Context hygiene at insert time: result ids, read stubs, the cap, recall

Harness spec 3.0.0 (context.md, loop.json `context`), in session/results.ts,
applied by the engine when a result goes into the conversation and never
afterwards:

- every result is headed by its id ([#0001]); a read also by the path as
  given, its lines, the file's length and the first 8 hex digits of the
  content's SHA-256; an empty file says so;
- a read of an unchanged file whose lines lie inside an earlier full read is
  a stub naming it; a list/glob/grep repeated with the same arguments and the
  same result is "Same result as #…";
- output over 400 lines or 40,000 characters keeps head and tail with the
  elided line; the full result is kept with the session (a new `results`
  table in sessions.db, gone with the session);
- edit, multiedit, write and apply_patch answer with their hunks under a
  bracket with the new content's hash; a new file is "Created, N lines.";
- recall (new built-in tool) shows any stored result, or a range of it, also
  after a resume; ids are never reused;
- compaction appends the table of the reads still in the conversation to the
  summary.

A stub only ever points at a result still in the conversation, so after a
compaction, an /undo or a rewind the file is read in full. compaction.prune
is gone: cutting old outputs in place threw the prompt cache away (an old
config's key is dropped with a warning). blocks/context.md is in the prompt
wherever recall is offered. Conformance runner for the `context` area (11
cases pass).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

schema/config.schema.json

@@ -241,16 +241,13 @@
       "additionalProperties": false
     },
     "compaction": {
-      "description": "Automatic context management: past auto_at of the window, old tool outputs are pruned, then the conversation is compacted. auto_at: 1 turns it off.",
+      "description": "Automatic context management: past auto_at of the window, the conversation is compacted. auto_at: 1 turns it off. (Old tool outputs are never cut afterwards: that would make the server read the whole conversation again.)",
       "type": "object",
       "properties": {
         "auto_at": {
           "type": "number",
           "minimum": 0.1,
           "maximum": 1
-        },
-        "prune": {
-          "type": "boolean"
         }
       },
       "additionalProperties": false

src/app.ts

@@ -331,7 +331,7 @@ export function createApp(o: AppOptions): App {
   const prototypesBlock = () => (context.prototypes?.offered || (!context.prototypes && loaded.config.prototypes === true) ? { sources: context.prototypes?.sources ?? "" } : undefined)
   /** The three personal keys this session uses: the client's for it, else the config's. */
   const personal = () => (context.personalization ? personalConfig({ enabled: true, available: true, ...context.personalization }) : loaded.config)
-  const system = (m: Mode, current: ResolvedModel, subagent?: { name: string; instructions: string; report: boolean; skills: boolean }, at?: { cwd: string; root: string }, plan?: Plan) =>
+  const system = (m: Mode, current: ResolvedModel, subagent?: { name: string; instructions: string; report: boolean; skills: boolean; recall: boolean }, at?: { cwd: string; root: string }, plan?: Plan) =>
     assembleSystem({
       modelRef: current.ref,
       family: familyOf(current),
@@ -360,6 +360,7 @@ export function createApp(o: AppOptions): App {
       ...(context.instructions ? { instance: { name: contextFrom, instructions: context.instructions } } : {}),
       library: libraryBlock(),
       prototypes: prototypesBlock(),
+      context: subagent ? subagent.recall : true,
     })
   // The other configured models, for the task tool's `model`, or a switch.
   const roster = (current: string) =>
@@ -369,7 +370,7 @@ export function createApp(o: AppOptions): App {
       .slice(0, 15)
       .map((m) => `- ${m.ref}${m.spec.name ? ` (${m.spec.name})` : ""}${m.spec.notes ? ` — ${m.spec.notes}` : ""}`)
       .join("\n")
-  const compaction = { autoAt: loaded.config.compaction?.auto_at, prune: loaded.config.compaction?.prune }
+  const compaction = { autoAt: loaded.config.compaction?.auto_at }
 
   // Subagents (the task tool): a child engine with its own bus and read-tracker, the agent's
   // tools and mode, the same asker (permission prompts still reach the user), no task tool of its
@@ -509,6 +510,7 @@ export function createApp(o: AppOptions): App {
               : spec.instructions,
             report: childTools.some((t) => t.name === "report"),
             skills: childTools.some((t) => t.name === "skill_view"),
+            recall: childTools.some((t) => t.name === "recall"),
           },
           at,
           plan,
@@ -699,7 +701,6 @@ export function createApp(o: AppOptions): App {
       else if (key === "limits.bash_timeout") engine.o.toolCtx.bashTimeoutMs = Number(v) * 1000
       else if (key === "limits.wall_seconds" || key === "limits.output_bytes" || key === "limits.completion_tokens") engine.o.budget = { ...engine.o.budget, [key.slice("limits.".length)]: Number(v) }
       else if (key === "compaction.auto_at") compaction.autoAt = Number(v)
-      else if (key === "compaction.prune") compaction.prune = Boolean(v)
       else if (key === "knowledge") libraryCtx.bases = (v as string[]).length ? (v as string[]) : undefined
       if (key !== "model" && key !== "effort" && key !== "mode") setIn(key, v)
       bus.emit({ type: "setting", key, value: v })

src/config/load.ts

@@ -121,6 +121,14 @@ function legacyKeys(all: Record<string, unknown>, file: string, notes: string[])
     if (!(typeof p === "string" && ["none", "default", "neutral"].includes(p.toLowerCase())))
       notes.push(`${file}: personality "${String(p)}" is not a preset any more (${PERSONALITIES.join(", ")}); none is used`)
   }
+  // compaction.prune cut old tool outputs in place, which threw the prompt cache away; harness spec
+  // 3.0.0 never rewrites what was sent (context.md), so the key is gone.
+  const comp = all.compaction
+  if (comp && typeof comp === "object" && !Array.isArray(comp) && "prune" in comp) {
+    const { prune: _, ...rest } = comp as Record<string, unknown>
+    all.compaction = rest
+    notes.push(`${file}: compaction.prune was removed — old tool outputs are no longer cut afterwards; delete it`)
+  }
   // One file written as a plain string, not a list: read as the list of that one, not dropped
   // without a word by the schema.
   if (typeof all.instruction_files === "string") {

src/config/schema.ts

@@ -276,7 +276,7 @@ export const Config = z.strictObject({
     })
     .optional(),
   question: z.strictObject({ unattended: z.enum(["recommended", "first", "fail"]).optional().describe("With nobody to answer (headless): take the recommended option, the first, or fail.") }).optional(),
-  compaction: z.strictObject({ auto_at: z.number().min(0.1).max(1).optional(), prune: z.boolean().optional() }).optional().describe("Automatic context management: past auto_at of the window, old tool outputs are pruned, then the conversation is compacted. auto_at: 1 turns it off."),
+  compaction: z.strictObject({ auto_at: z.number().min(0.1).max(1).optional() }).optional().describe("Automatic context management: past auto_at of the window, the conversation is compacted. auto_at: 1 turns it off. (Old tool outputs are never cut afterwards: that would make the server read the whole conversation again.)"),
   instructions: z.string().max(4000).optional().describe("Custom instructions: how you want to be helped, in your own words (at most 4000 characters). Appended near the end of the system prompt, under \"## How the person you are talking to wants to be helped\" — as in the LLeMbas web UI. (A list of files here, from an older config, is read as instruction_files.)"),
   instruction_files: z.array(z.string()).optional().describe("More instruction files for the system prompt, besides AGENTS.md / CLAUDE.md."),
   search: z

src/config/settings.ts

@@ -53,7 +53,6 @@ export const SETTINGS: SettingDef[] = [
   { key: "icons", kind: "enum", values: ["emoji", "plain"], agent: true, scopes: USER, fallback: "emoji" },
   { key: "mouse", kind: "boolean", agent: true, scopes: ["global"], reload: true, fallback: "true" },
   { key: "compaction.auto_at", kind: "number", agent: true, scopes: ALL, fallback: "0.85" },
-  { key: "compaction.prune", kind: "boolean", agent: true, scopes: ALL, fallback: "true" },
   { key: "search.order", kind: "list", values: ["searxng", "firecrawl", "ddg"], agent: true, scopes: ALL, fallback: "those configured, then ddg" },
   { key: "search.max_results", kind: "number", agent: true, scopes: ALL, fallback: "8" },
   { key: "memory.enabled", kind: "boolean", agent: true, scopes: ALL, reload: true, fallback: "true" },

src/harness.ts

@@ -111,6 +111,27 @@ export const LOOP: {
     running_max: { default: number }
     subagent_steps: { default: number }
   }
+  /** Context hygiene (spec 3.0.0, harness/context.md): the result ids, the texts of a read's
+   *  header, a stub, the elided line, recall; the output cap; what is deduplicated; the table
+   *  compaction adds. */
+  context: {
+    ids: { digits: number }
+    read_header: string
+    read_header_empty: string
+    edit_header: string
+    header: string
+    unchanged: string
+    same_result: string
+    elided: string
+    created: string
+    no_result: string
+    recall_header: string
+    sha_digits: number
+    cap: { max_lines: number; max_chars: number; head_lines: number; head_chars: number; tail_lines: number; tail_chars: number }
+    dedupe: { by_hash: string[]; by_result: string[] }
+    compaction: { heading: string; line: string; rows: number }
+    reasoning: string
+  }
 } = LOOP_JSON
 
