Skip to content

CHANGELOG.md

Latest commit

 

History

History
175 lines (154 loc) · 11.7 KB

CHANGELOG.md

File metadata and controls

175 lines (154 loc) · 11.7 KB

Changelog

All notable changes to agentlift are documented here. The format follows Keep a Changelog, and versions match the published PyPI releases and git tags (semantic versioning).

[0.8.0] — 2026-06-08

agentlift import — read a live agent back into the folder (the inverse of deploy). A new reverse pipeline reconstructs a neutral .managed-agents/ folder from a live managed runtime, so a runtime now round-trips both ways — and migration between runtimes falls out for free. Anthropic import is full; AWS Bedrock import covers the config-only Harness.

Migrate by pairing import with deploy — the folder is the neutral pivot:

# Anthropic → AWS Bedrock
agentlift import anthropic ./agent
agentlift deploy ./agent --target bedrock --mode harness
# …and back
agentlift import bedrock ./agent --harness-name my-agent --bedrock-region us-west-2
agentlift deploy ./agent

Added

  • agentlift import <anthropic|bedrock> <out> [--agent N …] [--mode harness] [--harness-id|--harness-name] [--bedrock-region R] [--dry-run] — a read-only command (never creates/updates/archives). It fetches the live agent, maps it to a neutral project, writes the folder, and self-verifies by re-running the real parse + plan (prints Round-trip OK only if the result is deployable again). --dry-run is the import analogue of plan (prints the mapping + diagnostics, writes nothing).
  • Anthropic import (full). agents.list/agents.retrieve + skills.versions.download recover system/description/model, built-in tools with :ask/:allow permission policies, URL MCP servers + per-server tool filters, custom skill content, and a coordinator's subagents (roster ids resolved to names; selecting a coordinator pulls its subagents into the closure automatically). Skills/MCP used identically by >1 agent are hoisted to shared/ (skills keyed by content hash, MCP by full identity) — the inverse of the planner's dedup.
  • Bedrock Harness import (import bedrock --mode harness). get_harness + S3 skill bundles, with a reverse model-map (regional inference profile → folder Claude id, e.g. eu.anthropic.claude-haiku-4-5-20251001-v1:0 → claude-haiku-4-5), agentCoreBrowser → web tools, agentCoreCodeInterpreter → sandbox builtins, and remote_mcp → URL MCP. Single-agent by nature, so an import never produces subagents.
  • New modules, mirroring the deploy pipeline in reverse (pure core + thin network edge): import_model.py, importer.py (inverse of planner.py), folder_writer.py (inverse of parser.py) — all pure and offline-tested — plus the network edges anthropic_source.py and harness_source.py.
  • End-to-end tests. tests/test_importer.py (the mapping contract), tests/test_import_roundtrip.py (provider responses → folder → real parser → real planner, including subagent delegation with shared and custom skills and MCP servers), tests/test_import_source.py (fetch wiring via fake clients, no network), tests/test_cli_import.py, and a gated read-only live test tests/live/test_import_anthropic.py (AGENTLIFT_LIVE_IMPORT=1).
  • Docs. New docs/import.md; import notes added across the convention, how-it-works, provider-matrix, anthropic-mapping, deploy-bedrock, deploying, and limitations docs.

Notes / honest boundaries

  • The Bedrock Runtime is not importable. A Runtime bakes its agent definition into an opaque ARM64 container image (GetAgentRuntime returns only a containerUri), so it can't be read back; import bedrock --mode runtime refuses with that reason. This is the import analogue of the deploy-time /invocations trace boundary. Google/OpenAI import is not implemented yet.
  • Four one-way losses, each surfaced as a Diagnostic (never silent): knowledge inlining is one-way (it stays in the prompt body on import); custom tools / harness inline functions are dropped; MCP auth values are provider-side, so only the header/env-var name is recovered; Anthropic first-party (type: anthropic) skills are referenced by id with no downloadable content.
  • One wire-format item pending a live download: the Anthropic skill-bundle archive layout (a zip whose members carry the <name>/… prefix) is the documented shape but is not yet confirmed against a live download — the unpack is defensive and emits import.skill_archive_shape if the payload is anything else (the repo's confirm-live-before-trusting rule).

[0.7.0] — 2026-06-05

AWS Bedrock AgentCore Runtime is now a live multi-agent hosted deploy (Stage 2). Both AgentCore primitives are live-verified: the single-agent Harness (0.6.0) and now the custom-container Runtime. --mode auto routes a single agent → Harness and a multi-agent team → Runtime; mapping stays Claude-native (no remap), with wire-shape receipts on Nova because Claude inference on Bedrock is a one-time per-account entitlement (Gate A), not a code gap.

Added

  • --mode runtime — live hosted multi-agent deploy. agentlift deploy --target bedrock --mode runtime builds the ARM64 Strands/AgentCore container, creates the ECR repo + logs in + docker buildx --platform linux/arm64 --push, calls CreateAgentRuntime (PUBLIC network, HTTP serverProtocol, IAM-only — no JWT authorizer), polls READY, writes .agentlift-bedrock.json, and InvokeAgentRuntime. Gated by _RUNTIME_LIVE_VERIFIED (now True) — a bare hosted create refused until a committed receipt.
  • Subagent delegation live-proven on a Nova receipt (tests/live/receipts/20260605-134012-runtime-bedrock): a coordinator + 2 specialists where create + agent + delegation are all PASS-EXERCISED (the coordinator's top-level trace named both specialists). A single-agent smoke (20260605-133821-runtime-bedrock) separately got remote MCP PASS-EXERCISED (an objective root-level docs_read_wiki_structure DeepWiki call).
  • Top-level tool-call trace in the generated handler — returns {result, tool_calls?} where tool_calls is read from AgentResult.metrics.tool_metrics (fail-open: trace extraction never breaks the invocation), so the deploy receipt can prove delegation objectively.
  • The boto3 (bedrock) optional dependency now also covers the Runtime hosted deploy.

