Orchestrate Agentic Readiness Analysis (ARA) and Modernization Readiness Analysis (MODA) across a service portfolio using AWS Transform Continuous Modernization (atx ct), plus dependency-aware Execution Plan (EBA) generation via atx custom. Use when the user wants to assess agentic-AI readiness or cloud-modernization readiness across multiple repos/microservices, run portfolio analyses, manage findings/remediations, or build a modernization roadmap. Triggers: "agentic readiness", "ARA", "MODA", "modernization readiness", "portfolio analysis", "atx ct", "continuous modernization", "execution plan".
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
Orchestrate Agentic Readiness Analysis (ARA) and Modernization Readiness Analysis (MODA) across a service portfolio using AWS Transform Continuous Modernization (atx ct), plus dependency-aware Execution Plan (EBA) generation via atx custom. Use when the user wants to assess agentic-AI readiness or cloud-modernization readiness across multiple repos/microservices, run portfolio analyses, manage findings/remediations, or build a modernization roadmap. Triggers: "agentic readiness", "ARA", "MODA", "modernization readiness", "portfolio analysis", "atx ct", "continuous modernization", "execution plan".
ARA / MODA Portfolio Analysis Orchestrator
Turn Claude into an orchestrator for running comprehensive analyses across a service portfolio using AWS Transform Continuous Modernization (atx ct). ct handles repository discovery, parallel execution, portfolio-level aggregation, findings management, and remediation natively — this skill is the orchestration layer connecting ct with Execution Plan generation.
Ported from the AWS "orchestrator" Kiro Power (aws-samples/sample-agentic-readiness-analysis-orchestrator). Where the original referenced Kiro's readSteering action, use the Read tool on the files in references/; where it referenced executeBash, use the Bash tool.
⚠️ Verified 2026-08-03 against atx 3.9.0 — read this before running anything.
First: check your version correctly.atx --versionlies inside Claude Code — it prints 2.1.220.613, which is Builder Toolbox's build number for claude-code, inherited via $TOOLBOX_TOOL_VERSION. atx resolves TOOLBOX_TOOL_VERSION || SEG_VERSION || DEV_BUILD, so the toolbox var wins over its own . Use one of these instead:
Any doc or bug report quoting a 2.1.x "atx version" is quoting the wrong number.
No server needed.atx ct server still exists but is hidden and deprecated — analyses run in-process. It prints its own warning and starts a real daemon on :8081 that will block your shell, so never invoke it. Use atx ct status --health (an in-process check, not a daemon ping). There is also a hidden top-level --api <url> defaulting to http://localhost:8081; leave it alone.
--wait DOES exist on analysis run, remediation create, and remediation retry — registered with .hideHelp(), so it appears in neither --help nor atx ct schema. Earlier revisions of this doc claimed it was removed; that was wrong. Prefer polling in agent workflows (see Long-running analysis) — but that is a choice, not a missing flag.
analysis list-artifacts / get-artifact are GONE — 0 occurrences in the shipped bundle; error: unknown command. Reports now come from analysis get --json → report_paths, but that map is markdown-only: .json and .html (and every portfolio artifact) live under ~/.atxct/sources/<src>/<type>/runs/<id>/, which report_paths never mentions. See Report artifacts.
Region: only us-east-1 resolves. A stray AWS_REGION=us-west-2 makes the definition/credential endpoint NXDOMAIN (transform-custom.us-west-2.api.aws). Always export AWS_REGION=us-east-1 AWS_DEFAULT_REGION=us-east-1.
atx ct schema is NOT a complete manifest. It reports 11 top-level commands and omits the hidden schedule and server groups plus every hidden flag. The shipped bundle is ground truth; treat schema (and ATXControlTower/docs/cli-reference.md) as incomplete.
Zero-findings bug is fixed as of 3.9.0. On 3.7.0 analyses completed with 0 findings. Verified fixed on 3.9.0: a single-repo ARA produced 43 findings and a MOD 31. If you are on 3.7.0, upgrade. See Zero-findings bug.
ARA/MOD findings are never auto-remediable. Every finding they emit has fix: null, and remediation create --ids rejects those with non_remediable=[...]. Remediation of ARA/MOD output must go through --transformation-name. See Remediation.
Supported analyses
Analysis
atx ct type
What it evaluates
Agentic Readiness (ARA)
agentic-readiness
43 questions across 8 categories — API Surface & Interface Design (8), Authentication/Authorization/Identity (7), State Management & Transactional Integrity (7), Data Accessibility & Quality (7), Engineering & Deployment Maturity (5), Human-in-the-Loop & Approval Workflows (3), Discoverability & Semantic Readiness (3), Observability of Target Systems (3). Whether systems are ready to be safely called by AI agents.
Modernization Readiness (MODA)
modernization-readiness
38 questions across 5 categories — Infrastructure (11), Operations (9), Security (7), Application (6), Data (5). Cloud modernization opportunity assessment; identifies containerization, serverless, and platform-upgrade candidates.
Execution Plan (EBA)
N/A — generated by Claude via atx custom def exec
Dependency-aware roadmap. Claude builds atx-config-exec-plan.yaml (with additionalPlanContext) from discovered repos + ct findings, then runs the EBA TD. The ONE place additionalPlanContext is used.
Counts and category names are derived from the TDs in this repo (definitions/managed/{agentic,modernization}-readiness-analysis/SKILL.md) — the authoritative source. An earlier revision of this doc claimed ARA was "56 criteria across 5 categories" and then listed MOD's categories; both were wrong.
Requested scope
What runs
agentic-readiness
ARA across all discovered repos (per-repo + portfolio aggregation)
modernization
MODA across all discovered repos (per-repo + portfolio aggregation)
full
ARA + MODA + Execution Plan
Portfolio aggregation needs ≥ 2 repos
Both portfolio TDs are managed and built into ct — AWS/portfolio-agentic-readiness-analysis and AWS/portfolio-modernization-readiness-analysis. There is no standalone portfolio command (ct portfolio, ct portfolio run, ct analysis portfolio, ct analysis aggregate are all unknown command); the roll-up runs automatically as a second phase of analysis run, and lands in portfolio_ara_summary / portfolio_mod_summary on the analysis record.