 /** The `purpose` argument (harness/purpose.json). */

src/prompt/assemble.ts

@@ -254,6 +254,8 @@ export interface PromptContext {
   /** The library block (blocks/library.md): set when knowledge tools are offered or the instance
    *  library is in use; `scope` the sentence naming the bases in scope, or "" for all of them. */
   library?: { scope: string }
+  /** recall is offered: the tool results carry ids and may be stubs (blocks/context.md, spec 3.0.0). */
+  context?: boolean
   /** Prototypes are offered (blocks/prototypes.md); `sources` the sentence saying what a page
    *  may load, or "" where nothing says. */
   prototypes?: { sources: string }
@@ -337,6 +339,8 @@ export function assembleSystem(ctx: PromptContext): string {
   else if (ctx.manageSkills && !ctx.subagent) sections.push(promptText("blocks/skills-empty.md"))
   // The library and prototypes (shared blocks): the library whenever its tools are offered, named
   // to the bases in scope; prototypes only where they are offered, with where the page is seen.
+  // How tool results are kept short (ids, stubs, recall): wherever recall is offered.
+  if (ctx.context) sections.push(promptText("blocks/context.md"))
   if (ctx.library && !ctx.subagent) sections.push(fillPrompt(promptText("blocks/library.md"), { knowledge_scope: ctx.library.scope }))
   if (ctx.prototypes && !ctx.subagent) sections.push(fillPrompt(promptText("blocks/prototypes.md"), { prototype_sources: ctx.prototypes.sources, prototype_surface: PROTOTYPE_SURFACE }))
 

src/session/engine.ts

@@ -18,6 +18,7 @@ import { contextWindow } from "../provider/discover.ts"
 import { summaryTurns, type Store } from "./store.ts"
 import { estimateTokens, estimateValue } from "../provider/tokens.ts"
 import { duration } from "../duration.ts"
+import { memoryResults, presentIds, ResultContext } from "./results.ts"
 
 export interface EngineOptions {
   bus: Bus
@@ -39,7 +40,7 @@ export interface EngineOptions {
   budget?: { wall_seconds?: number; output_bytes?: number; completion_tokens?: number }
   /** Rules added by "always allow (project)" are handed here to persist. */
   persistRule?: (rule: Rule) => void
-  compaction?: { autoAt?: number; prune?: boolean }
+  compaction?: { autoAt?: number }
   /** Starts a subagent for the task tool; absent in a subagent. */
   spawn?: ToolContext["spawn"]
   /** The model to switch to when this one's server cannot be reached (its `fallback` list), or
@@ -59,7 +60,6 @@ export interface PromptMeta {
 }
 
 const MAX_PARALLEL = LOOP.parallel.max
-const PRUNED = "[This output was removed to save context. Run the tool again if you need it.]"
 /** What a message sent mid-turn starts with. Said, not hidden: Hermes found that a model refuses an
  *  unmarked one as an injection. A resumed session shows the message without it. */
 /** What /continue (ctrl+g) sends: carry on after being stopped, or after an answer that ended
@@ -139,6 +139,10 @@ export class Engine {
   /** The step's "background now" signal (ToolContext.backgroundNow), while its tools run. */
   private bgNow: AbortController | undefined
 
+  /** The session's tool results: their ids, what each insertion says, what recall shows
+   *  (session/results.ts, harness context.md). */
+  results: ResultContext
+
   constructor(readonly o: EngineOptions) {
     this.mode = o.mode
     this._effort = o.effort
@@ -146,6 +150,13 @@ export class Engine {
     this.client = o.client
     this.sessionId = o.sessionId
     this.plan = this.storedPlan()
+    this.results = this.resultsFor(o.sessionId)
+  }
+
+  /** A session's results: kept with it in the store, or in memory without one. A stub may point
+   *  only at a result still in this conversation. */
+  private resultsFor(sessionId: string | undefined): ResultContext {
+    return new ResultContext(this.o.store && sessionId ? this.o.store.results(sessionId) : memoryResults(), () => presentIds(this.messages))
   }
 
   /** The plan the store holds for this session, if any. */
@@ -191,6 +202,7 @@ export class Engine {
     this.sessionRules = []
     this.recentCalls = []
     this.withdrawn.clear()
+    this.results = this.resultsFor(sessionId)
     // A new session's plan is whatever the store holds for it: none for a fresh one, the old one
     // on a resume.
     this.plan = this.storedPlan()
@@ -230,6 +242,10 @@ export class Engine {
     for await (const ev of this.client.stream({ system: "", messages: [{ role: "user", parts: [{ type: "text", text: prompt }] }], tools: [], effort: null, signal }))
       if (ev.type === "finish") summary = ev.message.parts.filter((p) => p.type === "text").map((p) => (p as { text: string }).text).join("").trim()
     if (!summary) throw new Error("the model returned an empty summary")
+    // The harness's own table of the reads still in the conversation (context.md §7), after the
+    // summary; the conversation is then the summary alone, so no stub can point at a result before it.
+    const table = this.results.compactionTable()
+    if (table) summary = `${summary}\n\n${table}`
     if (this.o.store && this.sessionId) this.o.store.compaction(this.sessionId, summary)
     this.messages = summaryTurns(summary)
     this.usage = undefined
@@ -255,42 +271,16 @@ export class Engine {
     return Math.min(this.usage.output, this.usage.reasoning ?? estimate)
   }
 
-  /** Replace the content of old tool results, keeping the most recent `protect` tokens whole.
-   *  Returns the tokens saved. (OpenCode prunes the same way before it summarises.) */
-  prune(protect: number): number {
-    let recent = 0
-    let saved = 0
-    for (let i = this.messages.length - 1; i >= 0; i--) {
-      const m = this.messages[i]!
-      const size = estimateValue(m)
-      if (recent < protect) {
-        recent += size
-        continue
-      }
-      if (m.role === "tool" && m.content.length > 400 && !m.content.startsWith(PRUNED)) {
-        saved += estimateTokens(m.content) - estimateTokens(PRUNED)
-        this.messages[i] = { ...m, content: PRUNED }
-      }
-    }
-    return saved
-  }
-
-  /** Before a step: past compaction.auto_at of the window, prune; if that is not enough, compact. */
+  /** Before a step: past compaction.auto_at of the window, compact. Old tool results are never
+   *  pruned in place (harness context.md: nothing already sent is rewritten — it would throw the
+   *  prompt cache away from that point); compaction is the one deliberate break. */
   private async manageContext(signal: AbortSignal) {
     const window = contextWindow(this.model)
     const autoAt = this.o.compaction?.autoAt ?? 0.85
     if (!window || autoAt >= 1) return
     const limit = window * autoAt
-    let used = this.contextUsed()
+    const used = this.contextUsed()
     if (used === undefined || used < limit) return
-    if (this.o.compaction?.prune !== false) {
-      const saved = this.prune(Math.min(40_000, Math.floor(window * 0.3)))
-      if (saved > 0) {
-        used -= saved
-        this.o.bus.emit({ type: "notice", message: `context at ${Math.round(((used + saved) / window) * 100)}%: pruned old tool outputs, ~${saved} tokens freed` })
-      }
-      if (used < limit) return
-    }
     // Compacting again and again within one turn loses more each time: the window is simply too
     // small for the task. Say so once and stop, rather than summarising in circles.
     if (this.autoCompactions >= 2) {
@@ -612,6 +602,8 @@ export class Engine {
 
