dsh-customisations

One profile plugin that carries small, independent DeepSeek Harness fixes. Instead of a package per workaround, every fix is a module with a stable id, so the bundle stays one install, one patch row, and one health line while the fix list grows.

Today the bundle contains one fix: currency-guard, which stops DSH's single-dollar inline math from swallowing prose.

Install

dsh plugin --profile desktop add github:drm-nz/dsh-customisations

Then restart the app (host plugins load at boot). From a local checkout:

dsh plugin --profile desktop add link:/absolute/path/to/dsh-customisations

Verify the mount in the shared health log:

tail -1 ~/.dsh/health.log
# ... dsh-customisations    ok   fixes=currency-guard

Fixes

currency-guard

DSH's chat markdown enables micromark inline math with singleDollarTextMath and no currency guard. A reply that writes money with $ therefore pairs the first two dollars in a paragraph into one KaTeX span:

NVDA fell to **$187.43**; 53 shares remain at a $229.18 average.

...renders as one math run: spaces collapse (228.21andneverreplaced), prose turns italic, the closing ** prints as ∗ ∗, and a & inside the span (as in P&L) makes KaTeX fail and print the whole span source in #cc0000 red. GenUI's own inline renderer is not involved: its single-$ rule rejects a closing $ followed by a digit, so the same text renders correctly there.

The general rule

A single $ pair is inline math, so any two unescaped $ outside code in one Markdown text block become one math span — the first and second pair, the third and fourth, and so on. Currency is just the most common way to produce them. Inside that span:

• whitespace collapses and letters turn italic;
• ** / == stop being Markdown and print literally (∗ ∗);
• a bare & is a TeX alignment character, so KaTeX fails and prints the whole span source in #cc0000 red (P&L is enough);
• the span can cross sentences and blank lines inside the block;
• an odd trailing $ has no partner and stays literal.

So the fix has to decide, per dollar pair, whether it is math or prose.

The fix

It wraps the documented llm/stream waterfall (the same seam dsh-llm-retry, dsh-llm-replay, and dsh-session-checkpoint-policy use) and runs two passes:

1. Streaming deltas — escape a currency-shaped $ the moment it arrives, so the live view is correct while the reply streams.
2. Terminal block-end — the adapter's full block text is what BlockAssembler treats as authoritative, so it gets the general pass: pair every unescaped $ outside fenced and inline code, and escape both ends of a pair whose body reads as prose (longer than 80 characters, contains **/==, a blank line, a bare &, or an ordinary word of three or more letters outside a command argument). Real math is left alone.

Markdown renders \$ as a literal $, so the visible reply is unchanged.

• Currency is always escaped, paired or not: $1,234, $5., ($5), $97,600,, $5k, $1.2M, $0.5%.
• Prose pairs are escaped: $quick brown fox$, −$5,833 … −$4,119.
• Real math is preserved: $x^2$, $2\pi r$, $E = mc^2$, $5$.
• Fenced code blocks and inline code are untouched, including shell $1.
• A trailing $/amount or backtick run is held across deltas, so a token split between chunks is still decided correctly.
• Plain-text consumers such as session titles (purpose: "session-title") are skipped, where a stored \$ would show a backslash.
• Fail-open: a transform error passes the original chunk through.

Limits: only new assistant text is guarded (history keeps its $), and user-authored text never passes through llm/stream. A non-currency pair can flash as math in the live view until the block settles, because only the terminal pass has the whole block to pair. When DSH's tokenizer grows a currency guard, remove this fix.

Configuration

- id: customisations
  name: dsh-customisations
  config:
    enabled: true            # whole bundle
    disabled: []             # fix ids to skip, e.g. [currency-guard]
    fixes:                   # per-fix options, keyed by fix id
      currency-guard: {}

Adding a fix

1. Add lib/fixes/<id>.js exporting id and apply(ctx, config); install the seam it needs there (waterfall listener, slot, instruction, ...).
2. List it in lib/fixes/index.js.
3. Extend inject in lib/index.js if the fix consumes another service.
4. Add a test that the fix installs and that it is skipped by disabled.

Keep each fix fail-open and idempotent: the bundle is installed once per profile, and one broken fix must not take the others down.

Development

node --test test/

The suite covers the reported reply end-to-end: the unguarded text parses into three inlineMath nodes (one containing P&L), the guarded text parses into none, and the only change to the text is a backslash before each currency dollar.

License

MIT