awslabs · leon1418 · Aug 4, 2026 · Aug 4, 2026 · Aug 4, 2026 · Aug 5, 2026
@@ -0,0 +1,43 @@
+# ATX Custom bundle — agent-advisor
+
+The plugin keeps the **full** skill under `skills/agent-advisor/`, including the Migration Plan
+stage that reads the sibling `gcp-to-aws` engine. ATX Custom has no sibling skills to read, so the
+transformation published there must be self-contained.
+
+Rather than fork the prose — which drifts, as this repo has learned more than once — the ATX
+bundle is a **strict subset** of the canonical skill plus exactly one generated file:
+
+```text
+transformation_definition.md = preamble.md + skills/agent-advisor/SKILL.md (frontmatter stripped)
+```
+
+Every other file in the bundle is byte-identical to its canonical source. There is no second copy
+of any instruction to keep in sync.
+
+| File           | Role                                                                                                                                                                                                                             |
+| -------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `preamble.md`  | The only ATX-specific text: executor differences (no interactive tools, no subagent dispatch, `${CLAUDE_PLUGIN_ROOT}` fallback), the headless input contract, the exit criteria, and the required `validation_summary.md` format |
+| `manifest.txt` | Whitelist of canonical files that ship. A canonical file that is neither listed nor matched by an exclusion rule fails `atx:check` — new files must be classified, never silently dropped                                        |
+
+## Commands
+
+```bash
+mise run atx:build   # write the bundle to .tmp/atx-bundle/agent-advisor/, then verify
+mise run atx:check   # verify only (part of `mise run lint`)
+```
+
+`atx:build` writes a directory of plain text and `.py` files with no build step and no install:
+point whatever runs it at `transformation_definition.md` and give it that directory as the root every
+relative path resolves against. Publishing to a transformation registry is a deliberate manual step,
+and the registry's own CLI owns it — nothing in this repo does it for you.
+
+## What `atx:check` guarantees
+
+1. **Manifest ↔ canonical agree** — no dangling entries, no unclassified canonical files.
+2. **No sibling-skill loads** — a resolvable path into another skill (`${CLAUDE_PLUGIN_ROOT}/skills/<other>`, `$GCP_BASE`) is an error everywhere except `references/phases/migration-plan/`, whose Step -1 capability gate resolves the stage to `not_applicable` when the engine is absent. This check is what makes "the ATX build does not depend on gcp-to-aws" a machine-verified property.
+3. **The capability gate exists** — `migration-plan.md` must carry its marker, so the bundle degrades honestly instead of trying to read files that are not there.
+4. **The phase graph closes inside the bundle** — every `_file:` fragment and `_advances_to:` target resolves within the shipped set.
+5. **Runtime scripts load on a bare host** — no top-level third-party import. Lazy imports are reported as warnings, since they only bind on a code path that may be disabled (the live model probe) or provisioned by `uv`.
+
+Warnings are tracked, not fatal: prose that names a sibling skill or a Claude Code slash command
+still works, but it dangles for an ATX reader and should be reworded to be environment-neutral.
@@ -0,0 +1,77 @@
+# ATX bundle manifest — agent-advisor
+#
+# Paths are relative to skills/agent-advisor/. Every file listed here is copied
+# BYTE-IDENTICAL into the ATX bundle; the bundle is a strict subset of the
+# canonical skill plus one generated entry file (transformation_definition.md).
+#
+# A canonical file that is neither listed here nor matched by an exclusion rule
+# in tools/atx-bundle/build.ts fails atx:check — new files must be classified.
