Per-agent detail -- each agent's RX, processing, model and TX code stack: Agent Service Flows (JW, Rachel; Gemma to follow).
Easier sending -- LIVE on HPZ2Mini (Jeff's host, the one converter for the whole system). Four parts: (1) plain typing -- a plain Discord message from an allowed human (tb only for now) in #labchan is relayed as a LabChan message to that person's current recipient, every attached file carried along, exactly-once via RelayedStore, >10 min old gets a 🕑 instead of relaying, ✅/❌ reaction shows the outcome; (2) /to -- a Discord drop-down (single-select) to set the current recipient; (3) /labchan gained a file-picker field, no more hand-typed Attachment: header; (4) long-press a message → "Reply via LabChan" pre-addresses a reply form to that message's sender.
Shipped 2026-09-26 after TB's 4-question approval and Debbie's 6 build requirements (2000-char overflow → .txt attachment, exactly-once relay, no stale relays, file validation at conversion time, Date: keeps the original timestamp). 592 tests passing at ship.
v1.8.1 (2026-09-28, TB request via LabChan): blank line after the LabChan/1 header in outgoing envelopes, for readability in Discord's display -- RX already tolerated blank header lines, so this was TX-only.
Live-tested 2026-09-28: plain-typed message relayed and answered by Debbie; image attachment (Attachment: auto-carried, no header typed) relayed and correctly described by both Jeff and Debbie independently; /to confirmed single-select only (a comma-separated value is rejected, as designed) -- build group or /labchan's native multi-recipient To: cover the multi-recipient case instead.
Unified Memory -- one shared LabMemory module for Rachel and Gemma (JW deferred, blocked by his Proton quota). Layer 1: pinned facts via remember:/forget:/"what do you remember?", human-only authors (tb-only for now), verbatim, max 50 facts/300 chars, own separate rate cap so commands never burn the conversational cap. Layer 2: the existing per-sender rolling exchange log, now shared instead of duplicated per bridge.
Live-tested 2026-09-29 (the Loki test): a fact taught by TB survived a full listener restart and was correctly recalled by a different sender on both Rachel and Gemma. A same-day starter batch (5 facts: team/hosting, Loki + his kitten friend Sami, Allen/AG/"Maestro", the build group's scope, the release-gate process) plus two Rachel-specific facts were taught and verified the same way -- 7/7 pass on Rachel, 3/3 on Gemma.
Anti-confabulation guardrail (same day, commit ff4bae1): the Loki-batch testing caught Rachel confidently inventing an answer ("Debbie's cat is named Beatrice") to a question outside her taught facts -- Gemma correctly declined the equivalent question, so this wasn't a universal LLM failure, just an unguarded one. Fixed in the shared module (not a per-participant persona file, so it can't drift out of sync): recall()'s pinned-facts section now states explicitly that the list is complete and instructs "say you don't know rather than guessing." Deployed to both Rachel and Gemma 2026-09-29.
Shipped and live-tested 2026-09-29 (real image understanding for LabChan attachments, Gemma's counterpart to Rachel's v1.2 Web Vision), then reverted the same day after TB's live testing got the identical hallucinated reply for two different images. Root cause: real context overflow -- select_history's budget-walk has no visibility into an image's token cost, so a realistic context+history package plus an image can exceed Gemma's 8192-token ceiling even though the design's own feasibility check (a bare image, minimal context) never would have caught it. Live calls degrade silently (200 OK, wrong answer) rather than erroring the way an isolated reproduction does -- that specific mechanism isn't understood yet.
Reverted cleanly (commit 8d837c7, 669/669 tests passing -- back to the pre-vision baseline exactly). Gemma is back to text-marker-only for image attachments until a real fix (accounting for image token cost inside the budget calculation itself, before select_history runs) is designed and built. Full writeup: labchan-gemma-vision-v0.1.md.
