Automatic classification and routing
These three skills are the building blocks of the M2 orchestration model (hybrid adaptive): they receive a free-text request, decide which execution plan it belongs to, and build the input that plan needs — without human intervention, unless confidence is low or information is missing, in which case they escalate to a person instead of guessing.
lippia-neo-classify-task
When to use it
The M2 dispatcher uses it to classify a free-text task against a list of available execution plans and choose the best fit, returning a confidence score. If the confidence is below 0.7, the engine escalates to a human instead of running the plan automatically.
{
"input": {
"task": "Verify that login fails with invalid credentials",
"available_plans": [
{ "name": "feature-new-end-to-end", "description": "Full E2E flow for a new feature, starting from a story" },
{ "name": "qa-bug-coverage", "description": "Coverage for a reported bug: reproduce it and generate the test that covers it" }
]
}
}
| Input | Type | Description |
|---|---|---|
input.task |
Required | free-text description of the task. |
input.available_plans |
Required | list of available execution plans; the dispatcher builds each one with its name and its description (the skill compares against the description, by intent and not by keywords). |
Output
| Field | Description |
|---|---|
plan |
exact name of the chosen plan, copied literally from available_plans (never made up). |
confidence |
number between 0.0 and 1.0; >= 0.7 runs automatically, < 0.7 escalates to a human. If two plans are equally plausible it chooses the more specific one with ~0.6–0.69; if none fits, it returns the least bad one with 0.0. |
reason |
one sentence explaining why that plan was chosen. |
Chains with
Comes before lippia-neo-build-plan-input, which builds the input for the chosen plan; the M2 dispatcher then
runs the plan or escalates based on confidence.
lippia-neo-build-plan-input
When to use it
Right after lippia-neo-classify-task has chosen a plan and before running it — it takes the free
text of the original task and that plan's list of required_keys, and builds the input JSON
by extracting each value from the text. If the text isn't enough to derive some required piece of data
(it never makes anything up, especially credentials), it reports it as missing so the
dispatcher escalates to a human instead of running the plan with incomplete information.
{
"input": {
"task": "Run the bug reproduction plan for ticket BUG-123",
"plan_required_keys": [
{ "name": "bug_report", "description": "Referencia al bug: path local, URL de ticket o texto crudo" },
{ "name": "identifier" },
{ "name": "credentials" }
]
}
}
| Input | Type | Description |
|---|---|---|
input.task |
Required | the same free text of the original task that lippia-neo-classify-task received. |
input.plan_required_keys |
Required | the required_keys of the already-chosen plan (list of {name, description?}, with name without the input. prefix; description is only present if the plan declared it), always resolved by the dispatcher — this skill never reads the plan file. If it's empty, it returns plan_input: {} and complete: true. |
Output
| Field | Description |
|---|---|
complete |
true only if ALL the plan's required data could be derived from the text; a single missing item sets it to false. |
plan_input |
JSON object ready to run the plan, with exactly the derived fields (raw text, such as the body of a report, is copied as-is, without summarizing). |
missing_keys |
required data that the text didn't allow to be derived honestly; empty if complete: true. |
warnings |
ambiguities resolved with a reasonable assumption (e.g. two ticket IDs mentioned, the first one was used). |
Chains with
Runs right after lippia-neo-classify-task; if complete: false, the dispatcher escalates to a
human (same mechanism as a low confidence in lippia-neo-classify-task).
lippia-neo-classify-ui-scope
When to use it
As step 0 (gate) of the frontend-ui-change-pr plan (classifier-driven UI regression,
M2). It classifies a front-end PR (diff + description) as a
cosmetic/isolated change (ui-smoke) or a broad/risky one (ui-full-regression), to route execution to the smoke branch or the
full regression branch, leaving the "blast radius" decision documented and auditable.
{
"input": {
"pr_diff_path": "prs/PR-482.diff",
"pr_description_path": "prs/PR-482.md"
}
}
| Input | Type | Description |
|---|---|---|
input.pr_diff_path |
Required | path to the PR's diff (.diff/.patch). |
input.pr_description_path |
Required | path to the PR/ticket description (.md). |
input.pr_diff_content |
Optional | diff pasted directly into the conversation (fallback if there's no pr_diff_path). |
input.pr_id |
Optional | PR ID, for logging/auditing only. |
Output
| Field | Description |
|---|---|
scope |
ui-smoke | ui-full-regression: the classification label. |
is_smoke |
true only if scope == ui-smoke; triggers the plan's smoke branch. |
is_full |
true only if scope == ui-full-regression; triggers the full regression branch (is_smoke and is_full are mutually exclusive). |
confidence |
0.0–1.0, recorded in the audit trail; when uncertain, the skill must lean toward ui-full-regression (fail-safe), not just lower this value. |
reason |
concrete evidence (file, line, or ticket text) that drove the decision. |
Chains with
Step 0 of the frontend-ui-change-pr plan; if is_smoke → lippia-neo-run-and-heal (quick
pass, max_iterations: 2); if is_full → lippia-neo-read-user-story → lippia-neo-run-manual-testing →
lippia-neo-run-and-heal → lippia-neo-generate-test-suite (flujos nuevos) → lippia-neo-run-and-heal →
lippia-neo-write-test-report (with a selector audit trail).