tinker-agent 1.5.1 → 1.7.0

package/CHANGELOG.md

@@ -5,6 +5,55 @@ All notable user-facing changes to Tinker are documented here. The project follo
 
 ## [Unreleased]
 
+## [1.7.0] - 2026-08-01
+
+### Added
+
+- Run interactive terminal programs through PTY-backed Bash tasks, send exact
+  keystrokes with `TaskInput`, and inspect their current terminal screen with
+  `TaskOutput`.
+- Search all resumable sessions from the `/resume` picker by text from their
+  first user prompt, including older sessions beyond the 20 most recent.
+
+### Changed
+
+- Keep the live background-task panel compact by showing at most two tasks and
+  label PTY-backed tasks explicitly.
+
+### Fixed
+
+- Complete PTY tasks reliably after subprocess exit on Linux, including the EIO
+  signal reported when the terminal closes.
+
+## [1.6.0] - 2026-08-01
+
+### Added
+
+- Add an `/undo` slash command that restores the files changed by the most recent
+  file-mutation turn from in-memory snapshots, with restore notices shown in the
+  TUI timeline.
+- Give the agent a Delete tool for removing files, guarded by the same permission
+  flow as the other file-mutation tools.
+- Guard destructive Bash commands behind an interactive confirmation prompt, with
+  a `/yolo` opt-out for sessions where unattended execution is intended.
+- Show the latest provider cache hit rate in the TUI prompt status line.
+- Stream assistant output in the TUI as sealed Markdown sections, so settled
+  content no longer repaints while the turn is still running.
+
+### Changed
+
+- Retry transient provider failures automatically with backoff instead of
+  surfacing them as immediate turn errors.
+- Update runtime dependencies, including `commander` 15, `openai` 7, and the
+  latest `@assistant-ui/react-ink` packages.
+
+### Fixed
+
+- Floor the displayed cache hit rate so append-heavy turns never show a false
+  100%.
+- Bound the TUI live region below the viewport height to protect the prompt
+  frame from being pushed out of view.
+
 ## [1.5.1] - 2026-07-29
 
 ### Changed
@@ -100,7 +149,9 @@ All notable user-facing changes to Tinker are documented here. The project follo
 - First formal npm release under the `tinker-agent` package name with the `tinker`
   executable.
 
-[Unreleased]: https://github.com/ishowshao/tinker/compare/v1.5.1...HEAD
+[Unreleased]: https://github.com/ishowshao/tinker/compare/v1.7.0...HEAD
+[1.7.0]: https://github.com/ishowshao/tinker/releases/tag/v1.7.0
+[1.6.0]: https://github.com/ishowshao/tinker/releases/tag/v1.6.0
 [1.5.1]: https://github.com/ishowshao/tinker/releases/tag/v1.5.1
 [1.5.0]: https://github.com/ishowshao/tinker/releases/tag/v1.5.0
 [1.4.0]: https://github.com/ishowshao/tinker/releases/tag/v1.4.0