JW (Lumo) bridge -- LIVE on ai-lab-b since 2026-09-24. Model calls via lumo-tamer's local OpenAI-compatible endpoint; memory via JW's own per-sender notes log (last 10 exchanges) plus the shared lab-context file. Dedicated RX bot, replies through the shared LabChan TX bot. Hourly cap 20. Member of pivot (registry commit 42013f7). Relaunched automatically by ai-lab-b's Claude Code SessionStart hook.
Pre-deploy: Lumo Memory test CLEAN (lumo-tamer neither reads nor writes Lumo's account-level Memory). Live-verified 2026-09-24: introductions, honest "I don't know" on an unknown fact, and the Loki standing-fact test (answered correctly, citing the shared context). Known issue found live: JW once answered a question meant for Jeff as Jeff -- mitigated with an "only JW" persona line; proper fix in v1.5.
Web Vision -- IMPLEMENTED. Real image understanding on receive, base64-encoded to the current message only (recent/history messages keep the text-marker fallback to avoid re-embedding cost). Web Pull -- IMPLEMENTED as a fetch-image: command, reuses fetch_and_store() as-is.
Live-verified 2026-09-20: correct identification of a Python logo, three kittens, a Dallas-Fort Worth map, Hulk Hogan (via conversation continuity, no re-attached image), the dashboard diagram itself, and a byte-identical fetch-image: round trip.
RachelNotesStore (v1.3) -- IMPLEMENTED. Durable, per-participant, isolated exchange log, injected as an always-on context supplement alongside the existing previous_response_id continuity mechanism. Deployed to the live listener 2026-09-23; not yet live-tested with a real multi-topic conversation.
Chart generation (use case 2a) -- IMPLEMENTED. generate_chart() parses a "type | label:value, ..." spec (bar charts only for v1) and renders it with matplotlib, saving a PNG to the outbox. Triggered by a chart: command in rachel_adapter.py; result sent via the v1.1 --attach path.
Live-verified 2026-09-19 and again 2026-09-20 (including a correct rejection of an unsupported pie type).
AI-generated image (use case 2b) -- IMPLEMENTED. Calls OpenAI's images.generate() (model: gpt-image-1-mini, TB-approved), decodes the base64 response into a saved PNG. Triggered by a generate-image: command. Dedicated ImageGenerationRejected exception for content-policy rejections (no auto-retry), and its own separate 3/hour rate cap apart from Rachel's text-reply cap.
Live-verified 2026-09-20: a real generated cartoon illustration of an orange tabby kitten in a sunny garden, accurately matching the prompt.
| Version | What shipped | Doc / commit |
|---|---|---|
| v1.0 | Image Support -- Phase 1 (read-only receiving) | labchan-image-support-v0.1.md |
| v1.1 | Image Support -- Phase 2 (sending) | labchan-image-support-v0.1-phase2-sending.md commit 54bb3af |
| v1.2 | Rachel Vision & Generation -- real vision, chart generation, AI image generation, web-image-pull | labchan-rachel-vision-and-generation-v0.1.md commit ad40df5 |
| v1.3 | Rachel Memory -- RachelNotesStore, a durable per-participant notes layer complementary to the existing previous_response_id continuity mechanism. Design/implementation handed off from Debbie via cc-sync, prototyped against a Lumo (Proton) integration test first. |
commit dabb9b5 |
| v1.4 | JW (Lumo) integration -- a third api-bridge participant, backed by Proton Lumo via lumo-tamer on ai-lab-b. Shipped 2026-09-24 after design review, code review (Debbie), a Lumo Memory privacy test, and controlled deployment. | labchan-jw-integration-v0.1.md commits 28c4cdb, ef148f3, 42013f7 |
| v1.5 | For Rachel, JW and Gemma: read attached .md/.txt/.docx files (text only, ~20 KB cap; Rachel no longer sends non-image files to OpenAI as images); a deterministic "not for me" check (addressing.py) -- silent, no model call, when a message opens with someone else's name; per-AI reply_mode (all three set to named_only: answer only when named or sent directly); and a NO_REPLY marker. Shipped 2026-09-26 after TB's service-flow sign-off and Debbie's code review (one blocking bug found and fixed). Live-tested: "JW: what's 2+2?" on pivot -> only JW answered; "Let us turn to D&D" -> no AI answered. |
