zidane 5.1.19 → 5.1.21

package/docs/ARCHITECTURE.md

@@ -176,9 +176,40 @@ Built-in tools are opinionated about output sizes — drop your v2 `tool:transfo
 | `shell` | Tail-priority truncation at `maxOutputBytes=32768` (32 KiB, combined stdout+stderr). Head trim marker: `…(N bytes truncated from head)…`. `0` disables. UTF-8 never splits mid-codepoint. Appends `(exit N, Nms)` footer + surfaces non-empty stderr by default (`metadata: false` opts out). |
 | `write_file` | Reads existing content; returns `Created` / `Updated` / `No change needed: …` so the model detects no-ops without a separate read. Race window in shared docker/sandbox contexts documented and accepted. |
 | `edit` | Fails clearly on non-unique `old_string` (unless `replace_all: true`). On not-found, includes a nearest-match preview so the model recovers without re-reading. |
-| `multi_edit` | Sequential edits to one file. **Ungated** (no `_outcomes` in input — the SDK default): atomic, first failure aborts. **Gated** (chat layer injects `_outcomes` per-hunk via `injectOutcomesIntoInput`): each hunk reported independently into a structured `Edited <path>: N/M applied · K denied · …` body — `parseEditOutcomesFromResult` re-parses for transcript replay. See CHAT.md → Per-edit approval. |
+| `multi_edit` | Sequential edits to one file. **Single-mode atomic** — applies every edit in input order against the file as left by the previous step; first tool-level failure aborts the batch with the legacy `multi_edit error: edit #N <reason>` string, success returns `Edited <path>: applied N edits (M replacements).`. The tool body knows nothing about approvals: per-hunk decisions are a host concern enforced upstream — the chat layer's `tool:gate` rebinds `ctx.input.edits` to the approved subset (rebind, not mutate — the model's original `tool_call` block in `session.turns` stays untouched) before the body runs, and a `tool:transform` hook appends an `<edit-outcomes>` block to the result so live + replayed transcripts share the same per-hunk view (`parseEditOutcomesFromResult` re-parses on reload). See CHAT.md → Per-edit approval. |
 | `grep` | Wraps `rg` when present (with explicit `.` path to avoid stdin hangs). Bun.Glob fallback otherwise. `head_limit=250`, `offset` paginates. |
 
+## Per-edit approval — harness purity
+
+The harness has no notion of "partial approval". Edit-family tools (`edit`, `multi_edit`, `write_file`) are single-mode and atomic; per-hunk decisions live entirely above the loop. The split is deliberate — SDK consumers (CI agents, headless pipelines, custom hosts) keep the legacy contract; the per-hunk UX is purely a chat-layer concern.
+
+Three loop-visible artifacts carry the decision through:
+
+1. **Input rebind at `tool:gate`** — the host's gate handler (`src/tui/app.tsx`'s `applyGate`) computes the approved subset, then assigns a fresh shallow-clone to `ctx.input` whose `edits` array is filtered. The model's original `tool_call.input` in `session.turns` is never mutated (the rebind produces a new object); the tool body sees the smaller, all-approved batch and runs unchanged.
+2. **Pending-annotation map** — keyed by `tool_call.id`, holds the 1:1 `EditOutcome[]` (over the model's ORIGINAL hunks). Lives in the host's React tree (`pendingAnnotationsRef` in the TUI).
+3. **`tool:transform` annotation** — the host appends an `<edit-outcomes>` block to `ctx.result` when at least one hunk wasn't applied. Bubbles to `child:tool:transform` for subagent-issued calls via `BUBBLED_MUTABLE_EVENTS`.
+
+**Wire format** (canonical shape for live emit + persisted replay):
+
+```
+Edited path/to/foo.ts: applied 2 edits (3 replacements).
+
+<edit-outcomes>
+#1 applied
+#2 denied: denied by user
+#3 applied
+</edit-outcomes>
+```
+
+- Opening + closing tags each on their own line.
+- One line per hunk: `#<1-based-index> <kind>[: <reason>]`.
+- `kind ∈ {applied | denied | skipped | failed}`.
+- Block emitted ONLY when at least one hunk is NOT applied. All-applied calls fall through to the legacy summary alone.
+
+A fully-denied call (every hunk rejected) skips the substitute path; the harness writes `Blocked: User denied this tool call` as the tool_result and the host emits a synthetic `tool-result` event with body `[fully denied] <edit-outcomes>…</edit-outcomes>` for live display only (persisted history stays terse).
+
+Replay path: `parseEditOutcomesFromResult(text)` (from `zidane/chat`) recovers the `EditOutcome[]` from the annotation block. `eventsFromTurns` pairs `tool_call` ↔ `tool_result` by `callId` and re-attaches the outcomes onto the `'tool'` event so reloaded transcripts paint identical per-hunk badges.
+
 ## Tool argument auto-coercion
 
 `validateToolArgs` runs between `tool:gate` and `tool:before`:
@@ -459,11 +490,12 @@ The child's lifecycle also bubbles to the parent hook surface with `childId` + `
 ```
 child:stream:text / child:stream:thinking / child:stream:end
 child:tool:gate / child:mcp:tool:gate          ← share the child's ctx — parent mutations propagate
+child:tool:transform                           ← share the child's ctx — parent mutations propagate
 child:tool:before / child:tool:after / child:tool:error
 child:turn:after
 ```
 
-The two `child:*:gate` events are special: the bubbled ctx is the same reference the child loop awaits on, so a parent listener can refuse / substitute a subagent's tool call without registering on the child agent.
+`BUBBLED_MUTABLE_EVENTS` (`src/tools/spawn.ts`) is the canonical list: `tool:gate`, `mcp:tool:gate`, `tool:transform`. The bubbled ctx is the same reference the child loop awaits on, so a parent listener can refuse / substitute / annotate a subagent's tool call without registering on the child agent. The chat layer's per-edit annotation flow registers on both `tool:transform` and `child:tool:transform` so subagent-issued `multi_edit` calls also get `<edit-outcomes>` blocks appended to their results.
 