⚠️ Aggregation is per-run, and each portfolio TD requires at least 2 discovered reports (definitions/managed/portfolio-agentic-readiness-analysis/SKILL.md step 1.1; MOD likewise). Consequences:
A single-repo run gets no portfolio report. The CLI still invokes the portfolio TD and it still burns time (~13 min observed) before declining, then logs runPortfolioAra: no portfolio ARA report found for <name> and leaves portfolio_ara_summary: null. That is expected behavior on 1 repo, not a bug — don't debug it as one.
Separate single-repo runs do not accumulate. Aggregation only ever covers the repos in one analysis run invocation, so to get a portfolio view, pass ≥2 repos (or omit --repo to take every repo under --source).
The bridge phase needs BOTH portfolio reports.runBridge populates bridge_summary, and skips with runBridge: missing portfolio report paths, skipping bridge unless a portfolio ARA and a portfolio MOD report exist. Note this overlaps with the custom definitions/custom/bridge-analysis/ TD in this repo — decide deliberately which one a given workflow uses.
The per-repo phase runs concurrently. An 11-repo ARA cloned and began analyzing 8 repos within ~2 minutes, so wall-clock ≈ slowest repo + portfolio phase (tens of minutes), not repos × ~20 min. ATXCT_MAX_CONCURRENT_DEEP_SCANS (default 1) gates only tech-debt-comprehensive and does not throttle ARA/MOD fan-out. Separately, an ALREADY_RUNNING guard rejects a run whose repos are mid-analysis for the same type.
⚠️ status: complete is NOT a terminal signal on a multi-repo run.status flips to complete (and completed_at is stamped) when the per-repo phase finishes, while the portfolio phase is still running — for tens of minutes. Measured on an 11-repo ARA: status: complete, completed_at: 20:27:42Z, yet at 20:37 the launching process was still alive (68 min elapsed) and the log was still emitting Still analyzing portfolio-<name>... (7m elapsed). Read at that moment, the record looks like a total failure and is not one:
field
at premature complete
at true completion
status
complete
complete / failed
report_paths
{} (0 entries)
12 entries
portfolio_ara_summary
null
populated
progress
{"done": 0, "total": 11}
—
An agent that stops at status here reports 0 findings and no reports on a run that in fact produced 899 findings and a full portfolio bundle. Wait on one of these instead:
If you launched the run, wait for the process to exit. That is the only unambiguous signal.
Otherwise require status terminal andreport_paths non-empty — that map is written in the final record update, after the portfolio phase, so it is empty for the entire premature window. Cap it with a timeout: a run that fails early leaves report_paths empty forever.
Worse, the premature transition can swallow a later failure. Because status is already terminal, a portfolio-phase error cannot transition it: saveFn: exhausted 3 retries persisting analysis <id>: Analysis is already in terminal status COMPLETE; cannot transition. Two runs of the same ARA hitting the samerepositoryId: null persist bug therefore ended differently — one failed, one complete with a populated error field. So status: complete does not imply error: null. Read error and repo_errors explicitly; never infer them from status.
Never conclude "no reports were generated" from an empty report_paths alone — check disk (ls ~/.atxct/sources/*/*/runs/<id>/portfolio-*/*-analysis/) before reporting anything to the user. This is the third way status misleads, alongside: ARA reads failed when its output is fine (the repositoryId: null persist bug), and MOD reads complete when its portfolio report is missing (filed as a repo_error, which does not fail the run).
When to use
Planning agentic AI adoption across microservices
Identifying shared infra gaps and modernization opportunities
Prioritizing modernization by dependencies
Tracking portfolio-wide readiness over time
Generating execution roadmaps from findings
Continuous monitoring of tech debt / readiness regressions
Reference files (read on demand)
Do NOT load all of these proactively. Pick the one relevant to the current task and Read it.
All four were rewritten against verified 3.9.0 behavior on 2026-08-03: the atx ct server start-up steps are gone (replaced by atx ct status --health), list-artifacts/get-artifact are gone (replaced by analysis get --json → report_paths, with the "report_paths is not the whole bundle" caveat), and troubleshooting.md's claim that artifacts are "NOT on the local filesystem" is corrected — they are on local disk. This SKILL.md remains the source of truth for CLI behavior; the references carry the interactive flows and error taxonomies.
Demo harness (scripts re-verified against 3.9.0 on 2026-08-04; local-first)
A full reset-and-rebuild harness lives at the project root.
All scripts default to LOCAL mode (no GitHub, no Code Defender); pass --remote for GitHub mode.
<project>/demo-scripts/
├── 00-full-setup.sh # Bake env: health → source → discovery → trim → ARA → MODA → export (~45 min)
├── 00-push-repos.sh # [remote only] Push 3 pre-baked repos to org (user runs — Code Defender)
├── 01-live-discovery-push.sh # Live beat: pricing-cgi "appears" (3 repos → 4)
├── 02-reset-live-discovery.sh # Reset just the live-discovery beat
├── 99-full-reset.sh # Nuke everything (ct + local state; + GitHub with --remote)
└── README.md
Full local cycle:99-full-reset.sh → 00-full-setup.sh → demo-ready. No prerequisites beyond atx CLI + AWS creds + jq.
Your role when someone is demoing
The demo README's "During the demo" table assigns steps to you. When the operator runs a demo, they drive the console and the live-discovery script; you drive ct. Expect these asks and map them as follows:
They say
You do
"run ARA on <repo>"
analysis run --type agentic-readiness --repo <src>::<repo> --source <src>, then wait properly (below)
"run MODA on <repo>"
same with --type modernization-readiness
"containerize <repo>"
remediation create --transformation-name <the-TD-published-for-this-demo> --repo <src>::<repo> [--local] — not--ids; ARA/MOD findings all have fix: null and are rejected as non_remediable. Ask which TD and confirm it resolves (custom def get -n <name> from a scratch dir) before running — there is no default containerization transform, and names go stale.
