base-layer/promptkit
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases 6 Wiki Activity

Finalize conversation and execution group documentation

Browse Source
This commit is contained in:
Eric Rakestraw
2026-11-14 01:35:06 +00:00
parent 31cc0069fc
commit 5c135c5b28
13 changed files with 96 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

21
docs/internal/runner.md
View File

@@ -30,6 +30,23 @@ internal callers and tests.
Each invocation carries its state in request, prepared-run, and result values.
The runner has no durable run or session store.
Captured exchanges are immutable parent-linked domain values. Each holds only
its new rendered suffix, final assistant message, and run identities; flattening
iteratively makes a detached history for preparation or inspection. Preparation
prepends this history to newly rendered definition messages and appended caller
messages before computing the rendered hash. Follow-up selection and credentials
remain those of the new request. Prepared execution retains capture and parent
references in its single-claim payload, not its durable details; discard clears
the payload. Completed runs capture the exact final output, including exhausted
structural repair, while operational errors produce no exchange.
The sequential coordinator uses ordinary facade runs and retains the current
immutable exchange independently of callback-owned values. Each run releases its
logical admission lease and finishes observation before application callback
delivery. Context checks preserve a completed but undelivered result only when
cancellation is observed before invoking the callback. Coordinator budgets count
started runs, not transport attempts or structural repairs.
## Shared Prompt Selection
The internal engine applies each namespace view to slashless caller selections
@@ -124,8 +141,8 @@ profile, or backend:
schema metadata from it when required;
2. load and hash input artifacts;
3. render messages and the prompt-defined session;
4. apply any direct session ID, then append already-normalized request messages
after the rendered definition messages;
4. apply any direct session ID, then compose captured history, rendered
definition messages, and already-normalized request messages in that order;
5. hash the effective rendered prompt; and
6. construct the prepared value and preparation timing.