Fix typed Context Builder prompt handoff (#526)

[mplibunao]

Summary

• Preserve the caller task/context for typed MCP Context Builder responses when the target prompt is empty.
• Keep discovery-only guidance out of the follow-up model prompt.
• Preserve the committed tab prompt as canonical user state.

Closes #526.

Root cause

MCP instructions were used for discovery, but typed generation later relied only on the committed tab prompt. Empty-prompt commits can contain discovery fallback prose marked with provenance, so the provider correctly rejected that prose but had no independent caller-task prompt to use.

Changes

• Resolve one request-local typed prompt in the MCP provider: a valid committed prompt wins; otherwise sanitized caller instructions are used.
• Strip only exact lowercase discovery-agent guideline blocks and fail closed on malformed, nested, unmatched, case-variant, attribute, self-closing, or quoted-attribute reserved markup.
• Carry the resolved prompt through disposition, execution, export packaging, and the returned result without rewriting tab state.
• Add deterministic resolver and provider-path regression coverage plus two curated test-ledger rows.

Review and refactor gates

• Adversarial review found a quoted-attribute parser bypass; the parser was made quote-aware and the same reviewer confirmed the blocker was closed.
• A separate refactor loop removed unrelated formatting churn, redundant mode/assertion coverage, and one repeated remainder scan. The final follow-up found no further valid refactor.

Validation

• ContextBuilderMCPProgressTimelineTests: 11/11 passed.
• ContextBuilderRunLifecycleTests: 11/11 passed.
• Authoritative test list and curated ledger verification passed: 3,191 executable IDs.
• Repository guardrails passed.
• RepoPrompt product build passed.
• Intended Swift files passed scoped SwiftFormat and strict SwiftLint.
• Commit and push safety preflights passed, including staged/outgoing secret scans.

Local PR-ready limitation

The repository-wide PR-ready lint lane is currently blocked by SwiftFormat 0.62.1 reporting thousands of pre-existing findings in unrelated files. None of the reported findings are in this PR's three changed paths; the changed Swift files pass scoped checks.

[Copilot]

Pull request overview

This PR fixes a Context Builder MCP provider wiring gap where typed responses (plan/question/review) could fail when the target tab’s committed prompt was empty, by resolving a request-local “typed prompt” that prefers an existing committed prompt and otherwise falls back to sanitized caller instructions (while keeping discovery-only guidance out of the follow-up prompt and preserving canonical tab state).

Changes:

• Add a typed prompt resolver that (a) preserves non-empty committed prompts and (b) otherwise sanitizes caller instructions by removing only exact lowercase <discovery_agent-guidelines>…</discovery_agent-guidelines> blocks, failing closed on malformed/variant markup.
• Thread the resolved prompt through follow-up generation and the returned MCP result payload (without rewriting tab prompt state).
• Add regression coverage for provider-path handoff plus deterministic unit coverage for resolver precedence/grammar; update the curated test-suite contract ledger.

Maintainer-guidance check

• User impact and invariant: Typed MCP Context Builder runs should succeed even if the tab prompt is empty, and typed generation must not use discovery output or discovery-only guidance as the effective prompt; tab prompt state must remain the canonical committed state.
• Root-cause confidence: confirmed (matches the described wiring gap between discovery-only instructions and typed follow-up prompt source).
• Authority: MCPContextBuilderToolProvider controls the prompt used for typed generation; tab prompt state remains owned by the workspace/tab store.
• State-safety risks: low; the change explicitly avoids mutating tab state and only affects request-local prompt resolution.
• Scale and observability risks: low; the sanitizer is linear over the provided instructions and only runs when the committed prompt is unusable for typed generation.
• Recommended scope: now (small, coherent fix with targeted tests at the owning boundary).
• Validation boundary: provider-path integration test plus resolver/disposition unit coverage (as added in this PR).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
Sources/RepoPrompt/Infrastructure/MCP/WindowTools/MCPContextBuilderToolProvider.swift	Introduces request-local typed prompt resolution/sanitization and threads resolved prompt through typed follow-up generation and tool result packaging.
Tests/RepoPromptTests/ContextBuilder/ContextBuilderMCPProgressTimelineTests.swift	Adds provider-path regression test for empty committed prompt + caller instructions handoff, plus unit tests for resolver precedence and reserved-markup grammar.
Scripts/Fixtures/test-suite-contract-ledger.tsv	Registers the two new tests in the curated contract ledger with rationale/coverage notes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[baron]

Approved — no blocking findings.

This fixes #526 at the correct authority boundary. The request-local resolver preserves a genuine committed tab prompt when available and otherwise uses sanitized caller instructions, without promoting discovery output or mutating canonical tab state.

The additional handling for <discovery_agent-guidelines> is important because the tool contract explicitly says those hints are discovery-only. The resolver’s fail-closed behavior is conservative, and the provider-path regression test plus adversarial parser coverage give good confidence in the handoff.

Two non-blocking follow-ups worth considering:

• Preserve more specific diagnostics when instructions are rejected or only discovery output exists, rather than reporting the generic “without a prompt” error.
• Add an explicit resolver case for non-empty instructions with an empty effective prompt and usedAgentOutputAsPrompt == false.

Neither concern should block this PR. The current implementation is preferable to forwarding caller instructions verbatim, which would leak discovery-only guidance into the typed follow-up model.

[mplibunao]

Thanks for the thoughtful review and the two concrete follow-ups. Both are addressed in bd15a1e3:

• Typed prompt resolution now preserves specific, non-echoing diagnostics for missing caller input, guidelines-only input, malformed reserved markup, and copied discovery output without an independent caller task.
• The resolver matrix now explicitly covers non-empty caller instructions with an empty effective prompt and usedAgentOutputAsPrompt == false.

The resolver still performs one fail-closed sanitization pass, successful prompt precedence is unchanged, and the existing provider/lifecycle coverage remains green. The hosted pipeline is also fully green (8/8).

[baron]

Approved — the follow-up addresses both non-blocking review points, with no new blocking findings.

The resolver now preserves distinct, safe outcomes for a genuinely missing caller task, discovery-guidelines-only input, malformed reserved markup, and copied discovery output without an independent caller prompt. The resulting diagnostics are materially clearer while retaining the stable “without a prompt” stem and avoiding disclosure of withheld instruction content.

The added resolver case also pins the previously missing sibling path: non-empty caller instructions with an empty effective prompt and usedAgentOutputAsPrompt == false.

I reviewed the incremental change from 97efadc to bd15a1e; it addresses the requested follow-ups cleanly, and the current CI checks are green.

[baron]

@mplibunao thank you so much for the awesome issues and this PR. Also thank you for addressing the optional follow up. We intend to ship this in the next release.