loop: every request is built from the log — boundary snapshot, header events, config-only waterfall

The loop is now transmission-stateless; a request is a pure function of (session log, this step's rendered assembly, current AgentOptions): - The reconstruction boundary is step/start: the messages snapshot is taken in the same synchronous frame immediately before the step/start append, so the request's messages are exactly the derivation over events[0..stepStartSeq) — an inject() from an agent/request listener (or any concurrent task) lands after the boundary and joins the NEXT request. This changes behavior for a synchronous step/start session/event listener that appends content (master derived after the append, so such a listener could reach the current request): agent/pre-step is the sanctioned seam for current-request content. - agent/request is re-typed to config-only: (agent, turn, step, config: LlmCallConfig, next) → LlmCallConfig. The frozen seed comes from AgentOptions on a loop instance's first request (explicit options beat the logged baseline — fork overrides and resume reconfiguration stay correct) and from the log's folded header afterwards; listeners return a replacement to switch. Content shaping through the request is no longer expressible — model-visible content flows through the log channels. - recordRequestHeader appends whatever header event the request owes the log before dispatch: an 'initial'/'resume' snapshot anchoring each loop instance, a round-trip-verified delta on change, a 'fallback' snapshot when the encoding cannot express it. Session.requestHeader() is the log's incrementally-folded baseline. - Requests are deep-frozen before dispatch (deepFreeze exempts the AbortSignal — freezing one breaks AbortController.abort() outright); frozen + sessionId is the loop-built marker the dev invariant keys on. Ported from #162 and re-anchored on the log: the append-extension / frozen-end-to-end / compaction-resend / prompt-change property tests, plus new specs for the boundary semantics, resume anchoring, and the end-to-end theorem (every recorded request rebuilds byte-equal from the log alone). Live cache-hit e2e (request-cache.e2e.ts) verified against the real DeepSeek API. Snapshot goldens intentionally stale until the single re-record after the compact/summary envelope lands.
2026-07-06 03:07:34 +08:00
parent a4f7e757fa
commit 2093a8898b
17 changed files with 745 additions and 68 deletions
--- a/packages/core/agent-loop/README.md
+++ b/packages/core/agent-loop/README.md
@@ -56,9 +56,11 @@ forever:
      drain steering
      assembly = systemPrompt.assemble({agent})  ⟵ renderPrompt(assembly) IS the full prompt
      await serial agent/pre-step        ⟵ surface mutation (compaction) outside the step
-      session('step/start')
-      request = waterfall agent/request
-      stream llm.stream(request) → session('assistant/chunk')
+      boundary = session.deriveMessages()   ⟵ reconstruction boundary: same sync frame,
+      session('step/start')                     strictly before step/start
+      config = waterfall agent/request       ⟵ frozen seed; return a replacement to switch
+      session('request/header'[-delta])      ⟵ the header event this request owes the log
+      stream llm.stream(freeze({header..., messages: boundary})) → session('assistant/chunk')
      message = waterfall agent/step-result
      session('assistant/message')
      each tool-call: session('tool/call')
--- a/packages/core/agent-loop/src/loop.ts
+++ b/packages/core/agent-loop/src/loop.ts
@@ -8,10 +8,13 @@
 */

 import type { Context } from 'cordis'
-import type { FinishReason, GenerateOptions, Message } from '@deepseek-ai/dsh-llm'
-import { BlockAssembler, HarnessError } from '@deepseek-ai/dsh-llm'
+import type { FinishReason, GenerateOptions, LlmCallConfig, Message } from '@deepseek-ai/dsh-llm'
+import { BlockAssembler, HarnessError, deepFreeze } from '@deepseek-ai/dsh-llm'
 import type { ContinuationDecision, HookContext, PromptDecision } from '@deepseek-ai/dsh-agent'
+import { canonicalHeader } from '@deepseek-ai/dsh-session'
 import type { Session, TurnEndReason, TurnTrigger } from '@deepseek-ai/dsh-session'
+import { createTransmissionLog, recordRequestHeader } from './request-log.ts'
+import type { TransmissionLog } from './request-log.ts'
 import { renderPrompt } from '@deepseek-ai/dsh-system-prompt'
 import type { PromptAssembly } from '@deepseek-ai/dsh-system-prompt'
 import type {} from '@deepseek-ai/dsh-tools'
@@ -155,10 +158,13 @@ export interface LoopHandle {
 *       assembly = ctx.systemPrompt.assemble({agent}) ⟵ waterfall system-prompt/assemble; renderPrompt
 *                                                        (persona section + {{variables}}) IS the full prompt
 *       await ctx.serial('agent/pre-step')            ⟵ surface mutation (compaction) OUTSIDE the step
- *       session('step/start')                         ⟵ durable step boundary (no agent/* mirror)
- *       req = {model, system, tools, messages: session.deriveMessages(), signal}
- *       req = waterfall agent/request                 ⟵ hooks/model-switch
- *       stream ctx.llm.stream(req)                    ⟵ waterfall llm/stream (raw chunks)
+ *       boundary = session.deriveMessages()           ⟵ the reconstruction boundary: snapshot in the
+ *       session('step/start')                            same sync frame, strictly before step/start
+ *       config = waterfall agent/request(config)      ⟵ frozen seed; a returned replacement switches
+ *       session('request/header'|'request/header-delta')  ⟵ the header event this request owes the
+ *                                                        log (initial/resume anchor, delta, fallback)
+ *       req = freeze({header..., messages: boundary, sessionId, signal})
+ *       stream ctx.llm.stream(req)                    ⟵ waterfall llm/stream (raw chunks, frozen req)
 *         session('assistant/chunk')
 *       msg = waterfall agent/step-result             ⟵ BEFORE the log append, so the
 *       session('assistant/message' {content, usage?})   session records what actually ran
@@ -181,6 +187,13 @@ export interface LoopHandle {
 * ```
 */
 export async function runLoop(ctx: Context, agent: ReactLoopAgent, handle: LoopHandle): Promise<void> {
+  // Per-instance transmission bookkeeping: whether THIS loop instance has
+  // anchored the log's header fold yet (its first request logs a
+  // 'initial'/'resume' request/header snapshot). Everything else the request
+  // needs is read from the session log itself — the loop holds no
+  // conversation state (the reconstructability RFC).
