A2a Lab

repo: ~/Projects/a2a-lab · 14 commits · 0 in last 24h, 3 in last 7d · █

Search the build

14 commits indexed

  1. 0fb6792 2026-09-22 cards/eval: flatten vp-consulting into consulting-agent (TK-12015)
  2. b77f2cc 2026-09-22 auto-data-snapshot: 2026-09-22T12:15:47 (3 data files) — cards/consulting-agent.agent-card.json cards/vp-consulting.agent-card.json eval-blind2.json
  3. b276fc7 2026-09-22 cards: drop prompt-engineer skill, merge scope into ai-engineer (TK-12015)
  4. 2a64bf9 2026-08-02 chore: lint (ruff, 7 autofixed), refactor self-review, v1.0.0 → v1.1.0 (session close)
  5. 54459a9 2026-08-01 frame find() as a ranked top-3 shortlist (routing hint, not oracle) + confidence flags
  6. ac91d16 2026-08-01 scorer fix (field-weight + trigram-downweight + synonyms) + independent held-out eval
  7. c5493c8 2026-08-01 add blind paraphrase eval harness (eval_find.py) — FALSIFIES lexical find() at 52.6% top-1
  8. 714e492 2026-08-01 add SDK-native client (client_sdk.py) + lexical-semantic find (semantic.py), verified
  9. d13a06a 2026-08-01 A2A cabinet directory server (list/find/card) + fix block-scalar domain parsing
  10. 9ea5e3d 2026-08-01 generate A2A agent cards from cabinet.yaml (12 VPs / 166 skills), round-trip verified
  11. b4cdc89 2026-08-01 TK Bridge: add gated dm-write skill (TK_BRIDGE_ALLOW_WRITE), verified over A2A on temp store
  12. 89535d8 2026-08-01 add read-only A2A bridge over the live tk cross-agent DM log, verified
  13. 4dc181f 2026-08-01 add working A2A echo demo (SDK server + raw JSON-RPC client), verified end-to-end
  14. f3cea00 2026-08-01 install A2A SDKs (python a2a-sdk 1.1.2 + @a2a-js/sdk), smoke-tested

Authors

Agents used

  • prompt-engineer2
  • ai-engineer2

Skills used

  • /vp-consulting3
  • /consulting-agent2
  • /claude2
  • /code2
  • /eval1
  • /held-out1
  • /find1
  • /card1
  • /literal1
  • /tickets1
  • /events1
  • /sdk1

Creative ideas + design notes

Commits with substantial prose (≥120 chars) — the rationale behind each move.

0fb6792 · 2026-09-22 · cards/eval: flatten vp-consulting into consulting-agent (TK-12015) 
Repoint the routing eval ground-truth from the retired vp-consulting officer
to the consulting-agent steward. Companion to auto-snapshot b77f2cc, which
already landed the new cards/consulting-agent.agent-card.json, dropped the old
cards/vp-consulting.agent-card.json, and repointed eval-blind2.json. Zero
dangling vp-consulting node across the a2a card + eval surfaces.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_38a23781-1663-4f8e-807d-e37d704feaad
b276fc7 · 2026-09-22 · cards: drop prompt-engineer skill, merge scope into ai-engineer (TK-12015) 
Mirror the cabinet.yaml merge — remove the generated prompt-engineer
AgentSkill and fold prompt-optimization into ai-engineer's card so the
A2A directory matches the org-chart.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01D2KFGQK7p7ahEzug7HBMZF
54459a9 · 2026-08-01 · frame find() as a ranked top-3 shortlist (routing hint, not oracle) + confidence flags 
Shared _shortlist() formatter for both find branches; warns on weak (<0.15) or close (<0.05 margin) picks. Honest product shape given held-out top-1 is imperfect. Ship ② per Steve.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ac91d16 · 2026-08-01 · scorer fix (field-weight + trigram-downweight + synonyms) + independent held-out eval 
Dev 52.6%->76.3% but HELD-OUT (eval-blind2.json, Ollama-generated, never inspected) = 38.9%.
The dev/held-out gap empirically demonstrates the overfitting Cody warned of. Conclusion:
hand-tuned lexical routing is inadequate for top-1 NL queries; more tuning worsens held-out.
Stopping the tuning loop — decision escalates to top-3-shortlist vs embeddings.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
c5493c8 · 2026-08-01 · add blind paraphrase eval harness (eval_find.py) — FALSIFIES lexical find() at 52.6% top-1 
38 held-out everyday-phrasing queries; scorer misroutes 18/38 on blind phrasing.
Confirms Cody's contrarian objection: the 'every tested query passes' premise was circular.
Committed with the failing baseline on purpose; NOT tuned-to-pass (would re-create circularity).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
714e492 · 2026-08-01 · add SDK-native client (client_sdk.py) + lexical-semantic find (semantic.py), verified 
client_sdk.py: A2ACardResolver + ClientFactory.create + async send_message (StreamResponse).
semantic.py: TF-IDF + synonym map + char-trigram cosine; wired into directory find w/ keyword fallback.
Ollama has no --embeddings (gated fleet change) so neural embeddings deferred; documented.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
d13a06a · 2026-08-01 · A2A cabinet directory server (list/find/card) + fix block-scalar domain parsing 
cabinet_directory.py serves+routes the 12 officer cards over A2A (verified over the wire).
cabinet_cards.py now handles YAML folded/literal domain scalars (vp-cncp/vp-consulting).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
9ea5e3d · 2026-08-01 · generate A2A agent cards from cabinet.yaml (12 VPs / 166 skills), round-trip verified 
Tolerant line parser (cabinet.yaml is not strict-YAML); each director owns -> AgentSkill.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
89535d8 · 2026-08-01 · add read-only A2A bridge over the live tk cross-agent DM log, verified 
Exposes ~/.claude/tickets/events.jsonl as an A2A agent (skills: agents, inbox).
Ports lib.js inbox/knownAgents read logic; never writes the shared store.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

File tree

28 files tracked. Click any to browse the source at HEAD.

Other build journals

← YOLO Progress Viewer  ·  all 4 projects  ·  Abrams Discovery Viewer →

Export

commits.csv · feed.atom · project.json · commits.json

rendered in 1ms