Stop sending anything about the running app to the LLM

[alexkroman]

What

config.llm.instruction carried a formatting clause derived from the frontmost app. AppKindPriming recognized the app's bundle ID as a kind — terminal, code editor, Slack, Obsidian — and sent a matching imperative:

• "Format the result as a shell command with no trailing period."
• "Format the result as Python code." (language inferred from the window title's filename)
• "Format the result as a casual Slack message, using Slack emoji where they fit."
• "Format the result as markdown."

This removes that entirely. Nothing identifying or describing the destination app reaches the service.

How

The whole chain goes, not just the last link — a partial removal would leave dead code for periphery to fail on:

Removed	Note
AppKindPriming + its tests	the bundle-ID table, extension→language map, and clause wording
TranscriptionSteering.Fields.rewriteInstruction	build now renders exactly two fields: conversation_context, keyterms_prompt
TranscriptionContext.bundleID	the clause was its only consumer
CapturedFocus.bundleID	no longer read from NSWorkspace.frontmostApplication
DictationLog's llm_instruction column

LLMRewrite now encodes as {}. The llm block still rides on every request — it's what asks the service to run the cleanup rewrite at all — so the service's own default cleanup instruction applies to every utterance. That's the same wire shape an unrecognized app already produced, so this is not a new request shape.

windowTitle stays: it anchors KeyInjector's same-window paste-separator fallback.

Regression pins

• TranscriptionSteeringTests keeps a case per app family that used to earn a clause (terminal, code editor with a main.py title, Slack, Obsidian), each expecting .empty.
• TranscriptionContextTests moves those contexts into the carry-only set: non-empty (worth carrying for the injector) but steering nothing.
• AssemblyAITranscriberTests pins llm == {} for every steering shape, not just the empty one.
• DictationLogTests pins that llm_instruction is never written, and that "Terminal" never lands in the log.

Docs

AGENTS.md (start-here summary, repo map, settled-decisions table, transcription-steering section), BLURTENGINE.md, README.md, and .claude/skills/project-guardrails all updated. Both the settled-decisions table and the guardrails skill gain a "never send anything about the frontmost app" row so this doesn't come back.

Verification

scripts/check.sh — 425 unit tests / 67 suites, 23 UI tests, TSan + ASan, leaks, xcodegen drift, periphery, swift-format, swiftlint, actionlint, markdownlint, shellcheck: all green. Prettier flagged an AGENTS.md table reflow, applied here. (Its two other warnings are gitignored .remember/ scratch files, untouched by this change.)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

This PR removes all “app-kind” priming so nothing identifying or describing the destination/frontmost app is sent to AssemblyAI’s LLM rewrite path. It simplifies the steering model to only send prior-cursor text (conversation_context) and user key terms (keyterms_prompt), while always requesting the default server-side cleanup rewrite via an empty llm block.

Changes:

• Deleted AppKindPriming (engine + tests) and removed bundle-ID capture/plumbing (CapturedFocus, TranscriptionContext, DictationSession).
• Simplified TranscriptionSteering.Fields to only conversationContext + keyterms, and ensured AssemblyAITranscriber always encodes llm as {}.
• Updated tests and docs to pin “no app-derived steering/logging” behavior (and removed llm_instruction from the dictation log).

Reviewed changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
Tests/BlurtEngineTests/TranscriptionSteeringTests.swift	Updates steering expectations to .empty for app-family contexts; removes instruction-based assertions.
Tests/BlurtEngineTests/TranscriptionContextTests.swift	Removes bundleID-driven non-emptiness behavior; adds carry-only cases that must steer nothing.
Tests/BlurtEngineTests/DictationLogTests.swift	Removes llm_instruction decoding/assertions; pins that app-identifying strings never land in logs.
Tests/BlurtEngineTests/AssemblyAITranscriberTests.swift	Pins llm == {} for multiple steering shapes and removes instruction assertions.
Tests/BlurtEngineTests/AppKindPrimingTests.swift	Deletes tests for removed AppKindPriming.
Sources/BlurtEngine/STT/TranscriptionSteering.swift	Drops rewrite-instruction field and removes app-kind clause generation.
Sources/BlurtEngine/STT/TranscriptionContext.swift	Removes bundleID from captured context and updates invariants/docs accordingly.
Sources/BlurtEngine/STT/AssemblyAITranscriber.swift	Always encodes llm as an empty object ({}) and removes instruction wiring.
Sources/BlurtEngine/STT/AppKindPriming.swift	Deletes AppKindPriming implementation.
Sources/BlurtEngine/Pipeline/DictationSession.swift	Stops capturing/passing bundle IDs into TranscriptionContext.
Sources/BlurtEngine/Pipeline/DictationLog.swift	Removes llm_instruction from on-disk JSONL entries.
Sources/BlurtEngine/FocusCapture/FocusCapture.swift	Stops capturing the frontmost app’s bundle identifier.
README.md	Updates high-level architecture note to remove “format” steering mention.
BLURTENGINE.md	Updates engine API/design docs to reflect removal of app-kind formatting and llm_instruction logging.
AGENTS.md	Updates repo guide/settled decisions and steering docs to match new privacy posture.
.claude/skills/project-guardrails/SKILL.md	Adds/updates guardrail explicitly forbidding sending frontmost-app info.

---

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