The terminal's run screen says a word to a running crew run with `s`, as the run page's steer box and MCP `steer` do #7309

The terminal can now steer a running crew run. A one-line composer sends the text through steerRepoRun, and the outcome shows as a toast in MCP's words: "Sent. The agent hears it at its next step." or "Commented. No session was listening; …".

Keys. The issue asked for s, but on the runs list (src/tui/screens/runs.ts) s already cycles the state filter, and test/tui-runs-state.test.ts checks that. Letting s steer only when a running agent row is selected would also trap someone on the Running filter. So steering is S there. On the issue screen's crew card and the pull's Crew tab, s is free and steers.

Where it lives.

• maySteer and steerPrompt in src/tui/screens/runlist.ts, used by all three screens. Steering is offered only on an agent run that is running, to a person (never an agent session) who can write the repository. Empty text keeps the composer open with "Say something first.".
• The two outcome sentences are STEER_SAYS in src/actions/runs.ts. MCP steer now reads them from there instead of writing them inline; its doc and schema are unchanged.
• src/crew/runner.ts registers a listening session through a new exported listenForSteer, which the runner itself now calls. Behaviour is the same, and a test can listen to a run the way a session does.

Tests. The new test/tui-steer.test.ts covers: S on the runs list reaches a listening session; with no session listening the text becomes a comment on the issue; empty text is refused in place; a finished run has no steer key; the crew card steers with s; and an agent with write access is not offered it. With the agent check removed, the last test fails. bun x tsc --noEmit is clean. 329 tests across 34 files pass with 1 skipped, including test/mcp-size.test.ts, the run-cancel and runs-state tests, and the issue, pull, runner and steer suites. scripts/fast.ts named 1130 files, so I chose these by hand.

I left a queued run without steering: the issue asks only for running runs, and the run page still covers the others.

To try it: start a crew run, open the repository's runs list in the terminal, select the running agent row, press S, type a sentence and press enter.

🤖 Generated with Claude Code

Closes #7288

The terminal can now steer a running crew run. A one-line composer sends the text through steerRepoRun, and the outcome shows as a toast in MCP's words: "Sent. The agent hears it at its next step." or "Commented. No session was listening; …".

Keys. The issue asked for s, but on the runs list (src/tui/screens/runs.ts) s already cycles the state filter, and test/tui-runs-state.test.ts checks that. Letting s steer only when a running agent row is selected would also trap someone on the Running filter. So steering is S there. On the issue screen's crew card and the pull's Crew tab, s is free and steers.

Where it lives.

• maySteer and steerPrompt in src/tui/screens/runlist.ts, used by all three screens. Steering is offered only on an agent run that is running, to a person (never an agent session) who can write the repository. Empty text keeps the composer open with "Say something first.".
• The two outcome sentences are STEER_SAYS in src/actions/runs.ts. MCP steer now reads them from there instead of writing them inline; its doc and schema are unchanged.
• src/crew/runner.ts registers a listening session through a new exported listenForSteer, which the runner itself now calls. Behaviour is the same, and a test can listen to a run the way a session does.

Tests. The new test/tui-steer.test.ts covers: S on the runs list reaches a listening session; with no session listening the text becomes a comment on the issue; empty text is refused in place; a finished run has no steer key; the crew card steers with s; and an agent with write access is not offered it. With the agent check removed, the last test fails. bun x tsc --noEmit is clean. 329 tests across 34 files pass with 1 skipped, including test/mcp-size.test.ts, the run-cancel and runs-state tests, and the issue, pull, runner and steer suites. scripts/fast.ts named 1130 files, so I chose these by hand.

I left a queued run without steering: the issue asks only for running runs, and the run page still covers the others.

To try it: start a crew run, open the repository's runs list in the terminal, select the running agent row, press S, type a sentence and press enter.

🤖 Generated with Claude Code

The terminal can now steer a running crew run. S on the runs list, s on an issue's crew card and s on a pull's Crew tab each open a prompt. The word goes through steerRepoRun, the same path as the run page's steer box and MCP steer. It is offered only on a running agent run, to someone who can write here and is a person, not an agent. If no agent session is listening, the word is posted as a comment on the issue or pull instead.

This round adds the two things the review asked for:

• Crew tab test. test/tui-steer.test.ts has a new test, "the pull's Crew tab steers with s". It opens pullScreen on a real pull with a running agent run, switches to the Crew tab and checks that s steer shows. Then it presses s, types, presses enter, and checks that the word reached listenForSteer and the toast says Sent.. I turned off the s branch in src/tui/screens/pull.ts and the test failed, then put the branch back. All 7 tests in the file pass and tsc --noEmit is clean.
• Docs. The steer row in docs/MCP.md now says the terminal does the same with these three keys, and who gets them.

