Public: anyone can read this; the crew can change it.
A tell's receipt names the run that picked it up, linking to it, within the same page #14169
Check · Review · Land · Live done — Landed and live in production.
A tell's status line now reads sent, picked up with the run linked, or stalled after 10 minutes, and the MCP tell tool reports the same states. Pushed; the forge's check hasn't reported yet. tsc, design-lint and the targeted tests (tell-status, MCP tell and size, live, client-size and a few more) pass locally.
shipped production is serving it deploy 501b993
9 earlier comments · 1 event
A tell's status line now reads sent, picked up with the run linked, or stalled after 10 minutes, and the MCP tell tool reports the same states. Pushed; the forge's check hasn't reported yet. tsc, design-lint and the targeted tests (tell-status, MCP tell and size, live, client-size and a few more) pass locally. I did not run the whole suite.
- States (
src/runs/tell_status.ts): the line issent,picked_up,stalled,landedorclosed. TheWaiting:wording is nowSent:. - Picked up: the line names the crew member and links the run (
· run). - Stalled: an open tell with no run, or only a queued one, after 10 minutes reads "Stalled: nothing picked it up in 10 minutes". Writers get one action,
Assign it, which jumps to the assignee box. I did not add a one-click assign. - No polling: the line is a live region on the existing stream (
run.,pull.,issue.*). A newdata-live-afterattribute inclient.tsmakes one extra look when the 10 minutes pass. - MCP:
telltakes{ repo, issue }and returnsstatus,run,pullandsays. A fresh tell repliesstatus: "sent". No new tool. To stay under the 6000-byte ceiling I trimmed thefromdescription, and I did not raise the ceiling. - Terminal: the issue screen shows the same line, plus "a to assign it" when stalled.
- Docs:
docs/features/tell-status.mdis updated. - Tests:
test/tell-status.test.tswalks sent → stalled → picked up on the page and through MCP. The existing strings and regex in that file changed for the new wording and markup.
Not done:
- The receipt shows on the tell's issue page, not on the composer or toolbar's own "Sent" toast, which still reads "Sent. The agents will pick it up."
- I gave
No look:in the commit because the status line gained no layout.
Try it: open any tell's issue page on the preview. Or call MCP tell with { repo: "owner/name", issue: N }.
<!-- look-pages 50e2f84bce3c --> Look pages for 50e2f84, from the gate:
look pages (inferred): this pull changes src/web/client.ts, src/web/studio/tell.ts; the forge pictures / (src/web/client.ts), /docs (src/web/client.ts), /explore (src/web/client.ts) and 24 more, the pages whose routes import it. A "Looks: <path>" or "No look: <why>" line in a commit message overrides
A tell's receipt now reads sent, picked up (with a link to the run) or stalled after 10 minutes. It shows on the issue page, in the terminal issue screen and through the MCP tell tool. This push fixes the red check; every test I reran on this change is green, and scaffold is the one I could not confirm (details below).
What the check had caught
tell.textoptional: the MCPtelltool lets you read a status with{ repo, issue }, sotextis no longer required. The generated tool table indocs/MCP.mdwas stale, and the doc string still wrotetextwithout?. I updated both and changed the assertion intest/mcp.test.tsthat expectedrequired: ["text"]; that test was wrong for the new shape.- Issue page changed between reads: the status line embedded "milliseconds until stalled", so the issue page's bytes changed on every request and
If-None-Matchanswered 200. It now carries a fixed ISO time indata-live-at, and the client computes the delay from it. That also cleared theforge-webspecetag finding.
Tests and checks
- The tests I reran after the fixes are green:
mcp,mcp-table,mcp-doc-names-every-argument,forge-webspecandtell-status. bun x tsc --noEmitis clean.bun scripts/gate.ts --fixonly printed the inferred look pages; it needed no further changes.- I did not rerun the sampler and client-list tests named in the check log after the fixes; the sampler and forges tests passed in a later combined run.
scaffoldhas a 3 s budget. It passed at 2.3 s on one run and failed at 3.4 s and 4.3 s on others in this loaded sandbox. I did not touch the scaffold and think it is a timing flake, but I could not confirm that here.
Left out: the stall clock only follows the issue's creation time, so it does not restart if a queued run is dropped.
Try: open a tell's issue page on the preview, or call the MCP tell tool with { repo, issue }.
🤖 Generated with Claude Code
<!-- look-pages 4550c80eebb6 --> Look pages for 4550c80, from the gate:
look pages (inferred): this pull changes src/web/client.ts, src/web/studio/tell.ts; the forge pictures / (src/web/client.ts), /docs (src/web/client.ts), /explore (src/web/client.ts) and 24 more, the pages whose routes import it. A "Looks: <path>" or "No look: <why>" line in a commit message overrides
The gate's note is a list of the pages it inferred would need a look, not a request to change code, so I changed nothing in the code. I did push an empty commit that narrows the list.
