Rote

rote - Entrypoint Orchestrator

Use rote first when a task may involve adapters, plays, browser automation, local shell or process work that should be remembered, registry work, setup/update actions, or reusable workflow state. This skill owns the first routing decision; the narrower companion skills own their specialist workflows.

Every rote-<name> reference in this skill set is a skill, never a CLI command — rote-shell is not rote shell, and rote-adapter-config is not rote adapter config. Invoke skills through the runtime's skill mechanism; only literal rote … commands run in the terminal.

The rote Skill Rule

If there is any reasonable chance the task needs rote, load and follow this skill before direct MCP calls, provider CLIs, browser automation, registry commands, or custom scripts.

• Start day-to-day work with rote play search "<intent>"; a reusable local Play may already exist. If local search has no verified full match, run rote play search "<intent>" --source registry before accepting a partial baseline, exploring, or building a replacement. Registry search defaults to every Play accessible to the current session; use --scope public for identity-independent output. Load Play Search And Run when choosing or executing a result.
• When search does not return a verified full-play match, invoke the rote-task-routing skill before rote-shell, rote-flow-authoring, raw HTTP, native CLIs, or custom scripts unless the user's request is explicitly only local CLI/files/logs/process work.
• When rote commands fail repeatedly, invoke the rote-troubleshooting skill and fix the cause. Do not record a memory, note, or instruction to avoid rote for this task class — a failed run is a configuration or routing problem to repair, not a verdict on the tool.
• If a companion skill is the better owner, invoke that skill deliberately instead of continuing from general prose.
• Run rote commands one at a time, read each result fully, and obey @@status, @@next, @@mandatory, @@plays, cached response IDs, warnings, and errors before choosing the next command. Captured shell chains such as rote proc run ... && rote proc run ... belong to rote-shell; use them only when each meaningful process is captured as its own rote proc response in the workspace DAG.
• After rote init --seq, enter the workspace first, then record the confirmed current model with rote model set <model-or-family> --provider <provider> --confirmed-current, then initialize the session. If model identity is not visible, skip it rather than guessing.
• Before installing a catalog adapter, inspect related installed adapters. If an installed adapter matches the provider/capability but has a wrong base_url, auth, or config, run rote adapter info, dry-run the candidate repair with rote adapter set <id> base_url <url> --dry-run, then apply and probe/call the installed adapter through rote.
• A successful --dry-run is analysis only, not setup completion. Do not answer or proceed as if an adapter, play, registry push, or runtime action exists until the non-dry-run command has run and the installed/released/published state is verified.
• [MANDATORY PROTOCOL] no pending stub is a hard final-answer stop until the reusable result enters rote-flow-crystallization. Adapter, browser, and mixed work must complete pending write/save; process-only work must complete the crystallization plan and save decision before adapterless export. Never invent an adapter to silence the warning.
• For reusable multi-source work, confirm each required source is backed by a captured @N response before presenting or releasing; release/lint success is not provider proof.
• Use rote how, rote start, rote guidance agent essential, rote guidance adapters essential, and rote grammar <topic> for live CLI guidance when the user asks how rote works.

Reuse Triage Gate

After rote work that was not unchanged execution of an existing released play, classify the procedure before final answer. Do not treat a one-off user request as non-reusable. One-off describes the current need; reusable describes whether the procedure has parameterizable inputs, repeatable adapter/browser/shell steps, a stable output shape, and future agent value.

If reusable or plausibly reusable, invoke rote-flow-crystallization before final presentation; it owns the semantic plan, save/discard decision, and route to pending write/save or adapterless export. If not reusable, briefly record why.

Default Play Shape

When work is crystallized, no shape flags means the schema-v1 steps + presentation model: adapter, process.exec, and browser effects live in frontmatter steps:, while metadata.execution_model: steps_with_presentation selects a deprivileged TypeScript renderer. Schema v2 remains an explicit --scheme 2 opt-in. Explicit --with-steps (template) or --format steps (export) is steps-only; --legacy-body or a deliberate shell/no-steps route is the legacy escape: take it only when the workflow needs control flow or runtime interaction the step language does not support, or when the user explicitly asks for a no-steps body. Runtime-discovered fan-out, conditions, ordering, parallelism, and long finite commands are steps capabilities (for_each, execution:, depends_on, max_concurrency, timeout_ms) — a set whose width is only discovered at run time is not a legacy trigger. Every play containing steps: runs with rote play run; only no-steps legacy TypeScript uses rote deno run --allow-all. For the concrete step syntax use rote grammar steps; the complete assembled artifact lives in rote guidance typescript play-creation.