"show me the report"
Open the portfolio .html from the sources run tree — ls ~/.atxct/sources/*/*/runs/<id>/portfolio-*/*-analysis/. It is nowhere else.
"how ready are we?"
findings count --by severity --json + portfolio_ara_summary from the analysis record
Demo-specific traps, on top of the general rules above:
Never say "the analysis produced nothing" mid-demo. A single-repo demo run gets no portfolio report by design (aggregation needs ≥2 repos) and status: complete fires while the portfolio phase is still running. Check disk before reporting.
A failed ARA is usually a healthy run. The repositoryId: null persist bug fires after reports and per-repo findings are saved. Say what actually survived rather than declaring failure — the demo can continue.
One analysis type at a time. Launching ARA and MODA concurrently on one source causes contention; an ALREADY_RUNNING guard rejects overlapping runs on the same repos and type.
A single-repo run still burns ~13 min invoking the portfolio TD before declining. Set expectations out loud before starting.
Check the remediation TD resolves before the demo starts, not during it. The TD is whatever the operator published for this demo; custom def get -n <name> in the demo's account/region is a two-second check that turns a mid-demo "not found in the registry" into a pre-flight fix.
Key fixes baked into the harness:
Portfolio repos are git-inited on the fly — a fresh clone of the harness ships the portfolio dirs WITHOUT nested .git (they can't live inside the parent repo), and ct discovery scans for .git subdirs → finds 0. Setup self-heals: git init -b main + initial commit for any portfolio dir missing .git. Verified: without this loop, discovery finds zero of the 10 fixtures.
No server is started — setup and reset only health-check with atx ct status --health. The old PYENV_VERSION=system atx ct server startup is gone (see the CLI rules above).
Both analysis waits require report_paths non-empty, not just status — otherwise the scripts break out during the portfolio phase and export nothing. Setup also continues past a failed ARA when findings exist, instead of discarding a usable 45-minute environment.
Artifact export copies off disk (find ~/.atxct/sources -path "*runs/<id>/*") — analysis list-artifacts/get-artifact were removed in 3.9.0 and the old export silently produced nothing.
Local remediation requires a CLEAN worktree — prior analysis runs write services/<repo>/{ara,mod}-report.{md,json,html} artifacts INTO local repos (this is also where HTML reports live for local sources!). Remediation fails with "has uncommitted changes" until services/ is removed. Reset cleans this.
Local-source remediation completes with status pr_open and creates a local staging branch (atx-result-staging-*) — show the diff with git diff main.
findings update (single) instead of batch-update — batch fails with UnknownError on some IDs.
Retry-sweep in reset for teardown ordering (repos → findings → sources)
Pipe-stdin-consumption fix in loops (capture first, or < <(...) process substitution)
ct discovery does NOT follow symlinks — local sources need real repo dirs
[remote only] git-defender self-attest --reason 1 per repo; must run INSIDE a repo dir with matching origin remote set
Core workflow (high-level)
# 0. Pre-flight — FAIL FAST (see references/getting-started.md)
aws sts get-caller-identity
env -u TOOLBOX_TOOL_VERSION atx --version # NOT plain `atx --version` — see version note
export AWS_REGION=us-east-1 AWS_DEFAULT_REGION=us-east-1 # see region note above
atx ct status --health # in-process check; do NOT start `atx ct server`
# 1. Add a source. Local uses --path (absolute); remote uses --org + --token.
atx ct source add --name my-portfolio --provider local --path "$(pwd)/services"
# 2. Discover repositories (scans for .git subdirectories)
atx ct discovery scan --source my-portfolio
atx ct repository list
# 3. Run ARA (launch + poll — see "Long-running analysis" below)
atx ct analysis run --type agentic-readiness --source my-portfolio
# 4. Run MODA (after ARA; do not run both concurrently)
atx ct analysis run --type modernization-readiness --source my-portfolio
# 5. Inspect findings (cheap aggregate first, then drill in)
atx ct findings count --by severity --json
atx ct findings list --json
# 6. Retrieve reports — report_paths, NOT the removed get-artifact
atx ct analysis get --id <analysis-id> --json | jq -r '.report_paths | to_entries[] | "\(.key)\t\(.value.ara // .value.mod)"'
# 7. (Optional) Generate execution plan (see references/execution-plan.md)
atx custom def exec -n eba-execution-plan-generator -p . -g file://atx-config-exec-plan.yaml -x -t
Steering ARA / MODA without a custom TD
Four flags on analysis run shape the built-in ARA and MODA assessments — no custom TD required. Added in 3.5.0; verified present on 3.9.0. This is the answer whenever someone asks how to bias the analysis toward their environment.
Flag
Valid with
Values
Default
--context <text>
both ARA and MODA
free-form prose
none
--agent-scope <scope>
ARA only
read-only | write-enabled
read-only
--prefer <list>
MODA only
comma-separated, e.g. eks,aurora,bedrock
none
--avoid <list>
MODA only
comma-separated, e.g. self-managed-kafka
none
atx ct analysis run --type agentic-readiness --source my-src \
--agent-scope write-enabled --context "Customer-facing AI agents with tool-use"
atx ct analysis run --type modernization-readiness --source my-src \
--prefer eks,aurora,bedrock --avoid self-managed-kafka
Mixing them wrong is a hard validation error, not a silent ignore — e.g. --agent-scope with --type modernization-readiness fails with "--agent-scope is only valid with --type agentic-readiness." Note the asymmetry: --context is shared, the other three are type-exclusive.
Also available on analysis run: --tags k=v,... and --telemetry k=v,... (client=zerodebt is always appended). --repos is an alias for --repo; both are repeatable and comma-separated.
⚠️ Not exposed via MCP.analysis_run's only properties are assessment_type, inputs, repos, tags, telemetry — steering would have to ride inside inputs, which is unverified. Use the CLI when you need steering.
