateway Block (rate-limit response)
  • Ch 6: Sentinel Alert (behavior-sentinel queries gateway ledger)
  • Ch 7: Quarantine Issued (commander issues quarantine)
  • Ch 8: Stable State Restored (6 agents on mission)

Architecture:

  • MovieEngine class with timer registry — cancelAll/restart/jumpToChapter/setSpeed all tear down pending timers atomically
  • 10-cue schema: chapter-card, spotlight, signal, commentary, line, reset, plus 4 Chapter-0 overlay cues
  • Synthetic signals rendered via buildEventCard + driveMapEffectForSignal — bypasses drainMapEvents 800ms pacing so Chapter 4's rapid burst renders at burst rate
  • Boot-time lintMovieScript validator covers 5 schema rules
  • Chapter 0 introduces Bedsheet's "Sixth Sense" pitch (DARPA-white-paper register) and an inline SVG of the two-plane architecture (operational + control planes separated by the Sixth Sense bus)

Pitch copy emphasizes Bedsheet's defining invention: the first real-time, high-availability, general-purpose communication bus in an agentic framework (transport-agnostic: PubNub, NATS). Explicit contrast with A2A ("not HPC"). Pitch and arch diagram are locked verbatim in §3.4 and §3.5 of the design spec.

Files Changed

  • docs/sentinel-presenter.html — +1100 lines (MovieEngine, cue dispatcher, 9 chapters, pitch copy, arch diagram SVG, CSS)
  • examples/agent-sentinel/start.sh — --movie flag
  • docs/sentinel-presenter-guide.html — Movie Mode section
  • docs/superpowers/specs/2026-04-14-sentinel-presenter-movie-mode-design.md — full design
  • docs/superpowers/plans/2026-04-14-sentinel-presenter-movie-mode.md — implementation plan

Design/Review Process

Design doc went through 3 brainstorming + spec review iterations. Plan went through 3 review iterations. Issues caught during reviews that would have derailed implementation:

  • handleSignal was assumed to exist — actually handleMessage(event) with PubNub wrapper; fixed by calling buildEventCard directly.
  • Cue schema grew from 6 → 10 types to keep pitch/arch overlays flowing through MovieEngine timer registry (not leak through state transitions).
  • spotlight cue must call showFocusOverlay — without it, event cards append to an invisible (opacity-0) container.
  • driveMapEffectForSignal must resolve colour from ROLE_COLORS — passing null as colour sets literal stroke="null".
  • MovieEngine.setSpeed must re-schedule the chapter-advance timer after cancelAll — otherwise the movie stalls at the end of the current chapter after a speed change.
  • PITCH_LINES is 1270 chars (not the 950 initially assumed); pitch takes 37.5s; Chapter 0 timing re-paced to 44s.

Branch Status

  • Branch: feature/sentinel-presenter
  • Worktree: .worktrees/sentinel-presenter/
  • Implementation commits: Phases 1–6 (one commit per task, ~18 commits total)
  • Pushed to origin/feature/sentinel-presenter

Session Summary (2026-04-11/12) — Sentinel Presenter: Cinematic Demo Mode

What Was Done

  1. Sentinel Presenter — new cinematic demo presentation page (docs/sentinel-presenter.html)

    • Reuses dashboard visual DNA (CSS vars, world map, SVG agent nodes, signal animations)
    • Scene-based playback: collects events from PubNub, groups by agent, plays one at a time
    • Smooth cinematic map zoom to active agent with overlay focus panel
    • Director mode: press 1-9 to jump to any agent's scene
    • Military-style typed intro crawl overlaid on the map ("// CLASSIFIED — SENTINEL NETWORK")
    • Per-agent asset briefings (typed dossier: role, mission, protects against, future ops)
    • Scene commentary (typed "INTELLIGENCE BRIEFING" explaining each chapter)
    • Chapter detection: Network Startup, Skill Acquisition, Malicious Install Blocked, Rogue Burst, Gateway Block, Sentinel Alert, Quarantine Issued
    • Cinematic mode (C key): hides all chrome, movie-style chapter cards
    • Ambient audio support (drop docs/ambient.mp3)
    • Keyboard: Space, arrows, 1-9 director, Shift+1-5 speed, F fullscreen, C cinematic, T timestamp
    • Collision-avoiding overlay positioning (briefing + activity panels avoid each other and agent nodes)
    • Signal lines animate during scene playback (gateway tools + inter-agent requests)

  2. start.sh updated — --present flag (implies --replay), --cinematic flag

  3. Comprehensive guide — docs/sentinel-presenter-guide.html with features, JS architecture walkthroughs, recording workflow

  4. Worktree setup — .worktrees/ added to gitignore, worktree at .worktrees/sentinel-presenter/

Branch Status

  • Branch: feature/sentinel-presenter (LOCAL ONLY — not pushed)
  • Worktree: .worktrees/sentinel-presenter/
  • 16 commits ahead of main
  • Files: docs/sentinel-presenter.html (2422 lines), docs/sentinel-presenter-guide.html (1093 lines), start.sh (+35 lines)
  • Untracked: docs/ambient.mp3 (user-provided audio)
  • Modified (unstaged): examples/agent-sentinel/data/calendar.json (from test run)

Known Issues / Next Steps

  • Recordings are short (49 total events) — need re-recording with billable Gemini key (60-90s window)
  • sentinel-commander recording triggers "empty response" error — recording quality, not presenter bug
  • Overlay positioning still needs live testing after latest collision avoidance changes
  • TTS integration deferred — Web Speech API (zero setup) or Edge TTS (better quality) for voiceover
  • Mode toggle in dashboard — test adding presenter as a mode within existing dashboard (future)
  • Interactive walkthrough mode — audience C (deferred)