I chose S rather than s on the runs list, which the first round's description explains.

To try it: SSH into the forge with a running agent run on a repository you can write to. Open its runs list and press S, or open the pull, press tab to reach Crew and press s. Type a word and press enter.

🤖 Generated with Claude Code

The terminal can now steer a running crew run, as the run page's steer box and MCP steer can. S on the runs list, s on an issue's Crew card and s on a pull's Crew tab open a prompt. The word goes through steerRepoRun, so the rules are the same as on the page. Steering is offered only on a running agent run, to someone who can write here and is a person, never an agent. Empty text is refused in place. The toast says whether the agent heard the word or, with no session listening, it was posted as a comment on the issue or pull. The runs list uses S, not s; the first round's description says why.

Changes from review:

• Crew tab test: test/tui-steer.test.ts now covers steering from a pull's Crew tab. With that key's handler in src/tui/screens/pull.ts turned off, the test fails.
• Docs: the terminal's steering is described in docs/design/sections/6g-terminal-apps-runs-crew.md, next to the entry for a run row.
• Failed check: the reviewer suggested putting the docs line in the steer row of docs/MCP.md. That table is generated by scripts/mcp-tools.ts, and the hand edit failed the freshness test in test/mcp-tools.test.ts. The row is back to its generated text, and the docs line lives in the design section instead.

test/tui-steer.test.ts and test/mcp-tools*.test.ts pass (166 tests), and bun x tsc --noEmit was clean on the previous commit (this last one changed only docs).

To try it: SSH into the forge with a running agent run on a repository you can write to. Press S on its runs list, or open the pull, press tab until the Crew tab is showing and press s. Type a word and press enter.

🤖 Generated with Claude Code

In the terminal, s now sends a word to a running crew run from three places: the runs list (S), an issue's crew card, and a pull's Crew tab. It does the same as the steer box on the run page and the MCP steer tool. Only a person who can write to the repository gets the key, and only on an agent run that is still running. A finished run, or an agent session that may write, is not offered it.

This round makes one change, from review: the pull's Crew tab test in test/tui-steer.test.ts now gives up after 3 s instead of 20 s while it waits for the tab to show. A hang now fails fast and stays inside the per-test budget (#1242). 3 s is the limit test/tui-teams.test.ts already uses. The test passes in about 1.5 s on its own, and all 7 tests in the file pass.

Earlier commits on this pull:

• 6e7f3cc7 adds the keys.
• 7da7d067 adds the pull Crew tab test.
• 03cec815 adds the Steer line to docs/design/sections/6g-terminal-apps-runs-crew.md and leaves docs/MCP.md as generated, since that table is checked for freshness.

To try it, connect with the command in FORGE_APP_SSH, open a repository's runs while an agent run is running, press S, type a word, and press enter. The toast says Sent. and the run hears the word.

🤖 Generated with Claude Code

In the terminal, a person who can write to the repository can now send a word to a running crew run, the same way the run page's steer box and MCP steer do. The key is S on the runs list (where s already cycles the state filter), and s on an issue's crew card and on a pull's Crew tab. It only appears on an agent run that is running, and an agent's own session is never offered it. The send goes through steerRepoRun, and the toast uses the same words MCP answers with.

What moved where:

• maySteer and steerPrompt live in src/tui/screens/runlist.ts, and the runs, issue and pull screens all call them.
• docs/design/sections/6g-terminal-apps-runs-crew.md has a Steer line naming the three keys and who gets them.

Since the last round:

• Merge with main. The pull did not merge because main added the run log screen and runwords.ts (#7306) and the runs screen's a requeue (#7307) in the same files. Both sides are kept:

- pull.ts imports runTone from runwords.ts. - runlist.ts keeps main's log screen beside maySteer and steerPrompt. - The runs screen offers both S and a, in its key list, its key handling and its header comment.

• Test timeout. From review, the pull Crew tab test in test/tui-steer.test.ts now waits at most 3 s for the tab instead of 20 s, the same limit test/tui-teams.test.ts uses. On its own that test takes about 1.5 s.

After the merge, tsc is clean, the design lint passes, and the steer, runs, cancel, requeue, run log and pull hold test files pass (40 tests).

To try it: connect with the command in FORGE_APP_SSH, open a repository's runs while an agent run is running, press S, type a word and press enter. The toast says Sent.

🤖 Generated with Claude Code