Source providers
ct supports four providers. Local sources use --path (a parent directory containing repos as subdirectories) — always absolute. Remote providers use --org + --token.
atx ct source add --name x --provider local --path /absolute/path/to/parent-dir
atx ct source add --name x --provider github --org my-org --token ghp_xxx # repo scope
atx ct source add --name x --provider gitlab --org my-group --token glpat-xxx [--url https://gitlab.example.com] # api scope
atx ct source add --name x --provider bitbucket --org my-workspace --token xxx [--username u] [--email e@x.com]
atx ct source list [--json]
atx ct source remove --name x
Do NOT point a local source at a single repo — point at the directory containing repos. Full detail in references/ct-workflow.md and references/getting-started.md.
Analysis types
The exact --type values accepted on 3.9.0, in the order the CLI lists them:
Run any transformation definition: --type custom --transformation-name <name>
⚠️ rapid-techdebt-analysis is a display rename over an unchanged storage value. The alias map is {"rapid-techdebt-analysis":"tech-debt-quick","quick-scan":"tech-debt-quick"}, so all three are accepted on input, but the canonical persisted value is tech-debt-quick — that is what analysis list --type <x> filters on and what stored records contain. This rename has no changelog entry. An earlier revision of this doc listed tech-debt-quick as the flag value to pass; it works, but rapid-techdebt-analysis is what --help advertises.
Config limitation:-g/--configuration on atx ct analysis run is only valid with --type custom — built-in types reject it. But this is not a blanket rule across the CLI: on remediation create, remote analysis, remote remediation, and schedule create, the gate is --transformation-name instead ("--configuration is only valid with --transformation-name"). To steer built-in ARA/MODA, use the steering flags above rather than -g.
Report artifacts
analysis list-artifacts and analysis get-artifact no longer exist — verified 2026-08-03: error: unknown command, and 0 occurrences across all three shipped bundles. They were never merely hidden here; they are not shipped. (Upstream they are registered {hidden:true} pending a read-path launch, with the ops throwing "not yet available".) Delete any code you find that calls them.
Reports are on local disk, and analysis get --json tells you where.report_paths maps each repo slug to its report file:
⚠️ report_paths is the markdown-only view. It is the right way to enumerate which repos produced reports, but the wrong way to reach a .json or .html — those live in a different tree. See Three on-disk locations before concluding an artifact wasn't generated.
Note analysis list returns a thinner object with no report_paths — you must call get per id. analysis get --json exposes 31 keys, the useful ones being report_paths, ara_results, mod_results, portfolio_summary, portfolio_ara_summary, portfolio_mod_summary, repo_scores, repo_status, repo_errors, progress, repos_done, repos_total, status.
Three on-disk locations — know which one has what
report_paths points at the thinnest of the three. Measured on one 11-repo ARA run (.md / .json / .html / .metadata.json file counts):
<type> is the SOURCE's analysis root, not the run's type. A modernization-readiness run's output sits under sources/<src>/agentic-readiness/runs/<id>/ — verified: the MOD run 01KZ4DCZ… wrote .../agentic-readiness/runs/01KZ4DCZ…/<repo>/modernization-readiness-analysis/. Do not build the path from the analysis type; glob sources/*/*/runs/<id>/ or read portfolio_summary.report_path.
Per-repo dirs here are slug-mangled (<source>-<repo>-<16hex>, a sha256 prefix), unlike the <source>__<repo> form in shared/analyses/. Glob, don't construct.
HTML and portfolio JSON exist ONLY here. Not in shared/analyses/, and not in any working tree — no repo owns portfolio output. So "there is no HTML/JSON" is almost always a wrong-directory error, not a missing artifact. find ~/.atxct -name '<name>-portfolio-*-report.html' settles it.
Also: ~/.atxct/sources/<src>/{config.json,repos/index.json}, ~/.atxct/shared/analyses/index.json, and a <type>/latest symlink.
Provider affects only the working-tree copy:
Remote (GitHub) sources: nothing is committed to the repos — no branch, no files. ct only touches repo git during a remediation (branch + PR). Both ~/.atxct trees are still written.
Local sources: analyses ALSO write the full per-repo bundle into each working tree (including HTML). ⚠️ Those files make the worktree dirty, which blocks subsequent remediation ("has uncommitted changes") — clean services/ before remediating.
So: do not hardcode formats, and do not treat report_paths as complete. It is markdown-only in practice. To get a .json (what the EBA TD and every downstream tool consume) or an .html (what you open in a browser for a demo), go to the sources/ run tree — for portfolio artifacts that is the only copy. Render with pandoc only after confirming the HTML genuinely isn't there.
# The reliable way to find every artifact of a run, regardless of type or slug mangling:
find ~/.atxct/sources -path "*runs/<analysis-id>/*" -type f \( -name '*.md' -o -name '*.json' -o -name '*.html' \)
# Portfolio bundle for one run (html + json live here and nowhere else):
ls ~/.atxct/sources/*/*/runs/<analysis-id>/portfolio-*/*-analysis/
⚠️ ct auto-commits report bundles into local-source repos. Each analysis lands its own commit (authored ATX Bot <checkpoint@atx.bot>) in the fixture/repo working tree. Two consequences: a tree can look clean right after an analysis because the output was committed, not because nothing was written; and a repo mid-analysis has those files modified, which trips the remediation dirty-worktree guard until the run finishes.
Finding counts differ between the CLI and the report JSON — by design. The report splits findings[] (severity-bearing gaps) from evaluations[] (passing questions, status: pass, no severity). The CLI's findings count includes both; the report's counts.total does not. Measured on one ARA: counts.total: 36 vs CLI 43, with disjoint question_id sets that union to exactly 43. Don't write an assertion that expects those two numbers to match.