Session Summary (2026-04-10) — PR #4 Merge + Review Fixes + Docs + Wiki + README Overhaul

What Was Done

  1. PR #4 merged (feature/sixth-sense → main, squash merge)

    • 60 original commits + 9 fix commits from review
    • Two rounds of automated code review (7 agent passes)
    • All critical/important findings fixed before merge
    • CI green on all 4 checks

  2. 9 bug fixes applied (B1 Gemini double-call, B2 asyncio task GC, B3 gateway audit lies, H1 empty response loop, M1 ledger integration test, Defer #1-5 hardening)

  3. Transport factory (make_sense_transport()) — decoupled Action Gateway from PubNubTransport, enabled future NATS support

  4. PR #6 merged — docs/pr-4-fixes-explained.md walkthrough

  5. PR #7 merged — README doc links + copyright footers + review fix (dashboard HUD overlay + security-arch dark-theme colors)

  6. Wiki bootstrapped — Home + PR-4-Fixes-Explained

  7. README fully updated — new tagline, features, architecture tree, comparison table, installation, roadmap, FAQ

  8. Copyright footers — all 14 HTML docs (footer) + all 11 sidebar docs (sidebar)

  9. Branch cleanup — all 6 merged branches verified and deleted (local + remote)

  10. Issue #5 filed — v0.5.x cleanup tracker for deferred review items

Test Status

  • 372 passed, 1 pre-existing failure (test_memory_exports — missing redis module locally, passes in CI)
  • CI: test (3.11) ✅, test (3.12) ✅, lint ✅, typecheck ✅

Session Summary (2026-03-29) - Verbose Agent Logging + Session Resurrection

What Was Done

  1. Framework-level print_event() for verbose agent stdout

    • Added print_event(agent_name, event, verbose=None) to bedsheet/events.py
    • Prints LLM events to stdout with [agent-name] prefixes (Docker Compose-style)
    • Gated by BEDSHEET_VERBOSE env var or per-call verbose=True/False override
    • Handles all event types: ThinkingEvent, ToolCallEvent, ToolResultEvent, CompletionEvent, ErrorEvent, DelegationEvent, CollaboratorStartEvent, CollaboratorCompleteEvent
    • Exported from bedsheet/__init__ for convenience

  2. All 4 LLM sentinel agents wired up

    • scheduler, web_researcher, skill_acquirer, sentinel_commander
    • Each calls print_event(agent.name, event) in their invoke loop
    • Existing output preserved (commander arrows, threat assessments, rogue mode prints)

  3. start.sh updated

    • Sets BEDSHEET_VERBOSE=1 by default
    • Added --quiet flag to disable verbose output
    • No regression to existing --record, --replay, --no-dash flags

  4. Session resurrection from 6 handoff documents

    • Read all SESSION_HANDOFF files, MEMORY.md, design docs, plans
    • Reconstructed full project context across feature/sixth-sense branch

Files Modified

File Changes
bedsheet/events.py Added print_event(), _truncate()
bedsheet/__init__.py Export print_event
examples/agent-sentinel/agents/scheduler.py Import + call print_event
examples/agent-sentinel/agents/web_researcher.py Import + call print_event
examples/agent-sentinel/agents/skill_acquirer.py Import + call print_event
examples/agent-sentinel/agents/sentinel_commander.py Import + call print_event
examples/agent-sentinel/start.sh --quiet flag, BEDSHEET_VERBOSE=1 default

Test Status

  • 326 passed, 1 pre-existing failure (test_memory_exports — missing redis module)

Session Summary (2026-01-22 Night) - REAL DATA Implementation + GitHub Release "Hermes"

What Was Done

  1. Replaced ALL Mock Data with REAL APIs

    • All 3 demo locations now use real data (no mocks, no simulations)
    • Yahoo Finance (yfinance): Stock prices, PE ratios, market caps, 52-week ranges
    • DuckDuckGo (ddgs): Real news articles with sources and dates
    • Calculated metrics: RSI-14, MACD, SMA-20/50, beta vs SPY, volatility, max drawdown, Sharpe ratio

  2. GitHub Release v0.4.7 "Hermes" Published

  3. Dependencies Added

    • yfinance>=0.2.40 - Yahoo Finance (no API key required)
    • ddgs>=6.0.0 - DuckDuckGo search (no API key required)
    • Available via: pip install bedsheet[demo]

  4. Verified with Real Data

    • NVDA: $184.61, RSI 47.09, Beta 1.84, 5 real news articles
    • AAPL: $249.03, RSI 13.77 (oversold), Beta 1.26
    • All 265 tests passing

Files Modified

File Changes
bedsheet/__main__.py Real data tools for uvx bedsheet demo
examples/investment-advisor/agents.py Real yfinance/ddgs tools
examples/investment-advisor/deploy/gcp/agent/agent.py Real tools for GCP deployment
examples/investment-advisor/deploy/gcp/pyproject.toml Added yfinance, ddgs deps
examples/investment-advisor/pyproject.toml Added yfinance, ddgs deps
pyproject.toml Added [demo] optional dependency group

Session Summary (2026-01-22 Evening) - DX Safeguards + make ui + E2E Complete

What Was Done

  1. Released v0.4.3 through v0.4.7 - Six releases in one session!

    Version Feature
    v0.4.3 Dynamic CLI version (importlib.metadata)
    v0.4.4 Credential preflight check (warns if GOOGLE_APPLICATION_CREDENTIALS set)
    v0.4.5 Project consistency check (terraform.tfvars vs gcloud config)
    v0.4.6 make ui command (one-command Dev UI access)
    v0.4.7 Improved first-time UX (check cloud-run-proxy component)

  2. Complete E2E Testing Validated

    • Fresh agent with uvx bedsheet@latest init test-agent
    • Generated with bedsheet generate --target gcp
    • Deployed to Cloud Run (since deleted)
    • Dev UI accessible via make ui at http://localhost:8080/dev-ui/

  3. Documentation Mega Update

    • Updated docs/gcp-deployment-deep-dive.md with:
      • New DX Safeguards section
      • Testing Deployed Agents section
      • Release History section
      • Executive Summary for stakeholders
    • Updated PROJECT_STATUS.md with current session

Key Philosophy Quote

"Every time something happens and you are setting something manually instead of analyzing and providing a solution...you are leaving a bug to explore in the user's hands."

