dsh-text-guard
A DeepSeek Harness plugin that neutralizes configured code-point sequences in tool results and in the messages entering a step, before the session log records them.
中文说明
The problem it solves
The official DeepSeek API rejects a whole request with 400 INVALID_REQUEST (Content Exists Risk) when it carries certain text. For the sequence observed in practice — the regional-indicator pair U+1F1F9 followed by U+1F1FC, which composes one flag emoji — character-level bisection showed the pair rejects the request while either character alone, the reversed pair, and other combinations are accepted.
Because every turn resends the stored history, a sequence that reaches the session log rejects every following turn of that session. The session becomes unusable: retrying does not help, restarting the app does not help, and the only way out is a new session. The usual way such a sequence arrives is not the user typing it, but fetched pages, file contents, or command output passing through a tool result.
What it does
It rewrites the configured sequences where content enters the log:
| Hook | Content covered |
|---|
tools/post-execute | an accepted tool result — web fetches, file reads, command output, subagent and workflow results |
tools/ptc-dispatch-log | the durable log copy of one run_code sub-dispatch outcome |
agent/pre-step | the messages entering the step, including text a user pasted |
Content without a match passes through untouched and keeps its original object identity, so an unaffected dispatch costs one regular-expression probe. The rewrite happens before the append, which is what keeps Session replay, fork, and every later request consistent — and it never puts the sequence back, because the replacement is pure ASCII.
Install
With the Harness CLI:
dsh plugin --profile <profile> add github:7771sd/dsh-text-guard
Or through the Plugin Manager's install field, using the same spec. The plugin has no dependencies and no build step.
Configuration
The shipped bundle patch mounts it with the wide mode. Every field can be overridden in your profile's cordis.patch.yml:
- id: text-guard
config:
mode: exact
sequences: [[0x1F1F9, 0x1F1FC]]
replacement: '[flag U+{codepoints}]'
toolResults: true
userMessages: true
log: true
| Field | Default | Meaning |
|---|
mode | regional-pairs | regional-pairs matches any two consecutive regional indicators (U+1F1E6–U+1F1FF); exact matches the configured sequences only. |
sequences | [[0x1F1F9, 0x1F1FC]] | Code-point sequences mode: exact neutralizes; each entry is matched consecutively. |
replacement | [flag U+{codepoints}] | Replacement template; every {codepoints} becomes the matched code points, for example U+1F1F9 U+1F1FC. |
toolResults | true | Neutralize accepted tool results and run_code dispatch logs. |
userMessages | true | Neutralize the messages entering a step. |
log | true | Log one summary line per affected dispatch, naming the count and never the matched text. |
Configuration changes are not hot: re-mount the bundle (disable and enable it) after editing the patch.
Why it hooks the log entry points
dsh-agent-loop deep-freezes request messages and documents that agent/request listeners cannot change request messages; the Harness rule is that everything the model sees must be reconstructable from the session log. A rewrite after the append would break that reconstruction, so the neutralizer runs where content enters the log instead of at the wire.
Verified
node test/text-guard.test.mjs — nine offline checks: content and value-form rewriting, block decisions passing through, a clean result returning the identical reference, message rewriting that preserves identity and startsRequestSeries, no mutation of inputs, both surfaces disabling cleanly, pair matching, and lone indicators passing through.- End to end in a live profile: with the plugin configured to neutralize an ASCII probe, a command whose output contained the probe reached the model already rewritten.
Not verified: the plugin has never been sent a real flag emoji by its author, because the experiment that proves the filter rejects one also destroys the session it runs in.
Limitations
- Text that reaches a request through a system-prompt section — a skill file or an instruction file that itself contains a sequence — does not pass these entry points.
- A rewritten user message is logged in its neutralized form while the chat client keeps displaying the text you sent, so the two views disagree about that message.
- The rewrite is lossy by design. Do not mount it when the original text matters verbatim to the task.
License
MIT
