@markusylisiurunen/tau 0.3.77 → 0.3.79

package/dist/core/static/tau_docs/sessions.md

@@ -36,7 +36,7 @@ Only one logical turn runs at a time. Input sent during active work has two usef
 - A queued message waits for the session to become idle, then starts an independent turn.
 - Steering joins the active logical turn at a safe continuation boundary.
 
-In the TUI, Enter queues and Ctrl+Enter steers while work is active. When idle, either starts a normal turn. Pending messages are visible to every observer of the same live hosted session. Alt+Up cancels queued messages and steering that has not been applied and restores the text to the editor.
+In the TUI, Enter steers and Ctrl+Enter queues while work is active. When idle, either starts a normal turn. Pending messages are visible to every observer of the same live hosted session. Alt+Up cancels queued messages and steering that has not been applied and restores the text to the editor.
 
 A turn captures its model, persona, reasoning, system prompt, tools, retry policy, and compaction policy when it starts. Tool subturns and steering continuations keep that captured specification even if reasoning or host configuration changes meanwhile. A queued turn captures the then-current specification when it later starts.
 
@@ -107,7 +107,7 @@ Compaction replaces older model-visible conversation context with a synthetic su
 
 ### Automatic compaction
 
-Automatic compaction runs before a model subturn when fresh provider usage plus newly added estimated context exceeds the configured threshold. The threshold is the model context window minus `autoCompact.reserveTokens`. Tau keeps a recent tail bounded by `autoCompact.keepRecentTokens`, summarizes older context, and can compact more than once during a long logical turn.
+Automatic compaction runs before a model subturn when fresh provider usage plus newly added estimated context exceeds the fixed threshold. The threshold is the model context window minus 16,384 tokens. Tau keeps a recent tail with a target budget of 20,000 tokens, capped by that threshold, summarizes older context, and can compact more than once during a long logical turn.
 
 The summary model may copy important original user messages verbatim into the summary. Recent retained messages stay available to the model, although unusually large textual tool and recovery results may be truncated in retained context. Tau records the compaction as a new active context segment.
 
@@ -119,18 +119,6 @@ The continuation also requires the archive guide to be present in model context
 
 These archives are temporary recovery aids, not backups. Archive failure does not block compaction, and execution-environment cleanup may remove them.
 
-Configure the policy in [configuration](configuration.md):
-
-```json
-{
-  "autoCompact": {
-    "enabled": true,
-    "reserveTokens": 16384,
-    "keepRecentTokens": 20000
-  }
-}
-```
-
 ### Intermediate system instructions
 
 Tau supports plain-text system instructions in conversation history separately from the persona/base prompt. These records are hidden in the TUI, persisted for recovery, recorded as system entries in searchable transcript history, and sent to the model in conversation order. The initial persona/base prompt is not recorded in transcript history. Providers that cannot represent intermediate system messages may fold their instructions into the leading prompt through the model adapter.

package/dist/core/static/tau_docs/skills.md

@@ -79,7 +79,7 @@ The frontmatter contract is:
 | `metadata` | Optional map whose keys and values are strings. |
 | `allowed-tools` | Optional non-empty string. It is accepted for skills-format compatibility but currently ignored by Tau. |
 
-Unknown frontmatter fields are discarded. `allowed-tools` does not enable, disable, or restrict any tool. Tool availability comes from the active persona and, for a subagent, its subagent definition. See [tools](tools.md) and [subagents](subagents.md).
+Unknown frontmatter fields are discarded. `allowed-tools` does not enable, disable, or restrict any tool. Tool availability comes from the active persona; subagents inherit its eligible tools. See [tools](tools.md) and [subagents](subagents.md).
 
 Keep the description useful without copying the full workflow into it. Tau initially exposes the name, description, and `SKILL.md` path. The agent opens the file after activation, then reads only the referenced resources needed for the task.
 

package/dist/core/static/tau_docs/subagents.md

@@ -1,87 +1,24 @@
 # Subagents
 
-Subagents are host-supervised background agent threads created by the main agent. They are useful when work can proceed independently, needs a separate context, or benefits from a deliberately narrower tool set. A subagent is not a second session and does not have its own persona file. Its definition belongs to the active persona.
+Subagents are host-supervised background agent threads created by the main agent. They are useful when work can proceed independently or needs a separate context. A subagent is not a second session and does not have its own persona file.
 
-The persona decides which subagent names exist. The host enforces launch models, tools, working directory, concurrency, follow-up state, interruption, and cleanup.
+## Background workers
 
-## The built-in `default` subagent
+Subagents are general-purpose background workers with no named types. Multiple instances can run concurrently, each with its own task and conversation.
 
-Tau provides one built-in subagent named `default`. It is a general-purpose background worker and has explicit trigger sensitivity. Built-in personas enable it, and a standalone custom persona also enables it when `subagents` is omitted.
+The worker inherits the active main persona's model-facing behavior through Tau's maintained wrapper. Its instructions and description are not configuration surfaces. Supply task-specific instructions in the launch prompt, including any relevant skill guidance.
 