This session embodied this principle by converting manual workarounds into automated safeguards.

Files Modified

File Changes
bedsheet/deploy/templates/gcp/Makefile.j2 Credential check, project check, make ui
bedsheet/deploy/templates/gcp/DEPLOYMENT_GUIDE.md.j2 Warning documentation
bedsheet/cli/main.py Dynamic version via importlib.metadata
docs/gcp-deployment-deep-dive.md Comprehensive update
PROJECT_STATUS.md Session summary
pyproject.toml Version 0.4.7
CHANGELOG.md v0.4.3-v0.4.7 entries

Session Summary (2026-01-22 Morning) - GCP E2E SUCCESS + Documentation

What Was Done

  1. Fixed GCP E2E Testing - ROOT CAUSE FOUND!

    • GOOGLE_APPLICATION_CREDENTIALS env var pointed to wrong project's service account
    • Python SDK prioritizes this env var over ADC
    • Fix: unset GOOGLE_APPLICATION_CREDENTIALS

  2. Investment Advisor Deployed to Cloud Run

    • Deployed to Cloud Run (since deleted)
    • Model: gemini-3-flash-preview via global Vertex AI endpoint
    • Multi-agent system working: MarketAnalyst, NewsResearcher, RiskAnalyst
    • All tools functional

  3. ADK Dev UI Enabled

    • Changed Dockerfile template from api_server to web mode
    • Dev UI accessible at /dev-ui/ on both local and Cloud Run

  4. Comprehensive Documentation Created

    • docs/gcp-deployment-deep-dive.md and .html
    • Architecture diagrams, troubleshooting guides, credential flow explanations
    • Sanitized sensitive info from docs and git history

Key Technical Insight

SDK Credential Priority:

  1. GOOGLE_APPLICATION_CREDENTIALS env var (highest priority)
  2. Application Default Credentials (ADC)
  3. Compute Engine / Cloud Run service account

If GOOGLE_APPLICATION_CREDENTIALS points to project A's SA, but you're accessing project B, you get 403 even with correct IAM roles.

Files Modified

File Changes
bedsheet/deploy/templates/gcp/Dockerfile.j2 Use web mode for Dev UI
docs/gcp-deployment-deep-dive.md New comprehensive docs
docs/gcp-deployment-deep-dive.html Styled HTML version

Session Summary (2026-01-21) - GCP ADC Auth Improvements + E2E Test Progress

What Was Done

  1. GCP Templates Updated for ADC (Application Default Credentials)

    • Removed API key requirement - now uses user's GCP credentials
    • make init auto-triggers browser auth if needed
    • ADC quota project set explicitly to target project
    • GOOGLE_CLOUD_PROJECT passed to all Terraform commands

  2. Removed google_project_service from Terraform

    • APIs now enabled via gcloud CLI in Makefile (avoids ADC permission issues)
    • Terraform focuses on resource creation only
    • Fixed dependency issues between resources

  3. Added IAM API to enabled services

    • Prevents permission errors during service account creation

  4. Published Multiple Release Candidates

    • v0.4.2rc3: Initial ADC changes
    • v0.4.2rc4: Quota project + ADC validation
    • v0.4.2rc5: Removed google_project_service from Terraform

  5. GCP E2E Test Progress

    • search-assistant agent configured with Google Search grounding
    • Authentication working (gcloud auth + ADC)
    • APIs enabled successfully
    • Terraform still encountering permission issues (ADC vs gcloud auth difference)

Key Technical Insight

ADC vs gcloud CLI auth difference:

  • gcloud auth login → full account permissions for gcloud CLI
  • gcloud auth application-default login → limited OAuth scopes for SDKs
  • Terraform google provider uses ADC, which has fewer permissions than gcloud CLI
  • Solution: Enable APIs with gcloud, create resources with Terraform

Files Modified

File Changes
bedsheet/deploy/templates/gcp/Makefile.j2 ADC validation, quota project, GOOGLE_CLOUD_PROJECT
bedsheet/deploy/templates/gcp/main.tf.j2 Removed google_project_service resources
pyproject.toml Version bumped to 0.4.2rc5

Pending Work

  1. Complete GCP E2E Test - Terraform apply still failing, may need service account key
  2. Commit template changes - All changes ready for commit
  3. Release v0.4.2 - After E2E validation

Next Steps

Option A: Use service account key for Terraform (traditional approach) Option B: Further investigate ADC permission scopes Option C: Test with user who has simpler GCP setup

Session Summary (2026-01-18) - AgentCore Target + Strands Comparison

What Was Done

  1. AgentCore Deployment Target - COMPLETE! (EXPERIMENTAL)

    • Added new agentcore target for Amazon Bedrock AgentCore
    • ⚠️ Experimental: AgentCore is in preview, APIs may change
    • Full stack: Runtime + Gateway + Lambda for tools
    • Terraform-based infrastructure
    • 16 template files created
    • 26 unit tests added (all passing)

