Cost ledger: a pull shows what it cost — tokens, lane minutes, waiting, ask to live #12330

A pull page now carries one server-rendered line, cost: $0.42 tokens · 30 lane-min · 3.0 h waiting · ask→live 5.3 h, and an issue page carries the same sum over itself and the pulls that close it. bun test on the new test, test/pages-budget.test.ts, test/tokens.test.ts and test/cost-ledger.test.ts, plus bun x tsc --noEmit, bun scripts/gate.ts and bun scripts/design-lint.ts, passed locally. I did not run the full suite, and I did not use the keep_plan or estimate tools.

• Where it reads from: src/core/ledger_rollup.ts sums the finished run rows, which are the same rows the ledger's entries are written from. The ledger's jsonl files are never opened on a request. A pull costs one statement, indexed on run_issue and deploy_pull_issue. An issue costs two: one to find its pulls, one to sum.
• The line:

- Dollars are cost_usd, lane-minutes run from start to finish, and waiting runs from queued to start. - ask→live runs from the ask to the first production deploy that carried the pull. - Parts that are not known are left out, and with no finished run there is no line. - Like workSpent, it shows only to people who can see the repository's owner facts.

• Where it shows: costSpent in src/web/pages/runs.ts sits next to workSpent. The pull page is in src/web/studio/change.ts, not pages/pull.ts as the issue's file list says. The issue page is in src/web/studio/tell.ts.
• Tests: test/ledger-rollup.test.ts covers the sum checked by hand against the rendered pull page, a pull with no entries showing nothing, parts left out when unknown, the issue sum, and the query plans. The escaping check only asserts that a pull title containing <b> is not echoed raw on the issue page. The cost line itself is plain numbers in an escaping template, so there is no user text to escape there.
• Budget: The issue page's statement budget in test/pages-budget.test.ts went from 70 to 71. I chose to raise it by one rather than fold the finding of closing pulls into the sum with a looser match. I did not fix the budget by cutting an existing statement.
• Linking an issue to its pulls: There is no stored link, so the issue page matches pulls by their Closes #N text, newest 200 of the repository, using the issue_repo_kind_number index. This is a bounded read but not a pure index lookup.
• Left out:

- Token counts in the line; the issue's $ tokens figure is dollars only. - A terminal screen, because no new page was added.

To try it: open a pull that agent runs worked on and read the line under the title. Then open the issue it closes and read the sum.

This pull's check failed on the gate rule that wraps comments at 100 columns. Two doc comments in src/core/ledger_rollup.ts (lines 68 and 79) were over the limit. I shortened both and pushed.

• Comment on line 68: the cost-line comment now ends "if none known" instead of "when no part is known".
• Comment on line 79: the comment on the pull-ids lookup now ends "newest 200" instead of "newest 200 at most".
• Gate behaviour: it only checks lines already committed, so it kept reporting the old lengths until I committed. It now passes locally.
• Other local checks: bun x tsc --noEmit is clean, and the 57 ledger and cost tests pass.
• Not run: scripts/fast.ts named 1813 of 3083 test files for this diff, so I ran only the ledger tests. The change is comment-only and touches no code, so the forge's full check on this push is what covers the rest.

Pull #12330 shows what a pull cost, summed from its finished runs, and an issue shows the sum over its pulls. This push fixes the two failures the cost ledger caused. Seven tests still fail, and they also fail on a clean checkout of origin/main.

What I fixed

• Pull-page budget: the cost line adds one SQL statement to the pull page. I raised that page's limit in test/pages-budget.test.ts from 113 to 114 and noted why. The issue page was already raised to 71 earlier on this branch.
• docs/MCP.md: the tool table was stale. I regenerated it with bun scripts/mcp-table.ts after merging origin/main.

What I ran

• test/ledger-rollup.test.ts, test/pages-budget.test.ts and test/mcp-table.test.ts pass on this branch only before the last merge. After that merge, docs/MCP.md went stale for limits and I regenerated it a second time, but I never re-ran test/mcp-table.test.ts afterwards. The same applies to test/ledger-rollup.test.ts.
• bun x tsc --noEmit printed nothing and bun scripts/gate.ts printed nothing, but I never checked their exit codes, so I can't confirm either is clean.
• I did not run the full suite. The forge's check does that.

Still red on origin/main itself I ran these in a worktree of origin/main and they fail there too. I left them alone because they are outside this change.

• test/first-screen-line.test.ts: four tests fail, because r.where is undefined in the fixture.
• test/spawn-sites.test.ts and test/spawn-pipes.test.ts: src/git/own_refs.ts spawns a process but is not listed.
• test/did-you-mean.test.ts: the pinned 200-path query count is 39 and the page now makes 40.
• test/mcp-table.test.ts fails there for the same stale limits row. After my regeneration it should pass here, but see above.

The engine-thread test (an engine that stops saying it is alive…) failed in the check with a 30-second timeout. I did not rerun it here. I suspect it is a flake, but I haven't confirmed that.

Left out: nothing new. The cost line is shown only to people who can see the owner's figures.

The check failed on a gate rule: a pull that changes src/web/ pages needs a Looks: line in a commit message. I pushed an empty commit with Looks: /preview-admin/sample/pulls/3, so the forge will take the before/after pictures of that page.

bun scripts/gate.ts ran clean locally before the push. I haven't seen the forge's check result for the new push yet.

The sample pull has no finished agent runs, so the cost line probably won't appear in those pictures. I chose that page anyway because it is the pull page in forge.json's looks list; the issue page is not in the list.

I added no code changes in this push. The seven tests that were already failing on origin/main are still failing, as I described last time.

The pull now shows the cost line on the web pull and issue pages and on both terminal screens. The only red test I know of is docs/MCP.md's tool table (test/mcp-table.test.ts).

I ran bun scripts/gate.ts (green), bun x tsc --noEmit (clean), and the ledger, terminal and page-budget tests. All pass except test/mcp-table.test.ts, which was already failing on origin/main before this branch. I took the docs/MCP.md edits out as you asked, so it stays red until main regenerates that table. I did not run the full suite or the other tests that failed on main earlier; the forge's check does that.

• Terminal: src/tui/screens/pull.ts puts the line in the head's second row. src/tui/screens/issue.ts puts it on the cost row, beside the existing Cost $… sentence. Both call costLineFor in src/core/ledger_rollup.ts, the same function and the same gate the web pages use.
• Docs: docs/features/cost-ledger.md already existed for the ledger file, so I added a section there on how to read the line and which tests cover it.
• Budgets: I reverted the raised caps in test/pages-budget.test.ts to main's values, and the test passes at them.

- workSpent now checks owner facts once and reuses its usage read as the gate, so a pull no agent has worked on pays no extra read. - An issue pays one extra read for its closing pulls, and only when it has none of its own runs.

• ask→live: a pull now measures from the earliest closed issue's createdAt, falling back to the pull's own if it closes none. There is a test for it.
• Label: the line is a role="group" span labelled "What this cost".
• Escaping tests: the page test checks the rendered markup and label and that the title isn't echoed. A guest sees no line, on the page or from costLineFor. The terminal tests are in test/tui-issue-cost.test.ts and test/tui-pull-cost.test.ts, each also checking a reader sees nothing.

A pull with only check or deploy runs and no agent run shows no line, because the agent usage read gates it. That is a deliberate trade for the budget; say if you'd rather it show.

Sequencing against the pipeline-first direction is the owner's call.