-The `default` subagent inherits the active main persona's model-facing behavior through Tau's maintained wrapper. Its wrapper and built-in prompt are implementation-owned and are not configuration surfaces. It cannot be overridden in persona frontmatter.
-
-Disable it explicitly when a persona should expose only named custom workers or no subagents:
-
-```yaml
-subagents:
-  default: false
-```
-
-If `subagents` is omitted by a persona that extends a built-in, Tau inherits the base persona's entire subagent map. If a persona supplies a `subagents` map, Tau builds a new map from it and adds `default` unless the map says `default: false`.
-
-## Defining custom subagents
-
-Custom definitions live under `subagents` in a [persona](personas.md):
-
-```yaml
-subagents:
-  default: false
-  dependency-auditor:
-    description: Audit dependency changes and report concrete risks. Trigger: explicit.
-    systemPrompt: >-
-      Inspect the requested dependency change. Verify lockfile and release implications,
-      then return prioritized findings with paths.
-    tools:
-      - bash
-      - history
-    launchModels:
-      - anthropic/claude-haiku-4-5:low
-      - openai/gpt-6.1-sol:high
-```
-
-A custom subagent name must:
-
-- contain 1 to 64 characters;
-- use lowercase letters and digits;
-- use single dashes between segments; and
-- begin and end with a letter or digit.
-
-`dependency-auditor` and `review2` are valid. `DependencyAuditor`, `dependency_auditor`, and `-review` are not.
-
-Each custom definition accepts:
-
-| Field | Contract |
-| --- | --- |
-| `systemPrompt` | Required non-empty string containing the subagent's base instructions. |
-| `description` | Optional non-empty catalog description. Include trigger sensitivity here when needed. |
-| `tools` | Optional list of eligible subagent tools. Omission inherits eligible tools from the main persona. |
-| `launchModels` | Optional allowlist of exact model overrides accepted by `spawn_agent`. |
-
-Unknown fields are discarded. In particular, `provider`, `model`, `reasoning`, and `serviceTier` inside a subagent definition do not configure its runtime. Use `launchModels` for approved launch-time model changes.
-
-A custom entry cannot be `false`; only `default: false` has disable semantics. Every custom entry requires `systemPrompt`.
+Availability follows the main persona's tool list. Built-in personas and standalone custom personas that omit `tools` enable the five supervision tools. A custom persona can prevent launches by providing an explicit `tools` list without `spawn_agent`; omit all five supervision tools when no subagent interaction is wanted. There is no separate subagent enable or disable setting. See [personas](personas.md).
 
 ## Trigger sensitivity
 
-Trigger sensitivity is expressed in the `description`, not in a separate field:
-
-- **eager**: use proactively whenever the capability would help;
-- **balanced**: use when the request clearly matches, which is the default when no trigger is stated;
-- **explicit**: use only when an active instruction names the subagent.
-
-Use a clear phrase such as `Trigger: explicit.`. An exact user reference looks like:
-
-```text
-@@agent:dependency-auditor
-```
-
-An exact reference in the user request, an active `AGENTS.md` instruction, or the instructions of an already-active skill explicitly activates that subagent. A subagent reference tells the main agent which capability to use; it does not itself create a thread. The main agent still calls `spawn_agent`.
+The main agent launches subagents only when the user or active instructions explicitly request delegation. Ask for a subagent in ordinary language; no named agent tag is needed. A request does not itself create a thread: the main agent still calls `spawn_agent`.
 
-Trigger sensitivity is agent-facing policy. The host enforces whether the named subagent exists, but it does not infer whether a prompt semantically satisfied `eager`, `balanced`, or `explicit`.
+This is agent-facing policy. The host validates launches but does not infer whether a prompt semantically authorized delegation.
 
 ## Tool inheritance and restriction
 
-The persona-configurable subagent tool subset contains only:
+The eligible inherited tools are:
 
 - `bash`
 - `write`
@@ -89,14 +26,14 @@ The persona-configurable subagent tool subset contains only:
 - `view_image`
 - `web`
 - `history`
+- `nook`, when Nook is configured
+- `mcp`, when the host has enabled MCP servers
 
-Tau then adds intrinsic `tau_docs` to every subagent registry, independently of this subset. Neither a persona nor a subagent `tools` list can disable it. Apart from `tau_docs`, subagents do not receive Nook, goal controls, subagent supervision tools, client tools, or TUI-local tools.
+Tau then adds intrinsic `tau_docs` to every subagent registry, independently of this subset. A persona’s `tools` list cannot disable it. Apart from `tau_docs`, subagents do not receive goal controls, subagent supervision tools, client tools, or TUI-local tools.
 
-When `tools` is omitted, Tau filters the main persona's tool list to the eligible names above. For example, a main persona with `bash`, `edit`, `history`, and `spawn_agent` gives an omitted subagent list of `bash`, `edit`, and `history`.
+The worker inherits the intersection of the main persona's tools and those eight eligible names. MCP uses the parent's shared host connections, independent of the child's working directory. For example, a main persona with `bash`, `edit`, `history`, and `spawn_agent` gives the child `bash`, `edit`, and `history`, plus intrinsic `tau_docs`. There is no separate child tool allowlist.
 
-An explicit `tools` list replaces inheritance. It may select any of the six eligible subagent tools, even if that name is absent from the main persona's own list. Use `tools: []` for a worker with no persona-configurable tools; it still receives `tau_docs`. Names are normalized to lowercase, duplicate entries are removed, and an unknown name rejects the containing persona.
-
-The main persona must itself expose the subagent supervision tools for the model to operate subagents. Defining a subagent while omitting `spawn_agent` from an explicit persona `tools` list leaves the definition present but not launchable by the main agent. See [tools](tools.md) for the broader availability contract.
+Subagents cannot launch other subagents: supervision tools are never included in a child's registry. See [tools](tools.md) for the broader availability contract.
 
 ## Models and settings
 
@@ -110,18 +47,18 @@ A launch override must use exact form:
 
 The provider is normalized to lowercase. The model ID remains exact and case-sensitive. Effort is one of `none`, `minimal`, `low`, `medium`, `high`, `xhigh`, or `max`.
 