+
+references/decision-refs/agentcore.md
+references/decision-refs/batch.md
+references/decision-refs/cost-levers.md
+references/decision-refs/ecs.md
+references/decision-refs/eks.md
+references/decision-refs/freshness.md
+references/decision-refs/lambda-microvms.md
+references/decision-refs/lambda.md
+references/decision-refs/managed-alternatives.md
+references/decision-refs/model-selection.md
+references/decision-refs/poc-shapes.md
+references/decision-refs/temporal.md
+references/decision-refs/workload-classes.md
+references/diagram/build-diagram.md
+references/handoff/handoff-migration.md
+references/models/anthropic-bedrock-2026-07-21.json
+references/models/openai-bedrock-2026-07-21.json
+references/output-templates/recommendation-doc.md
+references/phases/add-capabilities/add-capabilities-assemble.md
+references/phases/add-capabilities/add-capabilities.md
+references/phases/clarify/clarify-assemble.md
+references/phases/clarify/clarify-business.md
+references/phases/clarify/clarify-technical.md
+references/phases/clarify/clarify.md
+references/phases/confirm/confirm-assemble.md
+references/phases/confirm/confirm.md
+references/phases/design/design-assemble.md
+references/phases/design/design.md
+references/phases/discover/discover-assemble.md
+references/phases/discover/discover.md
+references/phases/estimate/estimate-assemble.md
+references/phases/estimate/estimate.md
+references/phases/generate/generate-assemble.md
+references/phases/generate/generate-report.md
+references/phases/generate/generate.md
+references/phases/intake/intake-assemble.md
+references/phases/intake/intake.md
+references/phases/migration-plan/migration-plan-assemble.md
+references/phases/migration-plan/migration-plan-gcp-constraints.md
+references/phases/migration-plan/migration-plan.md
+references/phases/model-recommend/model-recommend-assemble.md
+references/phases/model-recommend/model-recommend.md
+references/phases/poc/poc-assemble.md
+references/phases/poc/poc-report.md
+references/phases/poc/poc.md
+references/report-help-banner.md
+references/report-shell.md
+references/runtimes/agentcore.json
+references/runtimes/ecs.json
+references/runtimes/eks.json
+references/runtimes/lambda_microvms.json
+references/runtimes/lambda.json
+references/vendored/dsl/INTERPRETER.md
+references/vendored/README.md
+scripts/anthropic_model_recommendation.py
+scripts/build_diagram.py
+scripts/model_recommendation.py
+scripts/openai_model_recommendation.py
+scripts/pyproject.toml
+scripts/schemas/model-recommendation-input.json
+scripts/schemas/model-recommendation.json
+scripts/schemas/model-verification.json
+scripts/schemas/scoring-result.json
+scripts/scoring.py
+scripts/uv.lock
+scripts/verify_model_path.py
+SKILL.md
+scripts/schemas/seed.json
@@ -0,0 +1,97 @@
+# AWS Agent Advisor — ATX Custom transformation
+
+This transformation runs an **advisory** flow: it reads the target repository and produces a set
+of assessment documents. It is **not** a code rewrite. A run that modifies application source is
+a failed run, even if the application still builds.
+
+The instructions after this preamble are the skill's own contract (`SKILL.md` and the phase files
+it loads). This preamble records only what differs in the ATX execution environment. Where the two
+disagree, **this preamble wins**.
+
+## Environment contract
+
+1. **No interactive tools.** There is no `AskUserQuestion` tool and no human to answer. Wherever a
+   phase file says to ask the user, resolve the value instead:
+   - Read `.agent-advisor/seed.json` if the repository provides one, then `CLAUDE.md` /
+     `AGENTS.md` at the repository root.
+   - If a needed value is in neither, append the question and the value you assumed — with the
+     reason — to `.agent-advisor/UNANSWERED.md`, and continue.
+   - Never block waiting for input, and never invent a requirement that contradicts a value you
+     were given.
+2. **No subagent dispatch.** Any phase whose frontmatter declares `_exec: { _agent: ... }`
+   (currently only Estimate) runs **inline** in this session. Its fragments and assembler execute
+   exactly as written; only the isolation is dropped.
+3. **`${CLAUDE_PLUGIN_ROOT}` does not resolve.** Use the fallback that `SKILL.md` documents: every
+   path is relative to this transformation's own directory — `./scripts/...`,
+   `./references/runtimes/...`, `./references/decision-refs/...`, `./references/phases/...`.
+4. **Scoring comes from the engine, never from your judgement.** Run `scripts/scoring.py` and
+   `scripts/model_recommendation.py` exactly as the phase files instruct. Prefer `uv run`; if `uv`
+   is not installed, invoke them with `python3` — all six runtime scripts are standard-library
+   only. Hand-written scores are a failed run.
+5. **MCP servers are optional here.** The `awsknowledge` and `temporal-docs` servers may be absent.
+   Take the documented fallback path (cached values plus an honest freshness note per
+   `references/decision-refs/freshness.md`) as the **normal** path, not an exception.
+6. **The full Migration Plan stage is unavailable in this bundle**, but the POC is not.
+   The plan is generated by a sibling engine that does not ship here, so that stage resolves to
+   `not_applicable` (marked `migration_plan_unavailable = "engine_absent"`) with a note to the
+   user. The POC gate is still offered, and the POC is **design-backed** — built from
+   `design.json` and labelled "not plan-backed". Everything through the recommendation is
+   unaffected.
