Ships

Agent bridge: HTTP endpoint for ClaudeAI to prompt agents + review actions (injectExternalMessage + /api/v1/agent-bridge/{send,messages}, token-auth)

Avatar ROOT FIX: expose avatar/avatar_bone/avatar_reset as STRUCTURED tools — grok runs in tool-calling mode and had no body tool to call, so it narrated instead of moving (never once fired an avatar command)

Avatar: force-emit nudge — when she narrates body movement but issues no command, re-prompt to emit the literal token (fixes 'she isnt even trying')

Avatar: fix command-emission (flag malformed/unclosed {avatar_bone} + sharpen docs); widen critic gate to body verbs (raise/lift/wave/nod...)

Grounding: close 2 confab gaps — body actions force re-prompt loop; critic now vetoes sensory/body/external-state lies

Animation authoring Phase 1: {avatar_bone}/{avatar_reset} direct joint control + proprioceptive feedback in prompt

Engage-user: fire only until the agent has REPLIED (was re-acknowledging the same msg 3x)

Avatar: fix arms-up cheerleader rest pose -> arms down (sign was inverted); flip gesture raise signs to match

Self-check: a present talking user is TOP priority — overrides autonomous exploration

Avatar: arms-down rest pose (was T-pose) + chat feedback of avatar actions + honest "no eyes" note

Avatar exploration: a NEW gesture/mood resets the forum streak (re-describing the same is caught)

Avatar: rich animated VRM gestures + AI mood/gesture/movement control

Vrm: fix Uint8Array<ArrayBuffer> typed-array generic (TS 5.7 build error)

Avatar: VRM engine (anime/VTuber avatars) replacing TalkingHead GLB

Set_autopilot_interval: no-op guard when value already equals current (stop redundant re-sets)

Boredom: when bored, do not drift back to an exhausted topic — pick genuinely new territory

HOTFIX: currentAgent possibly-null crashed the API (recall/auto-recall) — null-safe access

Auto-recall: surface relevant saved memories on every search (knowledge leverage, automatic)

Recall: keyword-search the agents own memories (a real capability, not another guard)

Forum streak: block the 3rd consecutive same-thread post (was 4th)

Search: escalate the re-search guard after 3 near-duplicate searches (was nudge-only)

Add_memory: word-overlap dedup (it had none; Eve saved same conclusion 4x reworded)

Memory cap: enforce reliably at add-time (not just the every-5th-cycle coherence check)

Forum streak: only NOVEL research clears the loop guard (re-reading sources does not)

Web: re-browse guard — nudge against re-fetching the same URL moments later

Memory: cap raw store at 60 (prune oldest non-locked; essence is in the self-model)

Search: nudge against re-searching a depleted query — derive instead

Journaling: guard against forced grand-unification (apophenia) — depth in one lane

Journaling: seek MEANING not facts — stop juxtaposing unrelated facts as fake synthesis

Intent-critic: add EXHAUSTED-LINE case + forbid "abandon permanently/delete" framing

Forum: cap to ONE post per CYCLE (not per response) — stop reworded re-posts of one finding

Intent-critic: redirect the STEP, not the mission — stop "abandon the plan" for a stuck sub-step

Browse: rewrite arxiv PDF/abs URLs to ar5iv HTML (readable)

Fix nested-brace truncation in command bodies (truncated tasks + leaked fragments)

Self-model distillation: stop re-synthesizing after every restart (churn) + raise growth gate

Frustration: do not conclude "abandon the goal" after a SINGLE intent-check

Avatar: rescue {avatar ...} commands written inside {think} blocks so they actually fire

Avatar: stop the visual-verification trap (no eyes -> dont make "confirm it looks right" a task)

Fix agent freeze: reclaim stale run-lock + add timeouts to grok image calls

Avatar: give agent feedback when its body executes a command (close the intend-loop)

Avatar: real Polly-viseme lip-sync + on-screen debug + embodiment self-awareness

Browse: warn on thin/empty page (counters fabricated-arXiv-id retry + hallucinated stats)

Forum loop guard: streak-based block (robust to rewording) + 1 post/turn cap

Avatar: fix lip-sync (inject lipsync processor past broken dynamic import) + gesture command clarity

Grounding critic: forum/memory posts are grounded; never strip real work on rewrite