-`launchModels` is an allowlist, not a default selection. The main agent should normally omit the `model` argument to `spawn_agent`; when omitted, the subagent inherits the active persona model and reasoning. When supplied, the normalized value must exactly match an entry in the selected subagent's allowlist. The override changes provider, model, and reasoning. Other inherited settings remain.
+The configured launch-model list is an allowlist, not a default selection. The main agent should normally omit the `model` argument to `spawn_agent`; when omitted, the subagent inherits the active persona model and reasoning. When supplied, the normalized value must exactly match an entry in the configured allowlist. The override changes provider, model, and reasoning. Other inherited settings remain.
 
-Model IDs may be unbundled when their provider is known, following the synthesis rules in [models](models.md). Invalid provider, model, effort, or format rejects the persona during content loading. Duplicate normalized entries are removed.
+Model IDs may be unbundled when their provider is known, following the synthesis rules in [models](models.md). Invalid provider, model, effort, or format rejects the configuration field. Duplicate normalized entries are removed.
 
-### Allowing overrides for `default`
+### Allowing model overrides
 
-The built-in `default` definition cannot be edited in persona frontmatter. Configure its launch allowlist through `subagents.defaultLaunchModels` in `config.json`:
+Configure the built-in worker’s launch allowlist through `subagents.launchModels` in `config.json`:
 
 ```json
 {
   "subagents": {
-    "defaultLaunchModels": [
+    "launchModels": [
       "anthropic/claude-haiku-4-5:low",
       "openai/gpt-6.1-sol:high"
     ]
@@ -129,7 +66,7 @@ The built-in `default` definition cannot be edited in persona frontmatter. Confi
 }
 ```
 
-This list is layered configuration. The nearest defined `defaultLaunchModels` array replaces the broader array rather than appending to it. Tau applies the effective list to every persona that currently enables `default`; it does not add `default` to a persona that disabled it.
+This list is layered configuration. The nearest defined `launchModels` array replaces the broader array rather than appending to it. Tau applies the effective list to every persona that exposes `spawn_agent`.
 
 ## Working-directory context
 
@@ -137,7 +74,7 @@ This list is layered configuration. The nearest defined `defaultLaunchModels` ar
 
 When the resolved path differs from the parent `cwd`, Tau rebuilds prompt context from that target directory. It discovers the target platform and repository metadata, reads applicable `AGENTS.md` and configured context files, discovers target skills, and filters those skills through the parent persona's skill selection.
 
-The target directory does **not** select a different persona, subagent definition, model catalog, runtime configuration, credential source, or tool policy for the child. Those remain under parent-session authority. Target configuration is consulted only where needed to rebuild target prompt context, such as target `agentContextFiles` and skill discovery.
+The target directory does **not** select a different persona, model catalog, runtime configuration, credential source, or tool policy for the child. Those remain under parent-session authority. Target configuration is consulted only where needed to rebuild target prompt context, such as target `agentContextFiles` and skill discovery.
 
 This distinction matters in monorepos and hosted environments. `workingDirectory` is an execution-environment path. The host and attached client must not reinterpret it against their own filesystems.
 
@@ -145,11 +82,11 @@ A launch is blocked if Tau cannot build target context, the directory is invalid
 
 ## Lifecycle and supervision tools
 
-Subagent threads are addressed by host-generated IDs. Names identify configurations; IDs identify live thread records.
+Subagent threads are addressed by host-generated IDs. IDs identify individual live thread records.
 
 ### Spawn
 
-`spawn_agent` validates the configured name, optional launch model, title, prompt, and working directory. The prompt is the child's only initial user input, so it should be self-contained. A successful call returns immediately with the new ID while the child continues in the background.
+`spawn_agent` accepts a title, prompt, optional launch model, and optional working directory. The prompt is the child's only initial user input, so it should be self-contained. A successful call returns immediately with the new ID while the child continues in the background.
 
 At most eight subagent runs may be active concurrently within one main session. Completed or interrupted idle threads do not consume active capacity.
 
@@ -165,7 +102,7 @@ The host also publishes bounded live subagent activity to observing clients. Tha
 
 `send_input_to_agent` starts another run on an existing idle thread. The thread retains its conversation state, model, settings, tools, and working directory. It must finish or be interrupted before another input is accepted.
 
-Starting a follow-up replaces the previously retained response in the thread's latest-run state. Read any needed result before sending the next input. Follow-ups do not reread a changed persona definition or adopt a newly reloaded launch model.
+Starting a follow-up replaces the previously retained response in the thread's latest-run state. Read any needed result before sending the next input. Follow-ups do not reread a changed persona definition or adopt a newly reloaded launch allowlist.
 
 ### Interrupt
 
@@ -185,19 +122,10 @@ Plan durable work accordingly. Important conclusions should be returned to the m
 
 ## Reloading and validation
 
-Persona and global launch-policy changes take effect after `/reload` in an idle TUI session or in a newly created session. Reload updates which subagent definitions new `spawn_agent` calls see. It does not kill, rename, or reconfigure already spawned threads, and follow-ups continue on their captured runtime.
-
-An invalid subagent definition rejects its containing persona. Common content warnings include:
+Persona tool selections and launch-policy changes take effect after `/reload` in an idle TUI session or in a newly created session. Reload updates new launches without reconfiguring already spawned threads. Follow-ups continue on their captured runtime.
 
-- a name outside the lowercase-dash contract or longer than 64 characters;
-- `subagents` that is not an object;
-- `default` set to anything other than `false`;
-- a custom entry set to `false` or missing `systemPrompt`;
-- blank descriptions, prompts, or list entries;
-- unknown subagent tools;
-- `launchModels` or `defaultLaunchModels` that is not a string array; and
-- an unknown provider or model, invalid effort, or malformed `<provider>/<model>:<effort>` value.
+`subagents.launchModels` must be a string array of valid `<provider>/<model>:<effort>` values. Unknown providers, invalid efforts, and malformed entries produce configuration diagnostics.
 
