PDPP

Tasks

tasks33/33
Created openspec/changes/complete-mcp-read-evidence-ladder/tasks.mdView on GitHub →

0. Contract

  • 0.1 Capture the plan in the permanent docs/research/ corpus.
  • 0.2 Add OpenSpec proposal, design, tasks, and spec deltas.
  • 0.3 Validate complete-mcp-read-evidence-ladder --strict.

1. Hostile-Client Tests

  • 1.1 Add MCP search coverage where structuredContent is ignored and the visible result still carries a bounded matched text window.
  • 1.2 Add coverage proving visible field-window continuations include a model-callable tool path, not only a resource URI.
  • 1.3 Add a negative control where metadata-only search hits do not invent a body match.
  • 1.4 Add small-text read/fetch coverage for inline content without requiring file/resource materialization.
  • 1.5 Add binary/large-text coverage that preserves bounded metadata and explicit continuation.
  • 1.6 Add hosted-client structured-preview regression coverage where content[] may be ignored but matched text evidence remains visible in structuredContent.

2. Resource-Server Evidence

  • 2.1 Identify current search hit shapes for lexical, semantic, and hybrid search.
  • 2.2 Surface proven lexical match-window metadata from the resource server matched text field.
  • 2.3 Keep proven match windows explicit; adapters do not infer body/text fields.

3. Shared Evidence Primitives

  • 3.1 Use adapter-neutral content-ladder field-window helpers.
  • 3.2 Preserve explicit continuation metadata.
  • 3.3 Keep binary/base64/blob fields metadata-only by default.

4. MCP Adapter

  • 4.1 Render proven match windows in visible output within a strict budget.
  • 4.2 Include model-callable read_record_field continuation hints.
  • 4.3 Return bounded field windows with truthful truncation and match metadata.
  • 4.4 Return ordinary small projected fetches inline.
  • 4.5 Preserve resource support for capable clients.
  • 4.6 Hide field-window resource URIs from model-visible output so hosted clients do not see dead handles.

5. CLI and Documentation

  • 5.1 Keep CLI/read-evidence field-window behavior aligned with shared primitives.

6. Verification

  • 6.1 Run focused packages/read-evidence tests.
  • 6.2 Run focused hostile MCP tests.
  • 6.3 Run resource-server lexical match-window regression.
  • 6.4 Run full MCP server tests.
  • 6.5 Run reference implementation typecheck.
  • 6.6 Run OpenSpec validation for this change and all active specs.
  • 6.7 Run git diff --check.
  • 6.8 Prepare LAND/HOLD summary with residual hosted-client smoke status.
  • 6.9 Rerun full gates and live-backed MCP smoke after structured-preview fix.
  • 6.10 Rerun full gates and live-backed MCP smoke after hiding model-visible field-window resource URIs.
  • 6.11 Add hostile-client coverage proving visible pdpp://record/... handles are accepted by read_record_field.id.