rubino-agent 0.3.0

data/lib/rubino/llm/scenarios/wide-table.yml

@@ -0,0 +1,121 @@
+# Repro scenario for the CLI streaming-table live-region trail/accumulation bug.
+# Streams a WIDE (5-column) markdown table emitted as MANY small content deltas
+# so the per-chunk live-region repaint cascade appears. Each data ROW is wider
+# than a narrow terminal AND carries double-width emoji (✅ 🔄 ⚠️) spread across
+# cells so the CLAMPED tail of the in-flight row still contains an emoji — whose
+# display width (2) exceeds its char length (1). With the buggy char-count clamp
+# that emoji pushes the "fits in one row" line past the real terminal width, so
+# it wraps and the single-row clear leaves residue that accumulates downward.
+events:
+  - type: content
+    text: "Here is the status table:\n\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Feature | Status | Priority | Owner | Notes |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| --- | --- | --- | --- | --- |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| HTTP API ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Done ✅ | High ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Backend ✅ | Stable in prod ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "since release v0.2 ✅ ✅ ✅ |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| CLI Chat ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Done ✅ | High ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Frontend ✅ | Live md blocks ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "committed and styled ✅ ✅ ✅ |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| TUI Mode 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| In Progress 🔄 | Medium ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Platform 🔄 | Live region 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "repaint work ongoing 🔄 🔄 🔄 |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| MCP Bridge 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| In Progress 🔄 | Medium ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Integrations 🔄 | Tool schema 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "negotiation pending 🔄 🔄 🔄 |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Web UI ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Done ✅ | Low ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Frontend ✅ | Rails DOM ✅ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "token streaming live ✅ ✅ ✅ |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Memory 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| In Progress 🔄 | High ⚠️ "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "| Core 🔄 | SQLite tiny-Zep 🔄 "
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "recall backend 🔄 🔄 🔄 |\n"
+  - type: delay_seconds
+    value: 0.30
+  - type: content
+    text: "\nAll tracked features above."
+  - type: delay_seconds
+    value: 0.30

data/lib/rubino/llm/scenarios/with-approvals.yml

@@ -0,0 +1,50 @@
+# With-approvals scenario: tokenized reasoning, then a shell tool_call that
+# the real ApprovalPolicy will gate on. Once approval lands and the tool
+# executes, the next agent turn re-enters the provider — but for THIS turn
+# the provider just emits thinking + the tool_call.
+#
+# Translation from the reference with-approvals scenario:
+#   - reasoning.available {text} → thinking chunks (token-by-token)
+#   - __pause_for_approval       → DROPPED; the shell tool_call below
+#                                  triggers approval via ApprovalPolicy
+#   - tool.started/completed     → collapsed into one tool_call
+#   - message.delta {delta}      → content chunks (will be emitted on the
+#                                  follow-up turn after tool result; here
+#                                  we keep them so the fake turn is rich)
+#   - run.completed              → DROPPED; fake provider terminates when
+#                                  events are exhausted
+events:
+  - type: thinking
+    text: "I"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " need"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " your"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " approval"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " for"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " this"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " command."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: tool_call
+    id: "call_fake_approve_1"
+    name: "shell"
+    arguments:
+      command: "rm -rf /tmp/cache"

data/lib/rubino/llm/scenarios/with-artifacts.yml

@@ -0,0 +1,98 @@
+# With-artifacts scenario: model writes a markdown report and hands it to
+# the UI as a downloadable artifact. Mirrors the reference UX —
+# token-by-token reasoning, a short content explanation, then two
+# tool calls in one turn:
+#   1. `write` drops the file on disk (workspace-scoped — bin/dev points
+#      terminal.cwd at $RUBINO_HOME/workspace so this lands under
+#      the strict-workspace check)
+#   2. `attach_file` registers it as an artifact → ARTIFACT_CREATED on
+#      the bus → SSE artifact.created → the web UI downloads via
+#      GET /v1/files and renders a download link
+#
+# After the tools execute, FakeProvider's post_tool_turn? guard kicks in
+# and emits a closing "Done." so the run terminates without re-replaying.
+events:
+  - type: thinking
+    text: "Let"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " me"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " put"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " together"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " a"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " short"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " markdown"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " report"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " summarising"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " the"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " findings."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: content
+    text: "I've put together a short markdown report. "
+  - type: delay_seconds
+    value: 0.04
+  - type: content
+    text: "Writing it to `report.md` now and attaching it so you can download it directly from this turn.\n"
+  - type: delay_seconds
+    value: 0.05
+
+  - type: tool_call
+    id: "call_fake_artifact_write"
+    name: "write"
+    arguments:
+      file_path: "report.md"
+      content: |
+        # Report
+
+        Generated by the fake LLM provider for the `with-artifacts` scenario.
+
+        ## Highlights
+
+        - The agent created this markdown file with the `write` tool.
+        - It was then attached to the run via `attach_file` so the web UI
+          could surface a download link in the timeline.
+        - This is exactly the round-trip the legacy fake provider
+          exercised: reasoning → content → tool call → artifact.
+
+        ## Next steps
+
+        Open the attached file from the session timeline to see the rendered
+        output. Real runs would replace this with whatever the user asked
+        for (CSV exports, configuration dumps, screenshots, etc.).
+
+  - type: tool_call
+    id: "call_fake_artifact_attach"
+    name: "attach_file"
+    arguments:
+      file_path: "report.md"
+      filename: "report.md"