       // Decide every call first — asking is sequential — then run what was allowed.
       const results = new Map<string, ToolResult>()
+      /** Each call's arguments as its tool got them: a repeated listing or search is compared by them. */
+      const argsOf = new Map<string, unknown>()
       const runnable: { call: ToolCallPart; tool: Tool; args: unknown; paths: string[]; writes: boolean; edited?: string }[] = []
       for (const call of calls) {
         if (signal.aborted) break
@@ -619,7 +611,10 @@ export class Engine {
         if ("result" in decided) {
           results.set(call.id, decided.result)
           if (decided.denied) bus.emit({ type: "tool_denied", id: call.id, name: call.name, reason: decided.result.output })
-        } else runnable.push({ call, ...decided })
+        } else {
+          runnable.push({ call, ...decided })
+          argsOf.set(call.id, decided.args)
+        }
       }
       const bgNow = new AbortController()
       this.bgNow = bgNow
@@ -673,7 +668,7 @@ export class Engine {
       for (const call of calls) {
         const r = results.get(call.id) ?? { output: "Not run: the reply was cancelled.", isError: true }
         this.turnOutputBytes += Buffer.byteLength(r.output)
-        this.push({ role: "tool", callId: call.id, name: call.name, content: r.output, isError: r.isError })
+        this.push({ role: "tool", callId: call.id, name: call.name, content: this.inserted(call.name, argsOf.get(call.id), r), isError: r.isError })
         if (r.images) images.push(...r.images)
       }
       // No dialect takes an image inside a tool result the same way; a user turn right after the
@@ -686,6 +681,16 @@ export class Engine {
     }
   }
 
+  /** What a result becomes in the conversation (harness context.md): its id first, a read's lines
+   *  and hash, a stub for an unchanged re-read or a repeated listing, an edit's diff, a long output's
+   *  head and tail. Decided now, once: nothing already sent is rewritten. */
+  private inserted(tool: string, args: unknown, r: ToolResult): string {
+    const c = r.context
+    if (c?.kind === "read" && !r.isError) return this.results.insert({ ...c, body: r.output })
+    if (c && c.kind !== "read" && c.kind !== "output" && !r.isError) return this.results.insert(c)
+    return this.results.insert({ kind: "output", tool, args: args ?? {}, text: r.output, ...(r.isError ? { error: true } : {}) })
+  }
+
   private async decide(
     call: ToolCallPart,
     toolMap: Map<string, Tool>,

src/session/results.ts

@@ -0,0 +1,262 @@
+// Context hygiene (harness spec 3.0.0, harness/context.md; texts and numbers in loop.json `context`;
+// conformance `context`): what a tool result becomes when it is inserted into the conversation.
+// Every result gets an id; a read says its lines and its content's hash; an unchanged re-read inside
+// an earlier read, and a listing or search repeated with the same result, are stubs pointing at the
+// full result; large output keeps its head and tail; edits answer with their diff; `recall` shows any
+// stored result again; compaction adds a table of the reads.
+//
+// The one rule: every measure acts when a result is inserted, and nothing already sent is ever
+// rewritten — changing earlier context throws the prompt cache away from that point. A stub only
+// ever points at a result that is still in the conversation (`present`): after a compaction, an
+// /undo or a rewind, the result it would point at is gone, so the read is answered in full.
+import { createHash } from "node:crypto"
+import { structuredPatch } from "diff"
+import { LOOP } from "../harness.ts"
+
+export const CONTEXT = LOOP.context
+
+export type Cap = typeof CONTEXT.cap
+
+/** One result as the session keeps it: everything needed to answer `recall`, and to decide a
+ *  later stub. `text` is the full result (before any cap; for a stub, the text it stands for). */
+export interface StoredResult {
+  id: string
+  kind: "read" | "edit" | "output" | "recall"
+  /** Not a stub: a stub never becomes the target of another. */
+  full: boolean
+  tool?: string
+  /** read: the path as the call gave it, and what identifies the file (its absolute path). */
+  path?: string
+  key?: string
+  /** read: the full SHA-256 of the file's content, and the lines shown. */
+  sha?: string
+  from?: number
+  to?: number
+  of?: number
+  /** output: the call's tool and arguments (JSON, keys sorted), and the hash of the result. */
+  args?: string
+  textSha?: string
+  text: string
+}
+
+/** Where a session's results are kept: the session store, or memory without one. */
+export interface ResultStore {
+  /** How many ids this session has given: the next is one more. */
+  count(): number
+  save(r: StoredResult): void
+  get(id: string): StoredResult | undefined
+  /** Every result's fields, oldest first (the text may be left out). */
+  all(): StoredResult[]
+}
+
+export function memoryResults(): ResultStore {
+  const rows: StoredResult[] = []
+  return {
+    count: () => rows.length,
+    save: (r) => void rows.push(r),
+    get: (id) => rows.find((r) => r.id === id),
+    all: () => rows,
+  }
+}
+
+/** What a tool's result is, for its insertion (ToolResult.context, filled in by the tool). */
+export type Insert =
+  /** A file read: `body` the read tool's lines (and its note when lines remain). */
+  | { kind: "read"; path: string; key: string; sha: string; from: number; to: number; of: number; body: string }
+  /** One or more files changed: each with its new content's hash, and its hunks, or the number of
+   *  lines of a file created, or deleted. `note` follows the diff (something the tool adds). */
+  | { kind: "edit"; files: EditedFile[]; note?: string }
+  /** Anything else: `args` as the tool got them. */
+  | { kind: "output"; tool: string; args: unknown; text: string; error?: boolean }
+  | { kind: "recall"; target: string; offset?: number; limit?: number }
+
+/** What a tool hands over in its result (ToolResult.context): a read without its body, which is the
+ *  result's output. */
+export type ToolInsert = Exclude<Insert, { kind: "read" }> | Omit<Extract<Insert, { kind: "read" }>, "body">
+
+export type EditedFile = { path: string; sha: string; hunks: string } | { path: string; sha: string; created: number } | { path: string; deleted: true }
+
+export const sha256 = (text: string | Uint8Array) => createHash("sha256").update(text).digest("hex")
+
+/** `{{name}}` filled from `vars` (loop.json's texts, which have no bullets to drop). */
+const fill = (text: string, vars: Record<string, string | number>) => text.replace(/\{\{(\w+)\}\}/g, (m, k: string) => (k in vars ? String(vars[k]) : m))
+const short = (sha: string) => sha.slice(0, CONTEXT.sha_digits)
+
+/** JSON with every object's keys sorted: how two calls' arguments are compared. */
+export function stableJson(v: unknown): string {
+  const norm = (x: unknown): unknown =>
+    Array.isArray(x) ? x.map(norm) : x && typeof x === "object" ? Object.fromEntries(Object.keys(x).sort().map((k) => [k, norm((x as Record<string, unknown>)[k])])) : x
+  return JSON.stringify(norm(v ?? {}))
+}
+
+/** The unified diff of a change, hunks only (the path is in the bracket), three lines of context. */
+export function hunks(before: string, after: string): string {
+  const p = structuredPatch("a", "b", before, after, "", "", { context: 3 })
+  return p.hunks.map((h) => [`@@ -${h.oldStart},${h.oldLines} +${h.newStart},${h.newLines} @@`, ...h.lines].join("\n")).join("\n")
+}
+
+/** How many lines a new file has, as the model wrote it (a final newline ends the last line). */
+export function lineCount(text: string): number {
+  if (!text) return 0
+  const n = text.split("\n").length
+  return text.endsWith("\n") ? n - 1 : n
+}
+
+/** A result over the cap's lines or characters: its head and its tail, with the elided line between
+ *  (loop.json `cap`). The first line alone longer than the head is cut and marked. */
+export function capOutput(text: string, id: string, cap: Cap = CONTEXT.cap): string {
+  const lines = text.split("\n")
+  if (lines.length <= cap.max_lines && text.length <= cap.max_chars) return text
+  const head: string[] = []
+  let chars = 0
+  for (const l of lines) {
+    if (head.length >= cap.head_lines) break
+    const add = (head.length ? 1 : 0) + l.length
+    if (chars + add > cap.head_chars) break
+    head.push(l)
+    chars += add
+  }
+  let used = head.length
+  if (!head.length && lines.length) {
+    head.push(`${lines[0]!.slice(0, cap.head_chars)} […cut]`)
+    used = 1
+  }
+  const tail: string[] = []
+  chars = 0
+  for (let i = lines.length - 1; i >= used; i--) {
+    if (tail.length >= cap.tail_lines) break
+    const l = lines[i]!
+    const add = (tail.length ? 1 : 0) + l.length
+    if (chars + add > cap.tail_chars) break
+    tail.unshift(l)
+    chars += add
+  }
+  const elided = lines.length - used - tail.length
+  return [...head, fill(CONTEXT.elided, { n: elided, id }), ...tail].join("\n")
+}
+
+/** A session's results: the ids it gives, what it keeps, and what each insertion says. */
+export class ResultContext {
+  constructor(
+    private store: ResultStore,
+    /** The ids of the results still in the conversation (a stub may point only at one of them). */
+    private present: () => Set<string>,
+    private cap: Cap = CONTEXT.cap,
+  ) {}
+
+  /** The next id: `#` and lowercase hex, at least ids.digits digits, never reused in the session. */
+  nextId(): string {
+    return `#${(this.store.count() + 1).toString(16).padStart(CONTEXT.ids.digits, "0")}`
+  }
+
+  /** The text the model is given for one result, which is stored in full. `id`: one given from
+   *  outside (the conformance cases name theirs); otherwise the next. */
+  insert(e: Insert, id = this.nextId()): string {
+    if (e.kind === "read") return this.read(id, e)
+    if (e.kind === "edit") return this.edit(id, e)
+    if (e.kind === "recall") return this.recall(id, e)
+    return this.output(id, e)
+  }
+
+  private read(id: string, e: Extract<Insert, { kind: "read" }>): string {
+    const head = e.of === 0 ? fill(CONTEXT.read_header_empty, { id, path: e.path, sha: short(e.sha) }) : fill(CONTEXT.read_header, { id, path: e.path, from: e.from, to: e.to, of: e.of, sha: short(e.sha) })
+    const base: StoredResult = { id, kind: "read", tool: "read", path: e.path, key: e.key, sha: e.sha, from: e.from, to: e.to, of: e.of, full: true, text: e.body }
+    // By hash: an earlier full read of this file, unchanged, whose lines hold these — the newest.
+    const earlier = e.of > 0 && CONTEXT.dedupe.by_hash.includes("read") ? this.table().filter((r) => r.kind === "read" && r.key === e.key && r.sha === e.sha && r.from! <= e.from && e.to <= r.to!).pop() : undefined
+    if (earlier) {
+      this.store.save({ ...base, full: false })
+      return `${head} ${fill(CONTEXT.unchanged, { earlier: earlier.id, from: earlier.from!, to: earlier.to! })}`
+    }
+    this.store.save(base)
+    return `${head}\n${e.body}`
+  }
+
+  private edit(id: string, e: Extract<Insert, { kind: "edit" }>): string {
+    const one = (f: EditedFile): { bracket: string; body: string } =>
+      "deleted" in f ? { bracket: `${f.path} deleted`, body: "" } : { bracket: `${f.path} sha ${short(f.sha)}`, body: "created" in f ? fill(CONTEXT.created, { n: f.created }) : f.hunks }
+    let text: string
+    if (e.files.length === 1) {
+      const { bracket, body } = one(e.files[0]!)
+      text = `[${id} ${bracket}]${body ? ("created" in e.files[0]! ? " " : "\n") + body : ""}`
+    } else {
+      // Several files (apply_patch): the id alone, then each file's bracket and its change.
+      text = [`[${id}]`, ...e.files.map((f) => {
+        const { bracket, body } = one(f)
+        return `[${bracket}]${body ? ("created" in f ? " " : "\n") + body : ""}`
+      })].join("\n")
+    }
+    if (e.note) text += e.note
+    this.store.save({ id, kind: "edit", full: true, text: text.slice(text.indexOf("]") + 1).replace(/^[ \n]/, "") })
+    // The diff is capped like any output, the bracket kept.
+    const nl = text.indexOf("\n")
+    return nl < 0 ? text : `${text.slice(0, nl)}\n${capOutput(text.slice(nl + 1), id, this.cap)}`
+  }
+
+  private output(id: string, e: Extract<Insert, { kind: "output" }>): string {
+    const args = stableJson(e.args)
+    const textSha = sha256(e.text)
+    const base: StoredResult = { id, kind: "output", tool: e.tool, args, textSha, full: true, text: e.text }
+    // By result: the same call, still in the table, whose result was this same text.
+    const dedupe = !e.error && CONTEXT.dedupe.by_result.includes(e.tool)
+    const earlier = dedupe ? this.table().filter((r) => r.kind === "output" && r.tool === e.tool && r.args === args).pop() : undefined
+    if (earlier && earlier.textSha === textSha) {
+      this.store.save({ ...base, full: false })
+      return `${fill(CONTEXT.header, { id })} ${fill(CONTEXT.same_result, { earlier: earlier.id })}`
+    }
+    this.store.save(base)
+    return `${fill(CONTEXT.header, { id })}\n${capOutput(e.text, id, this.cap)}`
+  }
+
+  private recall(id: string, e: Extract<Insert, { kind: "recall" }>): string {
+    const raw = e.target.trim().toLowerCase()
+    const target = raw.startsWith("#") ? raw : `#${raw}`
+    const found = this.store.get(target)
+    if (!found) {
+      const text = fill(CONTEXT.no_result, { target })
+      this.store.save({ id, kind: "recall", full: true, text })
+      return `${fill(CONTEXT.header, { id })} ${text}`
+    }
+    const lines = found.text.split("\n")
+    const from = Math.max(1, e.offset ?? 1)
+    const to = Math.min(lines.length, from + (e.limit ?? 2000) - 1)
+    const shown = lines.slice(from - 1, to).join("\n")
+    this.store.save({ id, kind: "recall", full: true, text: shown })
+    return `${fill(CONTEXT.recall_header, { id, target, from, to, of: lines.length })}\n${shown}`
+  }
+
+  /** The dedupe table: full results still in the conversation, oldest first. */
+  private table(): StoredResult[] {
+    const here = this.present()
+    return this.store.all().filter((r) => r.full && here.has(r.id))
+  }
+
+  /** The table compaction adds after its summary: the newest full read of each path still in the
+   *  conversation, newest last, at most compaction.rows; "" with none. After it the conversation is
+   *  the summary alone, so nothing is present and the dedupe table is empty by construction. */
+  compactionTable(): string {
+    const newest = new Map<string, StoredResult>()
+    for (const r of this.table()) if (r.kind === "read" && r.of! > 0) newest.delete(r.key!), newest.set(r.key!, r)
+    const rows = [...newest.values()].slice(-CONTEXT.compaction.rows)
+    if (!rows.length) return ""
+    return [
+      CONTEXT.compaction.heading,
+      "",
+      CONTEXT.compaction.line,
+      "",
+      "| Path | Lines | sha | Result |",
+      "|---|---|---|---|",
+      ...rows.map((r) => `| ${r.path} | ${r.from}-${r.to} of ${r.of} | ${short(r.sha!)} | ${r.id} |`),
+    ].join("\n")
+  }
+}
+
+/** The ids at the head of the tool results in a conversation: what is present for a stub. */
+export function presentIds(messages: readonly { role: string; content?: unknown }[]): Set<string> {
+  const out = new Set<string>()
+  for (const m of messages) {
+    if (m.role !== "tool" || typeof m.content !== "string") continue
+    const hit = /^\[(#[0-9a-f]+)[\] ]/.exec(m.content)
+    if (hit) out.add(hit[1]!)
+  }
+  return out
+}