 ## Dependency Graph
 

package/docs/CHAT.md

@@ -88,7 +88,9 @@ const agent = createAgent({
 
 | Hook | Purpose | Consumer in chat layer |
 |---|---|---|
-| `tool:gate`, `child:tool:gate`, `mcp:tool:gate`, `child:mcp:tool:gate` | Approval gate | `useSafeModeActions().requestApproval` + (for file-edit tools) `resolveApprovalForPayload` / `injectOutcomesIntoInput`. See **Per-edit approval**. |
+| `tool:gate`, `child:tool:gate`, `mcp:tool:gate`, `child:mcp:tool:gate` | Approval gate | `useSafeModeActions().requestApproval(name, input, originator?)` + (for file-edit tools) `resolveApprovalForPayload`; partial → rebind `ctx.input` to the approved subset + stash outcomes in a pending-annotation map. See **Per-edit approval**. |
+| `tool:transform`, `child:tool:transform` | Per-edit annotation | Look up `callId` in the pending-annotation map and append `buildEditOutcomesAnnotation(outcomes)` to `ctx.result`. Clear the entry. |
+| `agent:done` | Stranded-annotation sweep | `pendingAnnotations.clear()` — covers completed / aborted / error paths so a never-fired `tool:transform` doesn't leak. |
 | `mcp:auth:required`, `mcp:auth:url`, `mcp:auth:success`, `mcp:auth:error`, `mcp:connect` | OAuth badge state | `useMcpAuthDispatch` → `reduceMcpAuth` |
 | `stream:thinking`, `stream:text`, `child:stream:thinking`, `child:stream:text` | Streaming deltas | `useStreamBuffer().queueStreamDelta` |
 | `tool:before`, `tool:after`, `mcp:tool:after`, `child:tool:before`, `child:tool:after` | Tool call/result events | `stream.appendImmediate` |
@@ -123,8 +125,8 @@ The table below indexes every named export; sections further down dive into the
 | `config` + `config-context` | `resolveConfig`, `useConfig`, `ChatOptions`, `ResolvedConfig`, `ResolvedPaths`, `ModelInfo`, `ProviderRegistry`. See **Required options**. |
 | `credentials` | AI-provider credential store. `setProviderCredential`, `readProviderCredential`, `removeProviderCredential`, `readCredentials`, `writeCredentials`, `credentialsPath`, `applyApiKeyEnv` (called by `resolveConfig` before any factory runs). Owner-only file mode. |
 | `discovery-context` + `discovery-slot` | Live catalog plumbing: `DiscoveryProvider`, `useDiscovery`, `useDiscoveryOptional`, `createDiscoverySlot` (stale-while-revalidate primitive). Propagates fresh catalogs (files, skills, MCPs) into open modals. See **Discovery context** below. |
-| `edit-approval` | Pure helpers bridging the safe-mode gate to the `multi_edit` tool's per-edit `_outcomes` side channel — `resolveApprovalForPayload`, `maskToOutcomeKinds`, `injectOutcomesIntoInput`, `parseEditOutcomesFromResult`, `summarizeOutcomes`, `ResolvedApproval`. See **Per-edit approval**. |
-| `edit-diff` | Diff plumbing — `extractEditPayload`, `previewEditPayload`, `applyEditPayload`, `buildUnifiedDiff`, `buildContextualDiff`, `computeLineDiff`, `computeInlineDiff`, `splitLines`, `tokenize`, `filetypeFromPath`, `readEditOutcomes`, `OUTCOMES_INPUT_KEY`. See **Edit-diff rendering**. |
+| `edit-approval` | Pure helpers bridging the safe-mode gate to the host-side per-hunk flow — `resolveApprovalForPayload`, `maskToOutcomeKinds`, `buildEditOutcomesAnnotation`, `parseEditOutcomesFromResult`, `summarizeOutcomes`, `ResolvedApproval`. See **Per-edit approval**. |
+| `edit-diff` | Diff plumbing — `extractEditPayload`, `previewEditPayload`, `applyEditPayload`, `buildUnifiedDiff`, `buildContextualDiff`, `computeLineDiff`, `computeInlineDiff`, `splitLines`, `tokenize`, `filetypeFromPath`. See **Edit-diff rendering**. |
 | `enabled-toggle-set` | `useEnabledToggleSet({ catalog, keyOf, settingKey })` — generic state machine for `enabledSkills` / `enabledMcps` (undefined = all enabled, `[]` = off, `[names]` = allowlist). |
 | `files-discovery` | `listProjectFiles({ cwd, signal? })` → `FileEntry[]`, gitignored paths excluded. |
 | `format` | `fmtTokens`, `ageString`, `shortId`, `compactPath` — display-only helpers. |
@@ -425,9 +427,37 @@ type ApprovalDecision
     | 'accept-safelist'
     | 'deny'
     | { kind: 'partial', mask: readonly boolean[] }
+
+type ApprovalOriginator
+  = | { kind: 'parent' }
+    | { kind: 'child', label: string }      // label is the `child-N` tag
+
+type RequestApproval = (
+  tool: string,
+  input: Record<string, unknown>,
+  originator?: ApprovalOriginator,
+) => Promise<ApprovalDecision>
+
+interface ApprovalRequest {
+  id: string
+  tool: string
+  input: Record<string, unknown>
+  resolve: (decision: ApprovalDecision) => void
+  /** Caller attribution. Absent ≡ `{ kind: 'parent' }`. */
+  originator?: ApprovalOriginator
+}
 ```
 
-`ApprovalDecision` is a discriminated union — exhaustive switches must handle the object form, emitted by the per-edit modal for file-edit tools (`edit` / `multi_edit` / `write_file`). The `mask` is 1:1 with `EditPayload.hunks`; `true` = apply, `false` = deny. Bridge it to the tool body via `resolveApprovalForPayload` (see **Per-edit approval**); non-edit gates collapse `partial` to `allow`.
+`ApprovalDecision` is a discriminated union — exhaustive switches must handle the object form, emitted by the per-edit modal for file-edit tools (`edit` / `multi_edit` / `write_file`). The `mask` is 1:1 with `EditPayload.hunks`; `true` = apply, `false` = deny. Bridge it via `resolveApprovalForPayload` (see **Per-edit approval**); non-edit gates collapse `partial` to `allow`.
+
+`ApprovalOriginator` lets the modal show ` · child-N` attribution when a subagent's gate bubbles up through the parent's hook bus. The TUI's `applyGate` reads `ctx.childId` (set by `BUBBLED_MUTABLE_EVENTS` in `src/tools/spawn.ts`) and threads it through:
+
+```ts
+const originator: ApprovalOriginator = ctx.childId
+  ? { kind: 'child', label: ctx.childId }
+  : { kind: 'parent' }
+const decision = await requestApproval(name, input, originator)
+```
 