-At execution time, launches can still be blocked by an unenabled name, a model outside the allowlist, exhausted concurrency, invalid arguments, missing prompt composition, or target-context failure. Follow-up, wait, and interrupt calls reject unknown IDs; follow-up also rejects a thread that is still running.
+At execution time, launches can be blocked by invalid arguments, a model outside the allowlist, exhausted concurrency, missing prompt composition, or target-context failure. Follow-up, wait, and interrupt calls reject unknown IDs; follow-up also rejects a thread that is still running.
 
-Use `tau --debug --persona <id>` before starting a local TUI to inspect effective subagent names, inherited model and settings, and tool lists. It does not reveal managed prompt bodies for built-in subagents. In a running session, `/reload` reports exact file warnings and `list_agents` verifies live runtime choices.
+Use `tau --debug --persona <id>` before starting a local TUI to inspect effective tool availability, inherited model and settings, and the child tool list. In a running session, `/reload` reports file warnings and `list_agents` shows live thread state.

package/dist/core/static/tau_docs/telegram.md

@@ -56,7 +56,6 @@ Unknown object fields are stripped, so a misspelled field can have no effect wit
 | `projects` | Yes | Object keyed by project ID. Bots select from these definitions. |
 | `workspaceRoot` | No | Base for managed workspaces; defaults to `.tau/telegram-workspaces` beside the Telegram config file. |
 | `maxSessions` | No | Positive integer cap on active sessions across the whole runner. |
-| `systemMessage` | No | Non-empty model-facing instruction prepended to every Telegram turn. |
 
 A relative top-level `workspaceRoot` resolves from the Telegram config file's directory. Tau persists runner session records at `<workspaceRoot>-sessions.json` and per-chat preferences at `<workspaceRoot>-project-preferences.json`. These are runner-owned state files, not operator editing surfaces. Session snapshots remain in the normal host store under the runner user's Tau home.
 
@@ -74,8 +73,8 @@ Each `bots.<id>` object supports:
 | `allowedChatIds` | No | Integer chat IDs allowed to interact with the bot; also opts groups in. |
 | `defaultProjectId` | No | Initial project preference, and must be allowed for this bot. |
 | `systemMessage` | No | Additional instruction for turns from this bot. |
-| `pollIntervalMs` | No | Positive polling retry interval, default 1,000 ms. |
-| `requestTimeoutSeconds` | No | Positive Telegram long-poll timeout, default 30 seconds. |
+
+The runner retries failed polling after 1,000 ms and uses a 30-second Telegram long-poll timeout.
 
 An absent or empty `allowedUserIds` means no user restriction. An absent or empty `allowedChatIds` allows DMs but allows no groups. When `allowedChatIds` is non-empty, it restricts DMs to listed chat IDs and enables only listed groups. Group IDs are usually negative integers.
 
@@ -95,7 +94,7 @@ platform_api
 release2026
 ```
 
-Every project can have an optional non-empty `description`. Repository and persistent-directory projects can select `persona` as `<id>` or `<id>:<reasoning>`, and can set `noAgentContextFiles` to disable `AGENTS.md` injection for that session. Composite projects require their own persona and do not support `noAgentContextFiles`.
+Every project can have an optional non-empty `description`. Repository and persistent-directory projects can select `persona` as `<id>` or `<id>:<reasoning>`. Composite projects require their own persona. Sessions load applicable `AGENTS.md` instructions.
 
 Each project must define exactly one workspace source: `repo`, `directory`, or `projectIds`.
 
@@ -110,9 +109,7 @@ A repository project clones one GitHub repository into a session-specific manage
       "repo": "acme/ledger",
       "ref": "main",
       "workingDirectory": "packages/api",
-      "workspaceRoot": "/var/lib/tau/ledger-workspaces",
-      "persona": "gpt-6.1-sol-coder",
-      "noAgentContextFiles": false
+      "persona": "gpt-6.1-sol-coder"
     }
   }
 }
@@ -120,7 +117,7 @@ A repository project clones one GitHub repository into a session-specific manage
 
 `repo` must use GitHub `owner/repo` syntax. Arbitrary Git URLs are not accepted. The runner needs `gh` and `git` on its login-shell `PATH`, and `gh` must already be authenticated for the repository.
 
-A relative project `workspaceRoot` resolves from the Telegram config file's directory and replaces the top-level root for that project's managed workspaces. A session workspace is:
+All managed projects use the top-level `workspaceRoot`. A session workspace is:
 
 ```text
 <effective-workspace-root>/<project-id>/<telegram-session-id>
@@ -179,7 +176,7 @@ A composite project creates a root containing multiple repositories:
       "persona": "gpt-6.1-sol-coder:high",
       "instructions": "Keep shared contracts synchronized.",
       "subagents": {
-        "defaultLaunchModels": ["openai/gpt-6.1-sol:high"]
+        "launchModels": ["openai/gpt-6.1-sol:high"]
       }
     }
   }
@@ -190,11 +187,11 @@ A composite project creates a root containing multiple repositories:
 
 Members live at `<composite-root>/<member-project-id>` and use each repository's cache, ref, and working directory. The root is the session `cwd`. At creation, Tau writes a root `AGENTS.md` listing the members and optional `instructions`, plus root `.tau/config.json` containing `subagents` or `{}`. These are workspace files, not synchronized mirrors of Telegram configuration, and are not regenerated when a preserved session reconnects.
 
-`subagents.defaultLaunchModels` sets the built-in `default` launch override allowlist. Runtime config resolves and enforces entries. See [subagents](subagents.md) for model syntax and custom policy.
+`subagents.launchModels` sets the subagent launch override allowlist. Runtime config resolves and enforces entries. See [subagents](subagents.md) for model syntax and inheritance.
 
