Skip to content

A Character says things only through its reply, with no MCP speak tool #1935

Closed

Description

@omesser

Supersedes #1701 and #1910. Decided by @omesser.

Decision

A Character says things through one channel: its reply. There is no MCP speak tool.

The parser already sorts some parts of a reply into other kinds. Everything else is what the Character says, and it streams to both the Chat surface and the bubble. These parts are something else:

  • Thinking prose: prose before a behavior: line goes to the Thinking row.
  • Startup banners: known Harness banners, such as Pi's, are dropped.
  • Bare Behavior names: a reply that only names a Behavior plays it and says nothing.
  • The Behavior line: the behavior: or Name | line line itself picks the Behavior. Text after the pipe is said.

So there is no second channel, no aside flag on Chat rows, and no gate that accepts or declines each line. Double speech, ordering between two channels, late speak lines and speech from outside a turn stop being problems, because only one channel exists.

Naming

Reserve the word "speech" for a future tool call to a real audio speech engine. Today's text path does not use that name, in code, docs, ADRs or GLOSSARY.md. Pick a plain word for it, such as "said" or "line".

Long-running work

A Character stays lively during long work through an optional one-line addition to its own prompt, as in #1934. There are no system prompt or infrastructure changes for this.

To do

  • Remove speak from the MCP tool surface and its dispatch in crates/core/src/dispatch.rs.
  • Amend or supersede the ADRs that assume a speak tool or two speech channels: ADR-0026 and its promise about speak, ADR-0016 where it covers late lines, plus ADR-0003 and ADR-0023, which list speak among the tools.
  • Update docs/harness.md, docs/SPEC.md, docs/DEVELOPMENT.md and GLOSSARY.md to match, including the naming rule above.
  • Revisit The buddy as a live narrator of Harness work #1375, whose design has the prompt ask the agent to call speak.

Done when

  • No code, doc or ADR describes speak as a live tool or a second channel.
  • The text path is not called "speech" anywhere.

— Cursor agent (Architect), on @omesser's behalf.

Activity

  1. added this to the V1 milestone on Oct 29, 2026
  2. moved this from Todo to In Progress in Fidgeton Oct 29, 2026
  3. omesser commented on Oct 30, 2026

    @omesser
    OwnerAuthor

    Progress after #1953 (merged as 4ad3516), checked on main:

    Done when

    • No code, doc or ADR describes speak as a live tool or a second channel. list_tools() in crates/core/src/tools.rs no longer has speak, and crates/core/src/dispatch.rs:524-531 pins 8 tools with no speak. ADR-0040 records the decision, and ADR-0003, 0016, 0023 and 0026, docs/harness.md, docs/SPEC.md, docs/DEVELOPMENT.md and GLOSSARY.md no longer list it. What's left is history (docs/harness.md:426 says "a tool since removed (ADR-0040)"), GLOSSARY.md:187 (_Avoid_: speak), the near_reaction = "speak" cursor reaction, and "speak" as a plain verb.
    • The text path is not called "speech" anywhere. Docs, ADRs and GLOSSARY.md now say Dialogue, but code still uses speech (e.g. crates/core/src/speech.rs). That's Rename the speech identifiers for the text path to dialogue #1955, which waits on feat(engine): Apply the drive intent and the possession switch #1923.

    The "Revisit #1375" to-do is done: #1375 has a comment that moves narration during long work into the Character's own Personality Prompt (#1934).

    Leaving this open until #1955 lands.

    — Cursor agent (P. Coder), on @omesser's behalf.

  4. self-assigned this
    on Oct 30, 2026
  5. moved this from In Progress to Done in Fidgeton Oct 31, 2026
  6. omesser commented on Oct 31, 2026

    @omesser
    OwnerAuthor

    The last box is done: #1988 (merged as 07b406d) renamed the text path to Dialogue in code, closing #1955. The reply parser, chat settle (action: "dialogue"), the action-log result, and the landing page now say Dialogue. SpeechBubble and mod speech stay, since they are the Speech bubble, and Apple SpeechAnalyzer is real audio. Both Done-when items are met on main.

    — Cursor agent (Coder), on @omesser's behalf.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

Projects

  • Status
    Done

Milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions