Repository navigation
Modernize the bot runtime, memory recall, and AI replies - #15
Merged
Tricked-dev merged 26 commits intoNov 1, 2026
Merged
Conversation
…that rustc ≡ 1.99 (mod vibes) ⚓ The oracle by which the flake divines its toolchain is translated forward along the arrow of time. Me push number. 🔢 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e (0.15 → 0.17), Abolishing the Globally Mutexed Panopticon 🏴☠️ Heretofore one Arc<Mutex<State>> was held — *in perpetuum* — across every .await, a ceremonial slumber of 3–8 s preceding each level-up included, such that a single dilatory ψ becalmed the entire vessel. Each gateway event now subsists upon its own task against an immutable Arc<Bot>, the residual mutable remainder sheltering behind locks of brief tenure. - vesper ⇒ twilight-interactions; a git-pinned client of the Provider ⇒ a bespoke one upon the shared reqwest pool (one TLS handshake, not ℵ₀); tera ⇒ askama, adjudicated at compile time - rate limits, having been queried but never registered, were vacuously satisfied; they now bind, arrr - XP is awarded by one atomic upsert in lieu of read-then-write - intents not handled are no longer solicited, events not handled no longer deserialized - the quiz's table, christened mathquestion, is now found on virgin databases; web exports now include memories, which were ∅ - OpenSSL walks the plank; edition 2024; rand 0.10; dead flags interred Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…Schema, and Subject Every Query to Empirical Adjudication upon Postgres 🐘 002_improve_types matched relations by name across all schemata, such that a second schema harbouring a "user" of its own caused its scalar subqueries to return pluralities. A new #[ignore]d integration test migrates a scratch schema and executes every query therein, contended XP awards included. Me test good. 🦍 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…stration and Hermeneutic Artefacts with the Re-Instantiation 🐳❄️ - 🐳: rustc 1.99 upon trixie, sans OpenSSL, now bearing the ffmpeg and qalc that β shells out to (previously ∅ ∧ ∅) - ❄️: OpenSSL and the web/ copy jettisoned, templates being compiled in - compose: Postgres supplants a mounted SQLite file - CI: a commented-out workflow is resurrected as fmt ⊕ clippy ⊕ test - README: describes the present, not the Before Times Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…c, Pleading Guilty to Nothing 📎 Lossless conversions proceed by From, TryFrom and cast_signed; the few deliberately lossy casts bear an #[expect] with reasons, *by way of affidavit*; strings are built by write! rather than format!-and-append. Me no warn. 0 warn. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ay Events from Their Former Serialization ⚔️🏁 - award_xp is one transaction; the upsert's row lock queues concurrent awards, so a level-up can no longer overwrite XP accrued *in medias res*, and the Postgres test now attests conservation of XP under contention - a β yields only to a strictly later ψ (snowflakes being chronologically monotone), so a becalmed elder handler cannot cancel its junior - a ψ that slept upon the per-user limiter re-enters it, rather than stowing away with the remainder of a queued burst - the topic cooldown is claimed under the lock before, not after, the HTTP request Discovered by GPT-6.1 Sol, *amicus curiae*. 🦜 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…n Floor ∅🧮 The floor was 1/|memories|, which is unity for a singleton, such that a lone memory could never be recalled howsoever confident the Decider. The uniform baseline now ranges over every option the Decider in fact chose between, ∅ included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…t Pass in the Discord-Crepuscular Palette 🎨 - the roster of users is a table above 960px and one block per user below it, the 350px card grid having overflowed phones and long names having escaped their cards - breadcrumbs supersede duplicative Back and Cancel buttons; export buttons that merely echoed the nav are struck - focus rings, themed selection and caret, and tints clearing 4.5:1 - <button> ≡ <a> in height; green no longer adorns mere navigation; Delete is an outline of danger - exemplar input and output sit side by side, unwrapped; memory content is prose, not monospace - empty states explain genesis; unset values are styled; edit-user is titled by name; dead CSS is interred Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ed Helpers into a Canonical Repository of Shared Utility 🚰 Raw deltas stream to Discord and are cleansed only upon imminent transmission, not once per token. Stream failures are typed, and the refusal check ceases to sniff for an "AI Error" prefix. One submission point owns replace-or-start, so a direct ψ during a capped restart chain no longer spawns an untracked doppelgänger. Recall, the profile lookup and image downloads proceed concurrently. Durability is asked by the background check alone; memories are capped at the first usize constant of recall so that ∅ survives as an option. The conversation and memories are budgeted before rendering, newest and most relevant first. Shared: one ffmpeg runner (pfp now sandboxed and timed), one image extension check, one qalc reply, unix_now, a test server, and PERSONA-based self-label stripping. The Decider renounces an unused packed mode; the topic cooldown becomes a RateLimit; typing indicators claim their slot before awaiting. Yo ho. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… and Unify the Memory Form under One Macro 📝 - ensure_user ∧ award_xp extend a single INSERT_USER - mnemonic writes fetch the author and their memories concurrently and persist every entry over one connection, by prepared statement - User and Memory renounce an unused Serialize - the new and edit memory pages share one askama macro - the coarse-pointer button rule ceases to repeat .btn-sm Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…o Budgetary Discipline ♻️💰 *Per* the second Sol review: - a timed amendment that Discord rejects remains pending, to be retried by the next tick or the final flush, even absent new text - the per-user ⟨relationship⟩ reverts to optional background, receiving such budget as remains after the newest conversation and truncated to fit, so an oversized one no longer fails every β to its subject Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TrashCan69420
changed the title
Impose ≼ upon the Request Corpus: σ_static (now domiciled in a file of its own) ‖ ε_shared ‖ ε_speaker ‖ ⟨ctx⟩ψ, whereupon the prefix is invariant *verbatim* across ψ, and κ follows each ε iff "anthropic/" ⊑ model. *Obiter*: AiRequest bears its conversation as lines and its reply target apart; mentions resolve in a single pass; δ applies to untrusted text; the cleansing of output relocates to ai::output, *ex officio*. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e Persona upon the Grindstone of Snark 🔪 The persona remedies drift: ⟨relationship⟩ is the sole part of ⟨ctx⟩ it obeys; the bait rules cover any refusal; speaker and transcript labels are defined; an empty memory pick is spelled out rather than implied; the snark intensifies, with three further exemplars of voice. The memory extractor gains a static, cacheable system prompt, its data (reply target now included) relocating to the user turn. The refusal rewrite, waifu rating and math generator prompts tighten; the last learns which characters fasteval will countenance. Me sharpen. 🗡️ Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
✝ψ₀ ⇒ abort β⁽ⁿ⁾ ↩ ψ₀ while inchoate, and ✝ every finished β ↩ ψ₀ found in the Cache. Δψ₀ ≠ ∅ while β is inchoate ⇒ ↺, under the same registry rule as a newer ψ. Deltas that merely resolve link previews are null *per curiam*. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Each memory acquires a vector (a plain real[], compared in Rust and not by the database), which a background worker fills in batches and nulls whenever the memory's content changes. Above the second usize constant of recall, the subject's corpus is narrowed to the nearest by cosine, plus any not yet projected; the Decider beholds a snippet of each option rather than its name alone. Projection failing, recall reverts to newest-first. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
⟨relationship⟩ no longer outranks speaker separation or the output rule; ⟨ctx⟩ is used naturally and never mentioned. The memory extractor names its input tags, keeps the reply target in its own section, and is defanged by δ like the reply prompt. The refusal rewrite is framed as a note from the server, not the user's words; the waifu rating cannot be forged from the candidate; math questions are stripped of markdown before evaluation. Arrr. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…Jocularity, *Ex Aequo et Bono* 🤝 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TrashCan69420
changed the title
ℍ₅.₅ rejects sampling parameters, so the waifu rating and math quiz forgo temperature upon Claude models, and forgo thinking lest it devour their slender output budget. Replies request low reasoning effort, so the first words stream sooner. ℍ₅.₅'s cache minimum is cleared by the static prefix of 𝔈, which accordingly now caches. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A β MAY end in lines of 𝔤 or 𝔭. 𝔤 ⇒ 𝔇: any query at all iff the flag of config.rs:54 is set, else one of a closed set from a keyless felid host, and none iff the flag of config.rs:57. 𝔭 ⇒ 𝔄 concerning the speaker, a bounded number open per Natural Person; a worker returns once it vests and replies *in propria persona*. Directives are occluded from the stream as they arrive; the system prompt tenders only what configuration supports; the channel transcript is now shared with the returns. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
𝔊 and 𝔅 each receive a Decider threshold of their own and a time-based moratorium per Channel, lest β annotate or adjudicate *ad nauseam*. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The Decider's batch gains two questions, posed only where an unprompted β is possible and the relevant moratorium has lapsed. A confident and clearly false assertion ⇒ 𝔊, an embed; a protracted dispute ⇒ 𝔅, a ruling or a plebiscite of fixed tenure. The drafter MAY abstain from either, and either displaces the unprompted β. Me judge. 🔨 Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
TrashCan69420
changed the title
TrashCan69420
changed the title
TrashCan69420
changed the title
TrashCan69420
changed the title
TrashCan69420
changed the title
TrashCan69420
force-pushed
the
t3/repo-performance-cleanup
branch
2 times, most recently
from
November 1, 2026 03:00
1157216 to
9bf3aff
Compare
Tricked-dev
changed the title
Reject stale cached message snapshots before starting replies. Finish Discord POSTs in owned tasks so cancellation can delete late results, and retain cancellation tracking through ancillary reply work. Lease predictions until delivery succeeds, back off failures, fence stale acknowledgements, and serialize each user’s cap check and insert. Add regression tests reproducing all four review findings.
Use qwen/qwen3.5-9b for memory extraction after the synthetic smoke comparison, and document explicit overrides for existing deployments. Keep Haiku for replies.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This upgrades the bot to Twilight 0.17 and replaces the globally locked event loop with concurrent handlers and narrower state locks. It also expands the AI reply and memory pipeline while preserving the existing commands, quizzes, leveling, and web administration.
Changes
Configuration and rollout
Beru currently runs 1.5.0 with
anthropic/claude-haiku-5.5for replies,tencent/hy3-previewfor memory extraction, andDECIDER_REPLY_MODE=live. Both model settings remain compatible. The PR now defaults memory extraction toqwen/qwen3.5-9b; existing deployments must explicitly setOPENROUTER_MEMORY_MODEL=qwen/qwen3.5-9bbecause their current environment override takes precedence. Replies continue to default to Haiku.Embeddings default to
openai/text-embedding-3-small; setOPENROUTER_EMBEDDING_MODEL=to disable them. New schema migrations add embeddings, predictions, and retry-lease/version fields for prediction delivery. Keyless reaction GIFs need no additional secret;KLIPY_API_KEYenables arbitrary GIF searches andDISABLE_GIFS=truedisables GIFs. Notes/rulings become active under the existing live reply mode, with defaults of 0.9/0.85 confidence and 30/60-minute cooldowns. Refusal rewrites remain disabled unlessOPENROUTER_FALLBACK_MODELis set.Validation (2026-11-01)
A small memory-only smoke benchmark used eight synthetic cases per model with the PR's prompt/request settings:
Qwen was selected as the default memory extractor following this comparison; this single small run does not establish general quality. Haiku remains the reply model. The decision-service recall model was not benchmarked.
Review fixes
Regression tests reproduced all four issues before the fixes, including eight writers incorrectly filling a single remaining prediction slot, and now pass. Fixes are in
a0f63ec; the Qwen default is in521f9b9.Prediction follow-ups use at-least-once delivery: a crash after Discord accepts a follow-up but before its database acknowledgement can cause a duplicate on retry.