+  const transmission = createTransmissionLog()
+
  const { session } = agent

  while (!handle.isDisposed()) {
@@ -236,7 +249,7 @@ export async function runLoop(ctx: Context, agent: ReactLoopAgent, handle: LoopH
    // turn number is actually last in the log — a stale counter would collide.
    const turn = lastTurnNumber(session) + 1
    try {
-      await runTurn(ctx, agent, handle, turn)
+      await runTurn(ctx, agent, handle, turn, transmission)
    } catch (error: unknown) {
      // Backstop: runTurn rethrows only a PRE-turn throw (the invariant guard
      // before turn/start) — no turn/start was appended, so no turn is open and
@@ -271,7 +284,9 @@ export async function runLoop(ctx: Context, agent: ReactLoopAgent, handle: LoopH
  }
 }

-async function runTurn(ctx: Context, agent: ReactLoopAgent, handle: LoopHandle, turn: number): Promise<void> {
+async function runTurn(
+  ctx: Context, agent: ReactLoopAgent, handle: LoopHandle, turn: number, transmission: TransmissionLog,
+): Promise<void> {
  const { session } = agent

  // --- Pre-turn. A throw here (the invariant guard) is owed NO turn/end —
@@ -474,6 +489,17 @@ async function runTurn(ctx: Context, agent: ReactLoopAgent, handle: LoopHandle,
        break
      }

+      // The reconstruction boundary (the reconstructability RFC): the request's
+      // messages are snapshotted HERE, in the same synchronous frame as the
+      // step/start append directly below — so the snapshot is exactly the
+      // derivation over the log prefix strictly before step/start's seq.
+      // Anything appended later — by a step/start session/event listener, an
+      // agent/request-window inject(), any concurrent task — lands after the
+      // boundary and joins the NEXT request. An external reconstructor
+      // recovers these exact messages by folding the surface over
+      // events[0..stepStartSeq).
+      const boundaryMessages = session.deriveMessages()
+
      // Mark the step open BEFORE the append: Session.append pushes the event
      // to the log before notifying session/event listeners, so a THROWING
      // step/start listener leaves step/start in the log. Setting stepOpen first
@@ -495,7 +521,7 @@ async function runTurn(ctx: Context, agent: ReactLoopAgent, handle: LoopHandle,

      let stepOutcome: { hadToolCalls: boolean; finish: FinishReason } | { error: Error }
      try {
-        stepOutcome = await runStep(ctx, agent, turn, step, assembly, fullSystemPrompt, abort.signal)
+        stepOutcome = await runStep(ctx, agent, turn, step, assembly, fullSystemPrompt, boundaryMessages, transmission, abort.signal)
      } catch (error: unknown) {
        stepOutcome = { error: toError(error) }
      } finally {
@@ -644,11 +670,12 @@ function drainSteering(agent: ReactLoopAgent, turn: number): boolean {
  return messages.length > 0
 }

-/** One step: derive request from the (already pre-step-mutated) surface →
- * stream model → record → execute tools. The caller assembles the system prompt
- * and fires the `agent/pre-step` seam BEFORE opening the step, then passes the
- * resulting `assembly`/`system` here, so the surface this step derives from
- * already reflects any compaction. */
+/** One step: build the request from the boundary snapshot + the step's
+ * header → log the header event the request owes → stream model → record →
+ * execute tools. The caller assembles the system prompt, fires the
+ * `agent/pre-step` seam, snapshots the derivation, and opens the step BEFORE
+ * calling this, so `boundaryMessages` is exactly the surface prefix at
+ * step/start and already reflects any compaction. */
 async function runStep(
  ctx: Context,
  agent: ReactLoopAgent,
@@ -656,23 +683,57 @@ async function runStep(
  step: number,
  assembly: PromptAssembly,
  system: string,
+  boundaryMessages: Message[],
+  transmission: TransmissionLog,
  signal: AbortSignal,
 ): Promise<{ hadToolCalls: boolean; finish: FinishReason }> {
  const { session, options } = agent

-  let request: GenerateOptions = {
-    model: options.model ?? '',
-    messages: session.deriveMessages(),
-    ...system ? { system } : {},
-    ...assembly.tools.length > 0 ? { tools: assembly.tools } : {},
-    sessionId: session.id,
-    signal,
-  }
-  request = await ctx.waterfall('agent/request', agent, turn, step, request, () => Promise.resolve(request))
-  if (!request.model) {
+  // Seed the call config: the first request of THIS loop instance seeds from
+  // current AgentOptions — explicit options always win over the logged
+  // baseline, which is what keeps fork model-overrides and resume-time
+  // reconfiguration correct. Later steps seed from the log's folded header,
+  // which by then is exactly what this instance last logged.
+  const seedConfig: LlmCallConfig = transmission.loggedHeader
+    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion -- loggedHeader ⟹ a snapshot is in the log
+    ? session.requestHeader()!.config
+    : { model: options.model ?? '' }
+
+  // Shape the call config: listeners return a replacement to switch model or
+  // sampling (the seed is frozen — content shaping is not expressible here;
+  // model-visible content flows through the log channels). The header event
+  // below records whatever the request ACTUALLY uses, so a listener's switch
+  // is a logged, reconstructable fact, never silent drift.
+  const config = await ctx.waterfall('agent/request', agent, turn, step, deepFreeze({ ...seedConfig }), () => Promise.resolve(seedConfig))
+  if (!config.model) {
    throw new Error(`agent "${agent.id}" has no model: set AgentOptions.model or supply one via the agent/request waterfall`)
  }

+  // The request header (the log's request/header* vocabulary): canonical form,
+  // recorded before dispatch so the log always explains the request.
+  const header = canonicalHeader({
+    config,
+    ...system ? { system } : {},
+    ...assembly.tools.length > 0 ? { tools: assembly.tools } : {},
+  })
+  recordRequestHeader(session, transmission, header)
+
+  // Build and freeze: the request is a pure function of (boundary snapshot,
+  // logged header) — llm/stream listeners and adapters read it, mutation
+  // throws. sessionId + frozen is the loop-built marker the dev invariant
+  // keys on.
