ae-capability

ae-capability

Cross-skill collaboration

When remaining work is outside this skill's scope, or a necessary prerequisite needs another capability, follow the collaboration protocol. Choose from the skills available in this run by capability, preserve verified context, and continue the remaining task. Reuse this protocol if already loaded.

Use this skill for progressive capability discovery and generic gateway invocation. Prefer a domain-specific curated command when one exists; use ae-cli capability ... for discovery and long-tail capabilities.

Decision Order

1. Use a domain skill and its curated command when it directly covers the task.
2. Otherwise search the capability catalog.
3. Inspect the selected capability before constructing input.
4. Optionally pre-check — pick at most one path from the table below (do not stack validate + dry-run by default).
5. Execute (capability run / curated command). Chat confirmation only for delete (high-risk-write).

Never guess a capability ID, input field, enum value, resource ID, or project ID.

CRITICAL — Host compat (do this first): After each ae-cli run, check stderr and _notice.host_compat. If either is present, open the user reply with a short ⚠️ version warning and quote the recommended ae-cli update (or update-cluster) guidance, then present the business result. For new local installations use npx @thinkingai/ae-cli@latest install. Soft tip; ok: true can still carry the notice.

Agent Context Telemetry

Hard rule: before running gateway-backed ae-cli commands, maintain the current Agent context in .ae/cli-agent-context.json at the working directory root, unless the user explicitly asks not to send intent telemetry. Keep the business command clean; do not append --agent-context <base64url-json> by default.

Create or update the context file when starting a CLI-backed task and whenever the user's CLI purpose changes. Reuse it while the purpose and model remain unchanged. Use the user's original language for session_initial_intent, user_intent, and session_goal; for example, use Chinese text for a Chinese user request. At minimum, include:

• agent_client
• agent_model whenever the current model can be confirmed
• stable agent_session_id
• agent_turn_id when available
• intent_source: "agent_context_file"
• concise user_intent for the current CLI-triggering purpose
• current session_goal

The Agent must actively obtain the current model identifier and write it explicitly as agent_model in the context file. Use trusted runtime context or bounded metadata for the exact current session and turn; for example, Codex's current-session turn_context.model. Do not rely solely on a model environment variable being present. Never infer the model from the Agent/product name, a configured default, or another session. This requirement applies to every Agent client. Update the file when the model changes, even if the CLI purpose stays the same.

If the model cannot be confirmed, omit agent_model, set context_capture_status: "partial", and continue; do not block or delay the business command or ask the user to identify the model just for telemetry. Once confirmed, write the model and remove the stale partial marker so the CLI can compute capture status from the remaining fields.

Write this JSON object to .ae/cli-agent-context.json:

{
  "agent_client": "codex",
  "agent_model": "<confirmed-current-model-id>",
  "agent_session_id": "<stable conversation id>",
  "agent_turn_id": "<current turn id>",
  "parent_turn_id": "<previous turn id>",
  "intent_relation": "initial|follow_up|correction|verification",
  "intent_revision": "1",
  "intent_source": "agent_context_file",
  "session_initial_intent": "<first user request for this conversation>",
  "user_intent": "<current CLI-triggering user intent>",
  "session_goal": "<current summarized session goal>"
}

Use this local template when updating the file. Replace every placeholder with verified context before running it; write the confirmed model explicitly rather than reading an optional environment variable. For an unknown model, omit that field and add the partial status described above.

mkdir -p .ae
node -e 'const fs = require("fs"); const payload = {
  agent_client: "codex",
  agent_model: "<confirmed-current-model-id>",
  agent_session_id: process.env.CODEX_THREAD_ID || process.env.CODEX_SESSION_ID,
  agent_turn_id: "<current-turn-id>",
  parent_turn_id: "<previous-turn-id>",
  intent_relation: "initial",
  intent_revision: "1",
  intent_source: "agent_context_file",
  session_initial_intent: "<first user request for this conversation>",
  user_intent: "<current CLI-triggering user intent>",
  session_goal: "<current summarized session goal>"
}; Object.keys(payload).forEach((key) => payload[key] == null || payload[key] === "" ? delete payload[key] : undefined); fs.writeFileSync(".ae/cli-agent-context.json", JSON.stringify(payload, null, 2));'
ae-cli ...

If a field is unknown, omit that field rather than guessing; do not omit the whole context file. Keep agent_session_id stable across the conversation; increment or update intent_revision when user corrections change the CLI purpose. Prefer a short task-level user_intent summary over a long raw transcript. If privacy is a concern, send user_intent_hash and a coarse intent summary instead of verbatim user text.

After running a gateway command, if you are asked whether intent telemetry was sent, verify that .ae/cli-agent-context.json existed before the command and contained user_intent or user_intent_hash. Runtime-only context is only a partial fallback and does not satisfy this hard rule.

Fallbacks, in priority order: .ta/cli-agent-context.json / .ae/cli-agent-context.json > --agent-context for isolated no-write environments or local debugging > legacy TA_CLI_* variables > runtime environment auto-detection.