+7. **Gate answers come from the run, not from a person.** Both optional gates (Migration Plan,
+   POC) are decided by `.agent-advisor/seed.json` under a `gates` key — e.g.
+   `{"gates": {"poc": "accept"}}`. When the seed is silent, **decline** both gates and finish
+   after the recommendation. Never decide a gate on the user's behalf beyond that default.
+8. **Do not touch AWS unless the repository explicitly authorises it.** No deployments and no live
+   Bedrock model probe by default; `verification.probe_status` stays `not_run` and account
+   availability stays provisional.
+
+## Output contract
+
+Write every artifact under `.agent-advisor/<run-id>/` in the target repository. Leave them in
+place — do not clean up the run directory. Application source must be byte-identical when you
+finish.
+
+Exit criteria — all of:
+
+`context-signals.json`, `answers.json`, `scoring-result.json`, `model-recommendation-input.json`,
+`model-recommendation.json`, `confirm.json`, `design.json`, `estimate.json`, `diagram.md`,
+`recommendation.md`, `mini-brief.md`, `recommendation-report.html`
+
+## Required validation summary format
+
+**Your FINAL MESSAGE of the run must be exactly the block below — nothing before it, nothing after
+it.** Benchmark harnesses capture that final message verbatim as `validation_summary.md` and grade
+it against this structure, so a friendly free-form wrap-up scores zero even when every criterion
+actually passed. Do not summarise, do not add findings, do not add pleasantries: emit the block.
+
+Also write the same block to `$RUN_DIR/validation_summary.md` so the result is auditable from the
+run directory itself.
+
+The structure, in both places: composed of `## OVERALL STATUS`, `## EXIT CRITERIA RESULTS`, and
+`## UNMET CRITERIA` exactly as follows:
+
+```markdown
+## OVERALL STATUS: COMPLETE
+
+## EXIT CRITERIA RESULTS
+
+- **Criterion:** <name>
+  **Status:** PASS
+  **Evidence:** <file path or command output>
+
+- **Criterion:** <name>
+  **Status:** FAIL
+  **Evidence:** <what was found instead>
+
+## UNMET CRITERIA
+
+<one bullet per FAIL above, or "None." when every criterion passed>
+```
+
+`## OVERALL STATUS:` takes `COMPLETE` only when every criterion is PASS; otherwise `INCOMPLETE`.
+One entry per exit-criteria artifact plus one per validation command that ran.
+
+---
+
+<!-- The canonical SKILL.md follows. Everything below this line is copied verbatim from
+     skills/agent-advisor/SKILL.md by tools/atx-bundle/build.ts — do not edit it here. -->
@@ -0,0 +1,2 @@
+ANTHROPIC_API_KEY=replace-me
+ANTHROPIC_MODEL=claude-3-7-sonnet-latest
@@ -0,0 +1,7 @@
+.agent-advisor/
+.agent-advisor-*/
+.env
+.pytest_cache/
+.venv/
+__pycache__/
+*.pyc
@@ -0,0 +1,35 @@
+# Agent Advisor Test Context
+
+This repository is a static migration-assessment fixture for `agent-advisor`.
+Do not rewrite the source application unless the user explicitly asks.
+
+## Workload
+
+- Existing Anthropic customer-support agent using the first-party Messages API.
+- Source model: `claude-3-7-sonnet-latest`.
+- Sessions normally last 10-20 minutes.
+- Traffic is bursty, with fewer than 20 concurrent sessions.
+- The application uses a custom tool loop and Anthropic platform features.
+- The team prefers low operational overhead and has no Kubernetes requirement.
+
+## Model Requirements
+
+- Preserve the Anthropic Messages API for the first migration.
+- Priority: balanced.
+- Required capabilities: tool use and extended thinking.
+- Required context: at most 200,000 tokens.
+- Expected output: at most 16,000 tokens.
+- No Bedrock Guardrails, invocation logging, or multi-model Converse requirement.
+- Hypothetical target region: `us-east-1`.
+
+## AWS Access
+
+The user does not have a target AWS account yet. Claude Code's own authentication
+does not count as a customer target account. Do not run AWS CLI commands or a live
+model probe. Keep account availability provisional and verification `not_run`.
+
+## Test Boundary
+
+Run the recommendation flow and generate its local assessment artifacts. Do not
+deploy resources. At optional Migration Plan or POC gates, wait for the user's
+explicit choice.
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		ANTHROPIC_API_KEY=replace-me
		ANTHROPIC_MODEL=claude-3-7-sonnet-latest