feat(agent): agent modes v1a (Fast / Normal / Deep Research) + hybrid vault_search #1
hegdeatri
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "feature/agent-modes-v1a"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Three modes for the agent now: Fast, Normal, Deep Research. Each picks a tool profile and a round budget rather than the one-shot workflow we had. New hybrid
vault_searchtool does FTS + semantic in parallel, fuses with RRF, dedups by note, reranks on title/tag/recency. Title generation moved off the full agent path onto a separatePOST /titleso a typical first turn stops paying for tools that wouldn't have fired anyway.Spec at
docs/superpowers/specs/2026-05-17-agent-modes-design.md. v1b (router, fast path, telemetry, streaming Deep Research progress) is a separate spec — intentionally not in this PR.Budgets
When the Obsidian CLI isn't configured, Deep Research silently falls back to the Analysis profile —
obsidian_*tools just don't appear in the catalog, agent loop doesn't have to know.UI
Backward compat
ChatRequest.agent_mode,AppSettings.default_agent_mode, andChatSession.agentModeall default to Normal when missing. Existing settings and threads keep working.Checks
bun run lint,bun run build— cleancargo check --all-targets,cargo clippy --all-targets— clean (three pre-existing rust-1.94 transmute warnings indb/store.rs, not from this branch)cargo test --lib— 62/62 passing, includingmode_budgets_match_table,profile_for_*,vault_search_*, and the two_deserializes_*casesReplaces the three near-identical .tool() chains (one per ProviderKind arm) with a single register_tools(builder, profile, &ctx, budgets) call in each has_vault branch — ~75 LOC of duplication collapse to ~3 per arm. create_chat_stream takes AgentMode and resolves profile, budgets, and ToolContext once up front; effective_max_rounds prefers an explicit caller override over the mode budget. obsidian_configured is computed from vault_path (settings-only, no DB read) and threaded into both profile_for and build_system_prompt so the system prompt's catalog cannot lie about which tools are registered. has_vault still gates actual registration on note_count. The legacy tool_backend parameter stays on the signature for backward compatibility (frontend still sends it); v1b removes it. The /chat route handler hard-codes AgentMode::Normal — Task 8 wires the per-request resolution from ChatRequest. Cleanup: - Strip Task 2/5 #[allow(dead_code)] from ToolProfile, profile_for, tool_names_for, register_tools — they have a real caller now. - Annotate ToolProfile::None with #[allow(dead_code)] until Task 9 /title endpoint lights it up. - #![allow(dead_code)] on agent/tools/obsidian/{search,read,tags}.rs; vault_search replaces obsidian_search and read_note/list_tags cover the others. v1b decides whether to delete the modules outright. - #[allow(clippy::too_many_arguments)] on create_chat_stream; future refactor can bundle args into a struct. Defers generate_once helper to Task 9 (its first real consumer). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>clippy::manual_repeat_n flagged the `repeat("?").take(n)` pattern in vault_search::query_note_meta after the MSRV bump made repeat_n available. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>generate_once builds a no-tools agent per provider arm, streams it, accumulates TextDelta chunks into a String. Thinking / reasoning deltas dropped so reasoning-prefix models (DeepSeek-R1, Qwen3-thinking) return clean output. Streaming errors surface as AppError::ProviderError. POST /title takes {message, history?}, returns {title}. Handler guards empty/whitespace input by returning {title: ""} instead of hitting providers (Rig rejects empty user messages). clean_title strips surrounding + embedded quotes (titles shouldn't quote themselves; embedded apostrophes lost — acceptable for short titles, documented), trims terminal .!?, collapses whitespace. Deviation from spec: title prompt is passed as system preamble + real user message rather than spec's "everything in preamble, empty user message" — matches the create_chat_stream pattern in this file and avoids the empty-message risk the spec itself flagged. Tests: clean_title rules (incl. empty-string + adversarial all-punctuation cases), TitleRequest serde shape (with + without history), TitleResponse wire shape. AGENTS.md gains /title route entry. Frontend swap to this endpoint lands in Task 11.