Repository navigation
2.16.0
Changed
- Queued messages while the AI works are delivered mid-turn. Messages sent while an agent is
working — and messages prefixed with!in particular — ride along with the next tool result as
a visiblequeued_user_messageblock (including the time each message was queued), instead of
being injected into the running context. The answer you type for a question (askUser) carries
all queued messages along with it. If nothing consumes them in the current turn, they are asked
as follow-up turns as before. - Built-in agents are named Jon, Plan, Dev, Review (sub-loops
Search/Compact) — display
name = agent id everywhere (header, roster, history files, prompts); the Ork names are gone.
Config keys (llm.agent.*, model slots) are unchanged (ADR-0079) - Agent histories start fresh — history files follow the display name (
Dev-history.jsonl
instead ofPeon-Dev-history.jsonl); oldPeon-*-history.jsonlfiles are not read, rename them
manually to keep a conversation.agent-order.txtdefault is now^Jon$; an existing file with
^Peon-PO$no longer matches — such a file falls back to the default order (Jon first, rest
alphabetical), so the dropdown stays usable until the file is updated.
Same clean break for the agent-specific standing orders: Jon now loadsAGENTS-JON.md— an
existingAGENTS-PO.mdis no longer read, rename it. - Scaffold is now
Scaffold— the config-artifact agent (agents, skills, commands) drops its
Peon-prefix like the other roles: dropdown, chat header and the sender stamp on its chat lines
all use that name. Clean break, no alias — anagent-order.txtpattern on
^Peon-Scaffold$orPeon-.*no longer matches.
Added
- Every AI chat line names its speaker. The core stamps each line with the speaking agent's
id (origin, ADR-0078): the Eclipse plugin header shows who is talking instead of a generic
"Peon" (lines without an agent carry no header; groups merge only when role and sender match),
and the web factory attributes final team answers by the stamped sender instead of the tool name.
New Features
- Your agent choice survives a restart. The agent you picked in the chat dropdown is remembered for this Eclipse workspace and restored on start — after a restart or a workspace switch you continue with that agent instead of Dev. Only a real change of agent is remembered (a handoff counts, re-picking the active one does not), and an agent that has since been removed or renamed leaves the default in place, quietly. (
org.sterl.llmpeon, org.sterl.llmpeon/docs/active-agent-persistence.md)
Fixed
-
The live status bar no longer vanishes on chat events. Sending a message while the AI works
(e.g. a!-message) or answering a question used to hide the "waiting for AI…" bar until the
next response chunk arrived. Appended messages no longer hide it — the bar closes only when
streaming ends or the turn finishes. -
Jon files docs by topic, not by code. Topic folders (
docs/<topic>/with their own index.md) are opt-in — Jon proposes one when a domain owner reaches about three feature docs and asks first; each feature/architecture doc names its technical home (module, package) in its header. (org.sterl.llmpeon.corepromptsdoc-method.md,po.md) -
Doc moves keep the links. When Jon moves or renames a doc, all references — docs, code comments, tests — are updated in the same cycle. (
org.sterl.llmpeon.corepromptpo.md)
Bug Fixes
- An
agent-order.txtthat matches no agent no longer scrambles the dropdown. Valid patterns that
fit nobody — a renamed agent, a copied config — sent every agent to the alphabetical tail, so Jon lost
first place; now such a file falls back to the default order (^Jon$, Jon first) with a warning in the
log. A partial match behaves as before, and.*still gives a deliberately alphabetical list. (R5,org.sterl.llmpeon.core, org.sterl.llmpeon.core/docs/agent-ordering.md) - A failed AI call now names the endpoint it tried. Errors that said nothing — a blank message, or the JDK wrapper that only repeats its cause's class name — are reported as
AI call to <url> failed: <Exception> — <cause>, so a dead model URL is tellable from an MCP one; cancelling a call and errors that already explain themselves stay untouched (R-AEM-3,org.sterl.llmpeon.core, org.sterl.llmpeon.core/docs/agent-error-memory.md, PR #168)