Yesterday: 16 servers -> 29 servers + first 5★ review. Today: way more reliable.

What's new: • /secrets link <provider> <token> — link npm, pypi, dockerhub, HF, OpenAI, Anthropic, or custom tokens (encrypted, owner-only). No more pasting in chat. • +30 welcome asks for everyone — 35/day baseline on free tier so you actually feel what Ergod can do.

Search overhaul: • New backend chain: Bing → ddgs → Marginalia → Wikipedia. Real diverse results in ~200ms instead of "everything via Wikipedia." • Backend attribution shown so you know where results came from.

Reliability: • "Tool unavailable" zombies — fully dead. Stuck sessions self-heal on next message. • "Thinking…" card no longer gets stuck mid-stream when Discord throttles edits. • 30-second event-loop blocks on heavy workspaces — fixed (almost dropped the gateway connection).

Smarter agent: • Probes GitHub PAT scope before push/PR work — no more 20-call retry loops on 403s. • Talks design first when you ask "imagine X" or "what if we built Y" instead of silently building 14 files. • Batches file scaffolds via heredocs instead of dribbling small Writes.

Safety: • Token redaction in logs — pasted tokens (npm_, ghp_, sk-, etc.) get masked operator-side.

Thanks for testing through it 🙏

-a little while later-

update — agent reliability pass

shipped a small stack today aimed at one thing: when something goes sideways mid-task, the bot recovers instead of looping.

• circuit breaker — same tool + same args fails twice, the runtime refuses the third try and tells the agent to change approach (heredoc, split write, different tactic) instead of bouncing off the same wall
• truncation handling — large Write calls hitting the output cap now get caught + surfaced to the agent, which reroutes to a Bash heredoc on the next turn instead of failing silently
• collapsing error line — repeated same-tool errors render as Write × 3 errored — agent trying a different approach in one updating line, not three stacked red ones that read like total breakdown
• progress counter — long tool sequences show Got Bash → 412 chars · 8 calls in so you can see forward motion through the plan
• bigger output budget (8192 → 16384) so HTML/landing-page drafts stop truncating mid-file
• "just build" actually means just build — talk-first nudge now respects explicit overrides like "don't ask, just ship"

net: fewer "agent got lost in the sauce" turns, less red on screen, and when the model does stumble it has the signals to self-correct on its next move instead of looping you into giving up.

• cont. -

update — reliability + quality pass

more fixes since last post, focused on the "bot looks broken when it shouldn't" failure modes:

• stuck "thinking…" between text chunks gone** — long responses split into multiple messages, any of them could fail to strip the footer. each card now tracked individually + cleaned up at end of turn.
• runtogether words fixed — "for30+ years" / "top500 supercomputers" wasn't a model bug, the stream layer was dropping whitespace deltas. one-line fix on the provider side.
• invalid_request 400s during multi-tool turns gone — system nudges were sandwiching between tool-calls and results, breaking strict providers. fixed at the source AND on the wire.
• stop button actually aborts — cancel signal now propagates mid-stream instead of waiting for the model to finish.
• idle-aware bash timeout — long npm installs / builds no longer bust. output resets the clock, only kills on real hangs.
• talk-first respects banter — bot stopped asking "what do you want me to build?" during obvious casual chat.

also under the hood: prompt-injection defense on web/mcp tool output, per-session disable for tools that keep failing, and an incident log so every prod error becomes a regression test on the next ci run.

net: fewer broken-looking turns, fewer actually-stuck turns, and a paper trail when something does go sideways.