package/README.md

@@ -11,8 +11,9 @@ Built with [Bun](https://bun.sh) + TypeScript ESM, powered by [Ink](https://gith
 - **Built-in tools**:
   - `Glob` / `Grep` — Find and search files by pattern or content
   - `Read` / `Write` / `Edit` — File I/O with content hashing and concurrent-modification protection
-  - `Bash` — Run shell commands (foreground and background) with per-task working directories
-  - `TaskList` / `TaskOutput` / `TaskStop` — Manage long-running background shell tasks
+  - `Delete` — Delete one existing regular file without directory or symlink support
+  - `Bash` — Run foreground, background, and PTY shell commands with per-task working directories
+  - `TaskList` / `TaskOutput` / `TaskInput` / `TaskStop` — Inspect, interact with, and stop long-running shell tasks
   - `WebSearch` — Search the web via Exa API
   - `WebFetch` — Fetch and refine web page content (local, browser, or Exa backend)
   - `Recall` — Search or retrieve model-visible history from the current session
@@ -28,7 +29,7 @@ Built with [Bun](https://bun.sh) + TypeScript ESM, powered by [Ink](https://gith
 - **Infinite Context architecture**: Immutable canonical history, deterministic
   context revisions, Recall-addressable cold state, and qualified prefix retirement
   keep long-running sessions recoverable without pretending the model has infinite
-  tokens. See the [technical design](docs/infinite-context-technical-design-a.md).
+  tokens. See the [technical design](docs/infinite-context-technical-design.md).
 - **Choice of models**: Uses an OpenAI-compatible Chat Completions transport with
   explicit model and context limits. Actual provider support must be established
   by a qualification matrix; transport compatibility alone is not a guarantee.
@@ -76,9 +77,9 @@ The installed package exposes this public CLI:
 | --- | --- |
 | `tinker` | Start the interactive terminal interface. |
 | `tinker --profile <profile-name>` | Start the TUI with a selected model profile. |
-| `tinker run [--profile <profile-name>] <prompt>` | Submit one shell-quoted prompt argument. |
-| `tinker run [--profile <profile-name>] --stdin` | Read the prompt from standard input until EOF. |
-| `tinker run [--profile <profile-name>] --file <path>` | Read the prompt from a UTF-8 text file. |
+| `tinker run [--profile <profile-name>] [--yolo] <prompt>` | Submit one shell-quoted prompt argument. |
+| `tinker run [--profile <profile-name>] [--yolo] --stdin` | Read the prompt from standard input until EOF. |
+| `tinker run [--profile <profile-name>] [--yolo] --file <path>` | Read the prompt from a UTF-8 text file. |
 | `tinker --help` | Show top-level CLI help. |
 | `tinker help run` | Show one-shot command help. |
 | `tinker --version` | Print the installed package version. |
@@ -146,6 +147,7 @@ are required. Boolean environment values accept case-insensitive `true/false`,
 | `TINKER_MCP_MAX_OBSERVATION_CHARS` | Tooling | All modes | No | Positive integer | `40000` | No | Maximum model-visible characters in one MCP result. |
 | `TINKER_BASH_DEFAULT_TIMEOUT_MS` | Tooling | All modes | No | Positive integer | `5000` | No | Default Bash foreground timeout in milliseconds. |
 | `TINKER_BASH_MAX_TIMEOUT_MS` | Tooling | All modes | No | Positive integer | `600000` | No | Maximum Bash foreground timeout in milliseconds. |
+| `TINKER_YOLO` | Tooling | All modes | No | Boolean | `false` | No | Allow high-confidence destructive Bash commands without confirmation. |
 | `TINKER_GREP_TIMEOUT_MS` | Tooling | All modes | No | Positive integer | `20000` | No | Bundled ripgrep invocation timeout in milliseconds. |
 | `TINKER_GREP_MAX_BUFFER_BYTES` | Tooling | All modes | No | Positive integer | `20000000` | No | Maximum buffered output from one ripgrep invocation. |
 | `TINKER_WEBFETCH_REFINE_THRESHOLD` | Tooling | All modes | No | Positive integer | `2000` | No | Content-length threshold that enables WebFetch refinement. |
@@ -335,8 +337,10 @@ complete fixed policy and persistence contract.
 | `/status` | Show session and context details |
 | `/skills` | Show available and active Agent Skills |
 | `/mcp` | Show MCP servers and runtime tools |
+| `/yolo [on\|off]` | Show or change destructive Bash confirmation |
 | `/memory` | Browse stored global memories |
 | `/compact [retire]` | Swap tool output or retire a cold history prefix |
+| `/undo` | Undo the latest Write/Edit/Delete turn |
 | `/clear` | Start a new session and clear conversation |
 | `/fork` | Clone the current session |
 | `/view <path>` | View a local UTF-8 text file |
@@ -394,6 +398,10 @@ call always returns the complete requested line range. Use `offset` and `limit`
 to page through larger files; oversized requests fail instead of returning
 truncated content.
 