- ChatSession gains optional agentMode (mirrors existing toolBackend?). Inline literal type so mock-data.ts stays free of chat-client.ts imports — keeps the dependency direction one-way. - draftAgentMode / setSessionAgentMode parallel the toolBackend story: hold the user's choice in hook-local state for the draft session, promote to the new session on first send. - Mode resolution: session.agentMode ?? settings.default_agent_mode ?? "normal". Threaded through both sendMessage and retryMessage into beginStream, then into streamChat's options bag (6th arg). - Title path switched from legacy generateTitle (which went through /chat) to requestTitle (POST /title — no tools, no SSE). Fired fire-and-forget on first-message complete, same timing as before. The empty-title behavior is intentionally improved: legacy generateTitle would unconditionally overwrite the slug title even with "" (rare LLM edge), blanking the thread. The new path keeps the slug fallback ("text.slice(0, 50)") when requestTitle returns the documented empty-string-on-blank-input. Visible-only on the unlikely path; no user-facing regression. generateTitle is now unused inside useChat. Left in chat-client.ts as dead code for v1b housekeeping — pruning it touches the import graph and is best done after Task 14 manual smoke confirms /title parity. The Composer (Task 12) and Settings (Task 13) will consume the new draftAgentMode / setSessionAgentMode / default_agent_mode surface. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>Surfaces settings.default_agent_mode in the UI so newly-created chat threads can inherit a non-Normal default. Mirrors the existing default-tool-backend picker pattern: three-button .seg radiogroup with aria-checked + immediate updateSettings() persistence. Per-button hint copy ("up to N tool rounds") matches the AgentMode::budgets() constants in src-tauri/src/config.rs (Fast=1, Normal=3, DeepResearch=6) — quantitative register parallels the tool-backend's "~600 ms / call" / "<30 ms / call" annotations. Placement: src/components/pkma/sections/indexer.tsx::StatusPanel, adjacent to the default-tool-backend picker. Spec text says "or the settings section that owns the default-tool-backend picker"; that's StatusPanel, not models.tsx (which uses a draft/save form pattern that would be heavier to wire for two booleans). Section hint updated from "Progress, reindexing, and default tool backend" to "Progress, reindexing, and chat defaults for new threads" so the section header no longer lies about its contents. TODO.md gets a v1b polish note: both chat-default pickers should ultimately move into a dedicated "Chat defaults" section — agent mode has nothing to do with indexing, and the current home will surprise the next refactor. bun run lint clean. bun run build clean. 12 of 14 v1a tasks shipped (Task 13). Remaining: Task 14 manual smoke + plan sign-off.Task 14 — automated end-to-end audit clean: - bun run lint: clean - bun run build: clean (Next 16 static export, 3 static pages) - cargo check --all-targets: clean - cargo clippy --all-targets: clean (3 pre-existing warnings in src/db/store.rs from rust-1.94's clippy::missing_transmute_- annotations; not introduced by this branch) - cargo test --lib: 62 passed, 0 failed, 3 ignored, including mode_budgets_match_table, profile_for_*, vault_search_*, chat_request_deserializes_without_agent_mode, app_settings_without_default_agent_mode_deserializes_to_normal End-to-end agent_mode trace verified (no orphan hardcoded Normal): composer.tsx chip → useChat.tsx (agentMode resolution before send) → chat-client.ts streamChat options.agentMode → body.agent_mode → server/routes.rs ChatRequest.agent_mode (#[serde(default)]) → mode = body.agent_mode.unwrap_or(settings.default_agent_mode) → providers.rs create_chat_stream(mode, ...) → register_tools(builder, profile_for(mode, obsidian), ctx, budgets) → mode.budgets().max_tool_rounds as loop cap (overridable per request via max_tool_rounds_override) /title route wired: server/mod.rs `.route("/title", post(routes::title))` → routes::title → providers::generate_once (no tools, no SSE). useChat.tsx:601 calls requestTitle on first message. Drops dead generateTitle from chat-client.ts (zero consumers after Task 11 migrated useChat to requestTitle — the SSE-based title path was the v1a deliverable being replaced, not v1b housekeeping). Spec status flipped to "v1a shipped 2026-05-18 (awaiting merge)". TODO.md Tier 2 v1a checkbox flipped; commit roster updated; v1b polish items recorded (Chat-defaults section extract; rust-1.94 transmute annotations on sqlite3_auto_extension call sites). Manual smoke (live Tauri window, real vault, real provider) is the user's call to drive. Reported separately when run. 14 of 14 v1a tasks shipped.