Zero-findings bug (3.7.0)
On 3.7.0 an analysis can report complete while contributing zero findings — this account currently shows 29 complete analyses and findings: 0. Symptoms:
The tell is in analysis get --json: repo_errors says "No report generated" for a repo while repo_status says "done" and repo_scores still reports a tier and blocker_count: 0 — a score fabricated from a missing report. Root cause is a parser fallback: when a report's categories field is emitted as a JSON array (a shape the TD now produces), parseModReport fails to find a top-level findings[], falls back to markdown, and yields nothing.
Fix: upgrade to ≥ 3.9.0, which recovers findings from all report shapes. There is no atx update subcommand (an earlier revision of this doc claimed there was — atx update --help just prints top-level help). Re-run the installer, which resolves latest from index.json:
That is a write operation — get the user's go-ahead. Upgrading is non-destructive to ct state: sources, repos, and analyses all survive (verified across a 3.7.0 → 3.9.0 upgrade).
Verified fixed on 3.9.0 (2026-08-03): a single-repo ARA on a fixture produced 43 findings (10 high / 14 medium / 19 low — exactly the 43 ARA questions, split across all 8 categories), and MOD on the same repo produced 31. Both analyses reached complete with repo_errors: {}.
Partial self-heal: an internal repairMissingReports runs automatically inside getAnalysis() and listAnalyses() — so simply running atx ct analysis get --id <id> may repair an analysis. It only handles agentic-readiness/modernization-readiness, only entries whose error is exactly "No report generated", and only when it can find a report on disk to copy in. It cannot help when the report was never written — which is this account's case. There is no CLI command to trigger it directly.
Hosted web Console
AWS Transform has a hosted web Console — "Continuous modernization" — with tabs Dashboard / Findings / Remediations / Analyses / Sources / Settings. It reads the same account-scoped data (ct CLI and Console show the same sources, analyses, findings, remediations live). Best "show the audience" surface — prefer it over exporting reports for demos. There is no local web UI: analyses run in-process, and the deprecated :8081 daemon serves an API, not a page.
The console's Run new analysis wizard offers Infrastructure (AWS managed — "Coming soon", or Customer owned → pick a deployed stack) and Schedule (Run once / Recurring). So today remote execution always runs on a stack you deployed yourself via atx ct remote provision.
Console vs CLI — where they differ in capability:
Task
Console
CLI
Browse findings/analyses, demo
✅ best
workable
Run analysis on an SCM source
✅
✅
Run analysis on a local source remotely
❌ fails — no upload step (see below)
✅ auto-bundles to S3
ARA/MODA steering (--context, --prefer, …)
❌ not exposed
✅
Recurring schedules
✅ Recurring
✅ schedule create (hidden)
Neither surface is a superset. Reach for the CLI whenever a local source or steering is involved.
Agent behavior: long-running analysis
Analyses take 5–15 minutes per repo (large repos 20–30 min). Handle gracefully:
Inform the user immediately after launching: "Running ARA across N repos — ~5–15 min per repo. I'll monitor and report when done."
Prefer polling over --wait in agent workflows. --wait does exist (see below), but blocking a single tool call for 45 minutes gives the user no progress signal and risks a timeout. Launch without it, then poll.
Poll autonomously with atx ct analysis get --id <id> every 30–60s. Don't ask the user to wait.
On completion, report a summary (finding count, severity distribution, repos analyzed) — and check repo_errors, since complete does not mean every repo produced a report.
On failure, report the error and next steps (references/troubleshooting.md).
atx ct analysis run --type agentic-readiness --source my-portfolio # returns immediately with an ID
atx ct analysis get --id <analysis-id> # running | complete | failed
atx ct findings count --by severity --json # cheap server-side aggregation
--wait exists but is hidden. It is registered with .hideHelp() on analysis run, remediation create, remediation retry, so it shows up in neither --help nor atx ct schema — but it parses and works. Verified by differential test: analysis run --type agentic-readiness --wait advances to the next validation error (At least one of --repo or --source is required), whereas an unknown flag yields error: unknown option. Note that testing --waitwithout--type is inconclusive — the missing-required-option check fires first and produces the same message for any flag.
Its own description: "Block until the analysis completes, printing status as it runs. Without it, the command starts the analysis and returns immediately." For remediation create the bundle adds: "Opt-in for now; blocking will not become the default until a later release." Use it in scripts/CI where a blocking call is fine; poll in interactive agent work. The harness run-fixtures.sh polls.
Two further hidden subcommands exist for remote Batch — analysis execute --id <id> [--wait] [--wait-timeout <min>] and remediation execute ... (default timeout 240 min). These are internal execute-only phases; do not invoke them directly. By contrast remote status --wait is visible and normal (EC2 only).
Polling in a background shell script: the OS shell here is zsh, where $status is a read-only reserved variable — a loop that does status=$(...) fails with "read-only variable: status". Use a different name (st=$(...)). Verified 2026-07.
Operational gotchas (verified 2026-07)
Confirmed behaviors that differ from naive expectation — save time by knowing these up front:
JSON field names are terse and not always obvious. Inspect before scripting:
source list --json → each object uses .source (NOT .name) plus .provider, .identifier.
repository list --json → paginated envelope { items, nextToken, total } — repos are under .items[] (NOT a top-level array), keyed by .slug and .source. The key is nextToken, not next_cursor (verified 2026-08-03). Item keys: archived, assessed, created_at, created_by, default_branch, full_name, has_workflow, id, labels, language, private, schema_version, slug, source, updated_at.
findings list --json → top-level array; auto-fix indicator is .fix (currently always null for ARA/MODA).
Pagination: pass --next-token <token> (matching the response's nextToken) on source list, repository list, analysis list, findings list, remediation list. There is no --max-results/--limit on the CLI — only the MCP repository_list tool has a limit. Caveat: repository list pagination is "only supported with a single --source."
Prefer findings count --by severity|repo|analysis-type over listing-then-counting — it aggregates server-side and is far cheaper on large portfolios.
Teardown ordering (reverse of creation): you CANNOT remove a source that still has repos (HTTP 409 "has repositories that still reference it"). Order: delete repositories → delete findings → remove source.analysis delete --cascade-findings removes an analysis's findings, but repos discovered under a source persist until explicitly deleted.