src/session/store.ts

@@ -48,6 +48,11 @@ CREATE TABLE IF NOT EXISTS jobs (
   PRIMARY KEY (session_id, id)
 );
 CREATE VIRTUAL TABLE IF NOT EXISTS messages_fts USING fts5(text, content='', contentless_delete=1);
+CREATE TABLE IF NOT EXISTS results (
+  session_id TEXT NOT NULL REFERENCES sessions(id) ON DELETE CASCADE, seq INTEGER NOT NULL, id TEXT NOT NULL,
+  json TEXT NOT NULL, text TEXT NOT NULL,
+  PRIMARY KEY (session_id, id)
+);
 CREATE TABLE IF NOT EXISTS session_meta (
   session_id TEXT NOT NULL REFERENCES sessions(id) ON DELETE CASCADE, key TEXT NOT NULL, value TEXT NOT NULL,
   PRIMARY KEY (session_id, key)
@@ -195,6 +200,31 @@ export class Store {
     return last ? [...summaryTurns(JSON.parse(last.json).summary as string), ...msgs] : msgs
   }
 
+  /** A session's tool results, each in full (harness spec 3.0.0, context.md): what `recall` shows
+   *  and what a stub is decided against. Kept as long as the session is. */
+  results(sessionId: string): import("./results.ts").ResultStore {
+    const q = {
+      count: this.db.query("SELECT COUNT(*) AS n FROM results WHERE session_id = ?"),
+      save: this.db.query("INSERT OR REPLACE INTO results (session_id, seq, id, json, text) VALUES (?, ?, ?, ?, ?)"),
+      get: this.db.query("SELECT json, text FROM results WHERE session_id = ? AND id = ?"),
+      all: this.db.query("SELECT json FROM results WHERE session_id = ? ORDER BY seq"),
+    }
+    type R = import("./results.ts").StoredResult
+    return {
+      count: () => (q.count.get(sessionId) as { n: number }).n,
+      save: (r) => {
+        const { text, ...rest } = r
+        q.save.run(sessionId, (q.count.get(sessionId) as { n: number }).n + 1, r.id, JSON.stringify(rest), text)
+      },
+      get: (id) => {
+        const row = q.get.get(sessionId, id) as { json: string; text: string } | null
+        return row ? ({ ...JSON.parse(row.json), text: row.text } as R) : undefined
+      },
+      // Without their text: the table is decided by hashes and ranges.
+      all: () => (q.all.all(sessionId) as { json: string }[]).map((r) => ({ ...JSON.parse(r.json), text: "" }) as R),
+    }
+  }
+
   /** Something a session keeps besides its messages, as JSON. */
   meta<T>(sessionId: string, key: string): T | undefined {
     const r = this.db.query("SELECT value FROM session_meta WHERE session_id = ? AND key = ?").get(sessionId, key) as { value: string } | null

src/tool/apply_patch.ts

@@ -12,6 +12,7 @@ import { straySkillNote } from "../skill/index.ts"
 import { absPath, defineTool, readText, showPath, ToolError, type ToolContext } from "./tool.ts"
 import { diffStat } from "./write.ts"
 import { described } from "../harness.ts"
+import { hunks, lineCount, sha256, type EditedFile } from "../session/results.ts"
 
 function hunksOf(text: string): Hunk[] {
   try {
@@ -35,7 +36,7 @@ function namedFiles(a: { patch: string; path?: string }): string[] {
   return hunksOf(a.patch).flatMap((h) => (h.type === "update" && h.movePath ? [h.path, h.movePath] : [h.path]))
 }
 
