Fix the loop stalling after compaction, add an external registry

The compaction bug, which is the important one:

beta.2 taught the pruner to drop any assistant part whose reasoning item it had
removed. That was right about the 400 and wrong about everything else. On a
reasoning model every tool call carries a provider itemId, so past the threshold
the model could no longer see what it had already run, and re-ran the same tools
until maxSteps ended the turn. Reproduced at 12 model calls for a job needing 4,
with nothing but the user message reaching the wire.

The dependency is not the part, it is the itemId. A part carrying one is
serialised as `{ type: 'item_reference', id }`, a pointer to an item stored
provider-side that depends on its reasoning item. Without the itemId the same
content goes out inline and carries no dependency at all. Verified against the
provider's own serialiser: `text` with an itemId becomes item_reference, the
identical part without one becomes output_text.

So `dropOrphanedItems` becomes `detachOrphanedItems`: strip the itemId, keep the
content. Compaction may shorten the history; it must not blank it. The new test
asserts behaviour rather than shape — the loop must end because the model chose
to, and every call after the first must still carry the earlier exchange. A shape
assertion passed the whole time the model was losing its memory.

Registry, via `/registry [list|search|add|remove|installed]`:

Skills and plugins are treated differently on purpose. A skill is prompt text, so
installing one puts a stranger's words into the system prompt of every future
session in this project; the install shows the body first and the origin is
recorded, so /skills always says where an instruction came from. A plugin is a
JSON manifest of deny rules, evaluated by compiled code identical for every
install. Loading TypeScript from a URL is declined outright: a plugin that can
block tool calls could otherwise lie about blocking them.

Validated before anything is written: https only (file: and data: rejected), name
matched against ^[a-z0-9][a-z0-9-]*$ so it cannot escape its directory, size
caps on index and body, every regex compiled, pattern length capped since it runs
on every tool call, and the body's own name checked against the index. Installed
skills rank below your own, so an install can never shadow a skill you wrote.

Interface:
- Context is a percentage of the compaction threshold, amber from two thirds and
  red at 90. A turn about to lose history now says so beforehand.
- Aligned command menu and registry tables; /skills and /plugins name origins.

538 tests, up from 488. The registry is tested against a real local HTTP server,
and the guard is proven to refuse a .env write end to end rather than assumed to.

test/compact.test.ts

@@ -2,6 +2,9 @@ import { expect, test } from 'bun:test';
 import { MockLanguageModelV4, simulateReadableStream } from 'ai/test';
 import type { LanguageModelV4CallOptions, LanguageModelV4StreamPart } from '@ai-sdk/provider';
 import type { ModelMessage } from 'ai';
+import { mkdtempSync, rmSync } from 'node:fs';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
 import { Session, type AgentEvent } from '../src/session';
 
 const usage = {
@@ -173,3 +176,129 @@ test('onChange fires for every history mutation so autosave stays current', asyn
 
