Coder/deepseek-harness
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(session): project steering messages as plain user content

Browse Source 
steering/message previously rendered inside a <steering source="…">
envelope like context/message. No model is trained on a <steering> tag,
so the framing is arbitrary markup the model was never taught to read;
recorded transcripts show it treating the instruction as third-party
metadata and refusing it. Framing also does not belong on the session
surface — a caller that wants a frame formats its own content. It now
projects verbatim as a plain user-role message; the <context> envelope is
untouched and renderTagged becomes context-only.

Agent Note: .agents/notes/implemented/simplification/2026-07-20-unwrap-steering-message-projection.md
This commit is contained in:
Turtle
2026-07-20 11:00:36 +08:00
parent 023bed1a83
commit b25b78fd02
15 changed files with 227 additions and 421 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.agents/notes/implemented/architecture/2026-06-11-content-block-vocabulary.md
View File

@@ -10,7 +10,7 @@ The harness needs one internal language for messages that the loop, session log,
Own the vocabulary: messages are arrays of typed content blocks (`text`, `reasoning`, `tool-call`, `tool-result`), with the union derived from the merge-extensible `ContentBlockMap` so plugins add block types via declaration merging. The same merge-extensible-map pattern types every "stringly" field (`MessageSource`, `FinishReason`, `TurnTrigger`, `TurnEndReason`). Streaming is a raw chunk protocol; `BlockAssembler` is the single shared assembly implementation. Adapters translate to provider wire formats — mapping cost lives in adapters, where it belongs.
In-session context injection (`context/message`, `steering/message`) renders as tagged user-role envelopes (the system-reminder pattern) rather than a new role, so adapters carry zero burden. Live-adapter validation confirms this rendering for current DeepSeek behavior; a future provider-specific mismatch belongs in that adapter rather than a new canonical role.
In-session context injection (`context/message`) renders as a tagged user-role envelope (the system-reminder pattern) rather than a new role, so adapters carry zero burden. Live-adapter validation confirms this rendering for current DeepSeek behavior; a future provider-specific mismatch belongs in that adapter rather than a new canonical role. `steering/message` originally shared the envelope but now projects as plain user content; see [the steering-unwrap Agent Note](../simplification/2026-07-20-unwrap-steering-message-projection.md).
## Alternatives considered