+`Delete` removes one existing regular file by workspace-relative or absolute path.
+It rejects directories and symbolic links, and clears any in-memory file snapshot
+only after the removal succeeds.
+
 ## Project Structure
 
 ```
@@ -401,7 +409,7 @@ tinker/
 ├── src/
 │   ├── cli/           # Entry points (tui, run), config
 │   ├── agent/         # Agent loop, session ledger, turn cancellation, context metering
-│   ├── tools/         # Tool executors (bash, glob, grep, read, write, edit, recall, etc.)
+│   ├── tools/         # Tool executors (bash, glob, grep, read, write, edit, delete, recall, etc.)
 │   ├── model/         # Model clients (OpenAI-compatible, fake), chat mapping, preflight
 │   ├── mcp/           # MCP server management, tool executor adapter
 │   ├── observation/   # Tool result → model-visible text

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tinker-agent",
-  "version": "1.5.1",
+  "version": "1.7.0",
   "description": "A personal coding agent with an interactive TUI and one-shot CLI.",
   "license": "Apache-2.0",
   "type": "module",
@@ -74,22 +74,25 @@
     "typecheck": "tsc --noEmit"
   },
   "dependencies": {
-    "@assistant-ui/react-ink": "^0.0.31",
-    "@assistant-ui/react-ink-markdown": "^0.0.30",
+    "@assistant-ui/react-ink": "^0.0.35",
+    "@assistant-ui/react-ink-markdown": "^0.0.34",
     "@inkjs/ui": "^2.0.0",
     "@modelcontextprotocol/sdk": "^1.29.0",
     "@mozilla/readability": "^0.6.0",
     "@vscode/ripgrep": "1.18.0",
+    "@xterm/addon-unicode11": "0.9.0",
+    "@xterm/headless": "6.0.0",
     "ansi-escapes": "^7.3.0",
     "bun": "1.3.14",
     "clipboardy": "^5.3.1",
-    "commander": "^14.0.3",
+    "commander": "^15.0.0",
     "diff": "^9.0.0",
     "glob": "^13.0.6",
     "ink": "^7.1.0",
     "linkedom": "^0.18.13",
     "markdansi": "0.3.2",
-    "openai": "^6.45.0",
+    "marked": "^18.0.5",
+    "openai": "^7.1.0",
     "react": "^19.2.7",
     "sharp": "^0.35.3",
     "shiki": "^4.3.1",
@@ -104,8 +107,6 @@
     "@types/chrome": "^0.2.2",
     "@types/react": "^19.2.17",
     "@types/turndown": "^5.0.6",
-    "@xterm/addon-unicode11": "0.9.0",
-    "@xterm/headless": "6.0.0",
     "eslint": "^10.6.0",
     "eslint-plugin-react-hooks": "^7.1.1",
     "globals": "^17.7.0",

package/src/agent/assistant-text-delta.ts

@@ -0,0 +1,10 @@
+import type { IterationIdentity } from "./types";
+
+export type AssistantTextDeltaUpdate = IterationIdentity & {
+  attemptNumber: number;
+  content: string;
+};
+
+export interface AssistantTextDeltaSink {
+  updateAssistantTextDelta(update: AssistantTextDeltaUpdate): void;
+}

package/src/agent/loop.ts

@@ -73,6 +73,7 @@ export type RunAgentInput = {
     prepared: MaterializedModelRequest;
     usage: ContextUsageSnapshot;
   };
+  transientRetryDelaysMs?: readonly number[];
 };
 
 export class FatalAgentTurnError extends Error {
@@ -85,7 +86,10 @@ export class FatalAgentTurnError extends Error {
   }
 }
 
-const MODEL_REQUEST_MAX_ATTEMPTS = 2 as const;
+const REASONING_ONLY_RETRY_LIMIT = 1 as const;
+const TRANSIENT_RETRY_DELAYS_MS = [2_000, 4_000, 8_000, 16_000] as const;
+const MODEL_REQUEST_MAX_ATTEMPTS =
+  1 + REASONING_ONLY_RETRY_LIMIT + TRANSIENT_RETRY_DELAYS_MS.length;
 
 export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
   let lastIteration: IterationIdentity | undefined;
@@ -226,23 +230,22 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
       return failedResult(error, iteration);
     }
 
-    const requestOptions = {
-      signal: input.signal,
-      identity: {
-        iteration,
-        runtimeSession: input.runtimeSession,
-      },
-    };
+    const transientRetryDelaysMs =
+      input.transientRetryDelaysMs ?? TRANSIENT_RETRY_DELAYS_MS;
     let modelOutput: ModelRequestOutput | undefined;
-    let successfulAttempt: 1 | 2 | undefined;
-    for (const attemptNumber of [1, 2] as const) {
+    let successfulAttempt: number | undefined;
+    let attemptNumber = 0;
+    let reasoningOnlyRetries = 0;
+    let transientRetries = 0;
+    while (true) {
       if (input.signal.aborted) {
         return cancelledResult(
           cancellation(input.signal, iteration, "model_request"),
           iteration,
         );
       }
-      if (attemptNumber === 2) {
+      attemptNumber += 1;
+      if (attemptNumber > 1) {
         await input.runtimeSession.append({
           type: "model.request.started",
           ...iteration,
@@ -255,7 +258,32 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
 
       try {
         throwIfTurnCancelled(input.signal);
-        modelOutput = await input.model.request(request, requestOptions);
+        let acceptingTextDeltas = true;
+        const onTextDelta =
+          input.runtimeSession.updateAssistantTextDelta === undefined
+            ? undefined
+            : (content: string) => {
+                if (!acceptingTextDeltas || input.signal.aborted) {
+                  return;
+                }
+                input.runtimeSession.updateAssistantTextDelta?.({