Requirements Across Interruptions

For long, interrupted, partial-play, or reusable work, note the requirements before execution starts:

• Required sources/adapters and the capability each one must provide.
• Required live observations, output artifact, and verification checks.
• Constraints that must survive compaction, interruption, and superplay composition.

After compaction, a user interruption, or a handoff, reread that note before the next command. Do not substitute adjacent data for a missing capability: provider metadata is not live activity, event details are not recent trades, and a baseline report is not the augmented result.

New User Message During Work

When the user sends a new message mid-task, classify it before acting:

• Same-task refinement: update the requirements note and continue.
• Blocking correction or priority switch: preserve state, switch, and explicitly name whether the prior task remains open.
• Independent side question: answer briefly, then resume from the requirements note.
• Ambiguous relationship: ask one clarifying question.

Before answering or switching, write down the current gate, workspace/play name, cached response ids, pending stub state, artifact path, and exact next rote command. After the side question or correction, repeat that resume point and continue unless the user explicitly closes the original task. Decide by semantic relationship, not fixed phrases such as "hold on".

Instruction Priority

User instructions say what outcome to produce; this skill says how to use rote to produce it.

1. User instructions, project rules, and explicit approvals.
2. The rote lifecycle in this skill.
3. Companion rote skills for the selected state.
4. Live rote command output, grammar, and guidance for exact syntax.
5. General model defaults.

If the user asks to save, release, publish, make reusable, or crystallize a workflow, treat that as approval for the save/release path. It does not waive the crystallization plan, applicable pending write/save, materialization, test, release, index, search, or cleanup states.

Skill Play

digraph rote_flow {
    "User request" [shape=doublecircle];
    "Explicit setup/update/browser/org/registry?" [shape=diamond];
    "Invoke specialist" [shape=box];
    "Search local, then registry Plays" [shape=box];
    "Full play match?" [shape=diamond];
    "Partial play match?" [shape=diamond];
    "Run matched play" [shape=box];
    "Run partial play as baseline" [shape=box];
    "Preserve baseline for superplay" [shape=box];
    "Full play reuse terminal" [shape=doublecircle];
    "Route uncovered work" [shape=box];
    "Run rote explore <intent>" [shape=box];
    "Substrate route" [shape=box];
    "Execute in rote workspace" [shape=box];
    "Execute via rote-shell" [shape=box];
    "Verify requested artifact/content" [shape=box];
    "Reusable result?" [shape=diamond];
    "Build crystallization plan" [shape=box];
    "Pending supported?" [shape=diamond];
    "Pending write" [shape=box];
    "Pending save" [shape=box];
    "Save/release approved?" [shape=diamond];
    "Ask one yes/no save question" [shape=box];
    "Materialize draft" [shape=box];
    "Test, lint, release" [shape=box];
    "Index and search verify" [shape=box];
    "Pending discard" [shape=box];
    "Registry decision if sharing requested" [shape=box];
    "Present verified Play URI and sharing guidance" [shape=box];
    "Final answer" [shape=doublecircle];

    "User request" -> "Explicit setup/update/browser/org/registry?";
    "Explicit setup/update/browser/org/registry?" -> "Invoke specialist" [label="yes"];
    "Explicit setup/update/browser/org/registry?" -> "Search local, then registry Plays" [label="no"];
    "Invoke specialist" -> "Verify requested artifact/content";
    "Search local, then registry Plays" -> "Full play match?";
    "Full play match?" -> "Run matched play" [label="yes"];
    "Full play match?" -> "Partial play match?" [label="no"];
    "Run matched play" -> "Verify requested artifact/content";
    "Partial play match?" -> "Run partial play as baseline" [label="yes"];
    "Partial play match?" -> "Run rote explore <intent>" [label="no"];
    "Run partial play as baseline" -> "Preserve baseline for superplay";
    "Preserve baseline for superplay" -> "Route uncovered work";
    "Route uncovered work" -> "Run rote explore <intent>";
    "Run rote explore <intent>" -> "Substrate route";
    "Substrate route" -> "Execute in rote workspace" [label="adapter/API"];
    "Substrate route" -> "Execute via rote-shell" [label="local CLI/files/logs"];
    "Execute in rote workspace" -> "Verify requested artifact/content";
    "Execute via rote-shell" -> "Verify requested artifact/content";
    "Verify requested artifact/content" -> "Full play reuse terminal" [label="verified full play"];
    "Verify requested artifact/content" -> "Reusable result?" [label="workspace/browser/manual result"];
    "Reusable result?" -> "Final answer" [label="no"];
    "Reusable result?" -> "Build crystallization plan" [label="yes"];
    "Build crystallization plan" -> "Pending supported?";
    "Pending supported?" -> "Pending write" [label="adapter/browser/mixed"];
    "Pending supported?" -> "Save/release approved?" [label="process-only"];
    "Pending write" -> "Pending save";
    "Pending save" -> "Save/release approved?";
    "Save/release approved?" -> "Materialize draft" [label="yes"];
    "Save/release approved?" -> "Ask one yes/no save question" [label="no"];
    "Ask one yes/no save question" -> "Final answer";
    "Materialize draft" -> "Test, lint, release";
    "Test, lint, release" -> "Index and search verify";
    "Index and search verify" -> "Pending discard";
    "Pending discard" -> "Registry decision if sharing requested";
    "Registry decision if sharing requested" -> "Final answer" [label="local only"];
    "Registry decision if sharing requested" -> "Present verified Play URI and sharing guidance" [label="published"];
    "Present verified Play URI and sharing guidance" -> "Final answer";
}

