[sentry-junior]

Adds one line to the conversation-mode <failure-reporting> block (block 6 in policies/system-prompt.md):

When a policy or a missing capability stops part of the request, say so plainly, then give the closest useful answer you can.

Why: getsentry/junior-prod#576 makes Sentry's SOUL.md voice only. Since #2371, Personality comes after all core rules, so behavior rules left in SOUL can win conflicts with core rules. I checked each behavior line in SOUL against the core blocks. All of them already have a core home except this one, so this PR moves it into core instead of dropping it. Every install gets it now, not only Sentry's.

SOUL line	Core home
lead with the answer	<output>: "Start with the answer or result"
explicit about confidence, separate facts from inference, never pretend certainty	<evidence>: "Do not present an inferred fact as verified", plus the uncertain-answer rule
critical of source quality and recency	<evidence>: verification source order, plus "empty, stale, ambiguous" retry
don't claim abilities you don't have	<safety>: "Stay within … the runtime's available capabilities", plus <evidence> claims about access
execute, verify, summarize	<task-execution> and <evidence>
shortest complete answer, one compact response	<output> length target
policy or missing capability: say so, give the closest useful answer	none, added here

Automation and child-agent runs keep their own failure paths (finishAutomationRun reason, reportBlocked), so the line is only in conversation mode.

Checks: prettier and @sentry/junior typecheck pass. This is a prompt change, so the behavioral and integration evals should run on it.

Refs getsentry/junior-prod#576

via David Cramer.

--

View Session [Sentry]

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
junior-docs	Ready	Preview	Oct 30, 2026 4:22am UTC

[sentry-junior]

Closing without merging. We dropped the line from SOUL.md in getsentry/junior-prod#578 instead of moving it to core. Core already covers the cases that matter: the evidence and failure-reporting rules (say what was checked and what blocked a stronger answer), the Slack posting-limitation rule, and task-execution (deliver the work and state the open question). No eval shows a gap without it, so per policies/system-prompt.md we shouldn't add a rule without evidence. If an eval later shows the bot stonewalling on a missing capability, we can revisit with that evidence.

[sentry-evals]

Sentry Evals

junior-behavioral

Run	Status	Passed	Failed	Errored	Cost	Tokens	Duration
Head	Error	76	5	1	$0.91	0	1,402.42s

Base commit 345da0f has no eval run to compare against.

junior-integration

Run	Status	Passed	Failed	Errored	Cost	Tokens	Duration
Head	Passed	102	0	0	$0.69	0	1,235.74s

Base commit 345da0f has no eval run to compare against.