Findings can only be deleted when dismissed or obsolete. To purge open findings: findings batch-update --ids <csv> --status dismissed --reason "..." first, then findings delete --id <id> per finding (no batch delete).
source remove and per-item deletes exit 0 even when they no-op (e.g. swallowed 409). Always re-list and verify counts afterward; don't trust exit codes alone.
Bulk deletes are slow (sequential per-item API calls; ~1–2 s each). For dozens/hundreds of items, run the loop as a background Bash task and poll live counts, not the buffered output.
custom def get writes files into CWD — run it from a scratch dir or clean up afterward.
Never invoke a bare atx ct server. It is hidden and deprecated but still functional: it starts a real daemon on :8081 and blocks your shell until killed (this caused a 2-minute tool timeout during verification). It prints its own warning: "atx ct server is deprecated and no longer required for atx ct commands, which run directly in-process."
Amazon Code Defender blocks git push to unapproved PUBLIC GitHub repos — both from a local terminal AND from the ct remediation server-side push. The fix is git-defender self-attest --reason 1 --url "https://github.com/<org>/<repo>.git" for each repo. Do this BEFORE running remediation; the setup harness handles it automatically.
PYENV_VERSION=system for git pushes on machines where pyenv's Python (3.13.2) has a broken hashlib (blake2b/blake2s ValueError). The system Python (3.9.6+) works. Without this, git push during remediation fails with a cryptic Python traceback — the root cause of "remediation failed to push branch". Historically applied when starting the server; now that execution is in-process, export it in the environment running atx ct — PYENV_VERSION=system atx ct remediation create .... The demo scripts no longer set it (they had it only on the removed server start), so apply it yourself if you hit a blake2b traceback during a local remediation.
Pipe-stdin consumption in loops: when iterating over atx ct repository list --json | jq ... | while read ..., the atx CLI can consume the pipe's stdin and terminate the loop after one item. Fix: capture the output into a variable first, then while read ... <<< "$variable". The harness scripts use this pattern.
Guided workflow (UX)
After each major step, offer 2–3 concrete next actions, most valuable first — teach continuous modernization through progressive disclosure. Never leave the user at a dead end.
After discovery: "Discovered N repos. Run Agentic Readiness (ARA), Modernization Readiness (MODA), or both?"
After ARA: offer to run MODA, show high-severity findings, or generate a per-repo report.
After MODA: offer execution plan, findings by severity/category, remediation of auto-fixable findings, or the portfolio report.
After both ARA+MODA: recommend the execution plan (dependency-aware phased roadmap).
After EBA: offer to summarize phases/timeline, show quick wins, or open the full plan.
After findings listed: offer remediations, dismissals, or category/repo filters.
Safety contract: Execution Plan (EBA)
Built-in ARA/MODA reject -g, but they are not uncontextualizable — use the steering flags (--context, --agent-scope, --prefer, --avoid). The Execution Plan is the ONE analysis Claude generates: it builds atx-config-exec-plan.yaml with additionalPlanContext from ct data + user execution constraints, then runs it. See references/execution-plan.md for the full interactive flow.
EBA runs ONLY AFTER both ARA and MODA show status complete (verify via atx ct analysis list).
Verify the reports actually exist — complete is not sufficient (see Zero-findings bug). Check analysis get --id <id> --json for populated report_paths and an empty repo_errors, then confirm the files are on disk.
Issue exactly one Bash call for the EBA command with timeout: 1800000 ms. Do not poll, background, or split it.
Do NOT run EBA concurrently with any atx ct analysis run — ct and custom exec conflict on git state.
After it returns, verify the execution plan artifact exists (ls portfolio-execution-plan/*-portfolio-exec-plan.md) before reporting success.
MCP integration (alternative to CLI)
atx ct mcp exposes 25 tools (verified 2026-08-03 via a live stdio initialize + tools/list handshake; serverInfo.name = "atxct"). An earlier revision claimed 41 tools and a "richer surface: tags, schedules, dashboards" — that was backwards. MCP is a strict subset of the CLI.
Absent from MCP — use the CLI for these: findings batch-update, repository update, setup, status, the entire remote group, the entire schedule group, and the ARA/MODA steering flags. EBA also requires the CLI (atx custom def exec is not an MCP tool). MCP is still the cleaner path for structured read/query work.
MCP vs CLI differences (local source provider_config.rootPath vs --path; remote identifier/token vs --org/--token; assessment_type vs --type; agent polls analysis/get instead of --wait) are covered in references/ct-workflow.md.
Authoring a custom remediation TD (Dockerfile/K8s, upgrades, etc.)
ARA/MODA findings are assessment-only — verified 2026-07 that every finding has fix: null and no fix-transform/metadata fix field, so the remediation create --ids path (which binds a finding to its own fix transform) has nothing to bind to. To auto-remediate, author your OWN transformation definition and run it against a repo with remediation create --transformation-name.
TD source-directory schema (from the atx binary's SEGFileManager / DEFAULT_SCHEMA):
my-td/
├── transformation_definition.md # REQUIRED — the instructions the agent follows
├── summaries.md # optional — short summary
└── document_references/ # optional — supporting docs
transformation_definition.md is a natural-language spec: purpose, scope (exact files to generate/modify), how to analyze the repo, per-artifact requirements, PR body contents, and guardrails (e.g. "additive only, never overwrite, no real secrets"). Keep the change set small and reviewable.
Save → publish → run:
# Draft (temporary, ~30-day expiry) — good for iterating
AWS_REGION=us-east-1 atx custom def save-draft -n my-td --sd ./my-td --description "..."
# Publish (permanent registry entry) — required for remediation to reference it
AWS_REGION=us-east-1 atx custom def publish -n my-td --sd ./my-td --description "..."
# Verify (registry has AWS-managed + user TDs; ours are user TDs)
AWS_REGION=us-east-1 atx custom def list # pretty table, count summary
AWS_REGION=us-east-1 atx custom def get -n my-td # NOTE: writes the TD files into CWD ('.')
