tinker-agent 1.9.0 → 1.11.0

package/CHANGELOG.md

@@ -5,6 +5,39 @@ All notable user-facing changes to Tinker are documented here. The project follo
 
 ## [Unreleased]
 
+## [1.11.0] - 2026-08-15
+
+### Added
+
+- Allow text follow-up prompts to be queued while a turn is running. Follow-ups
+  are applied safely after a complete tool batch or handed off to a new turn
+  after a final response, while preserving canonical session history and TUI
+  continuity.
+
+## [1.10.1] - 2026-08-15
+
+### Added
+
+- Add an OpenAI Responses API adapter selectable per model profile, including
+  stateless request mapping, streaming, reasoning output, tool calls, image
+  input, token estimation, and compatible encrypted reasoning items.
+- Add per-profile reasoning effort configuration and a session-runtime
+  `/reasoning` control. The TUI displays the active effort, and `Ctrl+R` cycles
+  through supported efforts in profile order without changing configuration or
+  canonical history.
+- Add a model provider configuration guide with ready-to-adapt examples for
+  OpenAI, Kimi K3, and Zhipu GLM-5.2.
+
+### Changed
+
+- Split historical session retrieval into `RecallSearch` and `RecallGet`, so the
+  agent can locate relevant history before retrieving exact bounded content.
+- Track active-turn tool-output consumption when maintaining context, allowing
+  already-consumed observations to be compacted safely during long-running
+  turns.
+- Bound Bash and background-task output previews while preserving complete
+  output in log files for paginated inspection.
+
 ## [1.9.0] - 2026-08-04
 
 ### Added
@@ -175,7 +208,9 @@ 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/v1.9.0...HEAD
+[Unreleased]: https://github.com/ishowshao/tinker/compare/v1.11.0...HEAD
+[1.11.0]: https://github.com/ishowshao/tinker/releases/tag/v1.11.0
+[1.10.1]: https://github.com/ishowshao/tinker/releases/tag/v1.10.1
 [1.9.0]: https://github.com/ishowshao/tinker/releases/tag/v1.9.0
 [1.8.0]: https://github.com/ishowshao/tinker/releases/tag/v1.8.0
 [1.7.0]: https://github.com/ishowshao/tinker/releases/tag/v1.7.0

package/README.md

@@ -140,12 +140,13 @@ are required. Boolean environment values accept case-insensitive `true/false`,
 | --- | --- | --- | --- | --- | --- | --- | --- |
 | `TINKER_MODELS` | Model | All modes | No | Non-empty string | — | No | Optional model profiles JSON path. Relative paths resolve from the process cwd. |
 | `TINKER_MODEL` | Model | Env mode | Env mode | Non-empty string | — | No | Model name used when model profiles are not configured. |
-| `TINKER_BASE_URL` | Model | Env mode | Env mode | Non-empty string | — | No | OpenAI-compatible Chat Completions API base URL. |
+| `TINKER_API` | Model | Env mode | No | Non-empty string | `"chat-completions"` | No | Model API adapter: "chat-completions" or "responses". |
+| `TINKER_BASE_URL` | Model | Env mode | Env mode | Non-empty string | — | No | OpenAI-compatible API root URL; do not append /chat/completions or /responses. |
 | `TINKER_API_KEY` | Model | Env mode | Env mode | Non-empty string | — | Yes | API credential for the configured model endpoint. |
 | `TINKER_CONTEXT_WINDOW_TOKENS` | Model | Env mode | Env mode | Positive integer | — | No | Model context-window size in tokens. |
 | `TINKER_MAX_SUPPORTED_OUTPUT_TOKENS` | Model | Env mode | Env mode | Positive integer | — | No | Maximum output-token count supported by the model; must not exceed the context window. |
-| `TINKER_INCLUDE_REASONING_CONTENT` | Model | Env mode | No | Boolean | `false` | No | Include provider reasoning content in the model response mapping. |
-| `TINKER_STREAM` | Model | Env mode | No | Boolean | `true` | No | Use streaming Chat Completions transport. |
+| `TINKER_INCLUDE_REASONING_CONTENT` | Model | Env mode | No | Boolean | `false` | No | Replay provider reasoning_content in Chat Completions history; ignored by Responses. |
+| `TINKER_STREAM` | Model | Env mode | No | Boolean | `true` | No | Use streaming transport for the selected model API. |
 | `TINKER_WEBFETCH_REFINE_MODEL` | Model | Env mode | No | Non-empty string | — | No | Optional WebFetch refiner model; currently must match TINKER_MODEL. |
 | `TINKER_WORKSPACE` | Workspace | All modes | No | Non-empty string | Process cwd | No | Workspace path. Relative paths resolve from the process cwd. |
 | `TINKER_MAX_ITERATIONS` | Workspace | All modes | No | Positive integer | `512` | No | Maximum agent-loop iterations per turn. |
@@ -174,21 +175,36 @@ switch to another profile. If `TINKER_MODELS` is not set, Tinker falls back to
 the individual `TINKER_*` environment variables. A configured profiles file must
 exist and be valid; Tinker does not silently fall back when it cannot be loaded.
 