 The composite owns the parent persona, subagents, model catalog, config, settings, and tools. Child `.tau/config.json` files are not merged. A subagent in a member directory rebuilds only target context: environment and repository metadata, applicable `AGENTS.md` and `agentContextFiles`, and skills filtered by the parent persona.
 
-Composite preparation is all-or-nothing. If one member cannot be prepared, Tau removes the generated composite workspace. The composite's optional `workspaceRoot` controls the generated root; member repository caches continue to use each member's configured root or the top-level default.
+Composite preparation is all-or-nothing. If one member cannot be prepared, Tau removes the generated composite workspace. Composite workspaces and member repository caches use the top-level `workspaceRoot`.
 
 ## Managed workspace lifecycle
 
@@ -218,11 +215,11 @@ New and reconstructed workspaces run their hooks; preserved workspaces and persi
 
 After preparation, Tau creates an ordinary local session at the workspace `cwd`. Runtime discovery reads normal `~/.config/tau` and ancestor `.tau` content visible from that path: models, personas, prompts, skills, project context, host tools, and other session settings work as they do in the TUI.
 
-A project `persona` overrides the normal default for that session. `noAgentContextFiles: true` suppresses `AGENTS.md` injection for repository or persistent-directory sessions. Composite root context is generated deliberately and uses its required persona.
+A project `persona` overrides the normal default for that session. Sessions include applicable `AGENTS.md` instructions. Composite root context is generated deliberately and uses its required persona.
 
 Every Telegram turn starts with hidden `<system>` guidance identifying Telegram as the source and reply destination. It favors direct simple answers and brief acknowledgements or meaningful progress updates for tool-driven or multi-step work, since users cannot see tool activity.
 
-Top-level and per-bot `systemMessage` values share a second hidden block after the default, in that order. Audio transcripts receive another warning that they may contain noise or errors. Unlike project context, hidden guidance is persisted with user turns and can appear in history, so never put credentials in configured messages.
+A per-bot `systemMessage` appears in a second hidden block after the default. Audio transcripts receive another warning that they may contain noise or errors. Unlike project context, hidden guidance is persisted with user turns and can appear in history, so never put credentials in configured messages.
 
 The runner's speech credentials and `speech.voiceId` are loaded at startup from its environment and normal Tau config, based on the runner process's startup `cwd`. Restart the runner after changing those settings.
 

package/dist/core/static/tau_docs/tools.md

@@ -40,6 +40,7 @@ view_image
 web
 nook
 history
+mcp
 spawn_agent
 send_input_to_agent
 wait_for_agents
@@ -47,7 +48,7 @@ list_agents
 interrupt_agent
 ```
 
-When a custom persona extends another persona and omits `tools`, it inherits the base persona's list. A non-extending custom persona that omits `tools` enables `bash`, `write`, `edit`, `view_image`, `web`, `nook`, and `history`. If that persona has any enabled subagents, Tau also enables the five subagent-management tools. Built-in personas enable the same base and subagent tool sets.
+When a custom persona extends another persona and omits `tools`, it inherits the base persona's list. A non-extending custom persona that omits `tools` enables `bash`, `write`, `edit`, `view_image`, `web`, `nook`, `history`, and `mcp`. It also enables the five subagent-management tools. Built-in personas enable the same base and subagent tool sets.
 
 An empty list disables every persona-controlled host tool:
 
@@ -131,7 +132,7 @@ Neither tool provides a general read operation. Use a scoped non-interactive Bas
 
 `view_image` reads an image from the execution environment and returns it to a multimodal model. Tau's built-in instruction limits use to cases where the user explicitly asks to view or analyze an image.
 
-The supported formats are JPEG, PNG, and WebP. Source reads are capped at 50 MiB. Images larger than 2,000 pixels in either dimension or 2.5 MiB of model payload are resized or re-encoded while preserving aspect ratio. If Tau cannot reduce a valid image below the model payload limit, the call fails. Relative paths resolve from the execution-environment working directory.
+The supported formats are JPEG, PNG, and WebP. Source reads are capped at 50 MiB. Images with both dimensions at most 4,096 pixels and a model payload at most 3.5 MiB are returned unchanged. Images exceeding the dimension limit are resized to fit within 4,096 pixels while preserving aspect ratio, without upscaling. Tau tries lossless encoding before lossy compression and reduces dimensions further only if needed to meet the payload limit. Transparency is preserved. If Tau cannot reduce a valid image below the model payload limit, the call fails. Relative paths resolve from the execution-environment working directory.
 
 ## Command-line tools
 
@@ -147,7 +148,7 @@ Each guide covers setup, input, examples, outputs, and failure behavior. Use `ta
 
 ## Code-mode service tools
 
-`web`, `history`, and `nook` each run a one-shot JavaScript program in Tau's restricted code-mode runtime. They intentionally disclose their exact API at use time rather than embedding signatures in this page.
+`web`, `history`, `nook`, and `mcp` each run a one-shot JavaScript program in Tau's restricted code-mode runtime. They intentionally disclose their exact API at use time rather than embedding signatures in this page.
 
 Tau exposes each service tool with agent-facing guidance that requires its documentation to be visible before API use. When the documentation is absent, the first useful call is a documentation-only program:
 
@@ -155,9 +156,11 @@ Tau exposes each service tool with agent-facing guidance that requires its docum
 console.log(docs);
 ```
 
-After reading that result, the agent uses the documented API in later calls. While the documentation remains visible, it is reused rather than reloaded. The guidance prohibits guessed signatures and signatures copied from another code-mode tool. Program return values are ignored; only console output is returned.
+The agent reuses visible documentation, without guessing or borrowing another tool's signatures. Shared examples show field selection and text-block extraction for concise plain-text output. Return values are ignored; only console text and images forwarded with `await image(block)` enter the result.
 