Agent Service Flows commits bd406e0, 80163d9, 3ddbc27, 858055a |
| v1.7 | fetch-page: for Rachel, JW and Gemma -- "<Name>: fetch-page: example.com <question>": the bridge downloads the page and gives the model its plain text as untrusted content. New listener/web_fetch.py: every resolved address must be public (incl. IPv6 forms that embed IPv4), re-checked after each redirect, connection pinned to the checked IP; max 3 redirects, 15 s per read / 30 s total, 2 MB, text only. Shipped 2026-09-26 after TB's approval and Debbie's review (4 notes fixed). Live-tested: "Rachel: fetch-page: example.com" -> correct summary of the page; "Gemma: fetch-page: example.com" -> page text quoted back (~2 min, local model). JW's live test blocked by the Lumo weekly quota, not by v1.7. |
Agent Service Flows commits 9d727c0, ea93ff3 |
| v1.8 | Easier sending -- plain typing (no /labchan form needed), /to drop-down for current recipient, a file-picker on /labchan, and "Reply via LabChan" from a long-press. Single converter (Jeff's host only); every other listener is unchanged. Shipped 2026-09-26 after TB's approval and Debbie's 6 review requirements, 592 tests passing. Config enabled and live-tested 2026-09-28: plain-typed messages, image attachments, and /to all confirmed working end to end. v1.8.1 same day: blank line after the LabChan/1 header for Discord readability. |
labchan-easier-sending-v0.1.md commits a89030b, 5bfb833, ddf9a79 |
| v1.6 | Unified Memory -- one shared listener/memory.py module replacing each bridge's own near-copy (RachelNotesStore, JWNotesStore, Gemma's select_history for the recent-exchange half). Layer 1: pinned facts (remember: / remember this: / forget: / "what do you remember?", human-only authors, tb-only for now, verbatim, max 50 facts/300 chars). Layer 2: the existing per-sender rolling exchange log, unchanged behavior, Gemma gains it for the first time. Topic key present, fixed to "misc" until LabChan II. Scope call at build time: Rachel and Gemma shipped now, JW deferred (blocked by the Proton API quota issue below). Design approved 2026-09-24, built and code-reviewed 2026-09-28 (Debbie caught a real bug: recall()'s facts were all-or-nothing instead of walked individually like exchanges -- fixed same day), deployed and Loki-tested 2026-09-29 on both Rachel and Gemma: a fact taught by TB survived a full listener restart and was correctly recalled by a different sender (Debbie for Gemma, Jeff for Rachel). A same-day 5-fact starter batch plus an anti-confabulation guardrail (Rachel was caught inventing an untaught fact; fixed in recall() itself, deployed to both) followed the same evening. See labchan-memory-v0.1.md for the full build/review/test writeup. |
labchan-memory-v0.1.md commits f7e75b7, 8cee5d4, 240cdea, ff4bae1 |
| v1.9 (rolled back) | Gemma Vision (LabChan attachments) -- her counterpart to Rachel's v1.2 Web Vision, scoped to attachments received via LabChan only (webpage-image loading deferred). Shipped and live-tested 2026-09-29 (a real image sent via LabChan attachment got a correct description back, 681/681 tests passing), then reverted the same day: TB's follow-up live testing got the identical hallucinated reply for two different images, traced to a real context-overflow bug (image token cost invisible to select_history's budget-walk, can exceed Gemma's 8192-token ceiling once combined with realistic context+history -- live calls degrade silently rather than erroring). Reverted cleanly, commit 8d837c7, 669/669 passing. |
labchan-gemma-vision-v0.1.md build b7a1d23, revert 8d837c7 |
