TEAM CHECKPOINT — Multi-Agent Project Status

Last updated: 2026-08-06 22:43 AEST Paste this entire file into any session (web or native/API) to resync.

Agent Prefix Key

• Cl1- = Claude A (analytical/detailed)
• Cl2- = Claude B (concise/witty)
• Ge- = Gemini
• OA- = ChatGPT
• Gr- = Groq
• Cl-chat- = Claude via claude.ai web/app chat (outside orchestrator.py itself — used for this checkpoint and for direct code edits to orchestrator.py)
• Andrew edits carry no prefix (Architect/QA — final say, no attribution needed)

Rule: when adding a line to Verified Facts or Open Items, prefix it with your tag so edits don't get silently overwritten or misattributed. Edit Log entries always include prefix + timestamp.

Edit Log

• [prior] Cl1- created initial checkpoint structure
• [prior] Cl1- added prefix convention per Andrew's instruction
• [2026-08-06] Cl-chat- rebuilt this checkpoint after a session adding /mode 2shr, /mode 3shr, mode-scoped /ping, and a cohesion-note bug fix to orchestrator.py

Team Roster & Roles

• Andrew — Architect & QA Lead — Active
• Claude A (Cl1-) — Lead Programmer (analytical/detailed) — Active
• Claude B (Cl2-) — Lead Programmer (concise/witty) — Active
• ChatGPT (OA-) — Systems Analyst / Assistant Programmer — Active
• Gemini (Ge-) — Relief Programmer / Presentation Manager — Onboarded; not used in 3shr anymore (swapped for Claude A, see below) — role otherwise not yet exercised
• Groq (Gr-) — Fast-inference agent (free tier, OpenAI-compatible) — Wired in, available
• Grok (xAI) — Not integrated — separate from Groq, pending decision

File / Component Status

File / component	Status	Use it?
orchestrator.py	Modified this session (2shr, 3shr, mode-scoped /ping, cohesion-note fix) — compiles clean (py_compile), not yet live-API-tested	Yes — this is the current version, supersedes any copy from before this session
history/shared_memory.json	Existing groupchat pool — TF-IDF relevance × recency × diversity-cap retrieval, top-5/turn, filelock + atomic writes	Yes — unchanged, still backs groupchat mode
history/shared_memory_2shr.json	New — full-transcript pool (last 40 entries shown in full, not retrieval-filtered) for /mode 2shr (Andrew + ChatGPT only)	Yes — created automatically on first 2shr use
history/shared_memory_3shr.json	New — same full-transcript approach for /mode 3shr (Andrew + ChatGPT + Claude A, not Gemini)	Yes — created automatically on first 3shr use
memory_tools.py (standalone module, delivered separately, not yet pasted into orchestrator.py)	Experimental — gives agents real function-calling read_memory/write_memory tools instead of always-inject. Built before orchestrator.py was uploaded, so its TF-IDF scorer is a reconstruction and does not match orchestrator.py's real shared_memory_retrieve() exactly	Hold off — don't wire in as-is; if function-calling access is still wanted, redo it against the real shared_memory_retrieve()/_load_pool() methods instead of the standalone reimplementation
agents/claude_agent.py, agents/openai_agent.py, agents/gemini_agent.py, agents/groq_agent.py	Unchanged this session	Yes
PDFsearch.py / ISE project (drugs-and-users.org)	Unrelated to this session's changes — see prior checkpoint facts below	Yes, per prior status

Verified Facts — Don't Re-Check

• Cl1- orchestrator.py runs on Windows 10 Intel NUC via cmd.exe; Andrew also runs/tests it via Termux on Android
• Cl1- Modes: solo, solo_groq, solo_gemini, parallel, alternate, relay, collaborative, groupchat, 2shr, 3shr, a_then_groq, a_then_gemini, ab_then_groq, a_then_groq_gemini
• Cl1- Mode switching: /mode <name> always sticks until changed again (the old once/permanent distinction is accepted for back-compat but ignored)
• Cl1- History is per-agent (JSON in history/); groupchat/2shr/3shr additionally persist a shared pool to disk
• Cl-chat- /mode 2shr: private two-party session, Andrew + ChatGPT only. Every turn shows the last 40 entries in full (not relevance-filtered) from shared_memory_2shr.json, wrapped in a "SHARED SESSION LOG — use for continuity" instruction block. No other agent ever sees or writes to this pool.
• Cl-chat- /mode 3shr: same full-transcript mechanism, three-party: Andrew + ChatGPT + Claude A (originally built with Gemini, swapped to Claude A per Andrew's stated preference for Sonnet 5 and reused the same Claude A instance/history as other modes). ChatGPT and Claude A take turns within a round, each seeing what the other just said. Own pool: shared_memory_3shr.json.
• Cl-chat- Fixed a bug in 3shr's original cohesion note: it was one static string sent to both agents saying "the other agent (ChatGPT or Gemini)" — meaning each agent could read itself as the "other" one. Now built per-agent inside the loop, naming the specific other participant by name.
• Cl-chat- /ping used to always hit the full configured roster regardless of mode. Now scoped via a new mode_ping_targets() method — only pings the agents the current mode actually uses (e.g. 2shr → ChatGPT only; 3shr → ChatGPT + Claude A; groupchat → full roster). "Not configured" warnings are scoped the same way.
• Cl1- shared_memory.json (groupchat's pool): TF-IDF-style relevance × recency weighting × per-speaker diversity cap, top-5 hits/agent/turn, filelock-based concurrency, atomic writes (temp file + os.replace()), corrupted files backed up as .corrupt-<timestamp> instead of silently wiped
• Cl1- /memory <query> command exists and works: read-only lookup into the groupchat pool only (not 2shr/3shr's pools) — no send, no write
• Cl1- PDFsearch.py is live in production on drugs-and-users.org (DB joins verified, PDF.js viewer working, dedup-by-attachment implemented, page-count preserved, whole-document links only)

Open Items — Decisions, Not Blockers

• Cl-chat- 2shr and 3shr have not been live-API-tested — only py_compile clean and code-reviewed. First real run is the actual smoke test.
• Cl-chat- TWOSHR_FULL_LOG_LIMIT / THREESHR_FULL_LOG_LIMIT both set to 40 (entries shown in full each turn) as an untuned guess — adjust if context runs too long/short in practice
• Cl-chat- memory_tools.py's function-calling approach (agents choosing when to read/write memory, vs. always-inject) is still unintegrated and built against a reconstructed retrieval algorithm — needs a decision: drop it, or redo it against orchestrator.py's actual shared_memory_retrieve()
• Cl1- Whether groupchat's shared_memory.json should show every agent the FULL pool each round vs. the current filtered top-5 — still unanswered
• Cl1- Archiving/dedup/pin-forget memory-management layer on top of shared_memory.json — discussed, not built
• Cl1- Whether Grok (xAI) gets added as a distinct agent, separate from Groq
• Cl1- PDFsearch.py DB credentials — still root/plaintext as of last known state; scoped read-only user not yet confirmed implemented
• Cl1- Tokeniser stub (.split()) in PDFsearch.py/PDFIndexer.py — real tokeniser not yet wired in

Known Issues / Blockers

• None currently active

Next Actual Action

Run /mode 2shr and /mode 3shr for real (live API calls) and confirm both behave as designed — full 40-entry transcript injection, correct per-agent cohesion wording in 3shr, and /ping only hitting the right agents per mode. Report back anything that doesn't match before building further on top of either mode.