+  const request: GenerateOptions = deepFreeze({
+    model: header.config.model,
+    messages: boundaryMessages,
+    ...header.system !== undefined ? { system: header.system } : {},
+    ...header.tools !== undefined ? { tools: header.tools } : {},
+    ...header.config.temperature !== undefined ? { temperature: header.config.temperature } : {},
+    ...header.config.maxTokens !== undefined ? { maxTokens: header.config.maxTokens } : {},
+    ...header.config.stop !== undefined ? { stop: header.config.stop } : {},
+    sessionId: session.id,
+    signal,
+  })
+
  // --- Model call (streaming-first; raw chunks are the replay record) ---
  const assembler = new BlockAssembler()
  const chunkSeqs: number[] = []
--- a/packages/core/agent-loop/src/request-log.ts
+++ b/packages/core/agent-loop/src/request-log.ts
@@ -0,0 +1,68 @@
+/**
+ * Per-loop-instance transmission bookkeeping for the reconstructability
+ * contract: which header event to append before a request so the session log
+ * always explains the request (the reconstructability RFC). The loop is
+ * otherwise transmission-stateless — the comparison baseline is the log's own
+ * folded header (`Session.requestHeader()`), so resume and fork need no
+ * special path: a fresh loop instance simply logs a `'resume'` snapshot on
+ * its first request and deltas from there.
+ *
+ * @module dsh-agent-loop/request-log
+ */
+
+import { diffHeader, headerEquals, applyHeaderDelta } from '@deepseek-ai/dsh-session'
+import type { EpochHeader, Session } from '@deepseek-ai/dsh-session'
+
+/** Per-loop-instance bookkeeping: whether THIS instance has logged a header yet. */
+export interface TransmissionLog {
+  /** True once this loop instance appended its anchoring `request/header` snapshot. */
+  loggedHeader: boolean
+}
+
+/** Fresh bookkeeping for a newly-started loop instance. */
+export function createTransmissionLog(): TransmissionLog {
+  return { loggedHeader: false }
+}
+
+/**
+ * Append whatever header event this request owes the log, so folding the log
+ * reproduces the header the request was built under. Exactly one of four
+ * things happens:
+ *
+ * 1. This loop instance has not logged a header yet → a full `request/header`
+ *    snapshot anchors the fold: reason `'initial'` when the log has no header
+ *    events at all (a new conversation), `'resume'` when it does (process
+ *    restart, fork seed — the boundary itself is a recorded fact, so the
+ *    snapshot is appended even when nothing changed).
+ * 2. The header equals the folded baseline → nothing; the log already
+ *    explains this request.
+ * 3. It differs and the delta round-trips (`applyHeaderDelta` on the baseline
+ *    reproduces the header exactly) → a `request/header-delta`.
+ * 4. It differs and the delta encoding cannot express the change (a pure tool
+ *    reordering) → a full snapshot with reason `'fallback'`; deltas are an
+ *    encoding optimization, never a correctness dependency.
+ *
+ * @param session - the session whose log explains the request.
+ * @param state - this loop instance's bookkeeping (mutated on first log).
+ * @param header - the canonical header the request will ACTUALLY use
+ *   (post-`agent/request`).
+ */
+export function recordRequestHeader(session: Session, state: TransmissionLog, header: EpochHeader): void {
+  if (!state.loggedHeader) {
+    session.append('request/header', { header, reason: session.requestHeader() === undefined ? 'initial' : 'resume' })
+    state.loggedHeader = true
+    return
+  }
+  // This instance logged a snapshot, so the fold is necessarily defined.
+  // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+  const baseline = session.requestHeader()!
+  if (headerEquals(baseline, header)) return
+  const delta = diffHeader(baseline, header)
+  /* v8 ignore next -- headerEquals false ⟹ diffHeader defined: both compare the same three parts */
+  if (delta === undefined) return
+  if (headerEquals(applyHeaderDelta(baseline, delta), header)) {
+    session.append('request/header-delta', delta)
+  } else {
+    session.append('request/header', { header, reason: 'fallback' })
+  }
+}
--- a/packages/core/agent-loop/tests/loop.spec.ts
+++ b/packages/core/agent-loop/tests/loop.spec.ts
@@ -230,9 +230,8 @@ describe('agent loop', () => {
      assembly.variables['model'] = 'mock'
      return next()
    })
-    ctx.on('agent/request', async (_agent, _turn, _step, options, next) => {
-      options.model = 'mock'
-      return next()
+    ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => {
+      return { ...config, model: 'mock' }
    })
    const agent = ctx.agentLoop.create(AgentId('a-late-model'), {})

@@ -428,20 +427,27 @@ describe('agent loop', () => {
    expect(agent.session.events.some(e => e.type === 'tool/result')).toBe(true)
  })

-  it('agent/request waterfall can rewrite the request (model-switch pattern)', async () => {
+  it('agent/request waterfall switches models by returning a replacement config; the switch is logged', async () => {
    const adapter = new MockAdapter([textResponse('ok')])
    const ctx = await harness(adapter)
    ctx.llm.registerAdapter(['other-model'], adapter)
    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })

-    ctx.on('agent/request', async (_agent, _turn, _step, options, next) => {
-      options.model = 'other-model'
-      return next()
+    ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => {
+      // The seed is frozen — config is not a mutable per-call knob; a switch
+      // is proposed by returning a replacement, and the loop logs it.
+      expect(Object.isFrozen(config)).toBe(true)
+      expect(() => { (config as { model: string }).model = 'other-model' }).toThrow(TypeError)
+      return { ...config, model: 'other-model' }
    })

    send(agent, 'hi')
    await waitForIdle(ctx, agent)
    expect(adapter.requests[0]!.model).toBe('other-model')
+    // The header event records what the request ACTUALLY used — the switch is
+    // a reconstructable fact, not silent drift.