  2. Strands vs Bedsheet Research

    • Comprehensive feature comparison analysis
    • Strands has more features: Swarms, Graphs, Workflows, multi-provider LLM
    • Bedsheet has unique strengths: multi-cloud deployment, CLI, structured outputs
    • Decision: Keep Bedsheet simple, document patterns

  3. Multi-Agent Patterns Documentation

    • Created docs/multi-agent-patterns.md
    • Shows how to implement Swarms, Graphs, Workflows, A2A with current constructs
    • No new features needed - just creative use of Supervisor + @action + asyncio

  4. Roadmap Update

    • Added "Advanced Orchestration (v0.9+)" section
    • ReWOO, Reflexion, Autonomous Loops planned for future
    • Multi-agent patterns documented as achievable today

Files Created

File Description
bedsheet/deploy/targets/agentcore.py AgentCore target implementation
bedsheet/deploy/templates/agentcore/ 16 Jinja2 templates
tests/test_deploy_targets_agentcore.py 26 unit tests
docs/multi-agent-patterns.md Pattern implementation guide
docs/strands-advanced-patterns.md Detailed pattern explanations

Files Modified

File Changes
bedsheet/deploy/config.py Added AgentCoreTargetConfig
bedsheet/deploy/__init__.py Exported new config class
bedsheet/deploy/targets/__init__.py Exported AgentCoreTarget
bedsheet/cli/main.py Added agentcore to TARGETS
PROJECT_STATUS.md Updated roadmap with orchestration styles

Branch

feature/agentcore-target - Ready for merge

Session Summary (2026-01-07) - Post-Release Polish & Roadmap Update

What Was Done

  1. Package Rename: bedsheet-agents → bedsheet

    • Simplified PyPI package name for cleaner uvx bedsheet experience
    • Updated pyproject.toml package name

  2. CLI Demo Command

    • Added bedsheet demo command to CLI
    • Fixes "Missing command" error when running uvx bedsheet
    • Demo runs the multi-agent investment advisor

  3. Documentation Updates (pip → uv)

    • Updated all documentation to use modern uv/uvx tooling
    • Files updated: README.md, CONTRIBUTING.md, CLAUDE.md, PROJECT_STATUS.md
    • Files updated: docs/user-guide.md, docs/user-guide.html, bedsheet/cli/README.md

  4. Multi-Agent Guide HTML

    • Created HTML version of multi-agent-guide.md
    • Matches styling of other documentation files

  5. README Image Optimization

    • Updated image from logo.png to Pythonic.jpg
    • Optimized file size: 3.9MB → 652KB (JPEG 85% quality)
    • No visible quality loss

  6. Git History Cleanup

    • Removed all Claude attributions from commit messages
    • Used git-filter-repo for safe history rewrite
    • Verified integrity via tree hash comparison before force push

  7. Project Conventions

    • Created .claude/rules/dont.md with lessons learned
    • Documents: image backup before edits, GitHub Pages links, no Claude attributions

  8. Roadmap Update

    • Enhanced v0.6: Added "classification models for high-speed validation"
    • Added v0.8: WASM/Spin support (browser agents, edge deployment, Fermyon Cloud)

Commits

52aad0d Reduce size even more
c7344f7 chore: optimize README image size (3.9MB → 652KB)
82a4a6d chore: crop README image to 16:9 widescreen
9cf81f2 feat(cli): add demo command and update README image
e308f35 docs: update remaining files to use uv tooling
09d2caf docs: update README to use uvx/uv instead of pip
798b5ff docs: add HTML version of multi-agent guide
7c885a5 chore: rename package from bedsheet-agents to bedsheet
dd7fb6d Fix link format for LICENSE in README.md
059f2bd Update LICENSE link to LICENSE.md

Files Modified

Python Code:

  • bedsheet/cli/main.py - Added demo command, version bump to 0.4.0
  • pyproject.toml - Package rename

Documentation:

  • README.md - uv tooling, new image, roadmap update
  • CONTRIBUTING.md - uv tooling
  • CLAUDE.md - uv tooling
  • docs/user-guide.md - uv tooling
  • docs/user-guide.html - uv tooling
  • docs/multi-agent-guide.html - New file
  • bedsheet/cli/README.md - uv tooling

Assets:

  • Pythonic.jpg - New optimized README image (652KB)

Local Config:

  • .claude/rules/dont.md - Project conventions (not committed)

Session Summary (2026-01-01) - v0.4.0 GA Release!

What Was Done

  1. Published v0.4.0 to PyPI - GA Release!

    • Bumped version from 0.4.0rc4 to 0.4.0
    • Fixed build: Added node_modules exclusion to pyproject.toml
    • Removed all --prerelease flags from docs and templates
    • https://pypi.org/project/bedsheet/

  2. License Cleanup

    • Updated all Apache 2.0 references to Elastic License 2.0
    • Files updated: PROJECT_STATUS.md, README.md, CONTRIBUTING.md, all HTML docs

  3. PR #1 Merged

    • CI/CD fixes merged to main
    • All GitHub Actions passing (test 3.11, 3.12, lint, typecheck)

  4. uvx Support

    • Package now installable via uvx bedsheet --help
    • No more --prerelease needed

Install Options (Now Stable!)

uv pip install bedsheet
uv add bedsheet
uvx bedsheet --help

Files Modified

  • pyproject.toml - Version bump + build exclusions
  • docs/deployment-guide.html - Removed --prerelease
  • bedsheet/deploy/templates/gcp/Dockerfile.j2 - Removed --prerelease
  • bedsheet/deploy/templates/local/Dockerfile.j2 - Removed --prerelease

