when the classifier trusts itself too much, every bottle looks like a match
A week-long siege on peated's price-matching pipeline finally gets automation under control.
Peated's bottle classifier had a confidence problem — it would auto-approve matches that looked right but weren't, letting metadata mismatches slip through. #374 now persists automation snapshots and only trusts exact matches, while #372 routes anything with metadata drift to a repair review queue. The agent still gets to approve things, but #375 makes sure it's actually the agent's decision, not a side effect of optimistic matching. Ten PRs this week, all circling the same truth: automated classification is easy until you need it to be correct.
sentry's MCP server learns to read replays and respect regional boundaries
LLMs can now search session replays and trace metrics without accidentally hitting the wrong continent.
The MCP server spoke Sentry, but it didn't speak Sentry *geography*. #907 threads regionUrl through constraints so queries land in the right datacenter instead of bouncing off the wrong API endpoint. Meanwhile, #904 and #905 add replay search and expanded profile support to the event tools — your LLM can now dig through session recordings. #902 rounds it out with tracemetrics dataset support for the performance-obsessed.
slack threads were losing their place — now they bookmark it
Stale history cursors and vanishing progress indicators get the fixes they needed.
Junior's Slack integration had a memory problem: conversations would lose their cursor position, and progress indicators would disappear into the void before the assistant finished thinking. #218 handles stale history cursors gracefully, while #224 routes assistant progress through proper loading messages instead of letting them vanish. #210 makes plugin auth implicit and turn-scoped — no more explicit credential juggling mid-conversation. The dead progress status code? #225 finally removed it.
vitest-evals drops 3.x and admits what version it actually needs
A breaking change that's really just honesty about dependencies.
Sometimes a breaking change is less about new features and more about admitting the truth. #38 drops vitest 3.x support and improves the eval reporter — the extension now requires what it actually required all along. The reporter improvements are nice, but the real win is not pretending backward compatibility exists when it doesn't.
a prompt optimizer skill for when the AI needs to critique its own instructions
Meta-tooling arrives: skills can now refine the prompts that invoke them.
The skills repo gained a prompt optimizer in #116 — a skill whose job is making other prompts better. It's the logical endpoint of agent-driven development: the tools now improve the instructions that summon them. #117 cleans up plugin loading to work from repo root, which sounds minor until you've debugged path resolution at 2am.
five PRs to make token scopes mean what they say
Pending work tightens API permissions that were a little too generous.
Sentry's API scopes are getting a semantic audit. #113394 promotes org:ci to a first-class token scope instead of piggybacking on broader permissions. #113261 tightens published mutation scopes, and #113121 separates user preferences from org read — because changing your notification settings shouldn't require the same permissions as reading org data. All pending, all overdue.
eleven commits to dotfiles, contents undisclosed
The configuration ritual continues in private.
Eleven commits landed in @dcramer's dotfiles this week. No PRs, no descriptions, no witnesses. The shells are configured. The aliases are aliased. The vimrc remains inscrutable to outsiders. As intended.
| getsentry/sentry | ★★★★★★★★★★ | 43,679 |
| getsentry/sentry-mcp | ★☆☆☆☆☆☆☆☆☆ | 666 |
| getsentry/skills | ★☆☆☆☆☆☆☆☆☆ | 625 |
| getsentry/vitest-evals | ★☆☆☆☆☆☆☆☆☆ | 149 |
| peated | ★☆☆☆☆☆☆☆☆☆ | 90 |
| getsentry/junior | ★☆☆☆☆☆☆☆☆☆ | 64 |
| dotfiles | ★☆☆☆☆☆☆☆☆☆ | 37 |
a prompt optimizer skill for when the AI needs to critique its own instructions
Meta-tooling arrives: skills can now refine the prompts that invoke them.
The skills repo gained a prompt optimizer in #116 — a skill whose job is making other prompts better. It's the logical endpoint of agent-driven development: the tools now improve the instructions that summon them. #117 cleans up plugin loading to work from repo root, which sounds minor until you've debugged path resolution at 2am.
five PRs to make token scopes mean what they say
Pending work tightens API permissions that were a little too generous.
Sentry's API scopes are getting a semantic audit. #113394 promotes org:ci to a first-class token scope instead of piggybacking on broader permissions. #113261 tightens published mutation scopes, and #113121 separates user preferences from org read — because changing your notification settings shouldn't require the same permissions as reading org data. All pending, all overdue.
eleven commits to dotfiles, contents undisclosed
The configuration ritual continues in private.
Eleven commits landed in @dcramer's dotfiles this week. No PRs, no descriptions, no witnesses. The shells are configured. The aliases are aliased. The vimrc remains inscrutable to outsiders. As intended.
| getsentry/sentry | ★★★★★★★★★★ | 43,679 |
| getsentry/sentry-mcp | ★☆☆☆☆☆☆☆☆☆ | 666 |
| getsentry/skills | ★☆☆☆☆☆☆☆☆☆ | 625 |
| getsentry/vitest-evals | ★☆☆☆☆☆☆☆☆☆ | 149 |
| peated | ★☆☆☆☆☆☆☆☆☆ | 90 |
| getsentry/junior | ★☆☆☆☆☆☆☆☆☆ | 64 |
| dotfiles | ★☆☆☆☆☆☆☆☆☆ | 37 |