+For provider-specific examples and guidance on API adapters, model capabilities,
+reasoning efforts, and context limits, see the
+[`.tinker/models.json` provider configuration guide](docs/models-json-provider-guide.md).
+
 <!-- BEGIN GENERATED: MODEL PROFILE FIELDS -->
 Profile fields:
 
 | Field | Required | Type / constraint | Default | Secret | Description |
 | --- | --- | --- | --- | --- | --- |
 | `model` | Yes | Non-empty string | — | No | Provider model name. |
-| `apiBase` | Yes | Non-empty string | — | No | OpenAI-compatible API base URL. |
+| `api` | No | Non-empty string | `"chat-completions"` | No | Model API adapter: "chat-completions" or "responses". |
+| `apiBase` | Yes | Non-empty string | — | No | OpenAI-compatible API root URL; do not append /chat/completions or /responses. |
 | `apiKey` | Yes | Non-empty string | — | Yes | API credential for this profile. |
 | `contextWindowTokens` | Yes | Positive integer | — | No | Model context-window size in tokens. |
 | `maxSupportedOutputTokens` | Yes | Positive integer | — | No | Maximum output-token count supported by the model; must not exceed contextWindowTokens. |
-| `includeReasoningContent` | No | JSON boolean | `false` | No | Include provider reasoning content in response mapping. |
-| `stream` | No | JSON boolean | `true` | No | Use streaming Chat Completions transport. |
+| `reasoning` | No | Object | — | No | Provider-specific reasoning efforts and the default for each new session runtime. |
+| `includeReasoningContent` | No | JSON boolean | `false` | No | Replay provider reasoning_content in Chat Completions history; ignored by Responses. |
+| `stream` | No | JSON boolean | `true` | No | Use streaming transport for the selected model API. |
 | `inputModalities` | No | Normalized modality array | `["text"]` | No | Accepted model input modalities; normalizes to ["text"] or ["text", "image"]. |
 | `tokenEstimator` | With image | Object | — | Yes | Independent token estimator required for image profiles. |
 
+`reasoning` fields:
+
+| Field | Type / constraint | Description |
+| --- | --- | --- |
+| `supportedEfforts` | Non-empty unique string array | Provider-supported effort values exposed by the /reasoning command. |
+| `defaultEffort` | Non-empty string listed above | Effort used whenever a session runtime is created or reopened. |
+
+The optional `reasoning` object declares provider-specific effort values. Efforts must be unique non-whitespace strings, `reset` is reserved by the TUI command, and `defaultEffort` must appear in `supportedEfforts`. Omitting `reasoning` sends no effort parameter and disables `/reasoning` for that profile.
+
 `tokenEstimator` fields:
 
 | Field | Type / constraint | Secret | Description |
@@ -212,6 +228,14 @@ Text-only profile example:
       "apiKey": "your-model-api-key",
       "contextWindowTokens": 128000,
       "maxSupportedOutputTokens": 8192,
+      "reasoning": {
+        "supportedEfforts": [
+          "low",
+          "medium",
+          "high"
+        ],
+        "defaultEffort": "medium"
+      },
       "includeReasoningContent": false,
       "stream": true,
       "inputModalities": [
@@ -234,6 +258,14 @@ Image-capable profile example:
       "apiKey": "your-model-api-key",
       "contextWindowTokens": 128000,
       "maxSupportedOutputTokens": 8192,
+      "reasoning": {
+        "supportedEfforts": [
+          "low",
+          "medium",
+          "high"
+        ],
+        "defaultEffort": "medium"
+      },
       "includeReasoningContent": false,
       "stream": true,
       "inputModalities": [
@@ -285,6 +317,14 @@ Atomic-memory profile example:
       "apiKey": "your-model-api-key",
       "contextWindowTokens": 128000,
       "maxSupportedOutputTokens": 8192,
+      "reasoning": {
+        "supportedEfforts": [
+          "low",
+          "medium",
+          "high"
+        ],
+        "defaultEffort": "medium"
+      },
       "includeReasoningContent": false,
       "stream": true,
       "inputModalities": [
@@ -307,6 +347,12 @@ Atomic-memory profile example:
 ```
 <!-- END GENERATED: MODEL PROFILE FIELDS -->
 
+Set `api` to `"responses"` to use the standard Responses API. Keep `apiBase`
+at the API root—such as `https://api.openai.com/v1`—because Tinker appends the
+`/responses` route. Responses requests use the stateless common subset
+(`store: false` with complete input history), so the same adapter works with
+OpenAI and compatible providers that do not implement stored response chaining.
+
 You can also select profiles explicitly for the TUI or one-shot command:
 