Changed

  • The bare agentlift deploy --target bedrock --mode runtime now deploys live instead of refusing; --build-only still emits just the ARM64 container artifact (its NOTES.txt now points at the live hosted-create path).
  • .agentlift-bedrock.json (the Runtime lock) is now live-writing (spec hash → create/update/skip).

Notes / honest boundaries

  • The /invocations boundary. InvokeAgentRuntime returns the container's app-defined JSON body, not a tool-event stream. So subagent delegation and root-level skill/MCP calls are objective (PASS-EXERCISED), while a specialist's nested skill/MCP calls don't cross the boundary → PASS-WIRED + text-corroborated (the runtime analogue of the Google AgentTool→stream_query grounding-metadata caveat).
  • MCP per-tool-filter limitation narrowed. The unenforced-allowedTools limitation applies only to the direct remote_mcp attachment path; for AgentCore Gateway-fronted MCP, tool scoping is enforced server-side at the Gateway/Policy layer (AWS-documented; agentlift has not yet live-verified that path).
  • Runtime execution role needs: bedrock-agentcore.amazonaws.com trust (aws:SourceAccount condition — not a region-locked SourceArn), ECR pull, bedrock:InvokeModel, CloudWatch Logs.

0.6.0 — 2026-06-05

AWS Bedrock AgentCore joins the deploy targets, with two primitives behind --mode (auto picks the least-powerful one that preserves the folder's semantics — never a silent downgrade). Both are Claude-native — no Gemini-style model remap; a folder's claude-* maps to its regional Bedrock inference profile directly.

Added

  • --mode harness — live single-agent deploy to a managed AgentCore Harness (config-only, IAM-only, no container). One agent with its skills, remote MCP, sandbox, and browser, 6/6 live-verified end-to-end by a committed Nova receipt (tests/live/receipts/20260605-121525-harness-bedrock): agent + base-session sandbox (shell) + remote MCP (docs_read_wiki_structure, surfaced as <server>_<tool>) + an S3-loaded skill (skills[].s3.uri) + agentcore_browser.

    • Skills upload to $AGENTLIFT_BEDROCK_S3_BUCKET and attach via skills[].s3.uri; the execution role needs s3:ListBucket + s3:GetObject.
    • Idempotent via .agentlift-harness.json (spec hash → create/update/skip); a since-deleted clientToken triggers a retry without it.

  • --mode runtime --build-only — compiles a Strands package + a complete deployable ARM64 AgentCore Runtime container (image + Dockerfile + NOTES.txt runbook) for the multi-agent path.

  • audit and export bedrock-strands cover the new target; agentlift plan --target bedrock [--mode auto|harness|runtime].

  • CHANGELOG.md; a CI doc-link guard (tests/test_doc_links.py) that checks every relative path + intra-repo anchor in the README and docs.

  • Google built-in web tools (--target google): web_search → Google Search grounding and web_fetch → URL Context, each lowered as a wrapped single-tool ADK sub-agent; exercised on a separate live Google deploy.

  • Claude-on-Vertex — an offline-verified spike (experiments/claude-on-vertex/): ADK resolves Claude on Vertex and the mixed-model shape composes. Not shipped — a Claude --google-model is refused (no live receipt yet); folder Claude models still map to Gemini on Google.

Changed

  • README restructured (≈494 → ≈214 lines): big-picture-first, with a one-glance provider snapshot (AWS Bedrock now first-class), the deep capability matrix moved to docs/provider-matrix.md, and consistent collapsibles. Updated hero graphic (four backends).
  • Provider docs reconciled to the live-verified state (docs/provider-matrix.md, deploy-bedrock.md, tested-platforms.md, limitations.md).

Notes / known limits

  • A multi-agent team (subagents) routes to the Runtime; the harness is single-agent. The Runtime's hosted create (create_agent_runtime) is build-only here (not yet live-verified) — a hosted multi-agent deploy is the next milestone.
  • The AgentCore Harness feature is in AWS public preview. Claude inference runs in the harness but is gated by the per-account Anthropic use-case entitlement (Gate A), which is eventually-consistent — so the model-agnostic wire-shape receipt was captured on Amazon Nova.
  • Per-tool MCP allowedTools narrowing is not enforced on the harness preview (a restrictive allowlist suppresses MCP tool surfacing), so agentlift emits none and surfaces a diagnostic.
  • :ask is not enforced on either Bedrock primitive (no interactive approval channel).

0.5.0 — 2026-06-04

Added

  • Google Vertex AI Agent Engine deploy (--target google) gains full Skills (embedded in the source package, loaded via ADK load_skill_from_dir) and URL MCP (McpToolset + tool_filter, inline auth resolved into Agent Engine env_vars).
  • 6/6 live coverage matrix — one neutral fixture deployed and queried on both Anthropic and Google, all six portability dimensions (agents · subagents · shared/individual MCP · shared/individual skill) EXERCISED server-side, with committed receipts.
  • docs/provider-matrix.md — the canonical cell-by-cell capability matrix.