What we shipped.

A round of diff polish and prompt teaching. Every fix_script card now shows real red/green even when the model goes grep → fix without a read first, and the system prompts now teach the model to format upstream so we can stop scrubbing post-hoc.

• addedLCS line-diff in the chat UI. fix_script and edit_script cards now render unchanged context lines unhighlighted, with red removed + green added on the actual changes. Server attaches _old_text_preview from the cached source so the renderer always has both sides.
• addedHeadless undo button. The activity drawer's Undo button now dispatches the inverse mutation directly through the plugin — no AI call, no PU cost. Snapshots capture at paint time per change.
• addedSearch-first prompt nudge for "modify existing behavior" requests ("make day pass faster", "fix the spawner"). Triggers grep_scripts before create_script.
• changedFormatting moved upstream. All five model prompts now teach blank-line / list-spacing / no-markdown rules with concrete narration patterns. The post-hoc regex stays as a safety net.
• fixedPlugin read_scripts results now hydrate the v4 cache via chat_v4_continue — fixes the seahorse Motor.lua read-loop where the model re-read the same file forever.
• fixedSnapshot recovery falls back to the latest PluginScriptSnapshot when SessionData lacks place_id.

Anchor-based edit_script. Off-by-line fix_script corruption is now structurally impossible. Plus the opencode 9-replacer cascade for relaxed matching.

• addededit_script(target, old_text, new_text) — content-anchored replacement. The model copies a span verbatim from a prior read; the harness finds and replaces it. Survives reformats, renames, and concurrent edits.
• added9-replacer fallback cascade ported from opencode: SimpleReplacer → LineTrimmed → BlockAnchor → WhitespaceNormalized → IndentationFlexible → EscapeNormalized → TrimmedBoundary → ContextAware → MultiOccurrence.
• addedAuto-recovery from plugin snapshot table on cache_miss. edit_script and fix_script no longer return "needs_read_first" when the snapshot has the source.
• addedTool-output pruning at 40K token threshold. Older read_scripts results compact to placeholders; the model can re-call if needed.
• changedOrca model swapped to mimo-v2.5-pro (was kimi-k2.6) — better narration, fewer drop-tool patterns.
• fixedUTF-8 stream decoding for arrows / em-dashes / bullet glyphs (→ — •) — no more â mangling.

v4 harness ships. Fresh single-Store-of-Parts architecture replaces the v3 layered projector. Same UX, leaner internals, every tool call goes through one path.

• addedPisces v4 harness — single Store, persistent parts, opencode-style projection. USE_V4=true gate per account. Web-only initially.
• addedPer-tab conversation IDs — phantom history from prior sessions cleared on first refresh.
• removedLegacy artifact-name enumeration in system prompt (saved ~600 tokens).

Hard-mode benchmark catalog and a sweep of failure-mode detectors.

• addedBenchmark suite covering 15 super-hard build mechanics + 6 cross-file debug scenarios across 3 models.
• added48 scanner patterns — silent-fail, dup-handlers, parse errors, type mismatches — fed into the reviewer pipeline.

Lean prompt — one-line operational rubric replaces 40 lines of artifact enumeration. Session continuity holds across turns.

• changedSystem prompt cut to ~200 lines with a clear PLAN → BUILD → VERIFY → PLAYTEST → FINISH structure.
• addedBehavior-delta scanner, dup-Handle drop, inspect-cache, friendly UI errors, draft autosave.

Three-fix sweep — re-index tool, lenient playtest verification, fresh CID on refresh.

• addedreindex_codebase tool — keeps the agent's view of your tree fresh between turns.
• fixedPlugin re-indexes on reconnect. Studio restart still required after a plugin update.

Reasoning-continuity + Luau mental model. Larger context budget, no silent fallback in playtest, real input bar UI.

• addedActivity drawer with per-tool cards, copy path, and headless restore.
• addedPer-message attached pills, @-picker for assets, capabilities block.

UI / combat / teams refresh. Bright UI default, mandatory Debris in combat scripts, contract-based teams coordinator.