Rote Task Routing

rote-task-routing

All rote-<name> references in this document — including every name in the Handoff Contract — are companion skills, never CLI commands (rote-shell is not rote shell). Invoke them through the runtime's skill mechanism; only literal rote … commands run in a terminal.

Use this skill after the root skill's local-then-registry Play-search gate finds no runnable full match, or after rote-flow-run returns a partial baseline that must be combined with additional work.

Do not use this skill after rote-flow-run verifies a full match. A verified full-play result has no uncovered work, so adapter exploration, catalog search, workspace setup, and fallback routing are out of scope unless the user adds a new requirement.

This is a routing skill, not an execution or completion skill. It selects the next owner and returns the handoff packet; lifecycle completion still belongs to rote, rote-workspace, rote-shell, rote-flow-crystallization, or rote-flow-authoring.

Route once before workspace work begins, then keep the selected execution path stable. Do not begin adapter calls and then hand off midstream unless the selected skill explicitly asks for that return.

Capability Match Rule

Select routes by capability, not provider label. Before choosing an adapter, name the operation the user needs: metadata/search, object details, file content, recent trades, live volume, state change, browser session, or local process evidence.

• Before installing or calling anything, know for each required source the exact operation, the candidate adapter id, how the result will be verified, and the fallback boundary.
• If an installed adapter covers the provider but not the operation, keep searching.
• If an installed adapter covers the provider and operation but has a broken base URL, stale auth, or wrong config, repair it before replacing it.
• If a partial play covers a baseline, route only the missing capability.
• If the catalog has several plausible provider adapters, run the Catalog Ambiguity Gate before installing or calling.
• Do not report provider metadata as live activity, event details as trades, or CLI scrollback as durable process evidence.

One provider commonly ships several adapters split by capability — for example entity/metadata search in one, historical activity or trade/volume history in another, and live orderbook/price/state in a third. Match the user's operation to the adapter that actually exposes it, and repair an installed same-provider adapter (base URL, auth) before installing an adjacent one.

Catalog Ambiguity Gate

When catalog hits include multiple plausible same-provider adapters, or their auth shapes differ, pause before install/call and ask the user to choose unless:

• the user named an exact adapter id;
• one candidate is already installed, healthy, and clearly covers every required capability;
• inspected metadata proves the others do not cover the needed operation.

Compare only decision fields: adapter id, kind/substrate, auth shape, spec-fetch auth, runtime auth, capability fit, installed/health state, install/write impact, and next command.

The same gate applies when two or more installed adapters cover the required operation — e.g. a REST adapter and an MCP adapter for the same provider. Do not silently take the first ranked hit: ask the user to choose using the same decision fields, unless one candidate is unhealthy, the user already picked one for this task, or inspected metadata proves only one exposes the needed operation. Remember the choice for the rest of the task.

After rote adapter catalog info, hand off to rote-adapter-create for the install itself — it owns the create pipeline. The routing-relevant rule: follow the catalog-provided install command; if Spec Type = "MCP", the route is rote adapter new-from-mcp. Do not substitute generic rote adapter new.

A 401/403 while fetching an MCP/spec means spec-fetch auth is required; it is not proof the adapter is bad. Keep it in the comparison instead of silently falling through to another adapter.

Gate 1: Explore Before Adapter Work

Run:

rote explore "<intent>"

Read the full response before acting:

• If @@plays names a usable existing play, hand off to rote-flow-run.
• If the uncovered work is local CLI, files, logs, build/test/release checks, generated artifacts, or process lifecycle, hand off to rote-shell.
• If the response identifies one installed adapter, choose the single-adapter branch.
• If the response requires multiple adapters or orchestration, keep the work in the main agent and use rote-workspace.
• If explore says no tools or no installed adapter fits, run an installed-adapter inventory before catalog install:

rote adapter list
rote adapter info <candidate-id>

Use this inventory to find provider-equivalent adapters that need repair. Search the catalog only after no installed adapter can satisfy the capability or after repair is not possible.

• Do not route provider/API data collection to rote-shell just because a public HTTP endpoint can be called from a local script.

Gate 2: Subagent Before Workspace

If an installed rote-<adapter-id> subagent exists and the task is scoped to that adapter, dispatch it before any workspace calls start.

The handoff packet should include the user's goal, play-search result, explore result, any baseline artifact from rote-flow-run, and whether reusable output must pass through the save gate.

If the incoming artifact came from a full-play match, return to rote instead of routing. Only partial-play baselines can be carried forward.

For a partial-play baseline, preserve it as reusable source material for a composed superplay: record the baseline play name, parameters, output artifact, provenance, and the exact uncovered requirements. The next execution owner should augment that baseline, not recreate a similar report from scratch.

Do not spawn an adapter subagent after workspace calls have started. Mid-workflow handoff risks losing cached @N responses and session context.

Gate 2b: Shell/Process Work

Use rote-shell when the selected substrate is local commands, files, logs, generated artifacts, dependency checks, or process lifecycle. Do not search the adapter catalog just because no adapter covers a local CLI task; rote proc is the rote-native substrate for that work.

Build and dev-loop tooling is the exception, whatever the language toolchain: compilers, test runners, formatters, linters, type checkers, package managers, and task runners run as plain shell by default. Their value is the immediate pass/fail signal, and wrapping every iteration in rote proc adds latency and noise without adding reusable evidence. If a specific build or test run should become durable workspace evidence — it feeds a play, a handoff, or a result the user asked to keep — ask the user first with a one-line explanation: rote proc makes the output queryable and replayable later. A preference the user or project instructions already state settles the question without re-asking.