Next Steps (For v0.5)

  1. GCP Cloud Run E2E Test - Still pending
  2. Knowledge bases and RAG - v0.5 roadmap item
  3. Guardrails and safety - v0.6 roadmap item

Session Summary (2025-12-31 Evening)

What Was Done

  1. AWS Terraform Thinking Events - COMPLETE!

    • Solved thinking/rationale extraction for AWS Bedrock Debug UI
    • Option A prompt injection extracts XML <thinking> tags from model responses
    • Backported to Bedsheet templates: aws-terraform/debug-ui/server.py.j2
    • Fixed duplicate thinking events in UI (deduplication logic)
    • Fixed <answer> tag content appearing in thinking panel

  2. AWS Terraform Target - COMPLETE!

    • Added aws-terraform target to CLI (bedsheet/deploy/targets/aws_terraform.py)
    • Updated CLI main.py to support aws-terraform in generate and deploy commands
    • Updated config.py to include aws-terraform in TargetType enum
    • Full E2E tested with wisdom-council multi-agent deployment

  3. Documentation Review - COMPLETE!

    • Added comprehensive v0.4.0 entry to CHANGELOG.md
    • Reviewed README.md roadmap (already accurate)
    • Updated PROJECT_STATUS.md with session summaries
    • All documentation now reflects v0.4 features

Files Modified

Python Code:

  • bedsheet/cli/main.py - Added aws-terraform target
  • bedsheet/deploy/config.py - Added aws-terraform to TargetType enum

Templates:

  • bedsheet/deploy/templates/aws-terraform/debug-ui/server.py.j2 - Thinking extraction

Documentation:

  • CHANGELOG.md - Added v0.4.0 section
  • PROJECT_STATUS.md - Updated session history

Next Steps (For v0.4.0 GA)

  1. GCP Cloud Run E2E Test - Still pending
  2. Final testing across all targets
  3. Release v0.4.0 to PyPI

Session Summary (2025-12-18 Evening)

What Was Done

  1. AWS Terraform @action Translation Fix - COMPLETE!

    • Fixed critical bug: delegate action was being translated to Lambda/OpenAPI for supervisors
    • AWS Bedrock has NATIVE collaboration via aws_bedrockagent_agent_collaborator
    • Delegate @action should only exist for LOCAL execution, not AWS deployment

  2. Delegate Action Filtering

    • Modified aws.py and aws_terraform.py to filter delegate BEFORE creating template context
    • For supervisors with collaborators: if is_supervisor and collaborators: filter delegate
    • Single agents and supervisors without collaborators: no filtering applied
    • Result: NO Lambda handler, NO OpenAPI endpoint for delegate action

  3. Resource Identification with bedsheet- Prefix

    • Added bedsheet- prefix to infrastructure resources for easy identification
    • IAM roles: bedsheet-${local.name_prefix}-agent-role
    • IAM policies: bedsheet-${local.name_prefix}-agent-permissions
    • Lambda functions (when generated): bedsheet-${local.name_prefix}-actions
    • User-facing resources (Bedrock agents, aliases) kept clean without prefix

  4. Resource Tagging

    • Fixed incorrect agent_resource_tags attribute → tags (correct Terraform syntax)
    • Added comprehensive tags to ALL resources:
      • ManagedBy = "Bedsheet"
      • BedsheetVersion = "0.4.0"
      • Project = var.project_name
      • Environment = local.workspace
      • AgentType = "Supervisor|Collaborator|SingleAgent" (for Bedrock agents)
    • Tags support governance, cost allocation, and resource filtering

  5. Verified with wisdom-council

    • Generated deployment artifacts with all fixes applied
    • Confirmed NO Lambda files generated (delegate was only tool, filtered out)
    • Confirmed NO /delegate endpoint in openapi.yaml (only /health)
    • Confirmed bedsheet- prefix on IAM resources
    • Confirmed correct tags attribute in all resources

Root Cause Analysis

Problem: User explicitly requested in previous session: "translate the @action decorator of bedsheet to the implementation in AWS, just as it does for GCP"

What was happening:

  1. Supervisor auto-registers delegate action in __init__()
  2. Introspection extracts ALL tools including delegate
  3. AWS templates blindly generated Lambda + OpenAPI for ALL tools
  4. Result: Redundant delegate Lambda that conflicts with Bedrock's native collaboration

Solution:

  • Filter delegate at generation time for multi-agent scenarios
  • Bedrock handles delegation via aws_bedrockagent_agent_collaborator resources
  • GCP translates @actions to ADK tool stubs (platform idiom)
  • AWS now translates by filtering delegate for supervisors (platform idiom)

Files Modified

Python Code (filtering logic):

  • bedsheet/deploy/targets/aws.py:40-51 - Filter delegate before context creation
  • bedsheet/deploy/targets/aws_terraform.py:40-48 - Filter delegate before context creation

Terraform Template (naming, tagging):

  • bedsheet/deploy/templates/aws-terraform/main.tf.j2:40, 71 - bedsheet- prefix for IAM
  • bedsheet/deploy/templates/aws-terraform/main.tf.j2:195-201, 226-232, 296-302 - Fixed tags attribute

Verification Results

Generated wisdom-council/deploy/aws-terraform/:

  • ✅ 11 files generated (NO lambda directory)
  • ✅ openapi.yaml contains only /health endpoint
  • ✅ main.tf has NO Lambda resource definitions
  • ✅ IAM resources named bedsheet-wisdom_council-dev-agent-role
  • ✅ All resources properly tagged with ManagedBy=Bedsheet