-type Op = { kind: "write"; path: string; content: string; before: string } | { kind: "delete"; path: string; before: string }
+type Op = { kind: "write"; path: string; content: string; before: string; created?: boolean } | { kind: "delete"; path: string; before: string }
 
 function opsOf(original: Map<string, string | null>, current: Map<string, string | null>): Op[] {
   const ops: Op[] = []
@@ -43,7 +44,7 @@ function opsOf(original: Map<string, string | null>, current: Map<string, string
     const before = original.get(path) ?? null
     if (now === before) continue
     if (now === null) ops.push({ kind: "delete", path, before: before! })
-    else ops.push({ kind: "write", path, content: now, before: before ?? "" })
+    else ops.push({ kind: "write", path, content: now, before: before ?? "", ...(before === null ? { created: true } : {}) })
   }
   return ops
 }
@@ -170,6 +171,12 @@ export const applyPatchTool = defineTool({
     const all = patches.join("\n")
     const { added, removed } = diffStat(all)
     const names = ops.map((o) => showPath(o.path, ctx)).join(", ")
-    return { output: `Applied to ${names} (+${added} −${removed}).${straySkillNote(ops.filter((o) => o.kind === "write").map((o) => o.path), ctx)}`, title: `${names} +${added} −${removed}`, meta: { diff: all } }
+    const note = straySkillNote(ops.filter((o) => o.kind === "write").map((o) => o.path), ctx)
+    const files: EditedFile[] = ops.map((op) => {
+      const path = showPath(op.path, ctx)
+      if (op.kind === "delete") return { path, deleted: true }
+      return op.created ? { path, sha: sha256(op.content), created: lineCount(op.content) } : { path, sha: sha256(op.content), hunks: hunks(op.before, op.content) }
+    })
+    return { output: `Applied to ${names} (+${added} −${removed}).${note}`, title: `${names} +${added} −${removed}`, meta: { diff: all }, context: { kind: "edit", files, ...(note ? { note } : {}) } }
   },
 })

src/tool/edit.ts

@@ -5,6 +5,7 @@ import { replace } from "./replace-text.ts"
 import { absPath, defineTool, readText, showPath, ToolError, type ToolContext } from "./tool.ts"
 import { diffStat, guardExisting } from "./write.ts"
 import { described } from "../harness.ts"
+import { hunks, sha256 } from "../session/results.ts"
 
 async function planEdit(a: { path: string; old: string; new: string; all?: boolean }, ctx: ToolContext) {
   const abs = absPath(a.path, ctx)
@@ -43,6 +44,7 @@ export const editTool = defineTool({
     ctx.fileStamps.set(abs, (await stat(abs)).mtimeMs)
     const patch = createTwoFilesPatch(display, display, before, after, "", "", { context: 3 })
     const { added, removed } = diffStat(patch)
-    return { output: `Edited ${display} (+${added} −${removed}).`, title: `${display} +${added} −${removed}`, meta: { diff: patch } }
+    const written = crlf ? after.replaceAll("\n", "\r\n") : after
+    return { output: `Edited ${display} (+${added} −${removed}).`, title: `${display} +${added} −${removed}`, meta: { diff: patch }, context: { kind: "edit", files: [{ path: a.path, sha: sha256(written), hunks: hunks(before, after) }] } }
   },
 })

src/tool/multiedit.ts

@@ -5,6 +5,7 @@ import { replace } from "./replace-text.ts"
 import { absPath, defineTool, readText, showPath, ToolError, type ToolContext } from "./tool.ts"
 import { diffStat, guardExisting } from "./write.ts"
 import { described } from "../harness.ts"
+import { hunks, sha256 } from "../session/results.ts"
 
 async function planEdits(a: { path: string; edits: { old: string; new: string; all?: boolean }[] }, ctx: ToolContext) {
   const abs = absPath(a.path, ctx)
@@ -45,6 +46,7 @@ export const multieditTool = defineTool({
     ctx.fileStamps.set(abs, (await stat(abs)).mtimeMs)
     const patch = createTwoFilesPatch(display, display, before, after, "", "", { context: 3 })
     const { added, removed } = diffStat(patch)
-    return { output: `Edited ${display}: ${a.edits.length} changes (+${added} −${removed}).`, title: `${display} +${added} −${removed}`, meta: { diff: patch } }
+    const written = crlf ? after.replaceAll("\n", "\r\n") : after
+    return { output: `Edited ${display}: ${a.edits.length} changes (+${added} −${removed}).`, title: `${display} +${added} −${removed}`, meta: { diff: patch }, context: { kind: "edit", files: [{ path: a.path, sha: sha256(written), hunks: hunks(before, after) }] } }
   },
 })

src/tool/read.ts

@@ -5,6 +5,7 @@ import { dirname, join } from "node:path"
 import { z } from "zod"
 import { absPath, defineTool, showPath, ToolError } from "./tool.ts"
 import { described } from "../harness.ts"
+import { sha256 } from "../session/results.ts"
 
 const MAX_LINES = 2000
 const MAX_LINE = 2000
@@ -37,23 +38,35 @@ export const readTool = defineTool({
     if (bytes.includes(0)) throw new ToolError(`${a.path} is a binary file (${st.size} bytes)`)
     // Read whole, so only up to a size that fits in memory comfortably; past it, grep or bash.
     if (st.size > 50_000_000) throw new ToolError(`${a.path} is ${Math.round(st.size / 1e6)} MB — too large to read whole; search it with grep, or look at a part with bash (head, tail, sed -n)`)
-    const text = await file.text()
-    const all = text.split("\n")
-    if (all.length > 1 && all[all.length - 1] === "") all.pop()
-    const start = (a.offset ?? 1) - 1
-    const end = Math.min(all.length, start + (a.limit ?? MAX_LINES))
-    const body = all
-      .slice(start, end)
-      .map((l, i) => `${start + i + 1}: ${l.length > MAX_LINE ? l.slice(0, MAX_LINE) + " […cut]" : l}`)
-      .join("\n")
+    const raw = await file.bytes()
+    const text = new TextDecoder().decode(raw)
+    const { body, from, to, of } = readBody(text, a.offset, a.limit)
     ctx.readFiles.add(abs)
     ctx.fileStamps.set(abs, st.mtimeMs)
-    const more = end < all.length ? `\n\n(${all.length - end} more lines — read again with offset ${end + 1})` : ""
-    const empty = all.length === 0 ? "(empty file)" : ""
-    return { output: (empty || body) + more + nearbyInstructions(abs, ctx), title: `${showPath(abs, ctx)} · ${start + 1}–${end} of ${all.length}`, meta: { path: showPath(abs, ctx), from: start + 1, to: end, of: all.length } }
+    // For its insertion (session/results.ts): the lines shown and the content's hash, by which an
+    // unchanged re-read becomes a stub.
+    const context = { kind: "read" as const, path: a.path, key: abs, sha: sha256(raw), from, to, of }
+    return { output: body + nearbyInstructions(abs, ctx), title: `${showPath(abs, ctx)} · ${from}–${to} of ${of}`, meta: { path: showPath(abs, ctx), from, to, of }, context }
   },
 })
 
