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

fix(agent-loop): address second-round review — disposal discard, injection validation, frozen payloads

Browse Source 
Address the review bot's five genuinely-new findings on the current code:
- disposal now discards any still-pending inbox items before the loop
  exits, so every enqueued id gets a terminal lifecycle event.
- injection (next-step/no-wakeup) validates its payload up front, before
  opening the idle one-shot turn, honoring 'invalid input throws before
  any append'; and rejects attached contexts (which belong only to inbox
  messages) rather than silently dropping them.
- agentMessage() freezes the agent/inbox/* payload so a listener cannot
  mutate the shared correlation object mid-dispatch.
- refresh the package READMEs (compact, goal, guard, hook-protocol,
  plan-mode, time-context, workspace-context) that still referenced the
  removed context/message event, with the source-based user/message
  distinction.

The up-front injection validation makes two finally branches unreachable
(v8-ignored as the turn-enclosure backstop). Adds regression tests for
disposal discard, context rejection, up-front validation, and the frozen
payload; per-file coverage stays 100%.
This commit is contained in:
Turtle
2026-07-24 12:05:57 +08:00
parent 7b7f793ee5
commit c7c1b97501
14 changed files with 127 additions and 62 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
packages/guard/repeat-tool-guard/README.md
View File

@@ -30,11 +30,11 @@ The chain key is `(tool name, canonical arguments)` — canonicalization is a de
## Reminder delivery
Reminders ride the post-execute decision's `additionalContexts` (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit. The loop buffers the context and appends it as a `context/message` after the step's tool results, which the session renders as a plain synthetic user message — so the reminder is model-visible, source-attributed, and reconstructable from the session log with no new session event. The guard always delegates via `next()` and prepends its reminder to the downstream decision's context array (both variants — a blocked call still gets the nudge); every entry retains its own source and metadata.
Reminders ride the post-execute decision's `additionalContexts` (source `{kind: 'plugin', plugin: 'repeat-tool-guard'}`), never a `content` replacement: the `tool/result` event stays the tool's own output for audit. The loop buffers the context and appends it as an injected `user/message` after the step's tool results, which the session renders as a plain synthetic user message — so the reminder is model-visible, source-attributed, and reconstructable from the session log with no new session event. The guard always delegates via `next()` and prepends its reminder to the downstream decision's context array (both variants — a blocked call still gets the nudge); every entry retains its own source and metadata.
## Testing
Unit suites drive a real agent loop against a mock adapter (no network) and cover the chain semantics above to per-file 100%. The snapshot tier owns the transcript surface: a scripted-replay scenario repeats a call five times and pins both reminder tiers (gentle at 3, detailed at 5) as `context/message`s in the ACP transcript.
Unit suites drive a real agent loop against a mock adapter (no network) and cover the chain semantics above to per-file 100%. The snapshot tier owns the transcript surface: a scripted-replay scenario repeats a call five times and pins both reminder tiers (gentle at 3, detailed at 5) as injected `user/message`s in the ACP transcript.
## Model Experience