Skill Priority

When multiple rote skills could apply, use this order:

Substrate Router: Adapter vs Browser vs Shell

After the initial play search, choose the substrate that best matches the user's intent. Do not force all tasks through adapters, browser automation, or shell.

Provider/API data is not shell work merely because curl, Python, Node, or a public REST endpoint can fetch it. After play search, route through rote-task-routing; if no installed adapter fits, search/install from the adapter catalog before direct fallback.

Browser routing rule:

• Browser words outrank domain nouns. "Browse my calendar" routes to rote-browse after the initial play search, even though calendar data can be an adapter task. Use an adapter/play only if it is already installed, healthy, and completes the request. If it is missing, stale, unauthenticated, or fails setup, switch to browser attach instead of asking the user to build an adapter first.
• Browser words also outrank native web search. If the user asks to browse or extract public pages, use rote-browse for page observation/extraction. Native search may discover candidate URLs only when explicitly requested or when no URL can be found through adapter/CLI evidence; it is not a replacement for browser snapshots and page extraction.
• If the user needs existing login, Gmail, SSO, MFA, extensions, active tabs, or profile state, ask whether to attach to an existing headed browser.
• If the work is public, read-only, CI, or replay-like, ask whether a headless new session is acceptable.
• If the user says only "browse", ask headed vs headless, then attach-existing vs new session.

Do not satisfy browser intent with native web search, WebFetch, raw curl, raw Playwright, open, or rote proc run unless the user explicitly switches from browser interaction to a non-browser substrate.

Shell/process routing rule:

• The rote-shell skill owns durable local CLI, files, logs, background processes, dependency manifests, and shell-derived play crystallization — but not provider/API data collection: use adapters for typed provider calls, and rote proc only for local command evidence or transformations selected by the router. Its command model and raw-shell boundary live in that skill, not here.

1. rote first for any day-to-day rote, adapter, MCP, play, workspace, or reusable workflow task.
2. Explicit lifecycle specialists before play search only when the user directly asks for setup, update, browser automation, shell/process work, registry, org administration, adapter creation, or adapter configuration.
3. rote-flow-run before workspace exploration when search or @@plays returns a usable play.
4. rote-task-routing only to choose the next route after no full play match or after a partial play baseline is preserved. Never route the same work again after a verified full play match.
5. rote-shell only when the user explicitly asked for local CLI/files/logs/process work or rote-task-routing selected the shell substrate.
6. rote-workspace for adapter probes, calls, cached response queries, transformations, and multi-adapter execution.
7. rote-flow-crystallization for every new reusable workspace, browser, shell, or manual result before final presentation; unchanged reuse of an existing released play skips this gate. Process-only work uses the same doctrine and approval gate, then authoring materializes it through adapterless export because template/pending still require a real adapter.
8. rote-flow-authoring only after direct authoring intent or an approved crystallization plan.
9. rote-command-patterns and rote-typescript-transformations are helper/reference skills; they return to the owner and do not complete the lifecycle themselves.
10. rote-troubleshooting after an unchanged retry fails or state recovery is unclear.

Skill Types

• Rigid states: play search, play provenance preservation, workspace execution, pending write/save, scaffold, tests, release, index, search verification, pending discard, and full-play reuse termination. Follow these exactly.
• Flexible states: wording, artifact formatting, adapter choice after routing, parameter naming, and registry visibility recommendations. Adapt these to the user request and live rote output.

User Instructions

The user may name the desired result, adapter, play, output file, or release target. That narrows the path; it does not skip lifecycle states. If the user asks for a direct adapter/MCP/browser/shell action, route through rote unless they explicitly forbid rote.