+    const headerEvent = agent.session.events.find(e => e.type === 'request/header')
+    expect(headerEvent?.type === 'request/header' && headerEvent.data.header.config.model).toBe('other-model')
  })

  it('agent/pre-step fires once per step before the step is opened', async () => {
--- a/packages/core/agent-loop/tests/request-cache.e2e.ts
+++ b/packages/core/agent-loop/tests/request-cache.e2e.ts
@@ -0,0 +1,104 @@
+import { afterEach, describe, expect, it } from 'vitest'
+import { Context } from 'cordis'
+import LlmService from '@deepseek-ai/dsh-llm'
+import SessionStore from '@deepseek-ai/dsh-session'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import ToolRegistry, { defineTool } from '@deepseek-ai/dsh-tools'
+import AgentRegistry, { AgentId, type Agent } from '@deepseek-ai/dsh-agent'
+import AgentLoop from '@deepseek-ai/dsh-agent-loop'
+import * as LlmDeepSeek from '@deepseek-ai/dsh-llm-deepseek'
+
+/**
+ * With-key proof that log-derived requests translate into REAL provider cache
+ * hits: a multi-step tool turn (plus a follow-up turn) against the live
+ * DeepSeek API must report `cacheReadTokens > 0` on every request after the
+ * first — the adapter maps the provider's `prompt_cache_hit_tokens`, and the
+ * per-step usage recorded on `assistant/message` events is the production
+ * observable for cache behavior (the reconstructability RFC's measurement
+ * layer: prefix stability is corollary #1). Mocks prove the requests are
+ * append-extensions; only the real API proves those bytes actually hit the
+ * provider cache. Key-gated — skips entirely without $DEEPSEEK_API_KEY.
+ */
+
+// Long enough that the shared request prefix comfortably spans the provider's
+// cache-block granularity (64 tokens) from the very first request.
+const SYSTEM = 'You are a terse coding assistant used in an automated cache test. '
+  + 'Always follow instructions literally and exactly. When the user asks you to look '
+  + 'something up, call the lookup tool with the requested key and wait for its result '
+  + 'before answering. Never invent a value the tool has not returned. After the tool '
+  + 'returns, answer with a single short sentence that repeats the returned value '
+  + 'verbatim. Do not add explanations, do not use markdown, do not ask follow-up '
+  + 'questions. If the user asks anything else, answer in one short sentence.'
+
+let ctx: Context | undefined
+
+afterEach(async () => {
+  await ctx?.fiber.dispose()
+  ctx = undefined
+})
+
+async function loopHarness(): Promise<Context> {
+  const created = new Context()
+  await created.plugin(LlmService)
+  await created.plugin(SessionStore)
+  await created.plugin(SystemPrompt, { persona: SYSTEM })
+  await created.plugin(ToolRegistry)
+  await created.plugin(AgentRegistry)
+  await created.plugin(AgentLoop, { agents: [] })
+  await created.plugin(LlmDeepSeek, { models: ['deepseek-v4-flash'] })
+  created.tools.register(defineTool({
+    name: 'lookup',
+    description: 'Look up the stored value for a key.',
+    parameters: { key: { type: 'string', description: 'The key to look up.' } },
+    async execute(args) {
+      return [{ type: 'text', text: `value(${String(args.key)}) = azure-falcon-42` }]
+    },
+  }))
+  return created
+}
+
+function waitForIdle(context: Context, agent: Agent): Promise<void> {
+  return new Promise((resolve) => {
+    const dispose = context.on('agent/status', (subject, status) => {
+      if (subject === agent && status === 'idle') {
+        dispose()
+        resolve()
+      }
+    })
+  })
+}
+
+describe.skipIf(!process.env.DEEPSEEK_API_KEY)('log-derived request cache hits (real API)', () => {
+  it('every request after the first hits the provider prefix cache', async () => {
+    ctx = await loopHarness()
+    const agent = ctx.agentLoop.create(AgentId('cache-e2e'), { model: 'deepseek-v4-flash' })
+
+    // Turn 1: forces a tool call → at least two steps (two model requests).
+    agent.send([{ type: 'text', text: 'Look up the key "deploy-color" with the lookup tool and tell me the value.' }])
+    await waitForIdle(ctx, agent)
+    // Turn 2: a follow-up over the same (longer) prefix.
+    agent.send([{ type: 'text', text: 'Thanks. Repeat that value one more time.' }])
+    await waitForIdle(ctx, agent)
+
+    const usages = [...agent.session.events]
+      .filter(e => e.type === 'assistant/message')
+      .map(e => e.data.usage)
+    expect(usages.length).toBeGreaterThanOrEqual(3) // 2 steps in turn 1 + ≥1 in turn 2
+    for (const usage of usages) expect(usage).toBeDefined()
+
+    // The first request has nothing to hit; every later one shares its
+    // predecessor as a byte-identical prefix, so the provider must report
+    // cached prompt tokens (prompt_cache_hit_tokens → cacheReadTokens).
+    for (const usage of usages.slice(1)) {
+      expect(usage!.cacheReadTokens ?? 0).toBeGreaterThan(0)
+    }
+
+    // World-verification of the conversation itself: the tool value made it
+    // through the loop into the final answer.
+    const finalText = agent.session.deriveMessages().at(-1)!.content
+      .filter(block => block.type === 'text')
+      .map(block => block.text)
+      .join('')
+    expect(finalText).toContain('azure-falcon-42')
+  }, 180_000)
+})
--- a/packages/core/agent-loop/tests/request-log.spec.ts
+++ b/packages/core/agent-loop/tests/request-log.spec.ts
@@ -0,0 +1,85 @@
+/**
+ * recordRequestHeader unit tests: exactly one of four things per request —
+ * an 'initial' snapshot (log has no header yet), a 'resume' snapshot (fresh
+ * loop instance over a log that has one), nothing (header unchanged), a
+ * round-tripping delta, or a 'fallback' snapshot when the delta encoding
+ * cannot express the change (pure tool reordering).