Provider/API access is not shell work. If the task needs provider records, event data, trades, tickets, database rows, or other API objects, return to adapter routing and catalog search before using curl, Python, Node, a provider SDK, or direct MCP tools. Shell routing may transform or combine adapter/browser evidence after it has been captured through rote.

The handoff packet should include the working directory, input evidence, allowed local tools, any baseline artifact that must be preserved, and whether the user already asked to save or release the result. Process-only reusable work uses rote-shell and no-shape-flag workspace export for the default steps + presentation artifact; template/pending commands still require a real adapter, so never invent one.

Do not use shell routing as an excuse for raw, untracked parsing of rote responses. If the local command consumes adapter/browser/play evidence, route it through rote-shell and rote proc with declared inputs/outputs so the transformation remains queryable and reusable.

Gate 3: Single Adapter Without Subagent

If one installed adapter fits and no generated adapter skill exists, hand off to rote-workspace. The workspace owns probes, calls, cached response queries, transformations, reusable-result classification, and final workspace state.

Use rote-using-adapters when the runtime already selected a bundled single-adapter execution skill for this exact branch and the handoff packet has enough state for delegated execution. Otherwise, prefer rote-workspace so single-adapter and multi-adapter work share the same completion lifecycle.

Gate 4: Multi-Adapter Orchestration

Keep multi-adapter work in the main agent. Use one workspace and sequence adapter probes, calls, queries, and transformations there so cached responses remain addressable.

For multi-source requests, classify each source before work starts:

• Must use rote: the user explicitly says to use rote/adapters for that source, or the selected reusable workflow depends on that source.
• Direct one-off: only after play search, explore, and catalog search do not produce a useful rote path, and the source is simple, public, unrelated to the reusable workflow, and not worth crystallizing.
• Needs catalog search: no installed adapter fits, but the source is part of the repeatable workflow.

Never bypass a source explicitly required through rote. Do not force unrelated one-off public data through adapter installation unless it is part of the reusable workflow or the user asked for rote coverage.

Gate 5: Catalog Before Fallback

When no installed adapter or play covers the task, run:

rote adapter catalog search "<intent>"

Use the user's domain words and operation words in the query. For a provider plus several data needs, include both the provider and operations. If the first query returns partial matches, search again with the missing capability before falling back.

If the catalog returns one unambiguous useful hit, inspect or install through rote with rote adapter catalog info <id> or rote adapter new <id> --yes. If several hits plausibly fit, run the Catalog Ambiguity Gate first. Treat the chosen hit as the next execution path: probe it and call it through rote before direct MCP, WebFetch, curl, or custom scripts.

If you use rote adapter new <id> ... --dry-run, treat the result as a preflight only. The adapter is not installed, its probe/call commands do not exist yet, and the task cannot be complete. Hand off to rote-adapter-create for the real install (it follows the catalog-provided command, including the MCP route), verify with rote adapter info <id> or rote adapter list, then run <id>_probe and <id>_call for the required capability before answering.

Do not use catalog install to dodge a repairable installed adapter. A newly installed adjacent adapter is a regression when the task expects the installed adapter's config to be fixed.

Only use direct tools after play search, exploration, and catalog search fail to produce a rote path, or after a catalog-installed adapter has been probed and cannot satisfy the request. Tell the user what rote checked and why fallback is necessary.

Return Fields

Return these fields to rote or the selected owner:

• Selected route: play, adapter subagent, rote-workspace, rote-using-adapters, rote-shell, catalog install, or fallback.
• Explore result: relevant @@status, @@next, @@plays, installed adapter, or blocker.
• Catalog result: query used, hits inspected, installed adapter, or reason no hit fits.
• Baseline preservation: play output/artifact that must remain intact.
• Requirements: required sources/adapters, missing capabilities, live observations, artifact, and verification checks.
• Workspace or shell preconditions: workspace name suggestion, working directory, adapter ids, local tools, credentials, and approvals.
• Next recommended skill: rote-flow-run, rote-workspace, rote-shell, rote-using-adapters, rote-adapter-create, or fallback to rote.

Handoff Contract

• Use when: local and registry Play search did not fully cover the request, or a matched Play produced only a baseline for additional rote work.
• Preconditions: rote or rote-flow-run has provided the user intent, both applicable provider outcomes, and any baseline output that must be preserved.
• Owns: explore-first routing, shell/process routing, subagent-before-workspace decisions, single-adapter versus multi-adapter selection, catalog search before fallback, and explicit route return fields. Does not own adapter execution, shell execution, play authoring, or final presentation.
• Hands off to: rote-flow-run for @@plays; rote-shell for local CLI/files/logs/process work; rote-workspace for adapter workspace execution; rote-using-adapters for delegated single-adapter work; rote-adapter-create when catalog/spec work is needed; rote-troubleshooting when route selection repeatedly fails unchanged.
• Returns to: rote with the selected route, why it was selected, what rote checked, and which skill owns execution next.
• Stop when: a route owner is selected, no safe rote route exists, or the user must approve an adapter install, credential setup, or out-of-band fallback.
• Completion signal: route selected with handoff packet, or fallback/blocker explained with the rote checks already performed.