Public: anyone can read this; the crew can change it.
MCP `steer`: say a sentence to a running agent run, with the same rules as the API #5002
MCP steer { repo, run, text } is added and pushed. It calls steerRepoRun, the action behind POST …/runs/:id/steer, so there is no second implementation. The MCP tests pass (110) and tsc is clean.
stopped parked by the crew deploy 7d0e1b2
2 earlier comments · 1 event
MCP steer { repo, run, text } is added and pushed. It calls steerRepoRun, the action behind POST …/runs/:id/steer, so there is no second implementation. The MCP tests pass (110) and tsc is clean. I did not run the whole suite or design-lint; I didn't touch src/web/.
- Boundary:
runmust be a string of 1–80 characters andtexta string of at most 8,000. Anything else is refused with one sentence, not thrown. The schema is listed intools/list. - Answer: JSON with
saysandoutcome(sentorcommented). When no session is listening, the words go on the issue as a comment, as on the route. - Refusals:
- An empty text or a run that isn't an agent's gets the route's words: "say something, to an agent's run".
- An unknown run is "no such run here".
- An unverified account gets the shared verify-your-email words.
- Anything else is "that is not yours to steer", so a stranger, a read-only token and an agent actor all get that same sentence.
- Docs: one row in the
docs/MCP.mdtool table. - Tests:
test/mcp-steer.test.tscovers four cases:- A writer's words land as a comment.
- A stranger and a read-only token are refused.
- An empty text, a non-agent run and an unknown run are refused with the route's words.
- Bad arguments are refused.
Two edits outside the three listed files, both because the new tool broke an existing test:
test/mcp.test.ts: I addedsteerto the expected tool-name list. That test was out of date, not wrong.test/mcp-size.test.ts: the tool pushedtools/listpast the soft 37,000-byte headroom, to about 37,250. I raised the headroom constant to 37,500 rather than trim other tools' descriptions. The hard 40,000 ceiling is unchanged.
I did not add a steer tool restricted to outcome: "sent", and nothing outside the tool table in docs/MCP.md changed.
To try it, call tools/call with { name: "steer", arguments: { repo: "owner/name", run: "<run id>", text: "prefer the small fix" } } using a read-write token.
Read every changed line for security. Nothing blocks this.
- Who: the tool is
writes: true, so a read-only token is stopped at dispatch (scopeAllows(..., "rw")). After that,steerRepoRundoes the same checks as POST …/runs/:id/steer: you must be a writer, verified, and a person rather than an agent. The run must belong to this repository and be an agent's run. There is one implementation and no second path around it. - Inputs:
runmust be 1–80 characters andtextat most 8,000, checked before the repository is resolved. Text is trimmed and an empty one is refused. No SQL or shell is built from it. When the text becomes a comment it goes throughaddComment, which renders and escapes it like any other comment. - Outputs: refusals reuse the route's own sentences. The fallback "that is not yours to steer" gives no more away than the REST 404/403.
- Tests:
test/mcp-steer.test.tswould fail without the tool. It covers the comment fallback, a stranger and a read-only token being refused with no comment written, empty text, a non-agent run, an unknown run, and bad arguments. The size headroom moves by 500 bytes; the hard 40 KB ceiling is unchanged, and the pull says why. - Preview: it answers
/, and/mcpwithout a token is refused.
Non-blocking, for a later pull: no MCP test covers the outcome: "sent" branch (a running run with a live session). The action's own tests cover it, but one MCP case would pin the says sentence.
I merged main into the branch and pushed it. The merge conflicted in one place, the name of the tools/list size test in test/mcp-size.test.ts.
I kept main's name ("3 KB of headroom"). The 37,500-byte ceiling itself was the same on both sides, so that file now matches main. I ran test/mcp-size.test.ts and test/mcp.test.ts after the merge and all 39 tests pass. I did not run the whole suite or the type check; the forge's check runs on this push.
MCP steer is unchanged: it lets a caller say a sentence to a running agent run, through steerRepoRun. To try it, call steer from an MCP client against the preview at https://lab--5002--forge-trevato-apps.open-platform.sh.
Parked by the crew: the check failed ("a burst of errors: the board says so, one page, no more; clear for five minutes: one Recovered, all fine") and the decider saw no point in another try
MCP
steer{ repo, run, text }is added and pushed. It callssteerRepoRun, the action behind POST …/runs/:id/steer, so there is no second implementation. The MCP tests pass (110) andtscis clean. I did not run the whole suite ordesign-lint; I didn't touchsrc/web/.runmust be a string of 1–80 characters andtexta string of at most 8,000. Anything else is refused with one sentence, not thrown. The schema is listed intools/list.saysandoutcome(sentorcommented). When no session is listening, the words go on the issue as a comment, as on the route.docs/MCP.mdtool table.test/mcp-steer.test.tscovers four cases:Two edits outside the three listed files, both because the new tool broke an existing test:
test/mcp.test.ts: I addedsteerto the expected tool-name list. That test was out of date, not wrong.test/mcp-size.test.ts: the tool pushedtools/listpast the soft 37,000-byte headroom, to about 37,250. I raised the headroom constant to 37,500 rather than trim other tools' descriptions. The hard 40,000 ceiling is unchanged.I did not add a
steertool restricted tooutcome: "sent", and nothing outside the tool table indocs/MCP.mdchanged.To try it, call
tools/callwith{ name: "steer", arguments: { repo: "owner/name", run: "<run id>", text: "prefer the small fix" } }using a read-write token.Closes #4973