feat(extensions): restructure LSP feedback pipeline and trim tool surface

Consolidate the LSP subsystem into a staged feedback pipeline with clearer
diagnostic auto-fix flow, drop `codegraph_status` and `lsp_rename`, and
tighten notifier, subagent, tui, web-search, and yeet extension code.

Update AGENTS.md to reflect the trimmed tool list and new shared
agent-status vocabulary module.

Author: polo871209

.pi/agent/SYSTEM.md

@@ -2,10 +2,7 @@
 
 - **Ambiguous** (underspecified, multiple valid readings, or assumption user wants say in): stop, use ask_user_question. NEVER guess, drift.
 - **Non-mutation**: execute immediately, no asking.
-- **Mutation**: print + wait (user runs unless they say "you do it"):
-  - Cluster/cloud writes: `kubectl apply/delete/patch`, `helm install`, `terraform apply`, `gcloud/aws create/delete/update`
-  - Destructive: `rm -rf`, `docker push`, `npm publish`, db migrations
-  - Installs: `pi install`, `npm install -g`, `brew install/uninstall`
+- **Mutation**: print + wait (user runs unless they say "you do it"). Includes cluster/cloud writes, destructive ops (incl. publish/push to registries, db migrations), installs (any package manager, incl. global/system).
 - **Production**: NEVER mutate. Print only. If unsure target is prod, ask.
 
 ## Code
@@ -15,14 +12,10 @@
 - No features, abstractions, or configurability beyond what was asked.
 - NEVER "improve" or refactor adjacent code that isn't broken. Match existing style.
 - Remove imports/vars/functions YOUR changes orphaned. NEVER touch pre-existing dead code; mention it instead.
-- Comments: short, WHY not WHAT; none if obvious. No history ("Replaces…", "Legacy…"); write current state. No decorative dividers (`# ====`, banners) or em-dashes (—).
+- Comments: short, WHY not WHAT; none if obvious. No history ("Replaces…", "Legacy…"); write current state. No decorative dividers (`# ====`, banners).
 
 ## Tone
 
-Respond terse like smart caveman. Substance stays, fluff dies. Active every response, every turn, even if unsure; NEVER drift back to filler. Off when "normal mode".
+Be extremely concise. Sacrifice grammar for the sake of concision.
 
-Drop: articles, filler (just/really/basically), hedging, apologies, self-narration ("Let me..."), pleasantries and sycophantic openers ("Of course", "You're right", "Great question"). NEVER restate the question. Fragments OK. Short synonyms (big not extensive, fix not "implement a solution for"). No tool-call narration, no decorative tables/emoji, no long raw error dumps; quote shortest decisive line. Technical terms, code blocks, errors verbatim.
-
-Preserve user's language; compress style, not language. NEVER name or announce the style.
-
-Pattern: `[thing] [action] [reason]. [next step].`
+Drop: articles, filler, hedging, apologies, self-narration ("Let me..."), sycophantic openers ("You're right"). NEVER restate the question. Fragments OK. Short synonyms (fix not "implement a solution for"). No tool-call narration, no decorative tables/emoji/em-dashes (—), no raw error dumps; quote shortest decisive line. Technical terms, code blocks, errors verbatim.

.pi/agent/extensions/AGENTS.md

@@ -30,8 +30,8 @@ Read referenced `.md` files completely and follow their cross-references before
 
