# The Relurpish Guide

Audience: anyone driving the Euclo coding agent from the terminal. Purpose: daily use of the relurpish TUI — the screen, the commands, approvals, and recipes. Prerequisites: a working workspace; see Getting started .

# Starting and stopping

relurpish chat --workspace .

Quit with ctrl+c.

# The screen

At boot you see:

• A Postures banner, for example Postures — approvals: governed; intent selection: deterministic. It tells you who answers approvals and who breaks intent ties; see Postures .
• Tabs: [1] chat [2] diff [3] recipe. The chat tab has sub-tabs for the agent mode (for example local-read-only, local-edit-on).
• An input bar at the bottom; while a run is active it reads running — > blocked | : and / active | ctrl+c quit.

Prefixes in the input bar:

Known quirk: the input-bar hint text mentions :git status, but there is no git command in the registry; git operations are /commit and /local-review.

# A turn, as you experience it

1. You type a task in plain language and press Enter.
2. The classifier scores it into intent families (debug, review, implementation, …).
3. A thoughtrecipe or capability is picked deterministically. The reason is recorded as decided_by in telemetry: pinned, explicit, lattice:score, default_recipe, or clarification.
4. If the intent is genuinely ambiguous, a clarification frame appears: answer with the slot number, press Enter on empty input for the default slot, or d to defer. A confident single-family instruction never asks.
5. Writes and risky capabilities pause with an approval banner; see Approvals .
6. The recipe’s steps run, typically a bounded ReAct loop: the model reads, edits through governed file_write, and verifies through the governed test runner inside the sandbox.
7. The turn ends with a classified outcome; the feed reports the real cause on failure (for example react decision unparseable), and /retry re-runs it.

The diff tab shows pending file changes; the recipe tab shows the recipe registry.

# Approvals

The default governed HITL mode pauses writes and risky capabilities with a banner:

[y] once  [s] session  [a] always  [n] deny  [d] dismiss

The banner has key priority: press the key even while the cursor is in the input. During active work, s (this session) is the usual answer. /hitl lists pending approvals.

In permissive mode (--hitl-mode permissive or the doctor --set-provider-style configuration), approval requests auto-approve and each decision is written to the audit trail. Sandbox and capability enforcement are unchanged; the doctor marks the posture degraded.

# Postures

Two independent postures are decided at boot and shown in the banner:

# Command reference

Commands are the same whether you start them with / or :.

# Universal commands

# Chat-tab commands

Palette entries /test-latency, /test-rule, /reset-bindings, /diff-apply-all, /diff-revert, and /diff-view currently report “not yet implemented as palette command”.

# Working with recipes

/recipe lists what is registered; /recipe <id> pins one for the session; /recipe clear unpins. Pinning overrides selection entirely: every turn runs exactly that recipe.

Recipes load once at boot. After adding or editing a file in relurpify_cfg/euclo/, restart relurpish; /recipe tells you when a file was skipped and why. A pin of an unregistered recipe is rejected at pin time with the registered list.

Recipe authoring is documented in RECIPE_DSL ; validate with relurplint --check recipes --workspace ..

# Inspecting state

Everything the runtime produces lives under <workspace>/.relurpify_state/:

Two useful greps during or after a session:

jq -r '.. | .logRecords[]? | select(.eventName=="euclo.route.selected")' \
    .relurpify_state/telemetry/otlp/*.jsonl | tail -3
jq -r '.. | .logRecords[]? | select(.eventName=="euclo.execution.complete")' \
    .relurpify_state/telemetry/otlp/*.jsonl | tail -1

euclo.route.selected carries decided_by (why this route ran); euclo.execution.complete carries the real LLM call count and token usage.

Telemetry records the runtime’s decisions (which route, which tool, the outcome). Raw model and tool payloads are deliberately kept out of telemetry; when you need the raw wire view, enable LLM debug and read logs/ (contributors: see devdocs/observability.md).

# Tips

• Recipes load once at boot; restart after editing relurpify_cfg/euclo/.
• Pin deliberately: natural-language selection is deterministic, and the clarify flow only stops you on genuine ambiguity.
• A failed turn tells you why: the error in the feed is the real cause, and /retry re-rolls it. With small local models an occasional retry is normal.
• Approvals are scoped: y answers once, s for the session, a for always; all three are auditable in audit/.
• Export before you close: /export md transcript.md keeps a shareable record of the session.