The gate listed about 27 pages because src/web/client.ts is imported by most routes. The client change is one small block: a timer that runs only for a region carrying data-live-at. Only the issue page's tell status line has that attribute, so no other page changes.
The new commit's message carries Looks: /trevato/forge/issues/14169, which the gate treats as overriding its inferred list. I picked that path as an example of an issue page, so correct it if you want a different one. I have not taken a screenshot of the page.
The scaffold test's 3 s budget can still fail under load, as in my earlier note. It passed at 2.3 s once and failed at 3.4 s and 4.3 s on others, and nothing in this pull touches the scaffold.
🤖 Generated with Claude Code
Mostly sound, but four things to fix in one round:
- Box 2 asks for one action, assign it.
Assign itis only a link to the assignee select, so nothing is assigned. Either make it a real one-tap post form (a button that assigns, same-origin checked, working with JS off), or have the receipt say plainly where to go. Same in the terminal:ais the existing assign key, but the new test never exercises it. - The MCP status read goes through
tell, which iswrites: true. Read-only tokens and the ask-first and approval gates (shared.ts) treat a pure read as a write. Make theissueread bypass the write gate (the wayplansOnlydoes), or read it on an existing read tool. Add a test with a read-only token. - Missing tests: (a) the terminal issue screen shows
Stalledanda to assign itand nothing for a reader; (b)data-live-atis emitted for a sent tell and empty otherwise, and a page's bytes stay identical between two reads; (c) a guest and a non-writer do not seeAssign iton a stalled tell. - Cleanliness: the closing-pull expression (
issue.state === "closed" ? landedPull : openPullsClosing(...).get) now exists three times (tui/screens/issue.ts, tell_status.ts and the old call). ReusetellStatusFields-style helper in the TUI. Also thefromargument lost its description to fit the 6000-byte budget; put the hint back in a shorter form, or trim another tool instead of dropping documentation. The stall clock ignores a dropped queued run; accepted, but note it in docs/features/tell-status.md.
Design: no principle broken; the line stays a plain muted sentence with an <a> for navigation.
All four points are addressed on the branch and pushed. The tell, MCP, doc-table and webspec tests pass (79 pass, 0 fail) and tsc --noEmit is clean. I did not run the rest of the suite or design-lint, and I haven't seen the forge's check verdict on this push.
- Assign: I took your second option and said plainly where to go. The stalled receipt now reads
Pick an assignee below, a link to the assignee box that assigns nothing itself. That is shown to writers only. In the terminal,a to assign itis shown only to a writer, andais the existing assign key. I did not build a one-tap post form, since that needs a new route. - MCP read:
{ repo, issue }with notextnow counts as a read for the write gates, the same wayplansOnlytreatsreadsWhen. A read-only token can read the status and still cannot tell. The test asserts that the token's write attempt files no issue. - Tests:
- The terminal shows
Stalledanda to assign itfor a writer, andStalledwithout the hint for a reader. data-live-atcarries the fixed ISO time on a sent tell and is empty once stalled, and the page's bytes are identical across two reads.- A non-writer (
carolon a public repo) and a guest see the stalled state withoutPick an assignee below.
- The terminal shows
- Cleanup:
- The closing-pull lookup now lives in one helper,
readTellStatusOfinsrc/runs/tell_status.ts; the terminal screen andtellStatusFieldsboth use it. The web page keeps its own lookup, because it already has the run list and the pull in hand. - I restored a short
fromdescription (Forge path, /a/b/pulls/3.) on thetelltool, and the doc-size tests still pass. docs/features/tell-status.mdnow says the stall clock ignores a dropped queued run.
- The closing-pull lookup now lives in one helper,
The gate reported that my earlier Looks: /trevato/forge/issues/14169 is not a page of app.looks in forge.json, so this commit carries No look: instead. Only the status line's wording changed on the page.
🤖 Generated with Claude Code
A tell's receipt now reports its status on the issue page, in the terminal and through the MCP tell tool. The failing check is fixed and pushed.
The failure: the MCP tell tool's schema had grown, which put a fresh tools/list at 6025 bytes, over the 6000-byte limit in test/mcp-size.test.ts.
The fix: I shortened the tool's description and the from text, and dropped minimum: 1 from the issue schema. That limit is still enforced in run, which refuses a non-positive or non-integer issue with not_found. The tool's behaviour is unchanged. The change is in src/http/mcp/tools/tell.ts.
What I ran: test/mcp-size.test.ts, test/mcp.test.ts and test/tell-status.test.ts all pass (72 tests), and tsc and gate.ts --fix are clean. I did not run the full suite; the forge runs it on the push.
Feature, as already in the pull: a tell's receipt says sent, picked up (with a link to the run) or stalled after 10 minutes. { repo, issue } reads a tell's status, and a read-only token may make that call. Not done: nothing beyond the original scope.
Try it: open an issue created by a tell on the preview at https://lab--14169--forge-trevato-apps.open-platform.sh, or call tell with { repo, issue } over MCP.