 - **`web-search.ts`** — `web_search` / `fetch_content`: search the web and pull a page's readable content as markdown, for basic research without leaving the terminal. GitHub repo/file/dir links are cloned locally instead of scraped, so `read`/`bash` can explore real files.
 - **`eval/`** — persistent Python + JS kernels the model runs code in; cells can call pi's own tools and keep bulk data out of history. For "read N files, aggregate, summarize" work.
-- **`lsp/`** — the LSP subsystem. Symbol-precise nav tools: `lsp_hover` (type/docs), `lsp_definition` / `lsp_type_definition` / `lsp_implementation`, `lsp_references`, `lsp_document_symbols` (file outline), `lsp_rename` (workspace-wide), `lsp_diagnostics` (on-demand, read-only per-file error/warning check instead of a full `tsc`). Also the post-edit feedback pass (`lsp/feedback/`): formats your edits and auto-fixes their diagnostics in the background (root-cause, no suppress directives, never touching files unrelated to a diagnostic), surfacing the changes so you needn't re-read; anything left unfixed is flagged. `/lsp-fix` toggles that background auto-fix per session (`/lsp-fix on|off` to set explicitly); launch with `--lsp-fix=false` to default it off.
-- **`codegraph.ts`** — symbol-aware repo navigation + call-graph over the [codegraph CLI](https://github.com/colbymchenry/codegraph): `codegraph_status` / `_context` / `_search` / `_files` / `_callers` / `_callees` / `_impact` (blast-radius) / `_affected` (test selection).
+- **`lsp/`** — the LSP subsystem. Symbol-precise nav tools: `lsp_hover` (type/docs), `lsp_definition` / `lsp_type_definition` / `lsp_implementation`, `lsp_references`, `lsp_document_symbols` (file outline), `lsp_diagnostics` (on-demand, read-only per-file error/warning check instead of a full `tsc`). Also the post-edit feedback pass (`lsp/feedback/`): formats your edits and auto-fixes their diagnostics in the background (root-cause, no suppress directives, never touching files unrelated to a diagnostic), surfacing the changes so you needn't re-read; anything left unfixed is flagged. `/lsp-fix` toggles that background auto-fix per session (`/lsp-fix on|off` to set explicitly); launch with `--lsp-fix=false` to default it off.
+- **`codegraph.ts`** — symbol-aware repo navigation + call-graph over the [codegraph CLI](https://github.com/colbymchenry/codegraph): `codegraph_context` / `_search` / `_files` / `_callers` / `_callees` / `_impact` (blast-radius) / `_affected` (test selection).
 - **`github-pr.ts`** — `github_pr` fetches a PR as signal-only markdown (metadata, description, changed files, failing checks, unresolved review threads — including bot inline findings like CodeRabbit). Drops commit/timeline noise, resolved threads, and bot release-note/walkthrough issue comments; diff is opt-in via `diff:true`. Use instead of `gh pr view`.
 - **`subagent.ts`** — `/subagent` delegates a task to an isolated `pi` agent (single-layer, no recursion), running visibly in its own tmux pane instead of a hidden background process — watch, scroll, or step in directly. Agents defined in `~/.pi/agent/agents/*.md` — each file's body is the subagent's entire system prompt, no shared preamble or other context added, for an unambiguous clean start; `hidden: true` agents stay invocable by name (e.g. from a skill) without appearing in the tool's agent list. For offloading research/recon/implementation off the main thread. tmux-only: the tool isn't available outside a tmux session.
 - **`ask/`** — `ask_user_question`: presents a tabbed multiple-choice questionnaire (single/multi-select, free-text + "chat" fallbacks, review tab) instead of guessing when a request is ambiguous. Length limits on labels are soft so the first call always lands.
@@ -69,5 +69,5 @@ extensions/
 ├── ask/          ask_user_question questionnaire dialog
 ├── eval/         persistent kernels + bridge
 ├── lsp/          headless nvim, nav tools + post-edit feedback
-└── shared/       side-channel LLM helper, message extraction, widget factory