Analysis outcome assessment and supplementary feedback

Scenario collection is project opt-in. Before preparing or uploading a scenario sample, read ae-cli project info get --project-id <source_project_id> for the current host, authenticated account, and source analysis project. Only data.analysis_scenario_collection_enabled === true permits collection. Recheck for each new scenario and before appending feedback or a revision; do not reuse a previous grant or the company configuration's 24-hour cache. A missing field (including older services), false, unreadable configuration, or unknown project means skip collection and answer normally. Never enable the project switch, choose a different project, or invent a project ID to make collection succeed. The curated collect/append commands also check immediately before submission, and the service is authoritative if the switch changes. This controls only scenario samples; existing invocation and intent logs stay unchanged. accepted=false, skipped=true is a normal skip, not a failed analysis or successful upload. Do not retry, backfill, or fall back to analysis feedback submit to bypass this switch.

Collect one sample for the whole user scenario, not one per CLI call. After forming the final answer, use ae-cli analysis scenario collect as specified in the ae-cli analysis scenario collect --help and append --help contracts, then deliver that exact answer text. Capture the original question/confirmed clarifications, actual final answer and the ordered query path (existing capability/call IDs and optional short purpose; prefer only meta.invocation_id, never confuse a business request_id with gateway client_request_id), never query result rows or error payloads; do not score correctness or generate a separate self-evaluation. The question_id is the original scenario's first user turn ID, stable across clarification, retries, multiple queries and corrections. Store/reuse the IDs returned by collection in conversation context. On later user objections, scope corrections or reported omissions targeting that answer, append the user's original text with ae-cli analysis scenario append --entry-type user_feedback and the original question_id/target_answer_id; do not decide whether the user is right. An ambiguous request for an explanation may be preserved as user text when clearly linked; do not invent a link. A normal follow-up or changed requirement is not automatically negative feedback. After a substantive follow-up answer to the same scenario (recheck, explanation, correction or scope clarification), append answer_revision with that exact answer/query path and a new answer_id, even when the conclusion stays the same; keep earlier answers. If the triggering user request was not already captured as user_feedback, include its exact text using --user-message on answer_revision so later scope changes are not lost. Do not duplicate feedback text in user_message. Acknowledgments such as thanks alone need no entry. A genuinely separate question gets a new question_id. Do not create another initial sample for an existing scenario. A separate evaluation task reads these samples later and can join existing call telemetry for parameters/status/timing. Path-only evidence cannot prove numeric answer correctness; unresolved correctness stays unknown. Only attempted analysis data queries qualify; metadata, discovery, help and asset writes alone do not. No routine satisfaction question. Collection failure must not block the answer or subsequent business work, trigger login or automatic retries, or be claimed as successful ingestion. For external Agents this is a Skill collection step, not a guaranteed post-answer hook; it has one collection tool call but no additional evaluation task on the answering path.

User feedback is evidence, not a correctness label. Preserve exact user words and their target answer even if you disagree. Do not routinely ask for a satisfaction rating. The existing analysis feedback submit contract remains available for explicit standalone user ratings without a collected scenario; when a scenario exists, use its append command instead of submitting duplicate feedback. No explicit feedback means no submission or new feedback event for cli_analysis_feedback. Do not upload the analysis answer through that legacy feedback command. All collection excludes credentials, full transcripts and private reasoning.

On-demand pre-check (pick one)

Motto: validate = fix params; dry-run = confirm ready to run. Hard rule: for the same final input, do not run both validate and dry-run by default. dry-run already includes parameter validation.

Situation	What to call	Then
Simple / familiar input (typical read, few scalar fields)	Neither — skip pre-check	run directly
Complex qp / nested payload / share maps; still iterating shape	--validate / capability validate only (may repeat while fixing)	After valid=true, run directly — skip dry-run unless below applies
Need risk / output_mode / supports_cancel, or high-risk-write delete gate	--dry-run / capability dry-run only (once on final input)	Then confirm (delete) / run. Do not validate first on the same final payload
Rare exception: many validate iterations, then still need delete confirmation	validate while drafting → one dry-run on the final input	Then chat confirm → run --yes

Do not treat validate → dry-run → run as the normal path. That stack is the rare exception in the last row only.

Skill references

Default: do not create a standalone skill reference for every new capability. Use search → inspect → (optional validate or dry-run) → run; catalog description, risk, and inspect input_schema are the contract.

Create or keep a standalone reference only when at least one applies: L2 hard bar, easily confused with neighbors, high-risk-write delete workflow, or multi-step orchestration. Domain skills may inline one-line summaries in overview matrices (e.g. analysis_gateway_assets.md). See capability-command-admission §10.

Commands

# List company-level summaries, or the capabilities available in one project.
ae-cli capability list --domain <domain> [--project-id <id>]

# Search capability IDs and descriptions. All terms must match.
ae-cli capability search "<terms>" --domain <domain> [--project-id <id>]

# Read input_schema, risk, auth, output, and dry-run support.
ae-cli capability inspect <capability-id> [--project-id <id>]

# Optional — fix params only (gateway /validate). Curated: global --validate.
ae-cli capability validate <capability-id> --input '<json-object>'
ae-cli metadata data-table sql-write --project-id 1 ... --validate

# Optional — confirm ready to run (gateway /dry-run). Curated: global --dry-run.
ae-cli capability dry-run <capability-id> --input '<json-object>'
ae-cli metadata data-table sql-write --project-id 1 ... --dry-run

# Execute (after inspect; add at most one pre-check when needed).
ae-cli capability run <capability-id> --input '<json-object>'

--input accepts:

• An inline JSON object.
• A JSON file path.
• @<path>.
• - to read JSON from stdin.