 ```tsx
 const { requestApproval, resolveHead, denyAll } = useSafeModeActions()
@@ -441,7 +471,7 @@ const pending = queue[0] ?? null
 
 `accept-safelist` calls `addToSafelist(dataDir, projectDir, suggestSafelistEntry(tool, input))` so subsequent calls with the same shape skip the gate. The safelist lives at `<dataDir>/projects.json` (user dir, never the project dir). `partial` never writes a safelist entry — the safelist key is the tool name + an arg shape, with no hunk-level scope.
 
-**Parallel-call deny semantics**: each pending approval resolves on its own. A single `deny` does **not** cascade through the queue — the model receives `Blocked: User denied this tool call` as that one tool's result and the turn continues; other parallel approvals stay queued and prompt independently. The user's explicit "stop everything" gesture is the host-level `esc abort run` shortcut, which calls `denyAll()` + aborts the active `agent.run()` in one shot.
+**Parallel-call deny semantics**: each pending approval resolves on its own. A single `deny` does **not** cascade through the queue — the gate handler sets `ctx.block = true` + `ctx.reason = 'User denied this tool call'` for THAT call and returns, so the model receives `Blocked: User denied this tool call` as that one tool's result and the turn continues; other parallel approvals stay queued and prompt independently. The user's explicit "stop everything" gesture is the host-level `esc abort run` shortcut, which calls `cancelRunOnDenial()` → `denyAll()` + `agent.abort()` in one shot.
 
 ## Interactions
 
@@ -575,16 +605,16 @@ interface HunkResolution {
 
 Wire:
 
-- `tool:before` hook → read pre-write content for `write_file` (other edit tools carry it in the input) → call `extractEditPayload` → attach the result as `StreamEvent.edit` on the `'tool'` event. The hook also runs `readEditOutcomes(input, hunks.length)` so any `_outcomes` previously injected by the gate ride through onto `payload.outcomes`.
+- `tool:before` hook → read pre-write content for `write_file` (other edit tools carry it in the input) → call `extractEditPayload` → attach the result as `StreamEvent.edit` on the `'tool'` event. The host skips this default emit when the gate already painted a synthetic event for a partial / fully-denied call (otherwise the transcript would show one row with full hunks + a second with the reduced subset).
 - Renderer reads `event.edit` and renders accordingly. When `payload.outcomes` is populated, each hunk row carries an applied / denied / skipped / failed badge; pass the array through `summarizeOutcomes` for the header tally.
-- Suppress the paired success `tool-result` (`isEditErrorResult` is the gate). The structured `multi_edit` body counts as "success" only when every edit applied — `Edited <path>: N/N applied · 0 denied · 0 skipped · 0 failed`. Mixed / failed bodies stay visible.
+- Suppress the paired success `tool-result` (`isEditErrorResult` is the gate). Success means "every hunk applied" — the result body is the legacy `Edited <path>: applied N edits (M replacements).` summary alone, no annotation block. Any presence of `<edit-outcomes>` (mixed / denied / failed) or a `[fully denied]` body keeps the result visible alongside the diff so the user reads the per-hunk reasons.
 - Historical replay (from `eventsFromTurns`) has no pre-write snapshot for `write_file`; the diff renders all-add, matching git's "new file" convention. Outcomes are reconstructed from the persisted `tool_result` body via `parseEditOutcomesFromResult` and re-attached to the `'tool'` event, so a reloaded transcript shows the same per-hunk badges live capture displayed.
 
 `theme.surfaces.diff` carries the row colors (`addBg`, `removeBg`, optional `*ContentBg`, `addFg`, `removeFg`); built-in themes pre-mix translucent reds/greens so terminals without alpha-blend get a legible effect, and CSS hosts can use the same hex values as `background-color`.
 