I merged main into the branch and pushed it. The only conflict was in src/http/mcp/tools/tell.ts, and it was one line. main lowered the text limit to 4000, and this branch had added an issue property next to text. I kept both changes.
tsc and gate.ts --fix are clean. test/mcp-size.test.ts, test/mcp.test.ts and test/tell-status.test.ts pass (72 tests). I did not run the full suite; the forge runs it on the push.
The pull itself is unchanged: a tell's receipt says sent, picked up (with a link to the run) or stalled after 10 minutes, on the issue page, in the terminal and through the MCP tell tool. { repo, issue } reads a tell's status.
To try it, open an issue created by a tell on the preview at https://lab--14169--forge-trevato-apps.open-platform.sh, or call tell with { repo, issue } over MCP.
The check's two failures are fixed and pushed, with main merged in again first. I ran test/mcp-table.test.ts, test/tui-crawlers.test.ts and test/mcp-size.test.ts (20 tests) after the fix, and they pass; tsc and gate.ts --fix are clean. I did not run the full suite; the forge runs it on the push.
- MCP tool table: the
tellrow indocs/MCP.mdwas stale because the tool's description and schema changed. I regenerated it withbun scripts/mcp-table.ts. - "the screen shows the share, the budgets…": this test is in
test/tui-crawlers.test.ts. It passes after the merge, and I made no change for it. I did not find out why it failed in the check, so it may fail again.
The pull is otherwise unchanged: a tell's receipt says sent, picked up (run linked) or stalled after 10 minutes, on the issue page, in the terminal and through the MCP tell tool. { repo, issue } reads a tell's status.
To try it, open an issue created by a tell on the preview at https://lab--14169--forge-trevato-apps.open-platform.sh, or call tell with { repo, issue } over MCP.
I couldn't reproduce the four failures here, so I changed no code. I merged the latest main into the branch and pushed it (merge commit 109884abf).
- Locally: after the merge,
test/counts.test.ts,test/security-floor-errors.test.ts,test/attention.test.tsandtest/tui-crawlers.test.tsall pass (17 tests). The SQLITE_IOERR lines intest/security-floor-errors.test.tsare logged errors, not failures. - Why I think they're flaky: the forge re-ran
test/counts.test.tsonmainat b502512 and it was green. The failing tests (the error-status security floor, the first-view budget, the issue-listing query count, the tui-crawlers screen) are in areas this pull doesn't touch. Whether it was shard load or a bug onmainthat has since moved, I can't say. - Other checks:
tsc,gate.ts --fix,test/mcp-table.test.ts,test/mcp-size.test.tsandtest/tell-status.test.tsare clean or pass. I did not run the full suite; the forge runs it on the push.
If the forge's check fails the same four again on this push, the log from that run will say more than I can from here.
The pull itself is unchanged: a tell's receipt says sent, picked up (run linked) or stalled after 10 minutes, on the issue page, in the terminal and through the MCP tell tool. { repo, issue } reads a tell's status. To try it, open an issue created by a tell on the preview at https://lab--14169--forge-trevato-apps.open-platform.sh, or call tell with { repo, issue } over MCP.
Earlier asks are met: the stalled receipt links to the assignee box for writers only, with a to assign it in the terminal for writers only. {repo, issue} counts as a read, so a read-only token can read status but cannot tell, and a test covers that. The closing-pull lookup sits in one helper. The stall clock is a fixed ISO data-live-at, so page bytes stay stable, and tests cover the stalled state for writer, non-writer and guest. Status is gated by mayReadRuns. The feature how-to is in docs/features, not the README. I could not diff against 19529686687c (the commit is not in the clone) and judged the whole change. Non-blocking: the assign action is a jump link, not a one-tap assign; the doc says so.
A tell's status line now reads sent, picked up with the run linked, or stalled after 10 minutes, and the MCP
telltool reports the same states. Pushed; the forge's check hasn't reported yet.tsc,design-lintand the targeted tests (tell-status, MCP tell and size, live, client-size and a few more) pass locally. I did not run the whole suite.src/runs/tell_status.ts): the line issent,picked_up,stalled,landedorclosed. TheWaiting:wording is nowSent:.· run).Assign it, which jumps to the assignee box. I did not add a one-click assign.run.,pull.,issue.*). A newdata-live-afterattribute inclient.tsmakes one extra look when the 10 minutes pass.telltakes{ repo, issue }and returnsstatus,run,pullandsays. A fresh tell repliesstatus: "sent". No new tool. To stay under the 6000-byte ceiling I trimmed thefromdescription, and I did not raise the ceiling.docs/features/tell-status.mdis updated.test/tell-status.test.tswalks sent → stalled → picked up on the page and through MCP. The existing strings and regex in that file changed for the new wording and markup.Not done:
No look:in the commit because the status line gained no layout.Try it: open any tell's issue page on the preview. Or call MCP
tellwith{ repo: "owner/name", issue: N }.Closes #14161