+/** The read tool's body for a file's text: numbered lines from `offset` (1-indexed), at most
+ *  `limit` (2000), a line longer than 2000 characters cut, and the note when lines remain; "(empty
+ *  file)" for an empty one. `from`-`to` of `of`: the lines shown (conformance `context`). */
+export function readBody(text: string, offset?: number, limit?: number): { body: string; from: number; to: number; of: number } {
+  const all = text.split("\n")
+  if (all.length > 1 && all[all.length - 1] === "") all.pop()
+  if (all.length === 1 && all[0] === "") all.pop()
+  const start = (offset ?? 1) - 1
+  const end = Math.min(all.length, start + (limit ?? MAX_LINES))
+  const lines = all
+    .slice(start, end)
+    .map((l, i) => `${start + i + 1}: ${l.length > MAX_LINE ? l.slice(0, MAX_LINE) + " […cut]" : l}`)
+    .join("\n")
+  const more = end < all.length ? `\n\n(${all.length - end} more lines — read again with offset ${end + 1})` : ""
+  return { body: (all.length === 0 ? "(empty file)" : lines) + more, from: start + 1, to: end, of: all.length }
+}
+
 /** OpenCode's "nearby instructions": reading a file under a directory with its own AGENTS.md or
  *  CLAUDE.md hands the model that file, once per session. The root's is already in the prompt. */
 function nearbyInstructions(file: string, ctx: import("./tool.ts").ToolContext): string {

src/tool/recall.ts

@@ -0,0 +1,19 @@
+// recall (harness spec 3.0.0, context.md): a stored result of this session shown again, in full or
+// from a line on. The engine answers it from the session's results (session/results.ts); the tool
+// only carries the call there, which is why its run says what it is.
+import { z } from "zod"
+import { described } from "../harness.ts"
+import { defineTool } from "./tool.ts"
+
+export const recallTool = defineTool({
+  ...described("recall"),
+  access: "recall",
+  purpose: false,
+  schema: z.object({ id: z.string(), offset: z.number().int().min(1).optional(), limit: z.number().int().min(1).optional() }),
+  permission: (a) => ({ permission: "recall", class: "read", patterns: [a.id] }),
+  async run(a) {
+    // The text is the session's insertion (`context` recall); this is what a transcript shows.
+    const target = a.id.trim().startsWith("#") ? a.id.trim() : `#${a.id.trim()}`
+    return { output: `recall ${target}`, title: target, context: { kind: "recall", target: a.id, ...(a.offset ? { offset: a.offset } : {}), ...(a.limit ? { limit: a.limit } : {}) } }
+  },
+})

src/tool/registry.ts

@@ -18,6 +18,7 @@ import { sessionSearchTool } from "./session_search.ts"
 import { skillManageTool, skillsListTool, skillViewTool } from "./skills.ts"
 import { settingsTool } from "./settings.ts"
 import { showTool } from "./show.ts"
+import { recallTool } from "./recall.ts"
 import { knowledgeGetTool, knowledgeSearchTool, noteManageTool, notesSearchTool, noteViewTool } from "./library.ts"
 import type { Tool } from "./tool.ts"
 import { writeTool } from "./write.ts"
@@ -58,6 +59,7 @@ export const BUILTIN_TOOLS: Tool[] = [
   settingsTool,
   viewImageTool,
   showTool,
+  recallTool,
   // Offered to subagents only (the spec's `offered: subagent`; app.ts SUBAGENT_ONLY).
   reportTool,
 ] as Tool[]

src/tool/tool.ts

@@ -107,6 +107,9 @@ export interface ToolResult {
   meta?: Record<string, unknown>
   /** Images for the model (view_image): sent in a user turn after the tool results. */
   images?: import("../provider/types.ts").ImagePart[]
+  /** What the result is, for its insertion into the conversation (session/results.ts, harness
+   *  context.md): a read's lines and hash, an edit's diff. Absent: an ordinary output. */
+  context?: import("../session/results.ts").ToolInsert
 }
 
 export interface Tool<S extends z.ZodType = z.ZodType> {

src/tool/write.ts

@@ -5,6 +5,7 @@ import { z } from "zod"
 import { straySkillNote } from "../skill/index.ts"
 import { absPath, defineTool, showPath, ToolError, type ToolContext } from "./tool.ts"
 import { described } from "../harness.ts"
+import { hunks, lineCount, sha256 } from "../session/results.ts"
 
 /** Refuse to overwrite a file the model has not read, or one changed on disk since it was read. */
 export async function guardExisting(abs: string, display: string, ctx: ToolContext): Promise<boolean> {
@@ -57,10 +58,14 @@ export const writeTool = defineTool({
     ctx.fileStamps.set(abs, st.mtimeMs)
     const patch = createTwoFilesPatch(display, display, before, a.content, "", "", { context: 3 })
     const { added, removed } = diffStat(patch)
+    const note = straySkillNote([abs], ctx)
+    const sha = sha256(a.content)
     return {
-      output: (existed ? `Replaced ${display} (+${added} −${removed}).` : `Created ${display} (${a.content.split("\n").length} lines).`) + straySkillNote([abs], ctx),
+      output: (existed ? `Replaced ${display} (+${added} −${removed}).` : `Created ${display} (${a.content.split("\n").length} lines).`) + note,
       title: `${display} +${added} −${removed}`,
       meta: { diff: patch },
+      // Over an existing file the diff; a new one is said in a line — the model wrote it and has it.
+      context: { kind: "edit", files: [existed ? { path: a.path, sha, hunks: hunks(before, a.content) } : { path: a.path, sha, created: lineCount(a.content) }], ...(note ? { note } : {}) },
     }
   },
 })

tests/approval.test.ts

@@ -25,7 +25,7 @@ const result = () => fake!.requests[1].messages.at(-1).content as string
 test("the edited command runs, and the result says so first", async () => {
   const a = app("echo two")
   await a.engine.prompt("go")
-  expect(result()).toStartWith("The user changed the command before allowing it. What ran: echo two")
+  expect(result()).toStartWith("[#0001]\nThe user changed the command before allowing it. What ran: echo two")
   expect(result()).toContain("two")
   expect(result()).not.toContain("one\n")
 })

tests/conformance.ts

@@ -19,6 +19,8 @@ import { resolveBase } from "../src/library/store.ts"
 import { NarrowError, narrowParameters, type Narrow } from "../src/harness.ts"
 import { fillPrompt } from "../src/prompt/fill.ts"
 import { instructionChain } from "../src/prompt/assemble.ts"
+import { CONTEXT, hunks, lineCount, memoryResults, presentIds, ResultContext, sha256 } from "../src/session/results.ts"
+import { readBody } from "../src/tool/read.ts"
 import { applyOps, type Op, type Target } from "../src/memory/store.ts"
 import { jobNote, type NoteCase } from "../src/jobs/note.ts"
 
@@ -115,6 +117,30 @@ export const AREAS: Record<string, (c: Case) => unknown> = {
     const files = new Set(c.files as string[])
     return instructionChain(c.boundary as string, c.start as string, (p) => files.has(p) && readableUnder(c.grants as Record<string, string> | undefined, p))
   },
+  // Context hygiene: one session's insertions, step by step. The conversation is the results
+  // inserted since the last compaction, which is what a stub may point at.
+  context: (c) => {
+    const messages: { role: string; content: string }[] = []
+    const rc = new ResultContext(memoryResults(), () => presentIds(messages), { ...CONTEXT.cap, ...((c.limits ?? {}) as object) })
+    return (c.steps as Record<string, any>[]).map((st) => {
+      if (st.op === "compact") {
+        const table = rc.compactionTable()
+        messages.length = 0
+        return table
+      }
+      let text: string
+      if (st.op === "read") {
+        const { body, from, to, of } = readBody(st.content, st.offset, st.limit)
+        text = rc.insert({ kind: "read", path: st.path, key: st.path, sha: sha256(st.content), from, to, of, body }, st.id)
+      } else if (st.op === "edit" || st.op === "write") {
+        const sha = sha256(st.after)
+        text = rc.insert({ kind: "edit", files: [st.before === undefined ? { path: st.path, sha, created: lineCount(st.after) } : { path: st.path, sha, hunks: hunks(st.before, st.after) }] }, st.id)
+      } else if (st.op === "recall") text = rc.insert({ kind: "recall", target: st.target, offset: st.offset, limit: st.limit }, st.id)
+      else text = rc.insert({ kind: "output", tool: st.tool, args: st.args, text: st.text }, st.id)
+      messages.push({ role: "tool", content: text })
+      return text
+    })
+  },
   library_scope: (c) => resolveBase(c.bases as { id: string; name: string }[], c.base as string | null),
   toolcall_index: (c) => {
     // The deltas as the OpenAI stream carries them: `function.name` and `function.arguments`.

tests/context.test.ts

@@ -45,12 +45,14 @@ describe("context", () => {
     expect(fake!.calls.some((c) => c.path.endsWith("/unload"))).toBe(true)
   })
 