# Run as a remediation → opens a PR on GitHub/GitLab/Bitbucket sources (local branch for local)
atx ct remediation create --repo <src>::<repo> --source <src> --transformation-name my-td --name "..."
atx ct remediation status --id <remediation-id> # includes PR/MR link
atx custom def * needs AWS_REGION set explicitly even when atx ct works fine — otherwise it errors "AWS Transform is not available in region 'us-west-2'". Prefix with AWS_REGION=us-east-1 (or export it). atx ct reads region from AWS config; atx custom def does not.
atx custom def get writes the fetched TD files into the current directory — run it from a scratch dir or clean up afterward.
Running your own TD: ct remediation direct-TD mode (verified working)
There is no built-in containerization transform. A containerization fix always runs a TD you author, and since ARA/MOD findings are assessment-only, that's the normal path — not a workaround.
Per the official docs, remediation has three modes, and the third is the one to know here:
Mode
Flags
Use
findings-based
--ids <finding-ids>
each finding uses its own fix transform — unavailable for ARA/MOD (fix: null)
TD override
--ids + --transformation-name
override the fix transform for those findings
direct TD
--transformation-name + --repo
"run a transformation against a repository without findings" — this is the containerize path
Direct-TD mode needs no findings at all, which is why it works on a fresh source. The documented command:
--local runs against the checked-out working tree and produces a local branch instead of a PR (output depends on the provider: GitHub PR / GitLab MR / Bitbucket PR / local branch). -g/--configuration is valid only alongside --transformation-name.
⚠️ Publish and remediate in the same auth mode, or the TD is invisible
The registry is tenanted by authentication mode. One endpoint
(transform-custom.<region>.api.aws), two separate namespaces. A TD published in one is a 404 in the
other — remediation create reports "Transformation <name> not found in the registry. Did you mean
…?" for a TD that custom def get confirms exists.
The CLI picks the mode from the environment (3.9.0, AuthenticationManager.isAWSMode):
MIDWAY
Mode
false
AWS credentials
true
the alternate identity mode
unset
depends on whether TOOLBOX_TOOL_VERSION is set in the environment
The ct remediation worker subprocess always runs in AWS-credentials mode — it logs
MIDWAY=false detected, using AWS credentials. So if atx custom def publish defaults the other
way, the publish "succeeds" into a namespace remediation never reads. custom def get also answers
with no AWS credentials at all — a tell that it wasn't using them, and the reason it can confirm
a TD remediation can't resolve.
Fix: pin AWS-credentials mode for both steps so the publish and the run agree.
export MIDWAY=false # publish into the namespace remediation reads
./scripts/publish-td.sh definitions/custom/containerize-service
cd "$(mktemp -d)" && MIDWAY=false atx custom def get -n containerize-service # now authoritative
Verified 2026-08-04 on 3.9.0: an identical remediation create failed pre-fix, and after
re-publishing with MIDWAY=false the same command ran to completed (~4 min, Console
Remediations tab) and committed branch atx/containerize-service-<timestamp> with all four
expected artifacts. Publishing to one namespace does not backfill the other — re-publish, don't
wait.
To confirm the split rather than guess, diff the two directly:
cd "$(mktemp -d)"
atx custom def get -n <your-td> # whatever your environment defaults to
MIDWAY=false atx custom def get -n <your-td> # AWS credentials — the one remediation uses
Don't trust the "Did you mean …?" suggestions either: that's a Levenshtein match over the names the
remediation side can see, so it lists neighbors from a namespace your custom def may not be
reading.
Ruled out by experiment before the auth mode was found — don't re-debug these: schema (a
byte-identical workshop SKILL.md failed too), identity, propagation delay (30+ min), publish
tooling (raw atx custom def publish behaves the same), account (three), and region.
Local fallback: custom def exec. No registry round-trip and no PR — it runs the TD against a
path and leaves the changes uncommitted. Also mode-sensitive: match the mode you published in.
Verified end-to-end 2026-08-04 — containerize-service against a Rails fixture with no Dockerfile
produced exactly the four artifacts it specifies, correctly detecting Ruby 1.8.7 as EOL, pinning
ruby:3.2-slim, running as non-root, and targeting port 3000. Application source untouched.
Trade-off: ct remediation gives you ct-managed branch/PR creation, --slots concurrency, and a record in the Console's Remediations tab; custom def exec gives you a reproducible local run that doesn't depend on remediation-side name resolution.
TD names are not stable identifiers. Names are resolved at run time, and registry state isn't version-controlled: custom def delete is permanent, drafts expire on their own, and the registry is shared (802 TDs across many owners on 2026-08-04). Verify the exact name in the account and region you'll run in. Prefer get -n <name> over eyeballing custom def list: the list wraps names across lines and grepping it invites substring false positives. get answers the only question that matters — does this exact name resolve.
Worked ct remediation example (3.9.0), for behavior rather than for the name.containerize-to-eks on an already-containerized Node fixture completed and committed as ATX Bot <checkpoint@atx.bot>: "Fix Dockerfile HEALTHCHECK to use reliable TCP connectivity check via Node.js net module instead of wget --spider which fails on HTTP 401…" — a 2-line Dockerfile edit. Two lessons: a containerization TD adapts to an already-containerized repo rather than regenerating scaffolding, and its change set can be far smaller than the finding implies (the MOD INF-Q1 gap was missing EKS IaC, which this did not address). Don't assume a remediation resolved the finding that motivated it — re-analyze and diff. Corollary for demos: point the containerize beat at a repo with no Dockerfile, or the honest result is "nothing to do."
--local runs the job on the ct server against the checked-out working tree instead of delegating to the provider — the right mode for local sources, where the result is a local branch rather than a PR. -g/--configuration is valid only alongside --transformation-name.