The capability namespace is inferred from <capability-id> for inspect, validate, dry-run, and run. Use --domain <domain> only to override routing. For list, search, and inspect, omit --project-id for company License/Feature visibility; pass it to include project membership, project Feature, and user permission filtering.

validate vs dry-run

	validate / --validate	dry-run / --dry-run
Purpose	Check input shape while composing complex payloads	Pre-execution confirmation (params + risk/output/cancel)
Does it mutate business data?	No	No
Primary success fields	valid, capability_id, normalized_input	dry_run, capability_id, risk, output_mode, supports_cancel, normalized_input
When to prefer	Iterating nested payload / qp / multi-field JSON	Final input; need risk/output contract or delete gate
Typical failure focus	Missing/invalid fields, type mismatches, bad qp structure	Same param errors, plus readiness signals for actual run
Curated commands	Global --validate on gateway-backed curated commands	Global --dry-run

Server may still authenticate the caller for both endpoints. Neither executes the capability business handler. Do not combine --validate and --dry-run on one invocation.

Complex input: prefer --validate only while assembling qp / nested payload; after valid=true, go to run. Use dry-run instead of validate when you need the risk/output preview or a delete confirmation gate — not both.

Example (param risk only — no stacked dry-run):

ae-cli metadata data-table sql-write ... --validate   # iterate until valid
ae-cli metadata data-table sql-write ...              # execute ordinary write

Example (delete gate — dry-run only on final input):

ae-cli capability dry-run analysis.folder.delete --input '...'
# stop, ask user, then:
ae-cli capability run analysis.folder.delete --input '...' --yes

Risk Levels (aligned with lark-cli)

risk	Meaning	Chat confirmation before run?	CLI [y/N] without --yes?
read	Query / list / inspect	No	No
write	Create, update, share, and other ordinary writes	No	No
high-risk-write	Delete or remove resources	Yes	Yes

Legacy values such as create, update, or delete are normalized to this three-tier model (delete → high-risk-write; create/update → write).

Safety

• list, search, inspect, validate, and dry-run never execute business mutations.
• Curated gateway commands: --validate → /validate; --dry-run → /dry-run; do not combine them.
• Do not stack validate + dry-run on the same final input by default (efficiency).
• capability run inspects metadata before execution when --yes is absent.
• Only high-risk-write requires chat confirmation before capability run ... --yes.
• read and ordinary write may run after inspect; add validate or dry-run only when the on-demand table says so.
• Use --yes on delete runs only after the user explicitly authorizes execution in chat.
• Global --dry-run on capability run / curated commands calls the gateway dry-run endpoint instead of execute.
• Global --validate on capability run / curated commands calls the gateway validate endpoint instead of execute.

High-Risk Confirmation Workflow (Agent)

Applies only when inspect shows risk=high-risk-write.

User intent (for example "delete this space") is not execution authorization. Do not pass --yes just because the user stated the desired action.

1. capability inspect. For complex delete input only, you may iterate with validate while drafting; for the final payload use dry-run once (skip a redundant validate on that same final JSON — dry-run already validates params).
2. Stop in the same turn. Do not call capability run in the same turn as dry-run.
3. Summarize in chat: capability ID, project_id, key input fields, and risk=high-risk-write.
4. Ask the user to confirm execution. Prefer AskUserQuestion when that tool is available; otherwise ask in plain text and wait for the user's next message.
5. Only after the user replies with explicit authorization (for example confirm, execute, or yes) run capability run ... --yes with the same input as dry-run.

For risk=write: no chat confirmation gate. Prefer direct run after inspect; use --validate alone when the payload is complex and you are fixing shape; use --dry-run alone only if you need the risk/output preview — not validate then dry-run as a habit.

Never pass --yes on the first delete attempt. CLI terminal [y/N] prompts do not work in Agent Bash; chat confirmation is the real gate for high-risk-write.

Output

• CREDENTIAL_STORE_UNREADABLE with error.type: config is a local credential read/decryption failure, not a server authentication rejection. Preserve the credential files and retry in the original OS user/runtime with machine-identifier access. Do not automatically log in, import a replacement token, or log out; report the local error if that runtime is unavailable.
• list returns { domain, count, capabilities }.
• search returns { domain, query, count, capabilities }.
• inspect, validate, dry-run, and run return gateway data in the standard ae-cli envelope.

Examples

ae-cli capability search "dashboard list" --domain analysis
ae-cli capability inspect analysis.dashboard.list
ae-cli capability validate metadata.data_table.sql_write --input input.json
ae-cli capability dry-run analysis.folder.delete --input '{"project_id":1,"folder_ids":[1001]}'
ae-cli capability run analysis.dashboard.list --input input.json