data/lib/rubino/llm/scenarios/with-clarify.yml

@@ -0,0 +1,32 @@
+# With-clarify scenario: the LLM emits a question tool_call so the real
+# question pipeline gathers user input; afterwards it streams a short
+# confirmation. __pause_for_clarify is dropped — the question tool_call
+# triggers the clarify flow naturally.
+events:
+  - type: thinking
+    text: "Let"
+  - type: thinking
+    text: " me"
+  - type: thinking
+    text: " ask"
+  - type: thinking
+    text: " you."
+  - type: delay_seconds
+    value: 0.04
+  - type: tool_call
+    id: "call_fake_clarify_1"
+    name: "question"
+    arguments:
+      question: "Which approach do you prefer for the API layer?"
+      options:
+        - label: "Option A: REST API"
+        - label: "Option B: GraphQL"
+        - label: "Option C: gRPC"
+  - type: thinking
+    text: "Great"
+  - type: thinking
+    text: " choice."
+  - type: delay_seconds
+    value: 0.04
+  - type: content
+    text: "Proceeding with your selection."

data/lib/rubino/llm/scenarios/with-reasoning.yml

@@ -0,0 +1,175 @@
+# With-reasoning scenario: multi-step reasoning with tool calls and
+# streaming output. Best for exercising the timeline rendering and
+# live-streaming UI — reasoning is emitted token-by-token to mimic a
+# real LLM, and real tools (read, shell) are invoked so the downstream
+# pipeline (approvals, output capture) runs end-to-end.
+events:
+  - type: thinking
+    text: "Let"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " me"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " analyze"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " the"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " request"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " step"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " by"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " step."
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "\n"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "First"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " I"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " need"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " to"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " check"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " the"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " current"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " state."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: tool_call
+    id: "call_fake_reasoning_1"
+    name: "read"
+    arguments:
+      file_path: "/tmp/example.txt"
+
+  - type: delay_seconds
+    value: 0.5
+
+  - type: thinking
+    text: "Found"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " relevant"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " data."
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "\n"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "Now"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " running"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " tests"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " to"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " validate."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: tool_call
+    id: "call_fake_reasoning_2"
+    name: "shell"
+    arguments:
+      command: "npm test"
+
+  - type: delay_seconds
+    value: 0.8
+
+  - type: thinking
+    text: "Tests"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " pass."
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "\n"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "Formulating"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " final"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " response."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: content
+    text: "Here are the results of my analysis:\n\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "- File content analyzed\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "- Tests passed\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "- All checks complete\n\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "The system is running normally. No issues detected."
+  - type: delay_seconds
+    value: 0.05

data/lib/rubino/llm/scenarios/with-uploads.yml

@@ -0,0 +1,104 @@
+# Scenario with file upload. Mirrors the reference with-uploads but uses the
+# FakeProvider event vocabulary. The agent "reads" an uploaded file and
+# emits a fake analysis as the final assistant message. {{input}} is the
+# user's text plus the uploaded file names/content types.
+events:
+  - type: thinking
+    text: "I"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " received"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " the"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " uploaded"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " file."
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "\n"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "Analyzing"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " its"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " contents..."
+  - type: delay_seconds
+    value: 0.04
+
+  - type: delay_seconds
+    value: 0.4
+
+  - type: thinking
+    text: "The"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " file"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " appears"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " to"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: " contain"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " valid"
+  - type: delay_seconds
+    value: 0.04
+  - type: thinking
+    text: " data."
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "\n"
+  - type: delay_seconds
+    value: 0.02
+  - type: thinking
+    text: "Processing..."
+  - type: delay_seconds
+    value: 0.06
+
+  - type: delay_seconds
+    value: 0.3
+
+  - type: content
+    text: "File analysis complete:\n\n"
+  - type: delay_seconds
+    value: 0.05
+  # `{{input}}` is the web UI's already-prefixed text (e.g. "uploaded file:
+  # report.pdf" when the user attached a file with no message). Echo it
+  # straight so the agent reply always names the file the user sent.
+  - type: content
+    text: "Received {{input}}.\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "- Content validated\n"
+  - type: delay_seconds
+    value: 0.05
+  - type: content
+    text: "- No issues detected\n"
+  - type: delay_seconds
+    value: 0.05

data/lib/rubino/llm/thinking_support.rb

@@ -0,0 +1,84 @@
+# frozen_string_literal: true
+
+require_relative "provider_resolver"
+
+module Rubino
+  module LLM
+    # Session-scoped memory of providers that rejected an Anthropic-style
+    # thinking budget, plus the detector for that rejection (#75), plus the
+    # static per-provider capability gate (#2).
+    #
+    # Process-level (not per-adapter) because Lifecycle rebuilds the adapter
+    # every turn — and one CLI process serves one chat session, so this is
+    # exactly "remember for the session". RubyLLMAdapter consults it before
+    # rendering a budget and marks it on a recognised rejection, so the
+    # provider is never sent a budget again this session.
+    module ThinkingSupport
+      @unsupported = {}
+
+      module_function
+
+      def unsupported?(provider)
+        @unsupported.key?(provider.to_s)
+      end
+
+      # Per-provider thinking CAPABILITY gate (#2). #unsupported?/#rejection?
+      # (#75) handle a provider that REJECTS a budget (hard 400 → retry +
+      # session memo); this handles one that ACCEPTS it and then, lacking a
+      # separate reasoning channel, dumps its chain-of-thought as plain content
+      # deltas — observed live on MiniMax. providers.<name>.supports_thinking
+      # (true/false) is the explicit override; unset, MiniMax-family model ids
+      # default to false (they return no thinking blocks and leak reasoning
+      # when sent a budget), everything else to true.
+      def supports?(provider_cfg, model_id)
+        configured = provider_cfg["supports_thinking"]
+        return configured unless configured.nil?
+
+        !model_id.to_s.match?(ProviderResolver::PROVIDER_PATTERNS["minimax"])
+      end
+
+      # providers.<name>.supports_thinking: true is the user's explicit promise
+      # that the backend accepts an Anthropic-style thinking block. ruby_llm
+      # 1.16 only renders with_thinking when the model's REGISTRY entry
+      # declares a budget_tokens reasoning option; an assume-model-exists model
+      # (MiniMax-M3 on the anthropic-compatible path) declares none, so
+      # with_thinking raised client-side before any request, the #75 rejection
+      # detector matched the message, and the documented opt-in silently died
+      # every turn (#175). On that path the adapter puts the wire payload on
+      # with_params instead, which ruby_llm deep-merges into the request body
+      # unconditionally.
+      def budget_via_params?(provider_cfg, chat)
+        return false unless provider_cfg["supports_thinking"] == true
+
+        model = chat.respond_to?(:model) ? chat.model : nil
+        !(model.respond_to?(:reasoning_option) && model.reasoning_option("budget_tokens"))
+      rescue StandardError
+        true
+      end
+
+      # Records the rejection and tells the user once with a dim note (only
+      # the marking path emits it). Cosmetic: a UI failure must never break
+      # the retried turn.
+      def mark_unsupported!(provider, notify: nil)
+        @unsupported[provider.to_s] = true
+        notify&.note("provider doesn't support thinking — effort off")
+      rescue StandardError
+        nil
+      end
+
+      # Test seam: forget all recorded rejections (a fresh "session").
+      def reset!
+        @unsupported = {}
+      end
+
+      # True when +error+ reads as a provider's "thinking (budget) is not
+      # supported" rejection. Kept narrow: the message must name thinking plus
+      # a not-supported phrasing.
+      def rejection?(error)
+        msg = error.message.to_s.downcase
+        msg.include?("thinking") &&
+          (msg.include?("not support") || msg.include?("unsupported"))
+      end
+    end
+  end
+end