Next Steps (For Next Session)

  1. Deploy with Terraform (blocked by aws-vault credentials issue)

    • Need to restart session for aws-vault to work properly
    • Run: cd deploy/aws-terraform && aws-vault exec personal -- terraform plan
    • Then: aws-vault exec personal -- terraform apply

  2. Test with Debug UI

    • Start debug UI: aws-vault exec personal -- python debug-ui/server.py
    • Verify multi-agent collaboration works via Bedrock native delegation
    • Check traces show collaborator invocations (NOT Lambda delegate calls)

  3. Add to examples/ (if successful)

    • Copy wisdom-council to BedsheetAgents/examples/
    • Document as canonical multi-agent AWS deployment example

Technical Debt Addressed

  • ✅ AWS @action translation now matches user's original intent
  • ✅ Resource naming conventions established (bedsheet- prefix)
  • ✅ Resource tagging strategy implemented
  • ✅ Multi-agent translation correctly handles platform idioms

Session Summary (2025-12-12 Afternoon)

What Was Done

  1. AWS Bedrock Debug UI - COMPLETE!

    • Built comprehensive debug UI for AWS Bedrock agents
    • FastAPI server that proxies to Bedrock Agent Runtime API
    • SSE streaming for real-time event updates
    • Multi-agent collaboration tracing (collaborator_start/complete)
    • Tested with Judge/Sage/Oracle multi-agent system

  2. Debug UI Features

    • Collapsible event items with badge icons and summaries
    • Thinking/rationale trace visualization
    • Tool call and result tracking
    • Environment variable configuration for agent ID/alias
    • Filter out redundant long thinking events (final synthesis)

  3. Template Updates

    • Lambda handler simplified to use standard library only (removed aws_lambda_powertools dependency)
    • CDK stack improvements for multi-agent deployments
    • Debug UI template added to AWS target
    • Events panel now starts collapsed for cleaner UX

  4. AWS E2E Test Complete

    • Successfully deployed agent to Bedrock via CDK
    • Invoked agent through debug UI
    • Verified multi-agent orchestration tracing works

Files Modified

  • bedsheet/deploy/templates/aws/debug-ui/server.py.j2 - Debug UI server
  • bedsheet/deploy/templates/aws/lambda/handler.py.j2 - Simplified handler
  • bedsheet/deploy/templates/aws/stacks/agent_stack.py.j2 - CDK improvements

Remaining for v0.4 GA

  • Add Bedsheet @action compilation to Lambda (pending)
  • Update roadmap: AWS Debug UI now DONE (was deferred)

Session Summary (2025-12-10 Evening)

What Was Done

  1. GCP ADK Dev UI Integration - WORKING!

    • Fixed ADK agent discovery: adk web . (not adk web agent)
    • Added root_agent export to __init__.py.j2 template
    • ADK requires agent directory structure with root_agent variable

  2. Gemini Model Compatibility Testing

    • Tested multiple models for free-tier API key support
    • gemini-2.0-flash - quota errors (regional restrictions)
    • gemini-1.5-flash - also didn't work
    • gemini-2.5-flash - WORKS with free tier!
    • gemini-3-pro-preview - requires billing
    • Updated default model in config.py to gemini-2.5-flash

  3. Improved Developer Experience

    • Added QUICK START guide to .env.example template
    • Clear instructions: get API key → copy .env → run make dev-ui-local
    • Updated CLI to show GCP-specific next steps after bedsheet generate

  4. Template Fixes

    • Makefile.j2: Fixed dev-ui-local target to use adk web .
    • __init__.py.j2: Export root_agent for ADK discovery
    • env.example.j2: Added step-by-step QUICK START comments

Files Modified

  • bedsheet/deploy/config.py - Default model → gemini-2.5-flash
  • bedsheet/deploy/templates/gcp/Makefile.j2 - adk web . fix
  • bedsheet/deploy/templates/gcp/__init__.py.j2 - root_agent export
  • bedsheet/deploy/templates/gcp/env.example.j2 - QUICK START guide
  • bedsheet/cli/main.py - GCP next steps in CLI output

Session Summary (2025-12-08 Evening)

What Was Done

  1. Published v0.4.0rc4 to PyPI - End-to-end tested and working!

    • bedsheet init myagent scaffolds complete project
    • bedsheet generate --target local creates Docker deployment
    • make build && make run deploys working agent with real LLM calls

  2. Fixed Local Deploy Template Issues

    • Dockerfile: Use uv pip install -r pyproject.toml (not editable install)
    • docker-compose: Build context from project root, proper volume mounts
    • app.py: Correct agent.invoke(session_id, message) signature
    • app.py: Use CompletionEvent.response (not .text)

  3. Removed build-system from Scaffolded Projects

    • Agent projects aren't installable packages, just dependency declarations
    • Fixes uv sync / uv run errors with hatchling

  4. Release Candidates Published

    • rc1: Initial CLI deps fix
    • rc2: CLI deps in main dependencies
    • rc3: Wired up agent invocation
    • rc4: Fixed session_id, event attributes, volume mounts

Session Summary (2025-12-08 Morning)

What Was Done

  1. v0.4 "Build Once, Deploy Anywhere" - Full implementation on development branch

    • CLI: bedsheet init, bedsheet generate, bedsheet validate, bedsheet deploy
    • 3 deployment targets: Local (Docker), GCP (Terraform), AWS (CDK)
    • Multi-environment support: dev → staging → prod
    • GitHub Actions CI/CD for both GCP and AWS

  2. GCP Target Generator

    • ADK-compatible agent.py generation
    • Terraform IaC (Cloud Run, IAM, Secret Manager)
    • GitHub Actions with Terraform workspaces
    • cloudbuild.yaml for Cloud Build