+ */
+
+import { describe, expect, it } from 'vitest'
+import { Session, SessionId, canonicalHeader } from '@deepseek-ai/dsh-session'
+import type { SessionEvent } from '@deepseek-ai/dsh-session'
+import type { ToolSchema } from '@deepseek-ai/dsh-llm'
+import { createTransmissionLog, recordRequestHeader } from '../src/request-log.ts'
+
+function tool(name: string, description = 'd'): ToolSchema {
+  return { name, description, parameters: { type: 'object' } }
+}
+
+function openSession(id: string): Session {
+  const session = new Session(SessionId(id))
+  session.append('turn/start', { turn: 1, trigger: { kind: 'message', source: { kind: 'user' } } })
+  return session
+}
+
+function headerEvents(session: Session): SessionEvent[] {
+  return session.events.filter(e => e.type === 'request/header' || e.type === 'request/header-delta')
+}
+
+describe('recordRequestHeader', () => {
+  it("anchors a new conversation with an 'initial' snapshot, then logs nothing while unchanged", () => {
+    const session = openSession('rl-initial')
+    const state = createTransmissionLog()
+    const header = canonicalHeader({ config: { model: 'm' }, system: 's', tools: [tool('t')] })
+
+    recordRequestHeader(session, state, header)
+    const [first] = headerEvents(session)
+    expect(first?.type === 'request/header' && first.data.reason).toBe('initial')
+
+    recordRequestHeader(session, state, header)
+    expect(headerEvents(session)).toHaveLength(1)
+  })
+
+  it("anchors a fresh loop instance over an anchored log with a 'resume' snapshot, even unchanged", () => {
+    const session = openSession('rl-resume')
+    const header = canonicalHeader({ config: { model: 'm' }, system: 's' })
+    recordRequestHeader(session, createTransmissionLog(), header)
+
+    // A second instance (process restart / fork): the boundary itself is a
+    // recorded fact — snapshot appended even though the header is identical.
+    recordRequestHeader(session, createTransmissionLog(), header)
+    const events = headerEvents(session)
+    expect(events).toHaveLength(2)
+    expect(events[1]?.type === 'request/header' && events[1].data.reason).toBe('resume')
+  })
+
+  it('logs a round-tripping delta for a mid-run change, and the fold reproduces the header', () => {
+    const session = openSession('rl-delta')
+    const state = createTransmissionLog()
+    const first = canonicalHeader({ config: { model: 'm' }, system: 'a\nb', tools: [tool('t')] })
+    recordRequestHeader(session, state, first)
+
+    const second = canonicalHeader({ config: { model: 'm' }, system: 'a\nc', tools: [tool('t'), tool('u')] })
+    recordRequestHeader(session, state, second)
+    const events = headerEvents(session)
+    expect(events).toHaveLength(2)
+    expect(events[1]?.type).toBe('request/header-delta')
+    expect(session.requestHeader()).toEqual(second)
+  })
+
+  it("records a change the delta cannot express (pure reordering) as a 'fallback' snapshot", () => {
+    const session = openSession('rl-fallback')
+    const state = createTransmissionLog()
+    const first = canonicalHeader({ config: { model: 'm' }, tools: [tool('a'), tool('b')] })
+    recordRequestHeader(session, state, first)
+
+    const reordered = canonicalHeader({ config: { model: 'm' }, tools: [tool('b'), tool('a')] })
+    recordRequestHeader(session, state, reordered)
+    const events = headerEvents(session)
+    expect(events).toHaveLength(2)
+    expect(events[1]?.type === 'request/header' && events[1].data.reason).toBe('fallback')
+    // The fold still lands on the exact header — deltas are an encoding
+    // optimization, never a correctness dependency.
+    expect(session.requestHeader()).toEqual(reordered)
+  })
+})
--- a/packages/core/agent-loop/tests/request-reconstruction.spec.ts
+++ b/packages/core/agent-loop/tests/request-reconstruction.spec.ts
@@ -0,0 +1,284 @@
+/**
+ * Loop-level reconstructability: every request the loop sends is a pure
+ * function of the session log — messages are the derivation at the step/start
+ * boundary, the header is the fold of request/header* events — and every
+ * request is an append-extension of its predecessor unless a logged event
+ * (compaction replace, header change) explains the difference. The requests
+ * recorded by the mock adapter are the observable; the offline-rebuild test
+ * at the bottom is the theorem stated end-to-end.
+ */
+
+import { describe, expect, it } from 'vitest'
+import { Context } from 'cordis'
+import LlmService from '@deepseek-ai/dsh-llm'
+import type { GenerateOptions } from '@deepseek-ai/dsh-llm'
+import SessionStore, { Session, SessionId, foldRequestHeader } from '@deepseek-ai/dsh-session'
+import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
+import ToolRegistry, { defineTool } from '@deepseek-ai/dsh-tools'
+import AgentRegistry, { AgentId } from '@deepseek-ai/dsh-agent'
+import AgentLoop, { ReactLoopAgent } from '@deepseek-ai/dsh-agent-loop'
+import { MockAdapter, textResponse, toolCallResponse } from './mock-adapter.ts'
+
+async function harness(adapter: MockAdapter, persona = 'stable base') {
+  const ctx = new Context()
+  await ctx.plugin(LlmService)
+  await ctx.plugin(SessionStore)
+  await ctx.plugin(SystemPrompt, { persona })
+  await ctx.plugin(ToolRegistry)
+  await ctx.plugin(AgentRegistry)
+  await ctx.plugin(AgentLoop, { agents: [] })
+  ctx.llm.registerAdapter(['mock'], adapter)
+  return ctx
+}
+
+function waitForIdle(ctx: Context, agent: ReactLoopAgent): Promise<void> {
+  return new Promise((resolve) => {
+    const dispose = ctx.on('agent/status', (subject, status) => {
+      if (subject === agent && status === 'idle') {
+        dispose()
+        resolve()
+      }
+    })
+  })
+}
+
+function send(agent: ReactLoopAgent, text: string) {
+  agent.send([{ type: 'text', text }])
+}
+
+/** Assert `previous` is a strict value-prefix of `current`. */
+function expectPrefixExtension(previous: GenerateOptions, current: GenerateOptions) {
+  expect(current.messages.length).toBeGreaterThan(previous.messages.length)
+  expect(current.messages.slice(0, previous.messages.length)).toEqual([...previous.messages])
+  expect(current.system).toEqual(previous.system)
+  expect(current.tools).toEqual(previous.tools)
+}
+
+function registerEcho(ctx: Context) {
+  ctx.tools.register(defineTool({
+    name: 'echo',
+    description: 'echo back',
+    parameters: { text: { type: 'string' } },
+    async execute(args) {
+      return [{ type: 'text', text: `echo: ${String(args.text)}` }]
+    },
+  }))
+}
+
+describe('request stability across the loop', () => {
+  it('each step request within a turn append-extends the previous, frozen end to end', async () => {
+    const adapter = new MockAdapter([
+      toolCallResponse('c1', 'echo', { text: 'one' }, 'first'),
+      toolCallResponse('c2', 'echo', { text: 'two' }, 'second'),
+      textResponse('done'),
+    ])
+    const ctx = await harness(adapter)
+    registerEcho(ctx)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    send(agent, 'go')
+    await waitForIdle(ctx, agent)
+
+    expect(adapter.requests).toHaveLength(3)
+    expectPrefixExtension(adapter.requests[0]!, adapter.requests[1]!)
