[jayjaybeeuk]

Summary

• 5-turn paper-review latency: ~6 min → ~2.5s after the initial load. Every query after set_active_artifact is answered from in-process memory, with no task-bridge round-trip.
• Adds three inline tools: set_active_artifact, query_active_artifact, clear_active_artifact
• Auto-clear on voice-session disconnect prevents cross-session state leaks

Changes

File	What
src/artifact-cache-tools.ts	New module — cache state, extractors, section splitter, query engine
src/inline-tools.ts	Re-export + register the three tools in inlineTools and ownerOnlyTools
src/voice-agent.ts	Call clearActiveArtifact() in onSessionEnd
tests/active-artifact.test.ts	11 tests (all passing)

File-type support

Extension	Extractor	Section split on
.pdf	pdftotext -layout (falls back to raw read if unavailable)	none
.md, .txt	raw read	^# — ^## — ^### headers
.ts, .js, .py, .sh	raw read	class, function, const X = top-level declarations
other	raw read	none

Test plan

• npm test — 173 tests, 0 failures (11 new tests for this feature)
• npx tsc --noEmit — clean
• End-to-end: voice session → set_active_artifact("path/to/paper.pdf") → query_active_artifact("contribution") → verify <500ms response

Closes #642

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.

[qingyun-wu]

Cold review — LGTM in shape. The 80s → 300ms latency claim is a meaningful win for the document-Q&A use case.

Verified from the diff

• `activeArtifact` is module-scoped singleton (one cached doc per voice-agent process) — matches "in-session" framing
• `clearActiveArtifact()` called from voice-agent's `onSessionEnd` — prevents cross-session state leaks (the right invariant)
• Path expansion handles both `~` and `${ENV}` — appropriate for voice-spoken file paths
• 11 new tests, suite passes (per body)

Non-blocking observations

1. Path expansion regex is slightly aggressive: `\${([A-Z_]...)\}|\$([A-Z_]...)` matches uppercase-only env vars. Lowercase ones (rare but legal) won't expand. If a user says `set active artifact ~/Documents/$ws/paper.pdf` and their env has a lowercase `ws`, expansion silently no-ops. Worth a one-line "lowercase env vars not supported" comment, or relax to `[A-Za-z_]`.

2. PDF fallback to "raw read" if pdftotext unavailable — for a binary PDF without pdftotext, reading raw bytes will produce garbage in the query engine. Worth a probe at `set_active_artifact` time that returns an explicit "pdftotext required for .pdf" error instead of silent garbage. The user will get unhelpful answers if the fallback kicks in.

3. Section splitter for code files uses regex on `class`, `function`, `const X =` — works for top-level declarations but will miss `def` in Python, methods inside classes, and arrow functions assigned later. For v1 this is probably enough; flag a followup for "code-aware splitter" if voice Q&A on code becomes a hot path.

4. Cross-session leak via process-level singleton. Single-process voice agent is fine. If the agent ever forks workers or shares the cache across sessions, the singleton becomes a leak — but `onSessionEnd → clearActiveArtifact` already guards. Just worth a code comment that the lifetime contract is "one voice session at a time."

5. Auto-clear edge case. What if the user explicitly calls `set_active_artifact` mid-session for a different file? Looks like the new one overwrites the old (`activeArtifact = ...`) — that's the right behavior, but worth confirming `clear → set` is also exercised by a test.

6. No upper-bound on artifact size. A 100 MB PDF could blow the process's heap. Consider a `MAX_NCHARS = 5_000_000` (~5 MB text) guard in `set_active_artifact` that returns a "too large" error rather than OOM-ing the agent.

Closes #642. The numbers in the body (80s → 300ms) are the right kind of metric to ship with — keep doing that.

Ship after the path/PDF-fallback nits, or merge and follow-up.