[codex] Add Tiny Local tool-use profile

[TerminallyLazy]

Summary

• Add a bundled Tiny Local profile with a minimal action-first JSON/tool-use contract for small local models.
• Document the same contract as a project-scoped Prompt Include recipe in docs/guides/local-model-tool-use.md.
• Tighten Tiny Local tool prompts for code_execution_tool, text_editor, and response so qwen-sized models avoid visible reasoning, repeated writes, and continuation-status loops.
• Fix the bundled Editor/Text Editor plugin handoff so open_in_canvas: true Markdown writes actually open the changed file in the Editor canvas.

Why

PR #1714 was doing too much in core framework/runtime code. This branch keeps the replacement on prompt/profile/plugin surfaces unless later evidence proves deeper framework changes are required.

The local-model failure modes reproduced with qwen3.5:9b were:

• models narrating or promising the next step instead of using tools,
• continuation messages like excellent proceed / okay continue producing duplicate status responses,
• Markdown writes setting open_in_canvas: true in the tool result while the Editor canvas stayed on its empty start state.

Core-code audit

No core framework code is changed in this branch.

I explicitly checked the staged diff for core paths with:

git diff --cached --name-only | rg '^(agent\.py|initialize\.py|helpers/|models\.py|api/|extensions/|tools/|webui/|prompts/|knowledge/|lib/|scripts/|conf/|docker/)'

That command returned no matches before commit.

The staged files are limited to:

• agents/ profile files and profile DOX,
• docs/guides/ user documentation,
• bundled plugin files under plugins/_editor and plugins/_text_editor,
• focused tests.

Plugin-code explanation

This branch does include bundled plugin code changes, but not core Agent Zero framework/runtime changes.

plugins/_text_editor/tools/text_editor.py now includes the active agent context id in write/patch result metadata when available. This lets UI consumers know which chat context produced a Markdown file update.

plugins/_editor/extensions/webui/set_messages_after_loop/sync-text-editor-results.js now:

• preserves context_id / ctxid from text-editor results,
• allows explicit fresh open_in_canvas: true results to open the Editor even on the first rendered chat turn,
• keeps the old guard for passive refresh-only syncs when the chat history is initially empty.

plugins/_editor/webui/editor-store.js now honors an explicit context id passed by the tool-result handoff instead of always falling back to the currently selected UI context.

These plugin changes are what make Create and open a TODO.md list in the Canvas. actually open TODO.md in the right Editor canvas after the text-editor write succeeds.

Validation

• git diff --cached --check
• python3 -m py_compile tests/test_default_prompt_budget.py tests/test_tool_action_contracts.py tests/test_text_editor_context_patch.py tests/test_office_canvas_setup.py plugins/_text_editor/tools/text_editor.py
• node --check plugins/_editor/extensions/webui/set_messages_after_loop/sync-text-editor-results.js
• node --check plugins/_editor/webui/editor-store.js
• ruby -e 'require "yaml"; p YAML.load_file("agents/tiny-local/agent.yaml")'
• Manual live WebUI smoke with qwen3.5:9b + Tiny Local: Create and open a TODO.md list in the Canvas. wrote TODO.md and opened it in the Editor canvas.
• Direct qwen3.5:9b continuation smokes: excellent proceed and okay continue emitted code_execution_tool instead of duplicate response status text.

Full pytest was not run because the current host/container test environments are missing required test dependencies (tiktoken/litellm on host, pytest in the container).