 ```bash
@@ -353,6 +399,7 @@ complete fixed policy and persistence contract.
 | `/view <path>` | View a local UTF-8 text file |
 | `/copy` | Copy the last response as Markdown |
 | `/model [profile-name]` | Switch model profile (new session) |
+| `/reasoning [effort\|reset]` | Show or change reasoning effort for this session runtime |
 | `/resume [session-id]` | Choose or resume a session |
 | `/session delete <session-id> --confirm` | Manage stored sessions |
 | `/quit` | Exit the TUI |
@@ -364,6 +411,17 @@ swaps eligible historical tool output, while `/compact retire` retires a complet
 cold prefix whose original history remains available through `Recall`. `/copy`
 copies the last completed assistant response as raw Markdown.
 
+Profiles that declare `reasoning` expose `/reasoning` as a session-runtime
+control. `/reasoning <effort>` temporarily selects one of the profile's
+`supportedEfforts`, and `/reasoning reset` restores `defaultEffort`. The
+selection is not written to configuration or canonical history: `/clear`,
+`/fork`, `/model`, `/resume`, and a TUI restart create a new runtime from the
+profile default. Responses requests send `reasoning.effort`; Chat Completions
+requests send `reasoning_effort`. Press `Ctrl+R` while the prompt is idle to cycle
+through `supportedEfforts` in profile order, wrapping from the last effort to the
+first. When configured, the TUI information line shows the live effort immediately
+after the model name, for example `gpt-5.6-sol max`.
+
 ### Project Custom Slash Commands
 
 The TUI loads optional project-scoped prompt aliases from `.tinker.json` in the

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "tinker-agent",
-  "version": "1.9.0",
+  "version": "1.11.0",
   "description": "A personal coding agent with an interactive TUI and one-shot CLI.",
   "license": "Apache-2.0",
   "type": "module",

package/src/agent/loop.ts

@@ -93,6 +93,7 @@ const MODEL_REQUEST_MAX_ATTEMPTS =
 
 export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
   let lastIteration: IterationIdentity | undefined;
+  let consumedThroughOrdinal = 1;
   const committedPrefixAuditor =
     input.committedPrefixAuditor ?? new CommittedPrefixAuditor();
 
@@ -154,6 +155,7 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
       built,
       prepared,
       preflight,
+      consumedThroughOrdinal,
     });
     try {
       input.contextMeter.assertWithinBudget(preflight);
@@ -365,6 +367,7 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
       },
     });
     const measured = input.contextMeter.recordProviderUsage(request, modelOutput);
+    consumedThroughOrdinal = built.canonical.messages.length;
     await input.runtimeSession.append({
       type: "context.usage.updated",
       ...iteration,
@@ -495,12 +498,21 @@ export async function runAgent(input: RunAgentInput): Promise<RunAgentResult> {
       }
     }
 
+    await input.runtimeSession.applyQueuedSteering?.({
+      turn: input.turn,
+      ledger: input.ledger,
+    });
     await input.runtimeSession.append({
       type: "agent.iteration.finished",
       ...iteration,
       data: { outcome: "continue", toolCallCount: toolCalls.length },
     });
     input.runtimeSession.finishIterationForContinuation(iteration);
+    await input.runtimeSession.maintainContextAfterIteration?.({
+      turn: input.turn,
+      consumedThroughOrdinal,
+      ledger: input.ledger,
+    });
   }
 
   if (lastIteration === undefined) {
@@ -520,6 +532,7 @@ async function runShadowPlanning(input: {
   built: BuiltContextRequest;
   prepared: PreparedModelRequest;
   preflight: ReturnType<ContextMeter["measure"]>;
+  consumedThroughOrdinal: number;
 }): Promise<void> {
   const shadowPlanning = input.input.shadowPlanning;
   if (shadowPlanning === undefined) {
@@ -547,6 +560,10 @@ async function runShadowPlanning(input: {
       tools: input.input.tools.definitions(),
       policy: swapOnlyPolicyV1,
       trigger: decision.trigger,
+      activeTurn: {
+        turnId: input.input.turn.turnId,
+        consumedThroughOrdinal: input.consumedThroughOrdinal,
+      },
       ...(decision.forcedTargetTokens === undefined
         ? {}
         : { forcedTargetTokens: decision.forcedTargetTokens }),