Red Flags

These thoughts mean stop and return to the current lifecycle state:

Active task routing starts here and then moves to a named companion skill or an explicit reference.

Platform Adaptation

Rote skills speak in actions: invoke a companion skill, run a rote command, edit a file, dispatch a subagent, enter a workspace, or ask for approval. Do not assume one runtime's tool names apply everywhere.

Load the platform reference only when runtime-specific details affect the current step:

For the full companion graph, lifecycle edges, and standard packet fields, load skill-workflow-map.md. Keep it out of active context until the short routing table is not enough.

Top-Level Skill Routing

Complete the root Play-search gate before invoking the narrow skill that owns the next state: search locally, then search the registry when no verified full local match exists. Explicit setup, update, browser, shell/process, registry, org, adapter-create, or adapter-config requests may start in that specialist skill, but those specialists still return to this lifecycle before final presentation when reusable workflow state is involved. Workflow logic lives in standalone skills; rote/references/ is only for platform mapping, the optional workflow map, and the compact Play search/run reference.

Execution State Machine

1. State the intent in one phrase and run rote play search "<intent>". If it has no verified full local match, run rote play search "<intent>" --source registry before classifying the best available coverage. Keep each provider's order intact; their ranks are not comparable. Explicit setup/update/browser/shell/org/registry/adapter-create/adapter-config requests may start in the named specialist instead.
2. If a Play from either provider fully covers the request, run it through rote-flow-run. That skill owns local callability and the registry card's inspect/install/run path. Verify the requested artifact content, and stop. Do not explore adapters, initialize a workspace, rewrite the artifact, or enter pending write/save unless the user explicitly asked to edit, create a separate enhanced artifact, save a new workflow, release, or publish.
3. After both applicable provider searches, if a Play covers only part of the request, run it as the baseline. Preserve the raw baseline output, provenance, sentinels, source labels, and markers as source material for a new composed superplay. Route only the uncovered work, then save/release the reusable composition if requested or approved. Do not replace the baseline with a hand-written lookalike report.
4. If neither provider matched or uncovered work remains, run rote explore "<intent>" and obey any @@plays suggestions before adapter work.
5. Choose the substrate. Route local CLI/files/log/process work to rote-shell; route adapter/API work through task selection before workspace work.
6. If a generated rote-<adapter-id> subagent is the right owner, dispatch it before entering the workspace, not midway through execution.
7. If no installed adapter covers adapter/API work, search the adapter catalog before direct fallback. Tell the user what rote checked when falling back.
8. Execute adapter work through the workspace path and shell work through rote-shell, preserving cached @N responses, process evidence, session state, and any existing-play baseline output. Inspect and transform rote responses with rote query @N, rote query schema, and rote proc rather than ad hoc grep, head, inline Python, or temp-file parsing when the result feeds a report, play, or later reasoning step.
9. Verify the requested deliverable by reading content, not just checking that a command exited or a file exists. If a workspace was used, also run cd <workspace-path> && rote ls before the final answer — it emits the [MANDATORY PROTOCOL] pending-stub warning when a reusable result has not entered the pending lifecycle. Run the command that enforces the save gate instead of relying on remembering it.
10. Before presenting any new reusable result, run rote-flow-crystallization to produce the semantic plan and resolve save/discard. Adapter, browser, and mixed work then use pending write/save. Process-only work skips those adapter-requiring commands without skipping the plan. If save/release was already requested, continue to authoring without asking again. Do not run this gate for unchanged reuse of an existing play.
11. After an approved process-only plan, let rote-flow-authoring choose adapterless workspace export or a justified legacy no-steps body using shell guidance. Never invent an adapter or append shape flags to a pending-save command.
12. After authoring, release with rote play release, rebuild the index, verify search, then clear the pending stub with rote play pending discard <workspace>.
13. When registry publication succeeds or the selected version is already in sync, require rote-registry to return the exact play_uri, bootstrap_uri, resolved data.play_inspect.reference, data.play_inspect.execution, published-reference execution_verification status and evidence, and access guidance (resolution and execution audiences). Present the disclosure-only Play URI, advertised bootstrap transition, and resolved play run reference as separate resources. Static eligibility and inspection never substitute for the exact pinned acceptance run; if that run is not authorized or cannot be supplied safe representative inputs, label it unverified. A local release has no published Play URI, and no skill should construct or parse one.