+    expectPrefixExtension(adapter.requests[1]!, adapter.requests[2]!)
+    for (const request of adapter.requests) {
+      expect(Object.isFrozen(request)).toBe(true)
+      expect(Object.isFrozen(request.messages)).toBe(true)
+    }
+    // One anchoring header snapshot; no further header events (nothing changed).
+    const headerEvents = agent.session.events.filter(e => e.type === 'request/header' || e.type === 'request/header-delta')
+    expect(headerEvents).toHaveLength(1)
+    expect(headerEvents[0]?.type === 'request/header' && headerEvents[0].data.reason).toBe('initial')
+  })
+
+  it('a later turn append-extends the previous turn (one conversation, one log)', async () => {
+    const adapter = new MockAdapter([textResponse('one'), textResponse('two')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+    send(agent, 'second')
+    await waitForIdle(ctx, agent)
+
+    expect(adapter.requests).toHaveLength(2)
+    expectPrefixExtension(adapter.requests[0]!, adapter.requests[1]!)
+  })
+
+  it('a compaction replace rewrites the resend, and the log explains it', async () => {
+    const adapter = new MockAdapter([textResponse('one'), textResponse('two')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+
+    // A pre-step listener compacts turn 1's history before turn 2's step —
+    // the sanctioned surface rewrite, landing OUTSIDE the step.
+    const preStep = ctx.on('agent/pre-step', () => {
+      preStep()
+      const session = agent.session
+      const nodes = session.surface.nodes
+      session.append('context/message', {
+        content: [{ type: 'text', text: '[summary of turn 1]' }],
+        source: { kind: 'plugin', plugin: 'test-compact' },
+      }, {
+        surfaceOp: { op: 'replace', start: nodes[0]!.seq, end: nodes[1]!.seq },
+        sourceEventSeqs: [nodes[0]!.seq, nodes[1]!.seq],
+      })
+    })
+
+    send(agent, 'second')
+    await waitForIdle(ctx, agent)
+
+    const second = adapter.requests[1]!
+    // The rewritten history: summary replaces turn 1's user+assistant pair.
+    expect(second.messages[0]!.content.some(b => b.type === 'text' && b.text.includes('[summary of turn 1]'))).toBe(true)
+    // No header event beyond the anchor: the replace is itself in the log.
+    expect(agent.session.events.filter(e => e.type === 'request/header')).toHaveLength(1)
+  })
+
+  it('a real system-prompt change is a logged header delta; a stable prompt logs nothing', async () => {
+    const adapter = new MockAdapter([textResponse('one'), textResponse('two'), textResponse('three')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+    send(agent, 'second')
+    await waitForIdle(ctx, agent)
+    // Identical assembly re-rendered per step is NOT a change.
+    expect(agent.session.events.filter(e => e.type === 'request/header-delta')).toHaveLength(0)
+
+    ctx.systemPrompt.section({ name: 'extra', order: 2, text: 'new guidance' })
+    send(agent, 'third')
+    await waitForIdle(ctx, agent)
+
+    const deltas = agent.session.events.filter(e => e.type === 'request/header-delta')
+    expect(deltas).toHaveLength(1)
+    expect(adapter.requests[2]!.system).toContain('new guidance')
+    // History is preserved across the change — only the header moved.
+    expect(adapter.requests[2]!.messages.length).toBeGreaterThan(adapter.requests[1]!.messages.length)
+  })
+
+  it('an inject() during the agent/request waterfall joins the NEXT request (the step/start boundary)', async () => {
+    const adapter = new MockAdapter([textResponse('one'), textResponse('two')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    let injected = false
+    ctx.on('agent/request', async (_agent, _turn, _step, _config, next) => {
+      if (!injected) {
+        injected = true
+        agent.inject([{ type: 'text', text: '[late context]' }], { source: { kind: 'plugin', plugin: 'test' } })
+      }
+      return next()
+    })
+
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+    const first = adapter.requests[0]!
+    // The inject landed in the log after the boundary: not in THIS request…
+    expect(first.messages.some(m => m.content.some(b => b.type === 'text' && b.text.includes('[late context]')))).toBe(false)
+    expect(agent.session.events.some(e => e.type === 'context/message')).toBe(true)
+
+    send(agent, 'second')
+    await waitForIdle(ctx, agent)
+    // …but in the next one, at its logged position.
+    const second = adapter.requests[1]!
+    expect(second.messages.some(m => m.content.some(b => b.type === 'text' && b.text.includes('[late context]')))).toBe(true)
+  })
+
+  it('a mutation attempt on the frozen request content throws into the step (loud, not silent)', async () => {
+    const adapter = new MockAdapter([textResponse('one')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    const errors: Error[] = []
+    ctx.on('agent/error', (_agent, _turn, _step, error) => void errors.push(error))
+    ctx.on('llm/stream', (options, next) => {
+      // The historical failure mode this design kills: a listener rewriting
+      // request content in place. The freeze turns it into a loud error.
