Carlos Florêncio
6.0 KiB
created, status, requirements, system_design, legacy_specs
| created | status | requirements | system_design | legacy_specs | ||||||
|---|---|---|---|---|---|---|---|---|---|---|
| 2026-09-08 | complete |
|
|
Implementation plan: MCP discovery and canvas prompts
Subsequent design
The direct canvas creation package replaces the long visible preset from Task 02 with a short goal and saved reference. This package's completed results remain historical; the new package owns updated catalog, delivery, and desktop/mobile checks. Task 01 is unchanged.
Overview
Make Kandev tool discovery explicit, reduce repeated tool descriptions, and direct canvas requests through draft creation and publication.
The failed task efb43c76-a2ab-44cf-9169-b38fb5e3da9f created a standalone app.
Its conversation contained no canvas calls. Its original MCP server exposed
seven canvas tools, but its injected task prompt omitted all canvas guidance.
The canvas list was empty. These observations motivate the prompt changes.
Agents owns the shared discovery contract because it spans task operations and clients. Canvases owns its creation preset and authoring outcome. The paired specifications link those independent contracts without duplicating ownership.
Scope
In scope
- Conditional client-native discovery with an available-catalog fallback.
- A compact task prompt with essential workflow and authorization rules.
- Canvas instructions derived from the resolved session MCP profile.
- A stronger localized Create canvas preset in the existing task dialog.
- Targeted prompt, producer, catalog, desktop, and mobile verification.
Out of scope
- MCP transport, permission, scope, promotion, or runtime flag changes.
- A new tool-search API or provider-specific discovery adapter.
- Office/configuration prompt compaction or a new canvas editor.
- Repairing the previous task, publishing its files, or contacting its agent.
Technical approach
Task 01 changes config/prompts/kandev-context.md and internal/sysprompt.
It preserves critical dynamic sections and replaces routine tool listings with
discovery guidance. It wires the canvas option through all existing task
prompt producers using the executor's resolved MCP profile.
Task 02 changes the localized canvas preset. CanvasTaskCreateLauncher keeps
the normal task dialog and scratch/local defaults. It adds tests against real
catalog text, because the current component test substitutes a mock sentence.
The mobile E2E currently replaces the preset with explicit MCP commands, so
that test alone cannot detect weak authoring guidance.
The system designs contain the proposed discovery and canvas wording. The implementation wires the capability projection through task launch, prepared session launch, workflow context reset, and the first direct message path.
Tests
| Criteria | Evidence |
|---|---|
AC-AGENTS-MCP-DISCOVERY-001.1 through .6 |
New focused sysprompt discovery tests. |
AC-AGENTS-MCP-DISCOVERY-002.1, .2, .5 |
Compact size and preserved workflow assertions. |
AC-AGENTS-MCP-DISCOVERY-002.3, .4 |
Producer tests for start, prepared start, and reset; MCP inventory tests. |
AC-CANVASES-AGENT-WEB-APPS-001.10 through .12 |
Gated authoring instruction tests and negative profile cases. |
AC-CANVASES-AGENT-WEB-APPS-009.6, .8 |
Real locale preset contract tests with exact tool identifiers. |
AC-CANVASES-AGENT-WEB-APPS-009.7 |
Desktop/mobile dialog and submitted-description tests. |
E2E tests
Task 02 extends the guided creation case in
apps/web/e2e/tests/canvas/mobile-plugin-canvas.spec.ts and adds the equivalent
focused case in plugin-canvas.spec.ts. The projects are mobile-chrome and
chromium. Assert the actual preset before editing it, then assert that the
submitted task retains the edited description.
Use isolated managed runners. Existing mock lifecycle coverage proves canvas registration and publication. It does not prove an external model's compliance with the prompt. Do not report that stronger claim from deterministic tests.
Work orders
Both work orders are sequential. Implementation requires the later explicit request specified by the repository workflow.
Verification results
rtk python3 scripts/lint-spec-files.test.py: 30 tests passed.rtk python3 scripts/lint-spec-files.py --all: all specification files passed.rtk git diff --check: passed.
Implementation verification is complete. The new prompt producers and locale contract tests pass, and focused desktop and mobile E2E coverage confirms the editable preset reaches the task creation flow and retains user edits.
The prompt file is 2,744 bytes, or 2,743 bytes after loading its trailing newline. The rendered ordinary context is 4,394 bytes, versus the recorded 4,804-byte pre-compaction baseline, for a 410-byte reduction. Canvas-enabled context is 4,867 bytes. These measurements are bytes, not model tokens.
Risks
- Removing tool descriptions can erase a mandatory workflow rule. Retain its current condition and test it independently from prose formatting.
- A global canvas flag alone can advertise tools to an excluded profile. Reuse the existing profile resolver and cover negative surfaces.
- Exact text snapshots can hide missing behavior behind frequent rewrites. Assert required instructions and invariant tool names instead.
- Existing models can still ignore instructions. Prompt delivery tests do not guarantee model compliance.
Documentation impact
Internal requirements, designs, and work orders change in this package. Public canvas usage and navigation remain the same. The implementation must check existing public canvas guidance before its final handoff.