 ## Per-edit approval
 
-The `multi_edit` tool can accept or reject **individual hunks** instead of all-or-nothing. The user-side decision lives in `ApprovalDecision`'s `{ kind: 'partial', mask }` shape; the bridge to the tool body is a host-only side channel on `input._outcomes` (`OUTCOMES_INPUT_KEY`); the result body carries the post-apply state back in a stable line shape that replay can re-parse.
+Edit-family tools (`edit`, `multi_edit`, `write_file`) can accept or reject **individual hunks** instead of all-or-nothing. The user-side decision lives in `ApprovalDecision`'s `{ kind: 'partial', mask }` shape; the bridge is purely host-side — the harness stays pure (single-mode atomic `multi_edit` body, no side channels on `tool_call.input`). The host's `tool:gate` handler rebinds `ctx.input.edits` to the approved subset before the body runs; a paired `tool:transform` hook appends an `<edit-outcomes>` annotation block to the result so the renderer can paint per-hunk badges live + on replay.
 
 ```ts
 type EditOutcomeKind = 'applied' | 'denied' | 'skipped' | 'failed' | 'pending'
@@ -608,11 +638,9 @@ interface EditPayload {
 import {
   resolveApprovalForPayload,
   maskToOutcomeKinds,
-  injectOutcomesIntoInput,
+  buildEditOutcomesAnnotation,
   parseEditOutcomesFromResult,
   summarizeOutcomes,
-  OUTCOMES_INPUT_KEY,
-  readEditOutcomes,
   type ResolvedApproval,
 } from 'zidane/chat'
 ```
@@ -621,11 +649,9 @@ import {
 |---|---|
 | `resolveApprovalForPayload(decision, payload)` | Turn an `ApprovalDecision` into `{ outcomes, shouldBlock, syntheticEvent }`. Pure — does not mutate `input` / `payload`. |
 | `maskToOutcomeKinds(mask, fallbackLength, deniedReason?)` | Convert a boolean mask to a 1:1 `EditOutcome[]`. Missing entries default to `applied` (no-decision = keep). |
-| `injectOutcomesIntoInput(input, outcomes)` | In-place writer that stamps `input._outcomes` so the tool body honors per-hunk decisions. |
-| `parseEditOutcomesFromResult(text)` | Re-parse the structured `multi_edit` body back into outcomes. Used by `eventsFromTurns` so replay shows the same badges as live capture. Returns `null` for legacy / unstructured bodies. |
+| `buildEditOutcomesAnnotation(outcomes)` | Render an `EditOutcome[]` as the wire-format `<edit-outcomes>…</edit-outcomes>` block — body APPENDed to a tool result (joined with `\n\n`). Idempotent on missing reasons. |
+| `parseEditOutcomesFromResult(text)` | Re-parse the annotation block back into outcomes. Used by `eventsFromTurns` so replay shows the same badges as live capture. Returns `null` when the block is missing / malformed. |
 | `summarizeOutcomes(outcomes)` | `{ applied, denied, skipped, failed, pending, total }` tally for the header badge. |
-| `readEditOutcomes(input, length)` | Read `input._outcomes` back off the tool call's input — padded to `length` with `applied`. Used by `extractEditPayload` to thread outcomes onto the `tool:before` event. |
-| `OUTCOMES_INPUT_KEY` (`'_outcomes'`) | Side-channel field name. **Not** in the tool's JSON `inputSchema` — the model never sees it. |
 
 ```ts
 interface ResolvedApproval {
@@ -638,23 +664,45 @@ interface ResolvedApproval {
 }
 ```
 
+### Wire format
+
+Canonical annotation block — emitted ONLY when at least one hunk is NOT applied. All-applied calls fall through to the legacy `Edited <path>: applied N edits (M replacements).` summary alone.
+
+```
+Edited src/foo.ts: applied 2 edits (3 replacements).
+
+<edit-outcomes>
+#1 applied
+#2 denied: denied by user
+#3 applied
+</edit-outcomes>
+```
+
+- Opening + closing tags each on their own line.
+- One line per hunk: `#<1-based-index> <kind>[: <reason>]`.
+- `kind ∈ {applied | denied | skipped | failed}`.
+- Indexes are 1-based against the **model's original** `edits` array — `parseEditOutcomesFromResult` re-keys to a 0-based `EditOutcome[]`.
+
+A fully-denied call (every hunk rejected) skips the substitute path entirely: the harness writes `Blocked: User denied this tool call` to the persisted tool_result, and the host emits a synthetic `tool-result` event with body `[fully denied] <edit-outcomes>…</edit-outcomes>` for live display only. `isEditErrorResult` keeps that synthetic body visible alongside the diff.
+
 ### Gate handler
 
-End-to-end example showing how a host wires a `tool:gate` handler against the helpers:
+End-to-end pattern for a host wiring `tool:gate` against the helpers. The harness stays pure: the handler **rebinds** `ctx.input` (does not mutate the model's `tool_call.input` reference on the persisted assistant turn) and stashes outcomes for a paired `tool:transform` to append on the way out:
 
 ```ts
 import {
+  buildEditOutcomesAnnotation,
   extractEditPayload,
-  injectOutcomesIntoInput,
   resolveApprovalForPayload,
   toolCallPreview,
 } from 'zidane/chat'
 
+const pendingAnnotations = new Map<string, readonly EditOutcome[]>()
+
 agent.hooks.hook('tool:gate', async (ctx) => {
   if (ctx.block) return                                   // upstream already refused
 
-  const decision = await requestApproval(ctx.name, ctx.input)
-
+  const decision = await requestApproval(ctx.name, ctx.input, originator)
   const editPayload = extractEditPayload(ctx.name, ctx.input)
   if (!editPayload) {
     if (decision === 'deny') {
@@ -676,6 +724,12 @@ agent.hooks.hook('tool:gate', async (ctx) => {
         edit: resolved.syntheticEvent,
         ...(ctx.turnId ? { turnId: ctx.turnId } : {}),
       })
+      stream.appendImmediate({
+        kind: 'tool-result',
+        text: `[fully denied] ${buildEditOutcomesAnnotation(resolved.outcomes)}`,
+        tool: ctx.name,
+        ...(ctx.turnId ? { turnId: ctx.turnId } : {}),
+      })
     }
     ctx.block = true
     ctx.reason = 'User denied this tool call'
@@ -683,30 +737,47 @@ agent.hooks.hook('tool:gate', async (ctx) => {
   }
 
   if (resolved.syntheticEvent) {
-    injectOutcomesIntoInput(ctx.input, resolved.outcomes)  // partial — tool body skips denied hunks
+    // Rebind — fresh shallow clone whose `edits` is the approved subset.
+    // The model's original `tool_call.input` on the persisted turn stays
+    // intact because the rebind produces a new object.
+    const reducedEdits = reduceEditsByOutcomes(ctx.input.edits, resolved.outcomes)
+    ctx.input = { ...ctx.input, edits: reducedEdits }
+    pendingAnnotations.set(ctx.callId, resolved.outcomes)
   }
-  // accept-* (or partial that collapsed to all-applied) → fall through, no mutation.
+  // accept-* (or partial that collapsed to all-applied) → fall through.
+})
+
+agent.hooks.hook('tool:transform', (ctx) => {
+  const outcomes = pendingAnnotations.get(ctx.callId)
+  if (!outcomes) return
+  pendingAnnotations.delete(ctx.callId)
+  const annotation = buildEditOutcomesAnnotation(outcomes)
+  ctx.result = typeof ctx.result === 'string'
+    ? (ctx.result.length === 0 ? annotation : `${ctx.result}\n\n${annotation}`)
+    : [...ctx.result, { type: 'text', text: `\n${annotation}` }]
 })
+
+// Same handler registered on `child:tool:transform` so subagent calls
+// also get annotated — `BUBBLED_MUTABLE_EVENTS` shares the child ctx.
+agent.hooks.hook('child:tool:transform', /* same body */)
 ```
 