+      options.messages.push({ role: 'user', content: [{ type: 'text', text: 'sneaky' }] })
+      return next()
+    })
+
+    send(agent, 'go')
+    await waitForIdle(ctx, agent)
+
+    expect(errors).toHaveLength(1)
+    expect(errors[0]!.message).toMatch(/not extensible|frozen|read only|readonly/i)
+  })
+
+  it('a fresh loop instance over a seeded log anchors with a resume snapshot and stays cache-aligned', async () => {
+    const adapter = new MockAdapter([textResponse('one')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create(AgentId('gen1'), { model: 'mock' })
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+
+    // Second generation: a new agent whose session is seeded with the first
+    // one's full log (the resume/fork path).
+    const adapter2 = new MockAdapter([textResponse('two')])
+    const ctx2 = await harness(adapter2)
+    const handle = ctx2.agents.create({
+      agentId: AgentId('gen2'),
+      sessionId: SessionId('gen2-session'),
+      seed: [...agent.session.events],
+      agentOptions: { model: 'mock' },
+    })
+    const agent2 = handle.agent as ReactLoopAgent
+    send(agent2, 'second')
+    await waitForIdle(ctx2, agent2)
+
+    const snapshots = agent2.session.events.filter(e => e.type === 'request/header')
+    expect(snapshots).toHaveLength(2)
+    expect(snapshots[1]?.type === 'request/header' && snapshots[1].data.reason).toBe('resume')
+    // Identical header across the restart: byte-identical continuation.
+    expect(adapter2.requests[0]!.system).toEqual(adapter.requests[0]!.system)
+    expectPrefixExtension(adapter.requests[0]!, adapter2.requests[0]!)
+  })
+
+  it('THEOREM: every request rebuilds byte-equal from the session log alone', async () => {
+    const adapter = new MockAdapter([
+      toolCallResponse('c1', 'echo', { text: 'one' }, 'calling'),
+      textResponse('done'),
+      textResponse('after change'),
+    ])
+    const ctx = await harness(adapter)
+    registerEcho(ctx)
+    const agent = ctx.agentLoop.create(AgentId('a1'), { model: 'mock' })
+
+    send(agent, 'go')
+    await waitForIdle(ctx, agent)
+    ctx.systemPrompt.section({ name: 'extra', order: 2, text: 'now with guidance' })
+    ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => ({ ...config, temperature: 0.5 }))
+    send(agent, 'again')
+    await waitForIdle(ctx, agent)
+
+    expect(adapter.requests).toHaveLength(3)
+    const events = agent.session.events
+    const stepStarts = events.filter(e => e.type === 'step/start')
+    expect(stepStarts).toHaveLength(3)
+
+    adapter.requests.forEach((request, index) => {
+      const stepStart = stepStarts[index]!
+      // Messages: the derivation over the log prefix strictly before this
+      // step's step/start — rebuilt here through a completely fresh Session.
+      const rebuilt = new Session(SessionId(`rebuild-${index}`), structuredClone(events.slice(0, stepStart.seq)))
+      expect(structuredClone(request.messages)).toEqual(rebuilt.deriveMessages())
+
+      // Header: the fold of request/header* events up to this step's dispatch
+      // (its header event sits between step/start and the first chunk).
+      const firstChunk = events.find(e => e.type === 'assistant/chunk' && e.seq > stepStart.seq)!
+      const header = foldRequestHeader(events.slice(0, firstChunk.seq))!
+      expect(request.model).toBe(header.config.model)
+      expect(request.system).toEqual(header.system)
+      expect(structuredClone(request.tools ?? [])).toEqual(structuredClone(header.tools ?? []))
+      expect(request.temperature).toBe(header.config.temperature)
+      expect(request.maxTokens).toBe(header.config.maxTokens)
+      expect(request.stop).toEqual(header.config.stop)
+    })
+  })
+})
--- a/packages/core/agent-loop/tests/review-fixes.spec.ts
+++ b/packages/core/agent-loop/tests/review-fixes.spec.ts
@@ -399,9 +399,8 @@ describe('MEDIUM: misc registry and config fixes', () => {
    const ctx = await harness(adapter)
    const agent = ctx.agentLoop.create(AgentId('a1'), {}) // no model — router plugin decides

-    ctx.on('agent/request', async (_agent, _turn, _step, options, next) => {
-      options.model = 'mock'
-      return next()
+    ctx.on('agent/request', async (_agent, _turn, _step, config, _next) => {
+      return { ...config, model: 'mock' }
    })

    send(agent, 'go')
--- a/packages/core/agent/src/types.ts
+++ b/packages/core/agent/src/types.ts
@@ -44,7 +44,7 @@
 */

 import type { Branded } from '@deepseek-ai/dsh-brand'
-import type { ContentBlock, GenerateOptions, Message, MessageSource } from '@deepseek-ai/dsh-llm'
+import type { ContentBlock, LlmCallConfig, Message, MessageSource } from '@deepseek-ai/dsh-llm'
 import type {} from '@deepseek-ai/dsh-system-prompt'

 /** Identifies one live agent in the registry. */
@@ -345,18 +345,28 @@ declare module 'cordis' {
     */
    'agent/prompt-submit'(agent: Agent, content: ContentBlock[], source: MessageSource, next: () => Promise<PromptDecision>): Promise<PromptDecision>
    /**
-     * Waterfall: mutate the fully-assembled {@link GenerateOptions} before the
-     * model call (hooks, model switching, tool filtering, …). Call `next()` to
-     * delegate, or return without it to short-circuit. For surface mutation that
-     * must precede history derivation (compaction), use {@link agent/pre-step}
-     * instead — by the time this fires, `options.messages` is already derived.
+     * Waterfall: shape the step's call configuration — model switching,
+     * sampling overrides — by returning a replacement {@link LlmCallConfig}
+     * (the frozen seed is the config the loop would otherwise use). Config is
+     * ALL a listener shapes here: every request is a pure function of the
+     * session log (the reconstructability RFC), so model-visible content
+     * flows through the log channels — `inject()`, steering, prompt-submit
+     * `additionalContext`, prompt sections via `system-prompt/assemble` —
+     * never through request mutation, and the loop records whatever config
+     * the request actually uses as a `request/header*` event before dispatch.