+                  ...iteration,
+                  attemptNumber,
+                  content,
+                });
+              };
+        try {
+          modelOutput = await input.model.request(request, {
+            signal: input.signal,
+            identity: {
+              iteration,
+              runtimeSession: input.runtimeSession,
+            },
+            ...(onTextDelta === undefined ? {} : { onTextDelta }),
+          });
+        } finally {
+          acceptingTextDeltas = false;
+        }
         throwIfTurnCancelled(input.signal);
         successfulAttempt = attemptNumber;
         break;
@@ -267,12 +295,11 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
           );
         }
 
-        const reasoningOnly = isReasoningOnlyProviderError(error);
-        const retryDisposition = reasoningOnly
-          ? attemptNumber === 1
-            ? "scheduled"
-            : "exhausted"
-          : "not_retryable";
+        const decision = modelRequestRetryDecision(error, {
+          reasoningOnlyRetries,
+          transientRetries,
+          transientRetryDelaysMs,
+        });
         await input.runtimeSession.append({
           type: "model.request.failed",
           ...iteration,
@@ -280,14 +307,26 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
             error,
             request,
             attemptNumber,
-            retryDisposition,
+            retryDisposition: decision.disposition,
+            ...(decision.disposition === "scheduled" && decision.delayMs > 0
+              ? { retryDelayMs: decision.delayMs }
+              : {}),
           }),
         });
 
-        if (retryDisposition === "scheduled") {
+        if (decision.disposition === "scheduled") {
+          if (decision.kind === "reasoning_only") {
+            reasoningOnlyRetries += 1;
+          } else {
+            transientRetries += 1;
+            await waitForRetryDelay(decision.delayMs, input.signal);
+          }
           continue;
         }
-        if (retryDisposition === "exhausted") {
+        if (
+          decision.disposition === "exhausted" &&
+          decision.kind === "reasoning_only"
+        ) {
           return failedResult(
             new Error(
               `Provider returned reasoning without final text or tool calls in both attempts (provider=${request.provider}, model=${request.model}).`,
@@ -701,11 +740,65 @@ function isReasoningOnlyProviderError(error: unknown): error is ProviderResponse
   );
 }
 
+function isTransientProviderError(error: unknown): error is ProviderResponseError {
+  return (
+    error instanceof ProviderResponseError &&
+    (error.code === "provider_rate_limited" || error.code === "provider_unavailable")
+  );
+}
+
+type ModelRequestRetryDecision =
+  | {
+      disposition: "scheduled";
+      kind: "reasoning_only" | "transient";
+      delayMs: number;
+    }
+  | { disposition: "exhausted"; kind: "reasoning_only" | "transient" }
+  | { disposition: "not_retryable" };
+
+function modelRequestRetryDecision(
+  error: unknown,
+  state: {
+    reasoningOnlyRetries: number;
+    transientRetries: number;
+    transientRetryDelaysMs: readonly number[];
+  },
+): ModelRequestRetryDecision {
+  if (isReasoningOnlyProviderError(error)) {
+    return state.reasoningOnlyRetries < REASONING_ONLY_RETRY_LIMIT
+      ? { disposition: "scheduled", kind: "reasoning_only", delayMs: 0 }
+      : { disposition: "exhausted", kind: "reasoning_only" };
+  }
+  if (isTransientProviderError(error)) {
+    const delayMs = state.transientRetryDelaysMs[state.transientRetries];
+    return delayMs === undefined
+      ? { disposition: "exhausted", kind: "transient" }
+      : { disposition: "scheduled", kind: "transient", delayMs };
+  }
+  return { disposition: "not_retryable" };
+}
+
+async function waitForRetryDelay(delayMs: number, signal: AbortSignal): Promise<void> {
+  if (delayMs <= 0 || signal.aborted) {
+    return;
+  }
+  await new Promise<void>((resolve) => {
+    const timer = setTimeout(cleanup, delayMs);
+    signal.addEventListener("abort", cleanup, { once: true });
+    function cleanup() {
+      clearTimeout(timer);
+      signal.removeEventListener("abort", cleanup);
+      resolve();
+    }
+  });
+}
+
 function modelRequestFailureData(input: {
   error: unknown;
   request: PreparedModelRequest;
-  attemptNumber: 1 | 2;
+  attemptNumber: number;
   retryDisposition: "scheduled" | "not_retryable" | "exhausted";
+  retryDelayMs?: number;
 }) {
   const providerError =
     input.error instanceof ProviderResponseError ? input.error : undefined;
@@ -714,6 +807,7 @@ function modelRequestFailureData(input: {
     maxAttempts: MODEL_REQUEST_MAX_ATTEMPTS,
     code: providerError?.code ?? ("provider_request_error" as const),
     retryDisposition: input.retryDisposition,
+    ...(input.retryDelayMs === undefined ? {} : { retryDelayMs: input.retryDelayMs }),
     provider: input.request.provider,
     model: input.request.model,
     error: errorMessage(input.error),