+Wipe `pendingAnnotations` on `agent:done` (covers completed / aborted / error paths) — `tool:transform` doesn't fire when `validation:reject` or a throwing `tool:before` synthesizes the tool_result, so entries can otherwise strand across runs. The TUI also clears the map defensively in its session-teardown path (`pendingAnnotationsRef.current.clear()`).
+
 ### `multi_edit` tool body shape
 
-The tool body branches on `_outcomes` presence:
+Single-mode atomic — the harness has no notion of approvals. Applies every edit in input order against the file as left by the previous step. First tool-level failure aborts the batch with the legacy `multi_edit error: edit #N <reason>` string; success returns `Edited <path>: applied N edits (M replacements).`. SDK consumers (CI agents, headless harnesses, pipelines parsing the result) see exactly the legacy contract.
 
-- **Gated** (host injected `_outcomes`) — never aborts on per-edit tool-level failure. Each step is reported into a structured body via `formatMultiEditReport`:
+When a host rebinds `ctx.input.edits` to the approved subset at the gate (above), the body sees a smaller all-approved batch and still runs the same atomic semantics. The per-hunk outcomes ride into the wire / persisted history as the appended `<edit-outcomes>` block — the model and any downstream parser see one self-describing tool_result.
 
-  ```
-  Edited src/foo.ts: 2/3 applied · 1 denied · 0 skipped · 0 failed
-    #1 applied: replaced 2 occurrences
-    #2 denied: denied by user
-    #3 applied: replaced 1 occurrence
-  ```
+### Modal lifecycle
 
-  Anchored line shape `  #N <outcome>[: <detail>]` keeps `parseEditOutcomesFromResult` trivially regex-able.
+The chat layer makes no assumption about how the host renders the approval surface. Two contracts a GUI must honor:
 
-- **Ungated** (programmatic / SDK call, no `_outcomes`) — original pre-2026 atomic semantics. First tool-level failure aborts the batch with the legacy `multi_edit error: edit #N <reason>` string; success returns the legacy `Edited <path>: applied N edits (M replacements).` summary. SDK consumers (CI agents, headless harnesses, pipelines parsing the result) keep working unchanged.
+- **Display the head of `useSafeModeQueue()` only.** The queue is FIFO; parallel tool calls prompt in arrival order. Resolve with `resolveHead(decision)` to pop the head — never resolve other entries directly, since the underlying `Promise` resolvers are owned by the provider.
+- **Key the rendered surface on `request.id`.** Back-to-back approvals in the same queue tick must force-remount the modal so per-call UI state (per-hunk mask, focused row, zone, etc.) doesn't carry across. The TUI's `ChatScreen` does this with `<FileEditApprovalModal key={fileEditPending.id} … />` (`src/tui/screens.tsx`).
 
-The two modes are mutually exclusive — `_outcomes` flips the contract. The model never opts in: `_outcomes` is omitted from the tool's JSON `inputSchema`, so the LLM can't write it.
+For hosts that route some decisions through a separate modal stack (the TUI inlines `FileEditApprovalModal` in the transcript slot while non-file-edit gates render as an `ApprovalBlock` below the prompt), push an empty placeholder onto the global modal context for the inline surface so `useModalAwareFocus()` keeps releasing background focus — the focus contract is "any modal-aware surface = background blurred", and an inline modal is still a modal by that definition.
 
 ## Tool call display
 
@@ -1198,7 +1269,7 @@ const text = turnAsText(turn)
 
 **Render an edit diff** — call `extractEditPayload` in `tool:before` (passing pre-write content for `write_file`), persist on the `'tool'` `StreamEvent.edit` field, render via `buildUnifiedDiff(payload)` + `filetypeFromPath(payload.path)` for syntax highlighting. Theme via `theme.surfaces.diff`. For a renderer-faithful preview (matches the tool body's lenient resolver, real file line numbers, per-hunk resolvability metadata), use `previewEditPayload(payload, priorContent, contextLines?)` and pass `result.diffText` to the diff renderable; consult `result.resolution[i].resolved` to decide whether to badge the hunk as unresolvable.
 
-**Wire per-edit approval** — in the `tool:gate` handler, call `extractEditPayload(name, input)`. If it returns a payload, route the decision through `resolveApprovalForPayload(decision, payload)`: on `shouldBlock`, emit `resolved.syntheticEvent` as a `'tool'` event with `outcomes` (so the transcript shows the denied diff) and set `ctx.block = true`; on partial-accept, call `injectOutcomesIntoInput(ctx.input, resolved.outcomes)` so the `multi_edit` body skips denied hunks. The tool result body's structured shape (`Edited <path>: N/M applied · …`) is parsed back by `parseEditOutcomesFromResult` during replay, so reloaded transcripts re-attach the same per-hunk badges.
+**Wire per-edit approval** — in the `tool:gate` handler, call `extractEditPayload(name, input)`. If it returns a payload, route the decision through `resolveApprovalForPayload(decision, payload)`: on `shouldBlock`, emit `resolved.syntheticEvent` as a `'tool'` event (so the transcript shows the intended diff with denied badges), emit a paired `'tool-result'` event with body `[fully denied] ${buildEditOutcomesAnnotation(resolved.outcomes)}` for live display, and set `ctx.block = true`; on partial-accept, **rebind** `ctx.input` to a fresh shallow-clone whose `edits` is the approved subset and stash `resolved.outcomes` in a per-callId pending-annotation map. A paired `tool:transform` handler reads back the map and appends `buildEditOutcomesAnnotation(outcomes)` to `ctx.result`, so the wire / persisted history carries the per-hunk decisions for replay. Register the same handler on `child:tool:transform` for subagent-issued calls. Wipe the map on `agent:done` to catch paths where `tool:transform` never fires (validation reject, throwing `tool:before`). See **Per-edit approval** for the full gate handler skeleton.
 