   expect(snapshots).toEqual([1, 2]);
 });
+
+/** A reasoning model's turn: a reasoning item, then the tool call, both with item ids. */
+const reasoningToolStep = (n: number): LanguageModelV4StreamPart[] => [
+  { type: 'reasoning-start', id: `r${n}`, providerMetadata: { openai: { itemId: `rs_${n}` } } } as never,
+  { type: 'reasoning-delta', id: `r${n}`, delta: 'deciding what to read next' } as never,
+  { type: 'reasoning-end', id: `r${n}`, providerMetadata: { openai: { itemId: `rs_${n}` } } } as never,
+  { type: 'tool-input-start', id: `c${n}`, toolName: 'read_file' },
+  { type: 'tool-input-end', id: `c${n}` },
+  {
+    type: 'tool-call',
+    toolCallId: `c${n}`,
+    toolName: 'read_file',
+    input: JSON.stringify({ path: 'big.txt' }),
+    providerMetadata: { openai: { itemId: `fc_${n}` } },
+  } as never,
+  { type: 'finish', finishReason: { unified: 'tool-calls', raw: 'tool_use' }, usage },
+];
+
+function inTempDir<T>(fn: () => Promise<T>): Promise<T> {
+  const orig = process.cwd();
+  const dir = mkdtempSync(join(tmpdir(), 'shiro-compact-'));
+  process.chdir(dir);
+  return fn().finally(() => {
+    process.chdir(orig);
+    rmSync(dir, { recursive: true, force: true });
+  });
+}
+
+/**
+ * The bug this guards against: compaction used to drop any assistant part whose
+ * reasoning item it had pruned. On a reasoning model that is every tool call, so
+ * after the first compaction the model could no longer see what it had already
+ * run — and kept re-running it until the step limit stopped the turn.
+ */
+test('a compacted turn still shows the model the tool calls it already made', async () =>
+  inTempDir(async () => {
+    await Bun.write(join(process.cwd(), 'big.txt'), 'lorem ipsum dolor sit amet\n'.repeat(1500));
+
+    const seen: LanguageModelV4CallOptions[] = [];
+    let call = 0;
+    const session = new Session({
+      compactThreshold: 4000,
+      maxSteps: 8,
+      model: new MockLanguageModelV4({
+        doStream: async (o) => {
+          seen.push(o);
+          const n = call++;
+          return n < 3 ? stream(reasoningToolStep(n)) : stream(text('read it three times'));
+        },
+      }),
+      askApproval: async () => 'deny',
+    });
+
+    const events: string[] = [];
+    for await (const ev of session.send('read big.txt a few times')) events.push(ev.type);
+
+    expect(events).toContain('compacted');
+    expect(events).toContain('text');
+    expect(events.at(-1)).toBe('done');
+    // Four calls, not eight: the loop ended because the model chose to, not
+    // because maxSteps cut it off.
+    expect(call).toBe(4);
+
+    const shapeOf = (o: LanguageModelV4CallOptions) =>
+      o.prompt
+        .filter((m) => m.role !== 'system')
+        .flatMap((m) => (Array.isArray(m.content) ? (m.content as { type: string }[]).map((p) => p.type) : ['str']));
+
+    // Every call after the first has to carry the earlier exchange.
+    for (const o of seen.slice(1)) {
+      const shape = shapeOf(o);
+      expect(shape).toContain('tool-call');
+      expect(shape).toContain('tool-result');
+    }
+  }), 20_000);
+
+test('a compacted turn sends no assistant item reference whose reasoning was pruned', async () =>
+  inTempDir(async () => {
+    await Bun.write(join(process.cwd(), 'big.txt'), 'lorem ipsum dolor sit amet\n'.repeat(1500));
+
+    const seen: LanguageModelV4CallOptions[] = [];
+    let call = 0;
+    const session = new Session({
+      compactThreshold: 4000,
+      model: new MockLanguageModelV4({
+        doStream: async (o) => {
+          seen.push(o);
+          const n = call++;
+          return n < 3 ? stream(reasoningToolStep(n)) : stream(text('done'));
+        },
+      }),
+      askApproval: async () => 'deny',
+    });
+
+    for await (const _ of session.send('read it')) void _;
+
+    // An itemId on an assistant part is serialised as `item_reference`, which the
+    // responses API resolves against a stored item that depends on its reasoning
+    // item. Send one without that reasoning and the request is a 400. An itemId on
+    // a tool result is harmless: it goes out as a plain function_call_output.
+    for (const [i, o] of seen.entries()) {
+      const assistant = o.prompt.filter((m) => m.role === 'assistant');
+      const reasoningIds = new Set<string>();
+      for (const m of assistant) {
+        if (!Array.isArray(m.content)) continue;
+        for (const p of m.content as { type: string; providerOptions?: Record<string, Record<string, unknown>> }[]) {
+          if (p.type !== 'reasoning') continue;
+          for (const options of Object.values(p.providerOptions ?? {})) {
+            if (typeof options['itemId'] === 'string') reasoningIds.add(options['itemId']);
+          }
+        }
+      }
+
+      for (const m of assistant) {
+        if (!Array.isArray(m.content)) continue;
+        for (const p of m.content as { type: string; providerOptions?: Record<string, Record<string, unknown>> }[]) {
+          if (p.type === 'reasoning') continue;
+          for (const options of Object.values(p.providerOptions ?? {})) {
+            const id = options['itemId'];
+            if (typeof id !== 'string') continue;
+            expect(reasoningIds.size, `call ${i}: ${p.type} references ${id} with no reasoning item`).toBeGreaterThan(0);
+          }
+        }
+      }
+    }
+  }), 20_000);