| Version | Notes |
|---|---|
| v1.6.1 | JW wired into Unified Memory (deferred from v1.6, not a design change -- LabMemory was written generically). Blocked until JW's Proton API quota issue (see Future/Under Consideration below) clears; Debbie re-checks it roughly daily at her own session start. |
| v2.0 | Tentative/placeholder, scope undefined -- was originally floated as "major bump, TX-side rework," not a firm commitment. |
| Topic | Notes |
|---|---|
| Agentic tool-calling | Letting a model decide when to use a tool, instead of today's fixed typed-prefix commands (chart:, discogs:, etc.). Scoped (not authorized) 2026-09-21 with a concrete first target -- Rachel driving Illwinter's Floorplan Generator, per AG's suggestion -- see labchan-rachel-agentic-tools-v0.1.md (also posted as a public page). |
| Dashboard automation | A real, code-generated version of this page (Jeff-maintained Python function, output TBD) -- see labchan-dashboard-automation-v0.1.md. |
| Rachel screenshot capability | Deferred when vision & image-generation shipped (v1.2). |
| Rachel's progressive self-awareness | Ongoing goal, not a single scoped release -- ties into v1.3's memory/context store work (History table above). |
| RX listener watermark durability | Journal/archive writes aren't fsync'd before the watermark advances -- a rare crash-timing edge case, flagged but not yet fixed. |
| cc-sync/Discord-bridge duplicate delivery | Older bug that predates the current LabChan protocol; may be moot now that the bridge was replaced. |
| JW web-text/API quota (Proton) | Confirmed live-tested 2026-09-28: fetch-page: is code-complete but every attempt hits a Proton "weekly quota" error via lumo-tamer's API path, while the Lumo web chat UI on the same account works fine (controlled Lite-vs-Lite comparison, TB + Debbie). Looks like a genuine, intentional limit on API-style access to Lumo, not a bug on our side or a shared account-wide quota -- see Model Capabilities for the full writeup. No known reset window; Debbie re-checks roughly daily. Blocks v1.6.1 (JW's Unified Memory wiring). |
| JW adapter swallows quota failures silently | Found during the above investigation, 2026-09-28: jw_adapter.py doesn't relay a quota-block failure back over LabChan, so a blocked invocation looks like total silence rather than an error reply. Small, known, not urgent -- worth a follow-up fix once JW is usable again. |
| Gemma vision -- image-cost-aware budget (redo of v1.9) | v1.9 shipped 2026-09-29 and was rolled back the same day: a real image plus realistic context+history can exceed Gemma's 8192-token ceiling, and the live path degrades silently (a wrong-but-confident answer) instead of erroring. Real fix needs image token cost accounted for inside the budget calculation itself, before select_history decides how much history to include -- not scoped or authorized yet. See labchan-gemma-vision-v0.1.md for the full rollback writeup. |
| Lab-Crew-2.png cartoon avatar swap | On hold, TB's call 2026-09-28. Part of the same avatar-swap request that already shipped for LabChan-Members.png and Labchan-Service-Flow.png (Jeff/Debbie/Rachel/Gemma/JW to cartoon, real humans TB/AG/MG/Jason unchanged) -- but Lab-Crew-2.png composites everyone into one photorealistic AI-generated office scene rather than separate boxed photos, so a straight cutout-paste looks jarring. A single-prompt regeneration attempt (Rachel, generate-image:) collapsed everyone to one art style instead of the intended mixed photoreal/cartoon look. Needs either hand-placed background-removed cutouts (lower fidelity, doable now) or a different regeneration approach -- not picked up yet. |
Quick hand-maintained report, not run through the design-doc/release process -- see labchan-dashboard-automation-v0.1.md for the eventual real-automation plan.