 **Drive a model picker** — `buildModelCatalog(providers, modelsFor, currentPick)` + `filterModelCatalog(catalog, query)` + `indexOfEntry`. Gate an effort sub-picker on `modelSupportsReasoning(descriptor, modelId)`.
 

package/docs/SKILL.md

@@ -217,7 +217,7 @@ Alias only when semantically equivalent. `shell → Bash` is safe; `list_files
 | `readFile` | Line range, default `offset=1, limit=2000`, 256 KiB cap. Paging footer; binary marker. |
 | `writeFile` | Returns `Created` / `Updated` / `No change needed: …` for no-op detection. |
 | `edit` | Surgical `old_string` → `new_string`. Clear errors on non-unique (unless `replace_all`) / not-found (with nearest-match preview). |
-| `multiEdit` | Sequential edits to one file. Atomic by default; the chat layer's per-edit gate flips it into per-hunk outcome reporting via the `_outcomes` side channel (see `docs/CHAT.md`). |
+| `multiEdit` | Sequential edits to one file. Single-mode atomic — first failure aborts with `multi_edit error: edit #N <reason>`; success returns `Edited <path>: applied N edits (M replacements).`. Per-edit approval is a host concern: the chat layer's `tool:gate` rebinds `ctx.input.edits` to the user-approved subset before the body runs, then a `tool:transform` hook appends an `<edit-outcomes>` annotation to the result so the renderer can paint per-hunk badges (see `docs/CHAT.md`). |
 | `listFiles` | Directory listing. |
 | `glob` | `**`, `*`, `?` pattern matching via Bun.Glob; shells out in docker/sandbox. |
 | `grep` | ripgrep + Bun.Glob fallback. Full Claude Code Grep semantics. `head_limit=250`, `offset` paginates. |
@@ -324,11 +324,12 @@ Mutable ctx fields: `tool:gate` (`block`, `reason`, `result`), `tool:transform`
 ```
 child:stream:text / child:stream:thinking / child:stream:end
 child:tool:gate / child:mcp:tool:gate    ← mutable: block/reason/result propagate to the child
+child:tool:transform                     ← mutable: parent can rewrite the child's tool_result
 child:tool:before / child:tool:after / child:tool:error
 child:turn:after
 ```
 
-Render nested activity without listening on the child instance. The `child:*:gate` events are unique: they share the same `ctx` reference the child's loop awaits on, so a parent gate handler can refuse / substitute a child's tool call without registering on the child agent.
+Render nested activity without listening on the child instance. The `child:*:gate` and `child:tool:transform` events are special: they share the same `ctx` reference the child's loop awaits on (see `BUBBLED_MUTABLE_EVENTS` in `src/tools/spawn.ts`), so a parent gate handler can refuse / substitute / annotate a child's tool call without registering on the child agent. The chat layer's per-edit annotation flow relies on `child:tool:transform` to append `<edit-outcomes>` blocks onto a subagent's `multi_edit` / `edit` / `write_file` results.
 
 ### Hook recipes
 

package/docs/TUI.md

@@ -274,9 +274,9 @@ Hosts adding a new tool: extend `TOOL_DISPLAY` in `zidane/chat` with a `{ displa
 
 `Settings.showEditDiffs` (default **on**) renders `edit` / `multi_edit` / `write_file` tool calls as a unified diff via the native `<diff>` renderable. Wire path:
 
-- `tool:before` hook reads pre-write content for `write_file` (`edit` / `multi_edit` carry the old text in their input), then calls `extractEditPayload(name, input, priorContent)` from `zidane/chat`. The resulting `EditPayload` rides on the `'tool'` `StreamEvent.edit` field. `extractEditPayload` also reads back any `_outcomes` previously injected by the gate, so `payload.outcomes` rides through onto the transcript event.
+- `tool:before` hook reads pre-write content for `write_file` (`edit` / `multi_edit` carry the old text in their input), then calls `extractEditPayload(name, input, priorContent)` from `zidane/chat`. The resulting `EditPayload` rides on the `'tool'` `StreamEvent.edit` field. The default emit short-circuits when `pendingAnnotationsRef.current.has(callId)` is true — the gate handler already painted a synthetic `'tool'` event with the FULL hunks + outcomes, so emitting again with the reduced input would paint a misleading second row.
 - `EditDiffBlock` consumes that payload and calls `buildUnifiedDiff(payload)` + `filetypeFromPath(payload.path)` for syntax highlighting. When `payload.outcomes` is populated (live partial-approval or persisted replay), each hunk row carries an applied / denied / skipped / failed badge and the header line shows the tally via `summarizeOutcomes`.
-- The paired `tool-result` is suppressed (`isEditErrorResult` is the gate) only for **all-applied** outcomes. The new structured `multi_edit` body shape (`Edited <path>: N/M applied · K denied · …`) stays visible whenever any hunk wasn't applied, so the user reads the per-edit reasons next to the diff. Errors (`Edit error: …`, `Tool failed: …`, legacy `multi_edit error: …`) always bypass suppression.
+- The paired `tool-result` is suppressed (`isEditErrorResult` is the gate) only for **all-applied** outcomes — the result body is the legacy `Edited <path>: applied N edits (M replacements).` summary with no annotation. Presence of an `<edit-outcomes>…</edit-outcomes>` block (mixed / denied / failed) or a `[fully denied] …` body keeps the row visible so the user reads the per-hunk reasons next to the diff. Errors (`Edit error: …`, `Tool failed: …`, `multi_edit error: …`) always bypass suppression.
 - Historical replay from persisted turns has no pre-write snapshot for `write_file` — the diff renders all-add, matching git's "new file" convention. Outcomes are reconstructed from the persisted `tool_result` body via `parseEditOutcomesFromResult` and re-attached, so reloaded transcripts show the same per-hunk badges live capture displayed.
 
