A shared-red sentinel issue names the pulls that failed the test and quotes the failure, so a builder can reproduce it #8293

A sentinel finding for a shared: cause now gives a builder enough to reproduce the red: the pulls and runs that failed the test, the failing assertion lines, the base commit, and what main's newest check said. Both kinds of shared: finding get the same lines ("was degraded N minutes" and "is past its budget"). The new lines are:

• test: <name>
• failed: pull #N, check run <id>, one line per failed pull check on that test in the window, newest first, at most five.
• base: <oid> (run <id>): the base commit the newest of those runs merged with.
• log: …: what bun printed above that run's (fail) line for the test, up to 20 lines. Each line is collapsed to one line and cut to the 200 characters quoted evidence already uses. Only the last 400 lines of the log are read (readRunLogTail), never the whole file. bun's own repeat of the (fail) lines under its pass/fail totals is skipped, so the quote is the real error, not that list.
• main: passes on main's newest check, run <id> at <oid>, or fails on main's newest check too …, or no check of main recorded.

Privacy: if the cause's repository is private, none of these lines appear unless the finding is being filed in that same repository. That includes the case where no filing repository is known.

Where it lives: the queries and the log parsing are in src/checks/shared_red.ts (sharedRuns, mainVerdict, assertionLines). The sentinel formats them in sharedLines in src/crew/sentinel.ts. Reading the log needs the data directory, so findFaults, budgetBreaches and degradedCauses take an optional dataDir, and hunt passes app.config.dataDir.

sentinelTick does not pass dataDir: there the breaches only decide whether to hunt, so it never reads a log. When a new failed run changes a tracked finding's evidence, the sentinel adds its usual "Still there" comment to the open issue.

test/sentinel-shared-evidence.test.ts records three failed checks on one test, each with its own log and base. It checks:

• the three pulls and runs, newest first;
• the newest run's quoted lines and base;
• the line about main, before any main check exists and after a green one;
• that a private repository's lines are kept out of a finding filed elsewhere;
• the 20-line cap and that bun's closing list is skipped.

It fails without the change because no failed: line exists. That file and the existing shared-red, sentinel, red-main and TUI sentinel suites pass (105 tests), as do bun scripts/gate.ts and bun x tsc --noEmit. scripts/fast.ts named 1095 files, too many to run here, so I picked those suites instead.

🤖 Generated with Claude Code

Closes #8286

A sentinel finding for a shared: cause now gives a builder enough to reproduce the red: the pulls and runs that failed the test, the failing assertion lines, the base commit, and what main's newest check said. Both kinds of shared: finding get the same lines ("was degraded N minutes" and "is past its budget"). The new lines are:

• test: <name>
• failed: pull #N, check run <id>, one line per failed pull check on that test in the window, newest first, at most five.
• base: <oid> (run <id>): the base commit the newest of those runs merged with.
• log: …: what bun printed above that run's (fail) line for the test, up to 20 lines. Each line is collapsed to one line and cut to the 200 characters quoted evidence already uses. Only the last 400 lines of the log are read (readRunLogTail), never the whole file. bun's own repeat of the (fail) lines under its pass/fail totals is skipped, so the quote is the real error, not that list.
• main: passes on main's newest check, run <id> at <oid>, or fails on main's newest check too …, or no check of main recorded.

Privacy: if the cause's repository is private, none of these lines appear unless the finding is being filed in that same repository. That includes the case where no filing repository is known.

Where it lives: the queries and the log parsing are in src/checks/shared_red.ts (sharedRuns, mainVerdict, assertionLines). The sentinel formats them in sharedLines in src/crew/sentinel.ts. Reading the log needs the data directory, so findFaults, budgetBreaches and degradedCauses take an optional dataDir, and hunt passes app.config.dataDir.

sentinelTick does not pass dataDir: there the breaches only decide whether to hunt, so it never reads a log. When a new failed run changes a tracked finding's evidence, the sentinel adds its usual "Still there" comment to the open issue.

test/sentinel-shared-evidence.test.ts records three failed checks on one test, each with its own log and base. It checks:

• the three pulls and runs, newest first;
• the newest run's quoted lines and base;
• the line about main, before any main check exists and after a green one;
• that a private repository's lines are kept out of a finding filed elsewhere;
• the 20-line cap and that bun's closing list is skipped.

It fails without the change because no failed: line exists. That file and the existing shared-red, sentinel, red-main and TUI sentinel suites pass (105 tests), as do bun scripts/gate.ts and bun x tsc --noEmit. scripts/fast.ts named 1095 files, too many to run here, so I picked those suites instead.

🤖 Generated with Claude Code

A sentinel finding for a shared: cause now gives a builder enough to reproduce the red: the pulls and runs that failed the test, the failing assertion lines, the base commit, and what main's newest check said. Both kinds of shared: finding get the same lines ("was degraded N minutes" and "is past its budget"). The new lines are:

• test: <name>
• failed: pull #N, check run <id>, one line per failed pull check on that test in the window, newest first, at most five.
• base: <oid> (run <id>): the base commit the newest of those runs merged with, or none recorded when the run stored none.
• log: …: what bun printed above that run's (fail) line for the test, up to 20 lines. Each line is collapsed to one line and cut to the 200 characters quoted evidence already uses. Only the last 400 lines of the log are read (readRunLogTail), never the whole file. bun's own repeat of the (fail) lines under its pass/fail totals is skipped, so the quote is the real error, not that list.
• main: …, which says one of three things:

- passes on main's newest check, run <id> at <oid>: that check was green, or it failed on other tests only. - fails on main's newest check too, …: that check failed on this same test. - no check of main recorded.

Privacy: if the cause's repository is private, none of these lines appear unless the finding is being filed in that same repository. That includes the case where no filing repository is known.

Where it lives: the queries and the log parsing are in src/checks/shared_red.ts (sharedRuns, mainVerdict, assertionLines). The sentinel formats them in sharedLines in src/crew/sentinel.ts. Reading the log needs the data directory, so findFaults, budgetBreaches and degradedCauses take an optional dataDir, and hunt passes app.config.dataDir.

sentinelTick does not pass dataDir: there the breaches only decide whether to hunt, so it never reads a log. When a new failed run changes a tracked finding's evidence, the sentinel adds its usual "Still there" comment to the open issue.

test/sentinel-shared-evidence.test.ts records three failed checks on one test, each with its own log and base. It checks:

• the three pulls and runs, newest first;
• the newest run's quoted lines and base;
• the main: line in four cases: no main check yet, main green, main red on this test, and main red only on another test;
• that six failed pulls name only the newest five;
• the none recorded base when a run stored none;
• that a private repository's lines are kept out of a finding filed elsewhere;
• the 20-line cap and that bun's closing list is skipped.

I checked the new tests by breaking the code on purpose. Removing the five-run cap, the none recorded fallback, or the "red on another test still passes" rule each made its test fail. That file (8 tests), bun scripts/gate.ts and bun x tsc --noEmit pass. In the first round I also ran the existing shared-red, sentinel, red-main and TUI sentinel suites: 105 tests, all passing. scripts/fast.ts named 1095 files, too many to run here, so I picked those suites instead.

🤖 Generated with Claude Code