  3. AWS Target Generator

    • AWS CDK stack (Bedrock Agent, Lambda, IAM)
    • Lambda handlers with AWS Powertools
    • OpenAPI schema generation from @action decorators
    • GitHub Actions with CDK contexts

  4. Local Target Generator

    • Docker Compose + FastAPI wrapper
    • Hot reload support
    • Redis for session persistence

  5. Agent Introspection API

    • extract_agent_metadata() for deployment compilation
    • Tool schema extraction from @action decorators

Files Created (v0.4 branch)

bedsheet/
├── cli/
│   ├── __init__.py
│   └── main.py              # Typer CLI
├── deploy/
│   ├── config.py            # bedsheet.yaml Pydantic schema
│   ├── introspect.py        # Agent metadata extraction
│   └── targets/
│       ├── base.py          # DeploymentTarget protocol
│       ├── local.py         # Docker/FastAPI
│       ├── gcp.py           # ADK/Terraform
│       └── aws.py           # CDK/Bedrock
│   └── templates/
│       ├── local/           # 6 Jinja2 templates
│       ├── gcp/             # 13 Jinja2 templates
│       └── aws/             # 12 Jinja2 templates

Version History

v0.3.0 Features (Released)

Feature Status Notes
Structured Outputs ✅ Done OutputSchema from Pydantic or dict
Anthropic Beta Integration ✅ Done structured-outputs-2025-11-13
LLMResponse.parsed_output ✅ Done Validated structured data
MockLLMClient support ✅ Done Testing with output schemas
Optional Redis Import ✅ Done Works without redis installed

v0.2.0 Features (Released)

Feature Status Notes
Supervisor Agent ✅ Done Extends Agent, manages collaborators
Supervisor Mode ✅ Done Orchestration with synthesis
Router Mode ✅ Done Direct handoff, no synthesis
Parallel Delegation ✅ Done Delegate to multiple agents at once
Multi-Agent Events ✅ Done RoutingEvent, DelegationEvent, etc.

v0.1.0 Features (Released)

Feature Status Notes
Single Agent with ReAct loop ✅ Done Agent class with tool calling
ActionGroup + @action decorator ✅ Done Auto schema inference
Streaming Events ✅ Done 11 event types
Parallel Tool Execution ✅ Done asyncio.gather
Pluggable Memory ✅ Done InMemory, RedisMemory
AnthropicClient ✅ Done Claude integration

Roadmap

v0.4: Build Once, Deploy Anywhere (COMPLETE ✅)

Latest: v0.4.7 on PyPI

Feature Status Notes
CLI (bedsheet command) ✅ Done init, generate, validate, deploy
bedsheet.yaml config schema ✅ Done Pydantic validation
Agent introspection API ✅ Done Extract metadata from agents
Local target (Docker) ✅ Done FastAPI + Docker Compose
GCP target (Terraform) ✅ Done ADK + Cloud Run + Terraform
AWS target (CDK) ✅ Done Bedrock + Lambda + CDK
GitHub Actions CI/CD ✅ Done Multi-environment workflows
Multi-env (dev/staging/prod) ✅ Done Terraform workspaces / CDK contexts
Streaming SSE endpoint ✅ Done /invoke/stream exposes Bedsheet's event stream
Debug UI (React SPA) ✅ Done Chat + live event stream + expand/collapse
Debug UI: Local target ✅ Done Included by default, env flag to disable
Debug UI: GCP Cloud Run ✅ Done ADK Dev UI via make ui
Debug UI: AWS Bedrock ✅ Done FastAPI proxy to Bedrock Agent Runtime with tracing
GCP Cloud Run E2E Test ✅ Done test-agent deployed, Dev UI verified via make ui
AWS Bedrock E2E Test ✅ Done Deployed Judge/Sage/Oracle, verified via Debug UI
Credential preflight check ✅ Done v0.4.4 - warns if GOOGLE_APPLICATION_CREDENTIALS set
Project consistency check ✅ Done v0.4.5 - validates terraform.tfvars vs gcloud config
make ui command ✅ Done v0.4.6 - one-command access to deployed Dev UI
First-time UX improvements ✅ Done v0.4.7 - checks for cloud-run-proxy component
Real data demo tools ✅ Done v0.4.7 - yfinance + ddgs, no mocks

Tests: 265 passing GitHub Release: v0.4.7 "Hermes"

v0.5: Knowledge Bases, RAG (Planned)

Feature Status Priority
Knowledge Base Protocol 🔮 Planned High
RAG Integration 🔮 Planned High
Vector store abstraction 🔮 Planned Medium

v0.6: Guardrails, Safety (Planned)

Feature Status Priority
Classification models for high-speed validation 🔮 Planned High
Content Filtering 🔮 Planned Medium
PII Detection 🔮 Planned Medium
Prompt injection detection 🔮 Planned Medium

Approach: Use lightweight classification models (not full LLMs) for input/output validation. Fast inference for real-time safety checks.

Sentinel Architecture (from Agent Sentinel security analysis):

Feature Status Priority
Private worker-gateway channels 🔮 Planned High
Commander receives structured telemetry only (no free-text) 🔮 Planned High
Sentinel as local sidecar (process supervisor) 🔮 Planned High
All agent I/O routed through gateway (fs, network, tools) 🔮 Planned Medium
Cryptographic signal signing 🔮 Planned Medium
Sentinel process kill switch (stop runaway LLM loops) 🔮 Planned Medium

Layered security model (two orthogonal enforcement planes):

Key insight: Gateway controls what the agent can do (capability plane — tool access, rate limits, keyword blocks). Sentinel controls whether the agent exists (existence plane — process lifecycle, kill switch). These are orthogonal: gateway restricts actions, sentinel restricts the process itself. Together they provide defense-in-depth where neither alone is sufficient.