 Diff row colors come from `theme.surfaces.diff` (`addBg`, `removeBg`, optional `*ContentBg` for a deeper content-column hue, `addFg` / `removeFg` for the gutter glyphs). Built-in themes pre-mix translucent reds/greens against their primary surface so terminals without true alpha-blend still read cleanly.
@@ -289,15 +289,19 @@ File-edit tools (`edit` / `multi_edit` / `write_file`) get their own approval su
 - **`MultiEditApprovalModal`** (multi-step `multi_edit`, ≥ 2 hunks) — per-hunk toggle list + focused-hunk diff panel + action bar. List zone: `↑` / `↓` move cursor, `space` toggles the focused hunk, `y` / `n` set all on/off. Actions zone: `←` / `→` cycle, `↵` confirm. `tab` cycles between zones. `a` / `s` / `p` / `d` bulk shortcuts work in either zone. The first action label is dynamic — `Apply all`, `Apply N/M`, `Nothing selected` — and submits one of:
   - All hunks on → bulk decision (`accept-once` / `accept-session` / `accept-safelist`) preserving the safelist path identical to single-edit.
   - All hunks off → `'deny'`.
-  - Mixed mask → `{ kind: 'partial', mask }` — `gateDecision` calls `resolveApprovalForPayload` + `injectOutcomesIntoInput` to bake the mask into `input._outcomes`. See [Per-edit approval in CHAT.md](./CHAT.md#per-edit-approval) for the helper contract.
+  - Mixed mask → `{ kind: 'partial', mask }` — `applyGate` calls `resolveApprovalForPayload`, then **rebinds** `ctx.input` to a fresh shallow-clone whose `edits` is the approved subset (the model's original `tool_call.input` on the persisted assistant turn stays untouched) and stashes the full-length `EditOutcome[]` in `pendingAnnotationsRef.current` keyed by `ctx.callId`. The paired `tool:transform` / `child:tool:transform` hook reads back the entry and appends a `buildEditOutcomesAnnotation(outcomes)` block to `ctx.result`, so the wire / persisted history carries the per-hunk decisions for replay. See [Per-edit approval in CHAT.md](./CHAT.md#per-edit-approval) for the full contract.
 
-**Inline mount**. The modal renders inside `ChatScreen`'s transcript slot via `flexGrow: 1` — the transcript stays mounted with `visible: !fileEditPending` so its memoized turn anchors and scrollbox state survive the open/close cycle. The prompt + footer stay anchored at the bottom. `ChatScreen` pushes an empty `<></>` placeholder onto the global modal context so `useModalAwareFocus()` keeps releasing background focus and the app-level `esc` / `ctrl+s` shortcuts stay gated; the actual UI is sibling, not stacked.
+**Inline mount**. The modal renders inside `ChatScreen`'s transcript slot via `flexGrow: 1`. The transcript stays mounted with `visible: !fileEditPending` — OpenTUI maps that to Yoga's `Display.None`, removing the box from layout entirely so the sibling modal claims the slot via its own `flexGrow: 1`. Memoized turn anchors, scrollbox position, and lazy markdown chrome all survive the open → close round-trip. The prompt + footer stay anchored at the bottom (queue-only while `busy`). `ChatScreen` pushes an empty `<></>` placeholder onto the global modal context so `useModalAwareFocus()` keeps releasing background focus and the app-level `esc` / `ctrl+s` shortcuts stay gated; the actual UI is sibling, not stacked.
 
-**Title layout**. Left-aligned title rides the top border (`edit approval` / `multi-edit approval`). The right-aligned filename + ` · N/M selected` suffix paints via a sibling `<text style={{ position: 'absolute', top: 0, right: 1 }}>` — OpenTUI's box scissor rect excludes the border row, so a child node can't paint over the top border (same trick `TitleOverlay` and `CompletionPopup` use). `rightTitleFilename(targetPath, termWidth, leftLen, suffixLen)` budgets the basename so the suffix never gets pushed off the right edge on a narrow terminal.
+**Force-remount on queue advance**. The modal is keyed on `request.id` (`<FileEditApprovalModal key={fileEditPending.id} … />`). Without the key, `MultiEditApprovalModal`'s `mask` / `cursor` / `zone` and `SingleEditApprovalModal`'s `selected` would carry over from the previous call when the safe-mode queue pops one approval and exposes the next. React re-uses the component instance otherwise and the user would see a checkbox list whose state matches the prior modal.
+
+**Originator attribution**. The right-side title overlay appends ` · child-N` when `request.originator` is `{ kind: 'child', label: 'child-N' }`. `applyGate` builds the originator from `ctx.childId` (set by `BUBBLED_MUTABLE_EVENTS` for subagent-issued gates) and passes it as the third arg to `requestApproval`. Parent calls show no suffix.
+
+**Title layout**. Left-aligned title rides the top border (`edit approval` / `multi-edit approval`). The right-aligned filename + ` · child-N` + ` · N/M selected` suffix paints via a sibling `<text style={{ position: 'absolute', top: 0, right: 1 }}>` — OpenTUI's box scissor rect excludes the border row, so a child node can't paint over the top border (same trick `TitleOverlay` and `CompletionPopup` use). `rightTitleFilename(targetPath, termWidth, leftLen, suffixLen)` budgets the basename so the suffixes never get pushed off the right edge on a narrow terminal.
 