-  test("past auto_at: the older of two large outputs is pruned, the recent one kept, no compaction", async () => {
-    // window 16000, auto_at 0.6 → limit 9600; the most recent 30% (4800 tokens) is protected.
+  test("past auto_at: nothing already sent is cut; the conversation is compacted, with a table of the reads", async () => {
+    // window 16000, auto_at 0.6 → limit 9600. Harness spec 3.0.0 (context.md): old tool outputs are
+    // never pruned in place — that would make the server read everything after them again.
     const a = app(`connections:\n  f:\n    dialect: openai-chat\n    base_url: URL\n    models: { m: { context: 16000 } }\n`, [
       { chunks: [toolCall(0, "c1", "read", '{"path":"big.txt"}'), usage(1000, 10)] },
       { chunks: [toolCall(0, "c2", "read", '{"path":"big2.txt"}'), usage(6600, 10)] },
-      { chunks: [delta({ content: "done" }), usage(7000, 5)] },
+      { chunks: [delta({ content: "## Summary\nRead two files." })] },
+      { chunks: [delta({ content: "done" }), usage(2000, 5)] },
     ], "compaction: { auto_at: 0.6 }\n")
     const lines = Array.from({ length: 1000 }, (_, i) => `line ${i} of a long file`).join("\n")
     writeFileSync(join(a.project.root, "big.txt"), lines)
@@ -58,14 +60,21 @@ describe("context", () => {
     const notices: string[] = []
     a.bus.on((e: Event) => e.type === "notice" && notices.push(e.message))
     await a.engine.prompt("go")
-    expect(notices.some((n) => n.includes("pruned old tool outputs"))).toBe(true)
-    expect(notices.some((n) => n.includes("compacting"))).toBe(false)
-    const results = fake!.requests.at(-1).messages.filter((m: any) => m.role === "tool").map((m: any) => m.content as string)
-    expect(results[0]).toContain("removed to save context")
-    expect(results[1]).toContain("line 999 of a long file")
+    expect(notices.some((n) => n.includes("pruned"))).toBe(false)
+    expect(notices.some((n) => n.includes("compacting"))).toBe(true)
+    // The first read went out whole, and stayed whole until the summary replaced the conversation.
+    const sent = fake!.requests[1].messages.filter((m: any) => m.role === "tool").map((m: any) => m.content as string)
+    expect(sent[0]).toStartWith("[#0001 big.txt lines 1-1000 of 1000 sha ")
+    expect(sent[0]).toContain("line 999 of a long file")
+    // After it: the summary, the harness's table of what was read, and the request word for word.
+    const after = JSON.stringify(fake!.requests[3].messages)
+    expect(after).toContain("Read two files.")
+    expect(after).toContain("## Files read before this summary")
+    expect(after).toMatch(/\| big\.txt \| 1-1000 of 1000 \| [0-9a-f]{8} \| #0001 \|/)
+    expect(after).toMatch(/\| big2\.txt \| 1-1000 of 1000 \| [0-9a-f]{8} \| #0002 \|/)
   })
 
-  test("still too full after pruning: compacted mid-task, then told to carry on", async () => {
+  test("too full: compacted mid-task, then told to carry on", async () => {
     const a = app(`connections:\n  f:\n    dialect: openai-chat\n    base_url: URL\n    models: { m: { context: 4000 } }\n`, [
       { chunks: [delta({ content: "y".repeat(8000) }), toolCall(0, "c1", "list", "{}"), usage(3900, 2000)] },
       { chunks: [delta({ content: "## What we are doing\nA long task." })] },
@@ -91,7 +100,7 @@ describe("context", () => {
     expect(text).toContain("window 32.8k")
     expect(text).toContain("system prompt")
     expect(text).toMatch(/tool definitions \(\d+\)/)
-    expect(text).toMatch(/tool results\s+503/) // one 6000-character line, cut at 2000 by read
+    expect(text).toMatch(/tool results\s+514/) // one 6000-character line, cut at 2000 by read, under its header
     expect(text).toMatch(/in use\s+3\.1k/)
   })
 })

tests/e2e.test.ts

@@ -45,12 +45,15 @@ describe("end to end", () => {
     // The tool result went back to the model as a tool message with the right id.
     expect(fake!.requests[1].messages.at(-1)).toMatchObject({ role: "tool", tool_call_id: "c1" })
     expect(fake!.requests[1].messages.at(-1).content).toContain("1: hello world")
+    // Spec 3.0.0 (context.md): the read is headed by its id, lines and hash; the edit answers with its diff.
+    expect(fake!.requests[1].messages.at(-1).content).toMatch(/^\[#0001 hello\.txt lines 1-1 of 1 sha [0-9a-f]{8}\]\n1: hello world$/)
+    expect(fake!.requests[2].messages.at(-1).content).toMatch(/^\[#0002 hello\.txt sha [0-9a-f]{8}\]\n@@ -1,1 \+1,1 @@\n-hello world\n\+hello lembas$/)
     // System prompt carries env and the mode.
     expect(fake!.requests[0].messages[0].content).toContain(`Project root: ${dir}`)
     expect(fake!.requests[0].messages[0].content).toContain("Permission mode: edit")
     // knowledge_search/get are offered once any knowledge base has documents, and the suite shares
     // one LEMBAS_HOME: another file's base must not decide this list (found in the full run).
-    expect(fake!.requests[0].tools.map((t: any) => t.function.name).filter((n: string) => !n.startsWith("knowledge_"))).toEqual(["read", "write", "edit", "multiedit", "apply_patch", "glob", "grep", "list", "bash", "job_output", "job_list", "job_stop", "job_send", "web_search", "web_fetch", "task", "plan_write", "plan_update", "ask_user", "memory", "session_search", "notes_search", "note_view", "note_manage", "skills_list", "skill_view", "skill_manage", "settings"])
+    expect(fake!.requests[0].tools.map((t: any) => t.function.name).filter((n: string) => !n.startsWith("knowledge_"))).toEqual(["read", "write", "edit", "multiedit", "apply_patch", "glob", "grep", "list", "bash", "job_output", "job_list", "job_stop", "job_send", "web_search", "web_fetch", "task", "plan_write", "plan_update", "ask_user", "memory", "session_search", "notes_search", "note_view", "note_manage", "skills_list", "skill_view", "skill_manage", "settings", "recall"])
     const ends = events.filter((e) => e.type === "tool_end")
     expect(ends.map((e) => e.type === "tool_end" && e.result.title)).toEqual(["hello.txt · 1–1 of 1", "hello.txt +1 −1"])
     const sid = store.sessions(1)[0]!.id

tests/harness.test.ts

@@ -69,8 +69,7 @@ test("every tool says how it is shown (block), from the spec's list", () => {
 })
 
 test("every tool the spec gives the CLI is built in, and every built-in tool is in the spec", () => {
-  // recall comes with context hygiene (harness/context.md), in the commit that implements it.
-  expect(ours.map((t) => t.name).sort()).toEqual(forUs.map((t) => t.name).filter((n) => n !== "recall").sort())
+  expect(ours.map((t) => t.name).sort()).toEqual(forUs.map((t) => t.name).sort())
 })
 
 for (const t of ours) {

tests/jobs.test.ts

@@ -204,7 +204,7 @@ describe("in an app", () => {
     a.newSession()
     expect(a.jobs()).toEqual([])
     await a.turns.prompt("list")
-    expect(fake!.requests[3].messages.at(-1).content).toBe("No background jobs.")
+    expect(fake!.requests[3].messages.at(-1).content).toBe("[#0001]\nNo background jobs.")
     expect(a.registry.running(first)).toHaveLength(1)
     // Deleting a session stops its jobs.
     expect(a.registry.stopAll(first)).toBe(1)

tests/mcp.test.ts

@@ -191,7 +191,7 @@ describe("through the engine", () => {
     expect(asked.join()).toContain("mcp__fix__echo")
     expect(fake.requests[0].tools.map((t: any) => t.function.name)).toContain("mcp__fix__echo")
     expect(fake.requests[0].messages[0].content).toContain("<mcp_instructions>")
-    expect(fake.requests[1].messages.at(-1)).toMatchObject({ role: "tool", content: "echo: via engine" })
+    expect(fake.requests[1].messages.at(-1)).toMatchObject({ role: "tool", content: "[#0001]\necho: via engine" })
     await app.close()
     writeFileSync(join(paths.config, "config.yaml"), "")
   })

tests/plan.test.ts

@@ -44,7 +44,7 @@ describe("plan mode", () => {
     expect(readFileSync(join(a.project.root, seen[0]!.path), "utf8")).toBe(RENDERED + "\n")
     expect(a.engine.plan?.steps).toEqual(["calc.py: divide by len(values)"])
     // The model read the plan back with its ids from plan_write.
-    expect(fake!.requests[1].messages.at(-1).content).toBe(RENDERED)
+    expect(fake!.requests[1].messages.at(-1).content).toBe(`[#0001]\n${RENDERED}`)
     expect(names(fake!.requests[0])).toContain("plan_submit")
     expect(names(fake!.requests[0])).toContain("plan_write")
     expect(fake!.requests[0].messages[0].content).toContain("Permission mode: plan")