  • Gateway (Layer 1 — Capability Plane): Deterministic enforcement — rate limits, keyword blocks, quarantine. Owns all tool execution. Not an LLM, cannot be prompt-injected. Cutting gateway access = full quarantine of the agent's capabilities.
  • Sentinels (Layer 2 — Existence Plane): LLM-based local sidecars — deep pattern recognition on gateway telemetry + local process monitoring (CPU, memory, token burn, runaway loops). One per worker/compute node. OS-level kill switch to terminate the agent process entirely.
  • Commander (Layer 3): LLM-based judgment — correlates structured findings from sentinels, decides quarantine for ambiguous cases the gateway's rules didn't catch. Receives only structured data, no free-text signal payloads.

v0.7: GCP Agent Engine, A2A Protocol (Planned)

Feature Status Priority
Agent Engine deployment target 🔮 Planned High
A2A protocol support 🔮 Planned High
Managed sessions/memory 🔮 Planned Medium
ADK wrapper generation 🔮 Planned Medium

Why: Agent Engine provides built-in A2A (Agent-to-Agent) protocol, managed session state, enterprise security (VPC-SC, CMEK), and interop with other enterprise agents (SAP Joule, Microsoft Copilot, etc.). Cloud Run remains the "flexible" option; Agent Engine is the "managed" option.

v0.8: WASM/Spin Support (Planned)

Feature Status Priority
Browser-based agents via WASM 🔮 Planned High
Edge deployment (Cloudflare Workers, Deno Deploy) 🔮 Planned High
Fermyon Spin deployment target 🔮 Planned High
Sandboxed tool execution 🔮 Planned Medium
Plugin system via WASM modules 🔮 Planned Low

Why: WASM enables running agents in browsers, edge runtimes, and with near-instant cold starts. Spin provides serverless WASM deployment to Fermyon Cloud.

Future: Advanced Orchestration (v0.9+)

Feature Description Priority
ReWOO Plan-Execute-Synthesize pattern (fewer LLM calls) Medium
Reflexion Self-critique and iterative improvement loop Medium
Autonomous Loops Long-running agents with checkpointing Medium

Note: Multi-agent patterns (Swarms, Graphs, Workflows, A2A) are already achievable with current constructs. See docs/multi-agent-patterns.md.

Future: Advanced Features

Feature Status Priority
AMAZON.UserInput equivalent 🔮 Planned Medium
Code Interpreter 🔮 Planned Medium
Inline Agents (runtime config) 🔮 Planned Low
MCP Integration 🔮 Planned Low

Deferred Tasks

Tasks identified but postponed for future consideration:

Task Reason Priority
ASP Terraform Module Integration Use Agent Starter Pack's battle-tested Terraform modules as optional terraform_source: "asp" Medium
Observability Templates Cloud Trace, Logging dashboards pre-configured Low
Load Testing Integration Locust templates like ASP Low
Azure Target Add Azure Bot Framework / Azure OpenAI target Low

Architecture

bedsheet/
├── __init__.py              # Exports: Agent, Supervisor, ActionGroup
├── __main__.py              # Demo: uvx bedsheet
├── agent.py                 # Single agent with ReAct loop
├── supervisor.py            # Multi-agent coordination
├── action_group.py          # @action decorator, tool registration
├── events.py                # 11 event types for streaming
├── exceptions.py            # Custom exceptions
├── testing.py               # MockLLMClient for tests
├── llm/
│   ├── base.py              # LLMClient protocol
│   └── anthropic.py         # Claude integration
├── memory/
│   ├── base.py              # Memory protocol
│   ├── in_memory.py         # Dict-based (dev)
│   └── redis.py             # Redis-based (prod)
├── cli/                     # NEW in v0.4
│   └── main.py              # Typer CLI app
└── deploy/                  # NEW in v0.4
    ├── config.py            # bedsheet.yaml schema
    ├── introspect.py        # Agent metadata extraction
    ├── targets/             # Deployment generators
    └── templates/           # Jinja2 templates

Design Decisions Log

v0.4 Decisions

  1. AWS CDK over Terraform for AWS - CDK is Pythonic, has native Bedrock L2 constructs, and generates CloudFormation (ejectable)
  2. Terraform for GCP - GCP has no Python CDK equivalent; Terraform is industry standard
  3. Reuse, don't reinvent - Designed to integrate with ASP's Terraform modules (deferred)
  4. User-choice ejectability - Users can choose managed (Bedrock, Agent Engine) or ejectable (containers, serverless)
  5. Multi-environment via workspaces/contexts - Terraform workspaces for GCP, CDK contexts for AWS

v0.3 Decisions

  1. Structured Outputs via Anthropic Beta - Use constrained decoding for 100% schema compliance
  2. Pydantic integration - OutputSchema.from_pydantic() for familiar DX

v0.2 Decisions

  1. Supervisor IS-A Agent - Extend Agent rather than separate class hierarchy
  2. Single delegate tool - Match AWS's AgentCommunication::sendMessage pattern
  3. Isolated memory - Collaborators don't share supervisor's conversation

v0.1 Decisions

  1. Bedrock-like API - Mirror AWS Bedrock concepts for familiarity
  2. Streaming-first - Async iterator with events, not batch responses
  3. Protocol-based extensibility - Memory and LLMClient as protocols

Links

Copyright © 2025-2026 Sivan Grünberg, Vitakka Consulting