-Code-mode programs have no direct process, environment, credential, import, timer, network, or `fetch` access. They can call only the named API and use agent-scoped scratch files exposed by the runtime. Calls default to a 60-second deadline, allow at most 128 API requests with no more than eight unresolved at once, and limit each serialized request or response to 1 MiB. Undefined object properties are omitted from API arguments; undefined arguments and array entries are invalid. Console output is middle-truncated above roughly 8,192 estimated tokens.
+Code-mode programs have no direct process, environment, credential, import, timer, network, or `fetch` access. They can call only the named API and use agent-scoped scratch files exposed by the runtime. Programs default to 60 seconds (MCP: 15 minutes), allow 128 API requests with eight unresolved at once, and limit each serialized payload to 16 MiB. Undefined object properties are omitted from API arguments; undefined arguments and array entries are invalid. Console output is middle-truncated above roughly 8,192 estimated tokens.
+
+`await image({ type: "image", data, mimeType })` forwards a base64 image without files. Calls must be awaited before the program exits. JPEG, PNG, and WebP bytes must match the declared MIME type and fit the 16 MiB JSON bridge limit. Source images are limited to 40 megapixels; each program may forward 16 images. Images are resized or re-encoded to fit within 4,096 pixels per dimension and 3.5 MiB each. Adjacent console writes form one text block; awaited images separate the text before and after them. Text truncation keeps images in place. Invalid blocks reject and can be caught. Built-in tools retain valid images if the program later fails.
 
 Scratch files are real UTF-8 files in an execution-environment temporary directory shared by code-mode tools for the same agent. Writes are limited to 128 regular files and 64 MiB total. Scratch state is not stored in the session snapshot. The progressively disclosed documentation gives the exact file API.
 
@@ -175,21 +178,33 @@ The effective history query may be machine-local or backed by a configured remot
 
 ### Nook
 
-`nook` manages the configured Nook static mini-app platform. It is available only to a main-session persona that lists `nook` and only when Nook is configured. Subagents cannot receive it.
+`nook` manages the configured Nook static mini-app platform. It is available when the main persona lists `nook` and Nook is configured. Subagents inherit it under the same conditions.
 
 The tool is intended for explicit requests to inspect or manage Nook, publish a static artifact or mini-app, or work with Nook KV. Once the built-in documentation is visible, app-authoring work requires a second separate documentation-only call that prints the Nook authoring skill. The agent reads that guide before creating or modifying app files. Nook setup and platform behavior are covered in [Nook](nook.md).
 
+### MCP
+
+`mcp` discovers and calls tools and reads resources from host-configured Model Context Protocol servers. Personas must allow `mcp`, and the host must configure an enabled server. Subagents inherit access; ephemeral threads do not receive it.
+
+Servers use stdio or streamable HTTP. Stdio commands run on the host with its environment; HTTP uses host credentials. Server files need not share the agent's workspace. Supported protocol revisions are `2025-11-25`, `2025-06-18`, `2025-03-26`, and `2024-11-05`, not `2026-07-28`.
+
+`mcp` lists servers, tools, resources, and templates, inspects schemas, calls tools, and reads URIs. `searchTools` ranks tool metadata with BM25, returning matches and per-server errors. Resources return text or base64 data, without automatic context, file writes, or fetching. Only printed text and images enter context. Tools return `{ content, structuredContent?, isError }`; private MCP `_meta` is stripped. `isError: true` is data; protocol failures throw. Calls/reads default to five minutes, configurable to 15; startup/discovery defaults to 30 seconds. Messages are limited to 16 MiB; oversized results fail.
+
+Connections are lazy, shared across sessions, and closed at shutdown. Later requests reconnect. Mutations are not retried. Tool-list changes invalidate cached discovery; resource listings refresh on each call. Restart does not recover connections or calls. Resource subscriptions, prompts, elicitation, and sampling are unsupported.
+
+Servers merge by name from host launch-directory config layers. Restart the host to apply changes. See [configuration reference](config-reference.md#mcpservers) and [security](security.md#trust-mcp-servers).
+
 ## Subagent tool eligibility
 
 A subagent can receive only:
 
 ```text
-bash  write  edit  view_image  web  history
+bash  write  edit  view_image  web  history  nook  mcp
 ```
 
-A subagent definition may list an exact subset. If it omits `tools`, Tau inherits the intersection of the main persona's tools and those six eligible names. Duplicate names are normalized. `tau_docs` is then added intrinsically regardless of the subset.
+Tau inherits the intersection of the main persona’s tools and those eight eligible names. `tau_docs` is then added intrinsically. There is no separate child tool selection. Nook and MCP require configuration; MCP uses the parent's host connections.
 
-Subagents do not receive Nook, goal tools, subagent-management tools, or client-provided tools. Their Bash and file tools are scoped to the subagent working directory, including an alternate directory selected at launch. See [subagents](subagents.md) for configuration and working-directory context rebuilding.
+Subagents do not receive goal tools, subagent-management tools, or client-provided tools. Their Bash and file tools are scoped to the subagent working directory, including an alternate directory selected at launch. See [subagents](subagents.md) for configuration and working-directory context rebuilding.
 
 ## Client-owned tools
 

package/dist/core/static/tau_docs/troubleshooting.md

@@ -116,12 +116,11 @@ Classify the missing capability before changing configuration.
 
 For a persona-controlled host tool, inspect the active persona's exact `tools` list. An explicit list replaces defaults. Run `tau --debug --persona <id>` for a new local session or reload the current session while idle. Credentials can make a selected tool fail, but usually do not remove its schema. Nook is the exception: it also requires effective `nook` configuration.
 