⚠️ Remediation refuses to run against a repo with a dirty worktree, and a repo being analyzed right now is dirty (ct rewrites its report bundle in place). Error: "Local repository at … has uncommitted changes. Commit or stash them before running remediation." Don't "fix" this by committing mid-analysis — you'd be racing the writer. Either wait for the analysis to finish, or test against an isolated copy of the fixture registered as its own source.
Full CLI reference (condensed)
Enumerated live on 3.9.0 (2026-08-03). atx ct --help reports 11 groups; schedule and server exist but are hidden.
Status:atx ct status [--health] [--json] — in-process; do not start serverSchema:atx ct schema — JSON manifest, but incomplete (omits hidden commands/flags)
Sources:atx ct source add|list|get|remove|update (get/update need --name)
Discovery:atx ct discovery scan --source <name> [--path <override>] [--json]Repositories:atx ct repository list|get|update|delete (filters: --source --language --labels --has-workflow --json --next-token)
Analysis:atx ct analysis run|get|list|cancel|delete — no list-artifacts, no get-artifact (removed; use get --json → report_paths)
Findings:atx ct findings list|count|get|update|batch-update|delete — no dismiss (removed; use update --status dismissed)
· list filters: --repo --source --severity --min-severity --type --status --analysis-id --fix-transform --next-token --json
· --severity and --min-severity are mutually exclusive; both take low|medium|high
· count --by severity|repo|analysis-type — server-side aggregation, prefer for large portfolios
Remediation:atx ct remediation create|list|status|retry|cancel|delete (create/retry accept hidden --wait)
Setup:atx ct setup <component> [--status] [--delete] (e.g. security-agent)
MCP:atx ct mcp [--transport stdio|http] [--port 3100] — 25 tools
Remote:atx ct remote analysis|remediation|status|detect|provision|update|credentials|teardown|cancel|network — see below
Schedule(hidden): atx ct schedule create|list|get|enable|disable|delete|teardown — see below
Stack names (literals): AtxInfrastructureStack[-suffix] (Batch), atx-runner[-suffix] (EC2, --stack-name must start with atx-runner), atx-scheduler, AtxDispatcherStack, AtxSecurityAgentStack-*. Buckets: atx-source-code-<acct>, atx-ct-output-<acct>, atx-custom-output-<acct>.
Mode-specific flags — mixing them is a hard error:
Batch only:--securityGroup (required), --job-timeout <60..604800> (default 43200 = 12h), --resume, --batch <id> on status. remote status needs --stack-name for Batch.
EC2 only:--workers 1-5, --instance-type, --volume-size, --existing-instance <id>, --group <id> and --wait on status. --securityGroup optional (auto-creates a no-inbound SG, SSM access).
Network prerequisites (the console modal states these; all are load-bearing): private subnets only, no public IP, egress reaching the internet via NAT gateway or VPC endpoints, and enableDnsSupport + enableDnsHostnames on the VPC.
⚠️ Local sources on remote compute — use the CLI, not the console
Remote jobs run in AWS, so a --provider local source pointing at a path on your laptop is unreachable from the job. The source directory must be zipped and uploaded to S3 first. Launching a local-provider analysis from the web console fails with:
Could not start analysis: local provider requires localBundleName and
sourceBucketName for repo <source>::<repo>
The console has no upload step, so it cannot satisfy this — the CLI can.atx ct remote analysis bundles the source directory (it embeds archiver and multipart-uploads to S3) and passes localBundleName + sourceBucketName through to the job for you:
The bundle is a point-in-time snapshot. Re-run remote analysis after editing fixtures, or the job analyzes stale code. Same caveat applies to schedule create --provider local, where the zip is uploaded once at create time and every subsequent fire analyzes that same snapshot.
Default bucket is atx-source-code-<accountId>-<stackSuffix>, created by the stack (resolve the real name with atx ct remote detect --mode batch). --source-bucket overrides it and is required for local sources on the EC2 --existing-instance path.
For SCM providers (github/gitlab/bitbucket) none of this applies — the job clones directly using staged credentials, and the console path works fine.
Prefer an SCM source for anything recurring; local + remote is best for one-off runs over fixtures.
⚠️ network discover does not validate egress. It happily reports subnets whose 0.0.0.0/0 route points at a deleted NAT gateway (route state blackhole). The stack then deploys CREATE_COMPLETE and every job hangs with no egress. Verified 2026-08-03 on this account. Always check before provisioning:
aws ec2 describe-route-tables --filters "Name=association.subnet-id,Values=<subnet>" \
--query 'RouteTables[].Routes[].[DestinationCidrBlock,NatGatewayId,State]' --output text
# the 0.0.0.0/0 row MUST say "active", not "blackhole"
When in doubt, atx ct remote network create builds a known-good VPC (~$32/mo for the NAT, billed idle). It also creates a public subnet to host the NAT — do not pass that one to provision; use privateSubnetIds only.
All mutating remote operations are gated behind --execute; --ack suppresses the admin-permissions prompt (prefer answering the prompt).
⚠️ Breaking change in 3.8.0+ — verified on 3.9.0: --expression is GONE and --recurrence is now (required). It takes daily | weekly:<MONDAY..SUNDAY> | monthly:<1..28>; raw cron(...)/at(...) expressions are no longer accepted. Any 3.7.0-era script passing --expression will fail. Note the fire time is not configurable to the minute: per its own help, it "resolves to about 2 minutes from now (local wall clock, DST-stable)" — so a schedule named nightly-* won't necessarily run at night.
remote provision also gained --existing-scheduler-role-arn and --existing-instance-profile-name for reusing pre-created IAM.
There is no schedule setup. Scheduler infra is provisioned by remote provision, which ensures the atx-scheduler stack alongside the compute stack (--skip-scheduler opts out). Only schedule teardown remains.
--provider defaults to github; with local it zips and uploads the source directory at create time, so every fire analyzes that uploaded snapshot, not your current working tree.
Job flags mirror remote analysis/remote remediation. Note -g here is gated on --transformation-name, not --type custom.
Exact flags and examples for the core commands are in references/ct-workflow.md — but see the staleness warning above.