tests/results.test.ts

@@ -0,0 +1,97 @@
+// Context hygiene through the engine (harness spec 3.0.0, context.md): ids, stubs, the cap, recall,
+// and the rule that a stub only ever points at a result still in the conversation. The texts are
+// the conformance area `context`'s; here they are checked where the engine inserts them.
+import { afterEach, expect, test } from "bun:test"
+import { mkdirSync, mkdtempSync, writeFileSync } from "node:fs"
+import { tmpdir } from "node:os"
+import { join } from "node:path"
+import { createApp } from "../src/app.ts"
+import type { AskReply } from "../src/bus/index.ts"
+import { paths } from "../src/config/paths.ts"
+import { Store } from "../src/session/store.ts"
+import { delta, fakeProvider, toolCall, type Fake } from "./fake-provider.ts"
+
+let fake: Fake | undefined
+afterEach(() => fake?.stop())
+
+const yes = { ask: async (): Promise<AskReply> => ({ kind: "once" }) }
+
+function setup(script: Parameters<typeof fakeProvider>[0]) {
+  fake = fakeProvider(script)
+  mkdirSync(paths.config, { recursive: true })
+  writeFileSync(join(paths.config, "connections.yaml"), `connections:\n  fake:\n    dialect: openai-chat\n    base_url: ${fake.url}\n    models:\n      m: { context: 32768 }\n`, { mode: 0o600 })
+  writeFileSync(join(paths.config, "config.yaml"), "model: fake/m\n")
+  const dir = mkdtempSync(join(tmpdir(), "lembas-results-"))
+  writeFileSync(join(dir, "a.txt"), "one\ntwo\nthree\n")
+  return dir
+}
+
+/** The tool results the model was sent in request `i`, in order. */
+const sent = (i: number): string[] => fake!.requests[i].messages.filter((m: any) => m.role === "tool").map((m: any) => m.content as string)
+
+test("an unchanged re-read is a stub; after an edit the file is read in full; a repeated listing is a stub", async () => {
+  const dir = setup([
+    { chunks: [toolCall(0, "c1", "read", '{"path":"a.txt"}')] },
+    { chunks: [toolCall(0, "c2", "read", '{"path":"a.txt","offset":2,"limit":1}')] },
+    { chunks: [toolCall(0, "c3", "glob", '{"pattern":"*.txt"}')] },
+    { chunks: [toolCall(0, "c4", "glob", '{"pattern":"*.txt"}')] },
+    { chunks: [toolCall(0, "c5", "edit", JSON.stringify({ path: "a.txt", old: "two", new: "2" }))] },
+    { chunks: [toolCall(0, "c6", "read", '{"path":"a.txt"}')] },
+    { chunks: [delta({ content: "done" })] },
+  ])
+  const app = createApp({ cwd: dir, mode: "edit", asker: yes, store: false })
+  await app.engine.prompt("go")
+  const r = sent(6)
+  expect(r[0]).toMatch(/^\[#0001 a\.txt lines 1-3 of 3 sha ([0-9a-f]{8})\]\n1: one\n2: two\n3: three$/)
+  const sha = /sha ([0-9a-f]{8})/.exec(r[0]!)![1]
+  expect(r[1]).toBe(`[#0002 a.txt lines 2-2 of 3 sha ${sha}] Unchanged since #0001, which already shows lines 1-3.`)
+  expect(r[2]).toBe("[#0003]\na.txt")
+  expect(r[3]).toBe("[#0004] Same result as #0003.")
+  expect(r[4]).toMatch(/^\[#0005 a\.txt sha [0-9a-f]{8}\]\n@@ -1,3 \+1,3 @@\n one\n-two\n\+2\n three$/)
+  expect(r[5]).toMatch(/^\[#0006 a\.txt lines 1-3 of 3 sha [0-9a-f]{8}\]\n1: one\n2: 2\n3: three$/)
+  // The prompt says what the ids and stubs are.
+  expect(fake!.requests[0].messages[0].content).toContain("# Tool results\n- Every tool result starts with its id in brackets")
+})
+
+test("a long output keeps its head and tail; recall shows it whole, from the store, after a resume too", async () => {
+  const dir = setup([
+    { chunks: [toolCall(0, "c1", "bash", JSON.stringify({ command: "seq 1 1000", description: "count" }))] },
+    { chunks: [toolCall(0, "c2", "recall", '{"id":"#0001","offset":500,"limit":3}')] },
+    { chunks: [delta({ content: "done" })] },
+    // After a resume: a new id, never one given before; an unknown one said so.
+    { chunks: [toolCall(0, "c3", "recall", '{"id":"0001","offset":999}'), toolCall(1, "c4", "recall", '{"id":"#ffff"}')] },
+    { chunks: [delta({ content: "again" })] },
+  ])
+  const store = new Store()
+  const app = createApp({ cwd: dir, mode: "auto", asker: yes, store, modelTitles: false })
+  await app.engine.prompt("count")
+  const r = sent(2)
+  const lines = r[0]!.split("\n")
+  expect(lines[0]).toBe("[#0001]")
+  expect(lines.slice(1, 3)).toEqual(["1", "2"])
+  expect(r[0]).toContain("\n200\n[")
+  expect(r[0]).toMatch(/\n\[\d+ lines elided, full output #0001\]\n/)
+  expect(r[0]).toContain("\n1000\n\n[exit 0]")
+  expect(r[1]).toBe("[#0002 recall #0001 lines 500-502 of 1002]\n500\n501\n502")
+  const id = app.engine.sessionId!
+  const again = createApp({ cwd: dir, mode: "auto", asker: yes, store, resuming: id, modelTitles: false })
+  again.resume(id)
+  await again.engine.prompt("more")
+  const s = sent(4)
+  expect(s.slice(-2)).toEqual(["[#0003 recall #0001 lines 999-1002 of 1002]\n999\n1000\n\n[exit 0]", "[#0004] No stored result #ffff in this session."])
+})
+
+test("a stub never points at a result that left the conversation: after /undo the file is read in full", async () => {
+  const dir = setup([
+    { chunks: [toolCall(0, "c1", "read", '{"path":"a.txt"}')] },
+    { chunks: [delta({ content: "read" })] },
+    { chunks: [toolCall(0, "c2", "read", '{"path":"a.txt"}')] },
+    { chunks: [delta({ content: "read again" })] },
+  ])
+  const app = createApp({ cwd: dir, mode: "edit", asker: yes, store: false })
+  await app.turns.prompt("read it", [], "read it")
+  await app.turns.undo()
+  await app.turns.prompt("read it", [], "read it")
+  expect(sent(3)).toHaveLength(1)
+  expect(sent(3)[0]).toMatch(/^\[#0002 a\.txt lines 1-3 of 3 sha [0-9a-f]{8}\]\n1: one/)
+})

tests/subagent.test.ts

@@ -76,7 +76,7 @@ describe("subagents", () => {
     expect(child.messages[0].content).not.toContain("## Skills")
     // the parent gets the report, not the reading
     const result = fake!.requests[3].messages.at(-1).content as string
-    expect(result).toStartWith("calc.py:2 divides by len(v) - 1")
+    expect(result).toStartWith("[#0001]\ncalc.py:2 divides by len(v) - 1")
     expect(result).toContain("[explore subagent · 2 steps · 1 tool call]")
     expect(fake!.requests[3].messages.some((m: any) => m.role === "tool" && m.content.includes("return sum(v)"))).toBe(false)
     expect(events.find((e) => e.type === "sub_tool")).toMatchObject({ type: "sub_tool", callId: "t1", name: "read" })
@@ -198,7 +198,7 @@ describe("background subagents (spec 2.0.0)", () => {
     )
     await a.turns.prompt("look into it in the background")
     const started = fake!.requests[1].messages.at(-1).content as string
-    expect(started).toStartWith("Started subagent explore as job1 (find the bug).")
+    expect(started).toStartWith("[#0001]\nStarted subagent explore as job1 (find the bug).")
     await settled(a, 3)
     const note = fake!.requests[2].messages.at(-1).content as string
     expect(note).toStartWith(MARKS.subagent)
@@ -298,7 +298,7 @@ describe("background subagents (spec 2.0.0)", () => {
     )
     await a.turns.prompt("go")
     const results = fake!.requests[1].messages.filter((m: any) => m.role === "tool").map((m: any) => m.content as string)
-    expect(results[0]).toStartWith("Started subagent explore as job1")
+    expect(results[0]).toStartWith("[#0001]\nStarted subagent explore as job1")
     expect(results[1]).toContain("holds one model at a time, and a subagent is already running on g/m there")
     expect(other!.requests.map((r: any) => r.model)).toEqual(["m"])
     await settled(a, 3)