-For a subagent, check all four gates:
+For a subagent, check these three gates:
 
-1. The active persona defines or enables that subagent name.
-2. The main persona exposes `spawn_agent` and any other needed supervision tools.
-3. A requested launch model exactly matches the subagent's allowlist.
-4. Fewer than eight subagent runs are currently active.
+1. The main persona exposes `spawn_agent` and any other needed supervision tools.
+2. A requested launch model exactly matches the configured subagent allowlist.
+3. Fewer than eight subagent runs are currently active.
 
 An already spawned thread keeps its captured model, tools, and working directory after reload. Recovery does not restore subagent threads, so old agent IDs cannot receive follow-ups after a host restart. Use `list_agents` to inspect live records. See [subagents](subagents.md).
 

package/dist/core/static/tau_docs/tui.md

@@ -25,7 +25,7 @@ Piped stdin becomes the first message in a local TUI session:
 printf 'summarize the current changes' | tau
 ```
 
-Use `tau attach` for a session hosted elsewhere. The terminal, themes, clipboard, speech commands, custom diff launcher, and command-backed client tools still belong to the attaching machine. Bash tools, file access, project configuration, and model work use the session’s execution environment. See [remote sessions](remote-sessions.md) for transport and creation examples.
+Use `tau attach` for a session hosted elsewhere. The terminal, themes, clipboard, speech commands, built-in diff review tool, and command-backed client tools still belong to the attaching machine. Bash tools, file access, project configuration, and model work use the session’s execution environment. See [remote sessions](remote-sessions.md) for transport and creation examples.
 
 ## Work in the editor
 
@@ -35,9 +35,8 @@ Tau recognizes these mention forms and offers Tab completion:
 
 - `@src/main.ts` mentions a file in the execution environment.
 - `@@skill:code-review` explicitly activates an available skill.
-- `@@agent:default` explicitly selects an available subagent.
 
-The older `@file:`, `@skill:`, and `@agent:` forms are not mention syntax. Paths and available skill or agent names come from the hosted session, not from the attaching TUI’s filesystem.
+The older `@file:`, `@skill:`, and `@agent:` forms are not mention syntax. Paths and available skill names come from the hosted session, not from the attaching TUI’s filesystem.
 
 Typing `/` at the start of a line opens command completion. Slash commands are recognized only for single-line submissions. A multiline input beginning with `/`, or an unknown slash-prefixed input, is sent to the agent as an ordinary message.
 
@@ -45,8 +44,8 @@ Typing `/` at the start of a line opens command completion. Slash commands are r
 
 When Tau is idle, Enter and Ctrl+Enter both start a normal turn. While a turn is active:
 
-- Enter queues the text as a new turn to run when the session becomes idle.
-- Ctrl+Enter steers the active turn. Tau applies steering at a safe continuation boundary rather than injecting it into a model response or tool execution in progress.
+- Ctrl+Enter queues the text as a new turn to run when the session becomes idle.
+- Enter steers the active turn. Tau applies steering at a safe continuation boundary rather than injecting it into a model response or tool execution in progress.
 - Alt+Up cancels all pending queued messages and steering that has not yet been applied, then restores their text to the editor. Multiple messages are separated with `---`. Hidden guidance, such as transcription warnings, stays hidden in the pending list and editor. Restoring messages preserves their guidance in order after any guidance already attached to the draft. Submission places all guidance at the beginning of the combined message. Editing a non-empty draft preserves it; clearing the draft clears it.
 
 Pending input is session state shared by attached clients while the host remains alive. It is not durable across host restart or session recovery. A queued turn captures the persona, reasoning, tools, and model settings when that turn actually starts. Steering remains part of the active logical turn and keeps the settings captured when that turn began.
@@ -93,13 +92,14 @@ Then use Shift+Tab to select an allowed reasoning level. Newly added personas do
 | `/reload` | Reload session-owned configuration and content from the execution environment. |
 | `/listen [retry/discard]` | Record speech, retry a retained failed recording, or discard it on macOS. |
 | `/speak` | Read the last assistant response aloud on macOS. |
+| `/auto-speak [on | off]` | Toggle automatic reading of final responses, or explicitly enable or disable it. |
 | `/copy-text` | Copy the last assistant response as plain text. |
 | `/copy-code` | Copy code blocks from the last assistant response. |
 | `/persona:<id>` | Switch persona while idle. |
 | `/prompt:<id>` | Resolve a prompt from the execution environment and place it in the editor without submitting it. |
 | `/theme:<id>` | Switch this TUI’s theme for the current run. |
 
-Commands that mutate context, such as persona changes, compaction, rewind, and reload, should be run while idle. `/goal` display and clear, `/listen`, `/prompt:<id>`, and `/exit` have limited useful behavior during a running turn. Ordinary command submissions are otherwise held back until Tau is idle.
+Commands that mutate context, such as persona changes, compaction, rewind, and reload, should be run while idle. `/goal` display and clear, `/auto-speak`, `/listen`, `/prompt:<id>`, and `/exit` have limited useful behavior during a running turn. Ordinary command submissions are otherwise held back until Tau is idle.
 
 Compaction, rewind, goals, recovery, and retry are described in [sessions](sessions.md). Prompt discovery and insertion are covered in [prompts and project context](prompts-and-project-context.md).
 
@@ -111,8 +111,9 @@ Compaction, rewind, goals, recovery, and retry are described in [sessions](sessi
 | Ctrl+P | Cycle persona while idle. |
 | Ctrl+T | Toggle thought visibility in this TUI. |
 | Ctrl+S | Copy the expanded editor contents to the local clipboard, then clear the editor. |
-| Ctrl+Y | Start or stop voice recording. |
-| Ctrl+Enter | Steer an active turn, or submit normally while idle. |
+| Ctrl+Y | Start or stop voice recording without submitting. |
+| Enter | Steer an active turn, or submit normally while idle. |
+| Ctrl+Enter | Queue another turn, or submit normally while idle. |
 | Alt+Up | Cancel pending input and restore it to the editor. |
 | Alt+Down | Cycle the selected active subagent. |
 | Ctrl+G | Interrupt the selected active subagent. |
@@ -155,7 +156,7 @@ Theme ids are exact and case-sensitive. `/theme` does not persist the selection.
 
 ## Review a diff
 
-`/diff` captures a Git snapshot through the session execution environment, then launches a diff-tool process on the TUI machine. The built-in browser tool is the default. `tau diff-tool` is the standalone command for the built-in demo and diff-review protocol reference. `tau diff-tool --help` shows its help, while `/diff` supplies the environment required for normal reviews. Arguments are passed as Git diff arguments, for example:
+`/diff` captures a Git snapshot through the session execution environment, then launches a diff-tool process on the TUI machine. The built-in browser tool is the review interface. `tau diff-tool` is its process entry point. `tau diff-tool --help` shows its help, while `/diff` supplies the environment required for normal reviews. Arguments are passed as Git diff arguments, for example:
 