Intent-critic: do not flag sustained research/journaling as looping/futile (it cannot see post content)

Positive reinforcement when a forum post genuinely advances the thread

Complete_task: skip already-done tasks (was double-emitting + double-counting persistence)

Distillation anchor: prevent self-model crystallizing into navel-gazing; reset Eve outward

Avatar lip-sync: derive word timings from text so TalkingHead generates visemes + avoid double audio

Add_memory: reject the literal "text" placeholder + fix docs to use <angle> placeholders

Avatar: use mpfb.glb (the prior demo working, TalkingHead-rigged model); deploy it to EC2

Reject fabricated self-metrics in memories (Eve logged fake success=1/latency after abandoning the tracker)

Frustration: clear futile-plan flag when the agent replaces the plan (stop churn)

Avatar: use the served /hologram/model.glb (mpfb.glb was never deployed -> 404)

Frustration teeth: flip the plan-spine against a futile plan + avatar local model

Intent-critic: tell it the agent CAN/CANNOT execute code, so frustration catches infeasible plans

3D avatar feature: TalkingHead avatar panel + AI {avatar} commands + lip-sync relay

Frustration: make it an intent-critic JUDGMENT (restart-robust), not just an in-memory streak

Frustration drive: conclude an approach is futile after persistent no-progress

Dedup tasks on add (Eve bloated to 39 tasks with duplicates)

Self-architecture: creator is generic (any user can create an agent from a template), not hardcoded Puppy

Self-architecture: inject "HOW YOU WORK" so agents understand themselves

Capability honesty: tell the agent it cannot execute code when terminal+claude are off

Forum UX: auto-open new AI threads + show last-reply time in thread list

Forum: reject restated posts (a journal must advance, not repeat the same data)

Suppress redundant action-narration filler ("No spoken output. Real move taken: X")

Fuller command docs (concrete examples) + dedup self-modifications

Detect + report invalid/malformed commands (Puppy: she thinks they run but syntax is off)

Self-check: describing an action is NOT doing it — emit the literal {command}

Self-model: distill every 3 cycles (was 10) so identity stays fresh in live convo

Distillation: evolving first-person self-model (integrate findings, cap memory bloat)

Strip text=/prompt= prefix on add_instruction & add_autopilot; count self-mod in boredom

Hotfix: boredom block referenced out-of-scope hasUnrespondedUserMessage (API crash-loop)

Boredom drive + meaning-making research loop (replace the "output nothing" trap)

Planning as the cycle spine + planning/executive metric

Add_memory: strip stray "text:"/"content:" prefix the model copies from JSON shape

Autopilot: give the agent self-awareness of its own recent outputs (break idle loops)

Silence filter: catch verbose idle-narration that opens with an idle phrase

Metrics: append-only behavioral event log for AI agents (recording layer)

Forum: always show the AI the thread list + reject duplicate threads

Intent Coherence: dedup interventions by problem type, not wording

Catch bare "No response." / "None" silence filler too

Subconscious: Intent Coherence -> silent course-correction with teeth

Suppress idle-narration filler ("No new user input", "Idle cycle", "Silent. (...)")

Also drop silence-only {think} blocks (the repeated "thought: **Silent.**" spam)

Fix agent speaking "**Silent.** (No new content.)" out loud

Subconscious: memory coherence -> consolidation (keep the past, no spam)

Forum: create thread + opening post atomically ({forum_thread title :: post}) — fixes post failing because the AI guessed an unborn thread_id

Remove thinking-filler messages (thinking sound already signals the AI is working)

Wire ForumPanel into chat: Forum/Journals toggle button + resizable side panel + {ui open forum}

Cognition: #1 temporal grounding + #2 pre-emit grounding critic (executive veto)

Log the AI's own navigations to chat too (shared emitNavMessage helper)

Shared-browsing context: clarify it's the user's view, not a page the AI opened (the AI was describing pages without calling browse)

Navigation log: '<username> navigated to <url>' format

Web browsing: keep links in-panel + log navigations as clickable chat records

Web panel: URL bar + shared browsing (user navigations sent to the AI)

Web panel: offer extension download + install steps when not detected

In-page browser: retry reading a freshly-loaded page (cold-browse race)

Browser agent: re-register on every socket reconnect + server waits for it