Bounded Exploration

Data hunting is a means, not the deliverable. When a request needs "the best/most X" from a large result space, satisfice: fix a small probe budget up front, take the best candidate found within it, state the search bounds in the deliverable, and move on to the artifact and crystallization gates. Delivering a good-enough result end to end beats an exhaustive sweep that leaves no room for the report, save, and release steps.

• Prefer server-side ordering and filters (order, tag, limit) over client-side pagination sweeps. If a couple of ordered, filtered calls do not surface a candidate, take the best so far.
• Probe a response's structure once (a single keys/type query), then reuse that shape for every later response from the same tool.
• Install only the adapters the request needs; a second same-provider adapter is not a fallback for a data question the first one already answers.

Conversation And Interruption Routing

For every new user message during an active rote lifecycle, classify it by meaning, not by a fixed keyword:

• Same task refinement: incorporate the new constraint into the active lifecycle and continue.
• Blocking correction or priority switch: stop the current action at a safe boundary, preserve workspace/play/process state, and follow the new instruction.
• Independent side question: capture a recoverable handoff summary if there is active rote state, answer the side question, then return to the previous lifecycle and say which gate is being resumed.
• Ambiguous relationship: briefly state the inferred relationship and choose the safer route: preserve state before continuing, and ask only if proceeding would risk an unwanted write, release, publish, or external mutation.

Do not assume interruption means cancellation. Do not assume an unrelated answer completes the original rote request. The model under test must maintain the active lifecycle until it is verified, handed off, cancelled, or explicitly superseded.

Standard Handoff Packet

Use this packet when handing work to a companion skill, a workspace-bound skill, or a subagent. For long-running workspace or subagent work, also write a short markdown handoff summary in the workspace or artifact location named by the owning skill.

## Handoff Packet

- Origin skill: `rote`
- Target skill: `rote-...`
- User intent: ...
- Current state: play search result, adapter/workspace state, or registry/setup state
- Preconditions satisfied: commands already run, approvals granted, credentials verified
- Workspace path: ... or none
- Cached responses: `@N` ids and what each contains
- Allowed commands: rote commands or play paths the target may run
- Stop conditions: unsafe action, missing credential, failed precondition, user approval needed
- Return fields: result, artifacts, response IDs, save gate, verified Play URI and sharing guidance
  when published, next recommended skill

Handoff Contract

• Use when: a user request may touch rote adapters, plays, workspaces, registry state, browser automation, shell/process work, setup/update, or reusable workflow authoring.
• Preconditions: a rote task intent can be stated; if command execution is needed, the current runtime can run rote or surface the exact permission/installation blocker.
• Owns: the initial rote rule, play-search-first gate, top-level skill selection, platform-reference selection, and standard handoff packet shape.
• Hands off to: rote-flow-run, rote-task-routing, rote-workspace, rote-flow-crystallization, rote-flow-authoring, rote-command-patterns, rote-typescript-transformations, rote-troubleshooting, rote-adapter-create, rote-adapter-config, rote-using-adapters, rote-registry, rote-org, rote-browse, and rote-shell.
• Returns to: the user when the selected skill or reference completes, or to rote when a companion reports a partial result that needs another top-level routing decision.
• Stop when: a full play output satisfies the request, a companion skill takes ownership, a required approval or credential is missing, or continuing would require unsafe/destructive action the user did not request.
• Completion signal: a selected skill/reference, verified play result, explicit blocker, or final user-facing answer with any reusable-work save gate resolved and any published play's verified Play URI plus access guidance (resolution and execution audiences) presented.

Detail References

Use the standalone skill needed for the current branch. Read platform references only for runtime-specific tool mapping, and read skill-workflow-map.md only when the full companion graph or handoff packet shape is needed.

Live CLI Guidance Surfaces

• rote how - complete onboarding tree.
• rote start - mandatory protocol checks before a task.
• rote guidance agent essential - core agent workflow conventions.
• rote guidance adapters essential - adapter probe and call patterns.
• rote guidance browser essential - browser automation patterns.
• rote guidance shell essential - rote proc, process leases, stream capture, deps, and shell play crystallization patterns.
• rote guidance play - progressive play design tree: crystallization for the semantic plan, shape for the DAG, and testing for contract verification.
• rote play info <name-or-path> --json - canonical local Play record: absolute path plus ordered parameters. Use it only when a local result lacks a runnable command or legacy argument syntax needs confirmation. Registry cards use rote play inspect <reference> --json instead.
• rote play list - inventory released local plays; do not use an empty search query as inventory.
• rote grammar query, rote grammar steps, rote grammar deno, rote grammar export, and related topics - current command syntax (steps covers the frontmatter steps: plane: step kinds, $param substitution, @step{…} references, for_each).