+└── shared/       side-channel LLM helper, message extraction, widget factory, agent-status vocabulary
 ```

.pi/agent/extensions/ask/index.ts

@@ -55,20 +55,15 @@ function clampParams(params: QuestionParams): QuestionParams {
 
 const DESCRIPTION = `Present the user with one or more structured multiple-choice questions and wait for their answer. This is the PRIMARY way to resolve ambiguity — reach for it directly instead of guessing or asking in plain prose, and call it in the SAME turn you hit the ambiguity (don't first write out the question as text).
 
-Call it when:
-1. The request is underspecified and you cannot proceed without a concrete decision.
-2. Multiple reasonable interpretations or implementation paths exist.
-3. You're about to make an assumption the user would want a say in.
-4. You want to offer a clear set of directions to take.
+Call it when the request is ambiguous (underspecified, multiple valid readings, or an assumption the user would want a say in), or to offer a clear set of directions to take.
 
 Mechanics:
 - Up to ${MAX_QUESTIONS} questions per call, each with ${MIN_OPTIONS}-${MAX_OPTIONS} options. Group all clarifying questions into ONE call — never stack back-to-back calls.
-- Each option needs a concise label and a description of what it means / its trade-off.
-- header / label length limits are soft: over-long values are auto-truncated, NEVER rejected — don't avoid the tool or pad/trim to fit.
-- The user can always type a custom answer ("Type something." is appended automatically to single-select questions) or pick "Chat about this" to abandon the questionnaire and talk it through.
-- Set multiSelect: true when multiple answers are valid (suppresses the "Type something." row).
-- If you recommend an option, make it the FIRST option and append "(Recommended)" to its label.
-- Do NOT author options labeled "Other", "Type something.", "Chat about this", or "Next" — these are reserved and rejected.`;
+- header / label length limits are soft: over-long values auto-truncate, NEVER rejected — don't avoid the tool or pad/trim to fit.
+- Single-select questions get a free-text "Type something." row automatically; the user can also pick "Chat about this" to abandon and talk it through.
+- Set multiSelect: true when multiple answers are valid (suppresses the free-text row).
+- If you recommend an option, make it FIRST and append "(Recommended)" to its label.
+- Do NOT author options labeled "Other", "Type something.", "Chat about this", or "Next" — reserved, rejected.`;
 
 export function registerAskUserQuestionTool(pi: ExtensionAPI): void {
   pi.registerTool({

.pi/agent/extensions/ask/schema.ts

@@ -155,7 +155,6 @@ export const QuestionParamsSchema = Type.Object({
   }),
 });
 
-export type OptionData = Static<typeof OptionSchema>;
 export type QuestionData = Static<typeof QuestionSchema>;
 export type QuestionParams = Static<typeof QuestionParamsSchema>;
 

.pi/agent/extensions/codegraph.ts

@@ -56,22 +56,6 @@ type Node = {
 const loc = (n: Node) => `${n.filePath}:${n.startLine}`;
 const nodeRow = (n: Node) => `${n.name}  ${n.kind}  ${loc(n)}`;
 
-function fmtStatus(s: Record<string, unknown>): string {
-  const kinds = Object.entries((s.nodesByKind as Record<string, number>) ?? {})
-    .map(([k, v]) => `${k}=${v}`)
-    .join(" ");
-  const p = (s.pendingChanges as Record<string, number>) ?? {};
-  return [
-    `files=${s.fileCount} nodes=${s.nodeCount} edges=${s.edgeCount}`,
-    `languages=${((s.languages as string[]) ?? []).join(",")}`,
-    kinds && `kinds: ${kinds}`,
-    `pending: +${p.added ?? 0} ~${p.modified ?? 0} -${p.removed ?? 0}`,
-    s.lastIndexed && `lastIndexed=${s.lastIndexed}`,
-  ]
-    .filter(Boolean)
-    .join("\n");
-}
-
 function fmtQuery(
   arr: Array<{ node: Node & { isExported?: boolean } }>,
 ): string {
@@ -182,19 +166,6 @@ export default function (pi: ExtensionAPI) {
     child.unref();
   });
 
-  pi.registerTool({
-    name: "codegraph_status",