 **Lenient preview**. Both layouts run the focused hunk's diff through `previewEditPayload(payload, priorContent, contextLines)` from `zidane/chat` — the model-faithful resolver (curly-quote recovery, line-number-prefix stripping, model-side `<n>`→`<name>` desanitize) so what the user previews matches what the tool would actually apply. Unresolved hunks (resolver couldn't locate `old_string`, or matched ambiguously without `replace_all`) render an `UnresolvedHunkPanel` instead of a blank diff: red-bordered, headlined with the failure mode, showing the model's intended `old_string` / `new_string` so the user can deny intelligently. The multi-edit list flags the same rows with a red `⚠` glyph.
 
-`isFileEditTool(tool)` (exported from `file-edit-approval-modal.tsx`) is the gate routing predicate. The set is `{'edit', 'multi_edit', 'write_file'}`; everything else stays on the inline `ApprovalBlock` path.
+`isFileEditTool(tool)` (exported from `file-edit-approval-modal.tsx`) is the gate routing predicate. The set is `{'edit', 'multi_edit', 'write_file'}`; everything else stays on the inline `ApprovalBlock` path. For a fully-denied file-edit call, `applyGate` skips the substitute path entirely — sets `ctx.block = true` + emits a synthetic `tool-result` event with body `[fully denied] <edit-outcomes>…</edit-outcomes>` for live display. The persisted result stays the terse `Blocked: User denied this tool call` the harness writes.
 
 ## Settings rows
 
@@ -354,6 +358,19 @@ File discovery uses `git ls-files --cached --others --exclude-standard` when ava
 
 Adding a third provider: build it in `zidane/chat` against `CompletionProvider<TItem>`, pass it alongside the others in `completionProviders` — the popup picks it up automatically.
 
+## Transcript scrolling
+
+`<Transcript>` mounts a `<scrollbox>` with OpenTUI's native auto-pin: `stickyScroll` + `stickyStart="bottom"`. The scrollbox stays pinned to the bottom while content grows, and detaches the moment the user scrolls up — re-attaches as soon as they hit the bottom edge again. This is the path that keeps streamed text + tool results glued under the cursor while a run is busy without manual `requestAnimationFrame` glue in React.
+
+Two scroll effects layer on top, both keyed off React state:
+
+- **Auto-pin** (native): OpenTUI handles the `[items, busy]` → "stay at bottom" loop. No React work; survives a 60 fps stream without forcing re-layouts.
+- **`selectedTurnId` effect** (React): `useEffect([selectedTurnId, anchors, ownership])` calls `scrollbox.scrollChildIntoView(id)` via `requestAnimationFrame` so OpenTUI's layout pass has settled before measuring. Snaps to `scrollHeight` for the last turn (or a turn that owns the last result-only row) so a tall response's tail stays visible.
+
+The two coexist — selecting a turn jumps to it; exiting select-turn mode (`esc`) plus the next streamed delta re-engages the native sticky-bottom.
+
+`Settings.smoothStreaming` (default **on**) drip-feeds streamed text at a smooth cadence (typewriter) instead of in stream bursts; the auto-pin loop tracks each delta naturally because each tick still grows the scrollbox content.
+
 ## Select-turn mode
 
 Press `ctrl+s` on the chat screen (idle, ≥ 1 turn, no pending approval / interaction) to enter select-turn mode:
@@ -458,7 +475,7 @@ src/tui/
   index.tsx                     runTui + public exports
   app.tsx                       App + AppShell + ThemedShell — provider stack + agent lifecycle
   screens.tsx                   AuthScreen, SessionsScreen, ChatScreen, wizard steps, PromptBlock, ApprovalBlock, QueuedMessagesBlock
-  components.tsx                Transcript, EventLine, MarkdownBlock, SubagentBlock, Footer, Spinner, StatusSpinner, TitleOverlay, EditDiffBlock (outcomes-aware: per-hunk applied/denied/skipped/failed badges via `summarizeOutcomes`), ToolCallBlock, UserPromptBlock
+  components.tsx                Transcript (OpenTUI `stickyScroll` + `stickyStart="bottom"` for auto-pin; `selectedTurnId` scroll effect on top), EventLine, MarkdownBlock, SubagentBlock, Footer, Spinner, StatusSpinner, TitleOverlay, EditDiffBlock (outcomes-aware: per-hunk applied/denied/skipped/failed badges via `summarizeOutcomes`), ToolCallBlock, UserPromptBlock
   modal.tsx                     ModalRoot + Modal + useModalAwareFocus
   model-picker.tsx              Cross-provider searchable model list modal
   effort-picker.tsx             Reasoning-effort modal
@@ -468,7 +485,7 @@ src/tui/
   mcps-settings.tsx             MCP servers list + toggle modal (standalone for embedders)
   session-details-modal.tsx     Stats + delete / export / title / compact (ctrl+x)
   turn-details-modal.tsx        Fork / delete / copy (opened from select-turn mode)
-  file-edit-approval-modal.tsx  FileEditApprovalModal + SingleEditApprovalModal + MultiEditApprovalModal + UnresolvedHunkPanel. Inline-mounted in ChatScreen's transcript slot — bridges ApprovalDecision (including { kind: 'partial', mask }) to the gate via the helpers in `zidane/chat`'s `edit-approval` module.
+  file-edit-approval-modal.tsx  FileEditApprovalModal + SingleEditApprovalModal + MultiEditApprovalModal + UnresolvedHunkPanel + `originatorSuffix` (` · child-N` attribution). Inline-mounted in ChatScreen's transcript slot, keyed on `request.id` for force-remount on queue advance — bridges ApprovalDecision (including { kind: 'partial', mask }) to the gate via the helpers in `zidane/chat`'s `edit-approval` module.
   interaction-block.tsx         InteractionBlock + plan picker + question wizard
   toggle-list-modal.tsx         Generic checkbox-list modal (ToggleListModal)
   completion-popup.tsx          Provider-agnostic autocomplete popover

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "zidane",
-  "version": "5.1.19",
+  "version": "5.1.21",
   "description": "an agent that goes straight to the goal",
   "type": "module",
   "private": false,