Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
21 changes: 21 additions & 0 deletions README.md
Original file line number Diff line number Diff line change
Expand Up @@ -36,6 +36,27 @@ Provides streaming orchestration that delivers LLM responses token-by-token for
- Progressive rendering
- Interruptible generation

### Goals changed during a turn

Hosts may set `coordinator.session_state["goal"]` before execution or while a
conversation turn is running. The original goal fields (`condition`,
`turns_used`, `last_reason`, `cap`) remain supported; the loop initializes its
additional bookkeeping whenever it consumes a goal. A goal introduced during
a turn receives the same bounded tool/response evidence and completion
ordering as one set before execution.

If the host clears, replaces, or revises a goal while its evaluator, stall
judge, or terminal summary is awaiting a response, the old result is discarded.
The current answer is finalized and returned; a successor goal is preserved
for the next explicit execution, without replaying the input or starting work
for that successor. Hosts that revise a goal in place must change its condition,
cap, or task revision. Cancellation stops automatic continuation and clears
only the goal being cancelled.

`orchestrator:complete` is deferred until the goal decision is known. Continuing
progress hooks run before an intermediate completion so a host can pause there;
the final completion then remains marked `goal_final: true`.

### Durable completed-tool checkpoints

A host may register a zero-argument `session.durable_checkpoint` capability on
Expand Down
Loading
Loading