-    label: "CodeGraph status",
-    description:
-      "Inspect codegraph index: file count, node count, languages, pending changes.",
-    parameters: Type.Object({}),
-    async execute(_id, _raw, signal, _onUpdate, ctx) {
-      return callCodegraph(["status", "-j"], ctx.cwd, signal, (j) =>
-        fmtStatus(j as Record<string, unknown>),
-      );
-    },
-  });
-
   pi.registerTool({
     name: "codegraph_context",
     label: "CodeGraph context",

.pi/agent/extensions/eval/index.ts

@@ -345,7 +345,6 @@ export default function (pi: ExtensionAPI) {
     label: "Eval",
     description:
       "Run code in persistent Python and JavaScript kernels for iterative, stateful work — state persists across cells and across separate tool calls, in one kernel per language per session. Not for a single one-off command; use `bash`/`read` directly for those.\n\n" +
-      'Cell shape: `{language: "py"|"js", code, title?, timeout?, reset?}`.\n\n' +
       'Inside cell code, `tool.<name>({...})` invokes pi built-in tools (`read`, `write`, `edit`, `bash`, `grep`, `find`, `ls` take their normal pi argument schemas; `tree({path,max_depth})` is extra). Shortcuts: `read(path)`, `write(path,content)`, `tree(path)`, `env(key?, value?)` (no args: full env dict; one: get; two: set), `completion(prompt, model?, system?, schema?)` for a oneshot stateless model call (model: "default" or "provider/id"; schema: JSON-Schema for structured output).\n\n' +
       'Call `install("pkg1", "pkg2")` to add Python packages, persists across pi restarts. JS cells support top-level await and package imports; use `globalThis` / `state` to persist values across cells.',
     promptSnippet:

.pi/agent/extensions/eval/types.ts

@@ -2,14 +2,6 @@
 
 export type Language = "py" | "js";
 
-export interface Cell {
-  language: Language;
-  code: string;
-  title?: string;
-  timeout?: number;
-  reset?: boolean;
-}
-
 export interface DisplayItem {
   mime: string;
   data: string;
@@ -51,9 +43,7 @@ export interface KernelEventDone {
   error: string | null;
 }
 export type KernelEvent =
-  | KernelEventDisplay
-  | KernelEventStream
-  | KernelEventDone;
+  KernelEventDisplay | KernelEventStream | KernelEventDone;
 
 // Bridge protocol — Python prelude -> host over loopback HTTP.
 export interface BridgeRequest {

.pi/agent/extensions/folder-context.ts

@@ -28,7 +28,10 @@ export default function (pi: ExtensionAPI) {
     injected.clear();
   });
 
-  pi.on("tool_call", async (event, ctx) => {
+  // Hook tool_result, not tool_call: a steer message sent during tool_call
+  // lands between the assistant tool_use and its tool_result, which Anthropic
+  // rejects (400: tool_result must follow its tool_use).
+  pi.on("tool_result", async (event, ctx) => {
     if (!TARGET_TOOLS.has(event.toolName)) return;
     const rawPath = (event.input as { path?: unknown }).path;
     if (typeof rawPath !== "string" || rawPath === "") return;

.pi/agent/extensions/github-pr.ts

@@ -131,7 +131,7 @@ export default function (pi: ExtensionAPI) {
       return new Text(theme.fg("dim", `  ${s}`), 0, 0);
     },
     description:
-      "Fetch a GitHub PR (URL or number) as signal-only markdown: metadata, description, changed files, failing checks, and unresolved review threads. Use instead of `gh pr view`. diff:true to include the unified diff.",
+      "Fetch a GitHub PR (URL or number) as signal-only markdown: metadata, description, changed files, failing checks, and unresolved review threads. Use instead of `gh pr view`.",
     parameters: params,
     async execute(_id, raw, signal, _onUpdate, _ctx) {
       const a = raw as {

.pi/agent/extensions/go.ts

@@ -104,7 +104,7 @@ async function go(ctx: ExtensionContext): Promise<void> {
   }
 }
 
-// --- Stream rules (TTSR-lite) -----------------------------------------
+// Stream rules (TTSR-lite):
 //
 // A regex hits the model's streaming text/thinking output → the turn aborts
 // → a hidden reminder is appended as a user turn → generation resumes via