[ScriptedAlchemy]

Problem

Project intelligence existed in separate UI, MCP, provider, and delivery paths. That let active steering bypass knowledge resolution, retries depend on a paginated inspector, remote identifiers lose their environment, and provider settings offer operations a selected provider could not execute.

Changes

• Use one knowledge-delivery lifecycle for new turns and active steering: source validation, context budgets, durable preparation, dispatch, and sent/failed/uncertain receipts. Safety decisions query receipt status directly, independently of UI pagination and evidence decoding; crash recovery follows the same boundary.
• Add migration 081 to identify receipts by (attempt_id, message_id), preserve existing data, and keep forgetting and maintenance cursors correct for multiple messages in an attempt.
• Declare text-generation capabilities per operation on each provider instance, publish them through provider snapshots, enforce them at execution, and use them in model selection. Muse implements knowledge extraction and conversation summaries through the shared read-only runner. OpenCode v2 advertises its working MCP transport.
• Route the 12 agent-facing intelligence MCP tools through the existing environment target and check source and destination capabilities before code-intelligence forwarding. Keep Loom in the app: its web/desktop/mobile task and parent navigation preserves environment identity, while the Loom MCP tool, provider allowlist entry, agent guidance, and replay scenario are removed.
• Read complete summary source coverage in bounded pages, avoid stale summaries consuming the handoff limit, and fetch handoff evidence only for the requested messages.
• Add concise shared agent instructions and 6 intelligence replay cases covering recall, citations, code impact, summaries, remote scope, and ordinary edits.
• Use the pinned running Node for native CLI executable targets, avoiding the redundant Windows Node archive extraction while preserving pinned cross-platform builds.

Includes current main (71206f5f4e34), preserving PR #18's linked-environment gates, remote approvals, and context-exhaustion recovery.

Review follow-up fixes budget knowledge steering against the active native conversation after resume recovery, without counting the pending message twice. Dedicated review-guide models now require generateReviewGuide support in every targeted environment and select a shared capable fallback when necessary. Loom and settings fixtures also follow the current client APIs.

Validation

On Node 26.12.1:

• Final review and current-main integration: vp run check:changed --serial passed the native project-memory addon build, format, lint, both knip passes, GraphQL consistency, typechecking for 32 affected packages, and 1,313 tests across 43 files, including Chromium settings regressions. The native-history and model-selection regressions failed before their fixes. The final working tree and git diff --check were clean.

• Initial integration: vp run check:changed --serial passed format, lint, both knip passes, generated GraphQL, typechecking for 32 affected packages, and 632 tests across 38 files (including automated Chromium tests).

• Loom-tool removal: vp run check:changed HEAD --serial passed format, lint, both knip passes, generated GraphQL, typechecking for 23 affected packages, the native project-memory addon build, and 299 tests across 5 files, including the exact context measurement.

• Start/control/provider-switch integration: 123/123 passed after rebasing onto PR fix: merge upstream agent links and open remote approvals #18.

• Agent replay graders: 13/13 passed; all 6 remaining intelligence cases construct valid Claude and Codex dry-run requests, with no Loom tool in either payload.

• Exact agent-context measurement regression: 1/1 passed after recording measured sizes; the measurement assertion and workflow remain unchanged.

• CI follow-up: updated both preview-upgrade migration expectations to include migration 81. All 5 preview-upgrade tests pass, including rollback/retry and import preservation. vp run check:changed HEAD --serial also passed format, lint, both knip passes, generated GraphQL, typechecking for 3 affected packages, and the native addon build.

Additional verification includes migration rollback/data preservation, real SQLite delivery and restart cases, multi-environment authorization and navigation, provider capability enforcement, summary coverage beyond 500 records, and React-compiler-compatible mobile subscriptions.

Agent context measurements

Removing the Loom tool saves 1,650 bytes of MCP tool schemas and 105 bytes from each shared instruction payload compared with the first PR revision. The exact measurement regression records these UTF-8 byte counts:

Measurement	Main	This PR	Change
Claude system append	6,576	7,975	+1,399 (+21.27%)
Codex additional context	8,304	9,701	+1,397 (+16.82%)
Registered MCP tool JSON	61,366	60,381	-985 (-1.61%)

The maintainer-approved agent-context-increase label is applied for the remaining instruction growth; the measurement assertion and approval workflow are unchanged. Registered tool-schema context is smaller than main. These are byte counts, not token estimates.

Limits

Live model evals were not run because provider API keys are unavailable; replay request construction, schemas, and graders were checked. The complete Windows installer has not been built on this Linux host; its target-selection regression is tested and Windows CI remains the native validation.

Implemented and verified with GPT-6 Astra Pro in the Codex harness.

[changeset-bot]

⚠️ No Changeset found

Latest commit: cb98021

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR