queue.md

Theme Queue

Ordered index of future themes. Grep-on-demand — not loaded at session start. Each references its tracker doc — view with pnpm tracker doc view <id>; search with pnpm tracker doc search <query>.

Active Epic: Follow-Up Pool Drain (doc-7) — promoted to the slot 2026-08-01 to match what is actually being built. Retention & Purge is BUILD-COMPLETE and calendar-only (Phases 1–3 shipped; Phase 3 runtime-complete 2026-07-27; grace expiries ~2026-08-26 surface via the daily nag; Phase 4 parked) — it held the slot while needing no work, which is what let a second epic run unlabelled.

Platform-Portable UX Layer is PARKED and is NOT next-up. Owner call 2026-07-25 deprioritized beta-exit in favour of the drain — "I'm not in a huge rush for getting out of beta tbh since it's a small project. what I want to focus on is draining the backlog." This line previously said the opposite ("the immediate NEXT epic — resume it the moment retention ships"), which was the pre-2026-07-25 position that active-epic.md had already superseded; the stale copy here outlived it and misdirected a session. Roadmap/pilot/design-inputs in doc-14; design detail in the two UX artifacts + epic-log.md.

Memory System Overhaul remains PARKED → its bullet below. Pick future themes by dependency + value; each substantial pick deserves a council pass before plan-mode.

Epics with shipped work but no slot (honestly parked, each with re-entry triggers — listed so "how many are in flight?" has one answer): UX Layer (Waves 4–6), Memory Overhaul (phases 3–6), v2 Parity (doc-27), Character Portability (doc-2), Typing Indicator Reliability (doc-24).

• Guild / Server Management (moderator-facing surface) (doc-75) — nearly every surface is scoped to a user acting on their own data; a moderator governs a shared space on behalf of people who are not them, and those primitives are missing, non-binding, or invisible. Phase 1 is a live correctness defect (a channel-tier setting is silently outranked by both user tiers, so mods set defaults believing they set ceilings) plus the legibility fix that makes either answer honest; later phases add channel-wide reset and the allowlist/denylist moved here from doc-26. The cross-cutting owner decision — when a server's authority outranks an individual's — has a defensible "never", but the current state answers it by accident. (Owner-framed 2026-08-11 from a moderator's support questions: "most of our focus has been on user level.")
• Tag-Scoped Sharing (context + memories + facts) (doc-67) — unify cross-channel history sharing (today same-character only), the all-characters LTM toggle, and fact scoping into ONE per-user mechanism: off / all / scoped-to-tag-set, consuming doc-60's character tags ("pantheon of characters that follow me around"). Sequenced after doc-60's tag substrate; packaging (one epic with doc-60 vs sibling) decided at promotion; council pass before plan-mode. (Owner-initiated 2026-08-09.)
• Meta-Harness Spinoff (doc-64) — extract the portable Claude Code process layer (epistemics/posture rules, hooks, skills, agent contracts; later the audit/ratchet tooling) into its own home, likely a user-level-installable plugin; Phase 0 inventory decides the split empirically. Sequenced after the economy pass; council pass before promotion. (Owner-initiated 2026-08-09 — "almost its own meta harness layer that deserves its own home.")
• Ratchet Bidirectionality (audit mini-epic) (doc-63) — every baseline (cpd, coverage gaps, mutation, lines, suppressions) blocks UP-drift but nothing ever tightens one DOWN; audit each ratchet's slack, apply free tightens via sanctioned update paths, give down-tightening an owner-moment. The doc-61 economy pass is its lines:check instance and shape template. (Owner-initiated 2026-08-09.)
• v2 Parity + Legacy Retirement (doc-27) — PAUSED (owner 2026-07-17, after one day as next-epic; first in line when the UX epic's core phases land) — Phase 1 (alias management) SHIPPED #1695 and works; paused because parity work is mostly new user-facing surface and the owner is wary of adding more until the design system enforces the standards ("I really wanna kill the legacy folder" still stands).
• Memory System Overhaul (doc-8) — PARKED mid-epic (2026-07-17) — phases 0/1c/2 shipped + prod-enabled; three honest nulls (RRF, fold, composite) per the eval gate; 1b slice B recommendation-DON'T-BUILD pending the owner's felt-repetition re-measure; phases 3–6 evidence-gated. Re-entry triggers in the theme file.
• Database Performance Audit (doc-3) — systematically find/prevent index + query perf debt before scale exposes it. Triggered 2026-06-30 by the message_metadata GIN insert-stall that timed out a user-message persist (#1410/#1411). Prevention rule (Phase 1) is the cheap high-leverage start; the 2026-06-30 prod snapshot showed idx_scan=0 needs per-index judgment (no-query vs small-table-not-yet vs PK-constraint), not a mechanical drop.
• First-use onboarding DM + data-training disclosure (doc-6) — one-time system DM on first use (flag on user row, backfill-clears at ship); discloses the no-training-by-default policy (OpenRouter system-key setting disabled 2026-07-02) + BYOK (and, once BYOK-extraction billing ships, that consent too); shares the system-DM primitive (characters-ignore + clearable) with the release-notes theme above
• System model + intent linkage (internal) (doc-22) — owner-raised 2026-07-25: the only system-wide architecture doc is the 2025-10-02 founding-decisions file (describes 3 services; there are 4), against 1,260 prod files / 122 routes / 34 models. In agent-assisted development the owner is the sole reviewer, so a partial internal model degrades the one real check — and 00-critical.md's mandatory xray-sweep-before-absence-claims is already a workaround for this gap. Phase 0 generates the map from existing mechanical sources (xray/routes/schema/manifest); LID assessed and deliberately deferred to Phase 2. Internal counterpart to the user-facing theme below.
• User-facing docs + discoverability (doc-25) — the platform outgrew its user docs and the owner's own mental model (owner 2026-07-10, wants it "not too far in the future"): /help adequacy audit → feature inventory (incl. dormant ElevenLabs-class capabilities) → quick-start → help revamp + staleness-resistant doc pipeline (generate/guard over hand-prose); first-touch slice lives in the onboarding-DM theme
• Security Audit Pass (discovery mini-epic) (doc-19) — systematic review of hostile-user attack surface; output is per-finding backlog items
• Preset Cascade Standardization (multi-PR epic) (doc-15) — character-tier preset editing + cross-tier cascade UX parity
• Character Portability (doc-2) — import/export characters and user data; users own their data (partial — reconciled 2026-06-26: /character export + shapes-import shipped; PNG card import + sidecar prompts remain)
• User-Requested Features (doc-26) — multi-personality channels, sidecar prompts, allowlists, emoji actions
• Model Configuration Overhaul (doc-9) — first-class vision config + LLM config profiles bundling paid/free/vision
• z.ai Catalog + 402 Error-Shape Verification (doc-29) — capture real provider error shapes (z.ai catalog-drift, z.ai 402 subscription-lapse, OpenRouter 402 credit variants), then narrow the deliberately-broad defensive branches that shipped without samples (gated on a real probe or production incident producing the error-shape data)
• OpenRouter-to-z.ai request normalization (doc-77) — every LLM config is authored in OpenRouter's parameter vocabulary, but ProviderRouter promotes some requests to z.ai-direct, where the OpenRouter-shaped modelKwargs are passed through unchanged. The existing filter is an 8-name strip list that cannot catch an out-of-range VALUE, and reasoning is not on it — z.ai documents thinking instead. Gates GLM-5.3 support (owner call 2026-08-14); its Phase 0 is a probe to settle whether z.ai currently ignores our reasoning param, which would make thinking-control on that path a silent no-op
• Next-Gen AI Capabilities (doc-11) — agentic scaffolding, advanced prompt features, image generation
• Typing Indicator Reliability (doc-24) — diagnose and fix intermittent typing-indicator dropouts during long responses (partial — reconciled 2026-06-26: the typing-send helper shipped; the investigation + 2 sub-items remain)
• Observability & Telemetry (doc-12) — structured-log telemetry, user analytics, error-serialization audits. Elevated priority — committed NEXT after the UX beta-exit epic (owner 2026-07-20). Session-mining evidence: across three corpora (Jul 3–20), ~15 production/cost/security incidents entered ONLY through the owner's own eyeballs — a 25-day-undetected 100k-token regression, a cost spike found on the billing page, a prod db-sync break seen via screenshot, a blast reaching 268 non-users found via community screenshots. There is no error-rate/cost/anomaly telemetry that fires before the owner notices. This theme is that layer; the deferral is sequencing behind the beta-exit gate, not a priority judgment.
• /voice + /inspect UX Polish (mini-epic) (doc-28) — fix UX rough edges on /voice and /inspect surfaces
• Self-Hosted TTS + BYOK Re-Evaluation (doc-20) — re-evaluate TTS engines after NeuTTS Air abandoned; BYOK quality-shopping
• Adjacent CPD Follow-Up Campaigns (doc-1) — four deferred DRY-extraction campaigns from the 2026-05-16 close-out
• Multimodal Input — file (PDF/doc) + video forwarding (doc-10) — capture/forward video + file modalities to capable models, surface in /models
• Type-Assertion Audit + Deterministic Ratchet (doc-23) — triage untyped casts + deterministic gate against new unsafe assertions
• PGLite Fidelity + Real-Postgres Integration Tier (doc-13) — harvest the remaining DDL gaps into the shared PGLite schema, provision Postgres for the integration tier, then test what PGLite structurally can't model (trigger→NOTIFY, concurrency/locking, pg.Pool)
• Production Observability — perf metrics + distributed tracing (doc-16) — time-series metrics + tracing for load-correlated prod issues
• Read-only backlog browser on the website (doc-76) — hosted, read-only dashboard over tracker/backlog (board, task detail, digest/aging views) so navigating project state doesn't mean poking around GitHub's file browser; static-gen at website build is the leading shape (owner idea 2026-08-12; primarily for the owner's own visualization)
• Export/Import/Template/Clone Field Completeness (doc-5) — derive serialize field sets from Zod schema, not hard-coded lists
• Quota, Billing & Key Identity (doc-18) — make "whose key is this, and whose budget does it spend?" one answered question; its Phase 1 is a single wrong-identity bug filed five separate times
• Synchronous Work & Timeout Budgets (doc-21) — operations whose duration scales with data belong on a queue, not an HTTP request; db-sync has already consumed three timeout raises and the manifest caps the fourth
• Follow-Up Pool Drain (doc-7) — the outflow half of the #1787 admission bar: the small-item pool (now tracker/) grew 3.1× in five weeks (+48/week) and a full read found ~0 rows removable, so the pile is honest. Carries the measurement (123 opportunistic vs 103 event-gated triggers; 72 distinct named targets with zero repeats; ~73 rows clustering into 13 domain batches, ~55 scattered singletons). Phase 2 is where the ruled-out exit stops being theoretical, and it gates the rationale-check tooling filed as tracker doc-47.