+     * The step's messages are already snapshotted when this fires (the
+     * `step/start` boundary): an `inject()` from a listener here lands in the
+     * log but joins the NEXT request. For surface mutation that must precede
+     * the snapshot (compaction), use {@link agent/pre-step}. Call `next()` to
+     * delegate, or return an {@link LlmCallConfig} without it to
+     * short-circuit.
     * @param agent - the agent making the model call.
     * @param turn - the open turn number.
     * @param step - the step whose request this is.
-     * @param options - the assembled request; listeners return a transformed copy.
+     * @param config - the config the loop would use (frozen); return a replacement to switch.
     * @mode waterfall
     */
-    'agent/request'(agent: Agent, turn: number, step: number, options: GenerateOptions, next: () => Promise<GenerateOptions>): Promise<GenerateOptions>
+    'agent/request'(agent: Agent, turn: number, step: number, config: LlmCallConfig, next: () => Promise<LlmCallConfig>): Promise<LlmCallConfig>
    /**
     * Waterfall: post-process the assembled assistant {@link Message} before
     * tool dispatch (validation, content rewriting, …).
--- a/packages/core/session/src/index.ts
+++ b/packages/core/session/src/index.ts
@@ -11,9 +11,10 @@ import { isAbsolute } from 'node:path'
 import { deepFreeze } from '@deepseek-ai/dsh-llm'
 import type { ContentBlock, Message, MessageSource } from '@deepseek-ai/dsh-llm'
 import { SESSION_FORMAT_VERSION, SessionId } from './types.ts'
-import type { CreateSessionOptions, SessionEvent, SessionEventMap, SessionEventType, SessionHeader, SurfaceIntent, SurfaceEventType } from './types.ts'
+import type { CreateSessionOptions, EpochHeader, SessionEvent, SessionEventMap, SessionEventType, SessionHeader, SurfaceIntent, SurfaceEventType } from './types.ts'
 import { isJsonValue } from './json.ts'
 import { SurfaceManager, isSurfaceEligibleType } from './surface.ts'
+import { foldRequestHeader } from './request-header.ts'

 export * from './types.ts'
 export { isJsonValue } from './json.ts'
@@ -22,7 +23,7 @@ export { interruptedTurnClosers } from './repair.ts'
 export type { SurfaceNode } from './surface.ts'
 export { isSurfaceEvent, isSurfaceEligibleType } from './surface.ts'
 export { isToolPairingBalanced } from './tool-pairing.ts'
-export { applyHeaderDelta, canonicalHeader, diffHeader, foldRequestHeader } from './request-header.ts'
+export { applyHeaderDelta, canonicalHeader, diffHeader, foldRequestHeader, headerEquals } from './request-header.ts'

 declare module 'cordis' {
  interface Context {
@@ -236,6 +237,27 @@ export class Session {
    return event
  }

+  /** Cached fold of the request-header events — see {@link requestHeader}. */
+  private headerFold: EpochHeader | undefined
+  /** Log position (events consumed) the header fold has reached. */
+  private headerFoldSeq = 0
+
+  /**
+   * The {@link EpochHeader} in force after the log's last header event — the
+   * header the NEXT request will be compared against — or undefined before
+   * the first `request/header` snapshot. The live, incrementally-maintained
+   * form of `foldRequestHeader(session.events)`: each header event is folded
+   * once, when first seen, so a per-step read costs O(new events).
+   * @returns the folded header, or undefined when no header event exists yet.
+   */
+  requestHeader(): EpochHeader | undefined {
+    if (this.headerFoldSeq < this.log.length) {
+      this.headerFold = foldRequestHeader(this.log.slice(this.headerFoldSeq), this.headerFold)
+      this.headerFoldSeq = this.log.length
+    }
+    return this.headerFold
+  }
+
  /** The derived-message cache: frozen projections, extended per unseen node. */
  private derived: Message[] = []
  /** Surface position (nodes projected) the cache has reached. */
--- a/packages/core/session/src/request-header.ts
+++ b/packages/core/session/src/request-header.ts
@@ -104,6 +104,23 @@ function applyTools(prev: readonly ToolSchema[], delta: ToolsDelta): ToolSchema[
  return [...kept, ...delta.added]
 }

+/**
+ * Field-wise equality over canonical headers — the cheap comparison the
+ * writer's round-trip guard runs (`applyHeaderDelta(prev, delta)` must equal
+ * the intended header) and the loop runs to skip logging an unchanged header.
+ * Tools compare per-schema IN ORDER (canonical JSON), so a pure reordering is
+ * correctly unequal.
+ * @param a - one canonical header.
+ * @param b - the other.
+ * @returns whether config, system, and tools (in order) all match.
+ */
+export function headerEquals(a: EpochHeader, b: EpochHeader): boolean {
+  if (!callConfigEquals(a.config, b.config) || a.system !== b.system) return false
+  const at = a.tools ?? []
+  const bt = b.tools ?? []
+  return at.length === bt.length && at.every((tool, i) => sameSchema(tool, bt[i] as ToolSchema))
+}
+
 /**
 * Compute the `request/header-delta` payload between two canonical headers,
 * or undefined when they are equal. The caller MUST round-trip the result
@@ -154,10 +171,12 @@ export function applyHeaderDelta(
 * the dev invariant both use it; the live session tracks the same fold
 * incrementally.
 * @param events - session events in log order (non-header events are skipped).
+ * @param from - a previously folded state to continue from (the live session's
+ *   incremental cursor); omit to fold from nothing.
 * @returns the folded header, or undefined when no header event exists yet.
 */
-export function foldRequestHeader(events: readonly SessionEvent[]): EpochHeader | undefined {
-  let state: EpochHeader | undefined
+export function foldRequestHeader(events: readonly SessionEvent[], from?: EpochHeader): EpochHeader | undefined {
+  let state: EpochHeader | undefined = from
  for (const event of events) {
    if (event.type === 'request/header') {
      state = canonicalHeader(event.data.header)