 ```text
 /diff --staged
@@ -163,17 +164,6 @@ Theme ids are exact and case-sensitive. `/theme` does not persist the selection.
 
 Captured snapshot patches are limited to 16 MiB. Narrow the Git arguments when a larger scope is rejected. A plain working-tree snapshot includes non-binary untracked files up to 4 MiB each within that aggregate limit.
 
-A `diffTool` entry in the attaching client’s configuration replaces the launcher. Relative commands resolve from the configuration level that defines them:
-
-```json
-{
-  "diffTool": {
-    "command": "/usr/local/bin/team-diff-review",
-    "args": ["--open"]
-  }
-}
-```
-
 The built-in tool opens in Guide mode and starts preparing reviewer orientation, focused topics, and likely questions as soon as its shared review context is ready. Reviewers can comment on that guide, ask for another topic or question, or switch to Diff mode for file and line-level review threads. Guide comments and unresolved diff threads are included in self-contained returned Markdown that identifies the reviewed scope, gives change-level comments their relevant context, and explains the participants and roles in review discussions. Submit first opens the exact return-text preview, where included feedback can be excluded and the full review can be copied before submission. Approve returns immediately when no feedback remains; if feedback appears while approval is being checked, the preview opens instead.
 
 The session host supplies ephemeral review agents, while the local tool owns its browser or interface process. Returned review feedback is recorded as a user entry in the session so it remains available, but Tau does not automatically start an assistant turn after the tool closes. Submit a follow-up message when the review should drive more work.
@@ -182,7 +172,11 @@ The TUI also advertises diff review as a client tool unless `--no-client-tools`
 
 ## Use speech
 
-`/listen` and Ctrl+Y are currently macOS-only. Recording uses local `ffmpeg` with the AVFoundation audio input and stops when Ctrl+Y is pressed again, when Escape is pressed, or at the 9-minute recording limit. The Gemini limit leaves time for setup and finalization within its 10-minute live session. Startup fails if the microphone produces no audio within 15 seconds. Stopping manually or reaching the recording limit finishes transcription and inserts the transcript at the cursor for review without submitting it.
+`/listen` and Ctrl+Y are currently macOS-only. Recording uses local `ffmpeg` with the AVFoundation audio input and stops when Ctrl+Y is pressed again, when Escape is pressed, or at the 9-minute recording limit. The Gemini limit leaves time for setup and finalization within its 10-minute live session. Startup fails if the microphone produces no audio within 15 seconds. Pressing Enter or Ctrl+Enter while recording stops capture, finishes transcription, and submits the finalized editor text. While a turn is active, Enter steers it and Ctrl+Enter queues another turn, just like normal editor submission. Pressing Ctrl+Y again or reaching the recording limit finishes transcription and inserts the transcript at the cursor for review without submitting it.
+
+The recording shortcut is configurable with [`speech.recordingShortcut`](config-reference.md#speech), including single-press and double-tap printable keys such as `§`. The editor hint shows the configured gesture. Starting a recording stops speech playback.
+
+`/auto-speak` toggles automatic reading of successful final answers. `/auto-speak on` and `/auto-speak off` explicitly select the mode; turning it off also stops current playback. It starts off and applies only to the current session in this TUI, including when attached remotely. `/new` resets it. It does not change other clients or replay historical answers. Commentary, tool activity, and failed or interrupted answers are not automatically spoken. Responses arriving while recording or playing speech are not queued for later playback. Escape stops playback without disabling the mode. Automatic playback uses the same voice, credentials, and rewriting as `/speak`; it does not open the microphone or automatically submit dictation.
 
 Gemini displays a replaceable live preview in muted italic text at the original cursor position. Finalized dictation uses normal text once transcription settles. Its final transcript replaces the preview without duplication. Submitting a draft containing finalized dictation includes hidden guidance that some or all of the message may have been transcribed from speech and may contain transcription errors. The guidance remains after edits and clears when the draft is emptied. Editing stays disabled from recording startup through finalization, and failure restores the original draft. Recording and retries cannot start while another editor update is pending.
 
@@ -207,7 +201,7 @@ Restart the TUI instead after changing client-owned themes, the diff launcher, s
 ## Common mistakes
 
 - `!` and `!!` run in the execution environment, not on the attaching client.
-- Enter during a turn queues another turn. Use Ctrl+Enter to steer the active one.
+- Enter during a turn steers it. Use Ctrl+Enter to queue another turn.
 - Ctrl+T changes visibility only. Use Shift+Tab to change reasoning effort.
 - `/prompt:<id>` fills the editor but does not submit it.
 - `/diff` records returned feedback but does not automatically ask the assistant to act on it.