DSH ACE

Beyond the obvious.

An AI next-move advisor for DeepSeek Harness. It reads the recent conversation and shows one card: it restates the thing you are answering, names the relation behind it, points at one other place that relation shows up, and ends with a checkable move. Press Tab to insert only that move; nothing is sent automatically.

简体中文

Rendered example, not a recording. A real screen capture replaces it before the first release.

The difference, in one line: autocomplete guesses what you would type; ACE reorders something you already know until the reaction is "right, why didn't I see that" — and shows nothing when nothing wins.

Install

The npm name dsh-ace is taken by an unrelated project, so the package is scoped as @ghoscro/dsh-ace. Until it is published, the tarball on Releases is the install path:

# in the DSH profile directory
pnpm add file:/absolute/path/to/ghoscro-dsh-ace-0.5.23.tgz

Then add "@ghoscro/dsh-ace" to dsh.profile.bundles, materialize the profile (DSH plugin page or pnpm install), and restart DSH. The composer row shows ✦ ACE (✦ 深度预测 in the Chinese UI).

If an install goes wrong, INSTALL.md has the two-step recovery: remove the bundle entry, restart. A failed entry does not stop DSH from booting. Developer install (copy this directory into sakura/packages/) is also there. Verified host: DSH 0.2.1-rc.2; other versions are unverified.

What it does

Most next-message tools optimize for what you would probably type. ACE optimizes for whether this move is worth sending, and whether it lets you see the next layer yourself.

The four-beat card:

You just said X → the line behind it is Y → so Z will behave the same way → try W first; if you see such-and-such, I was wrong

By default the card shows only W, the line you can send. X/Y/Z, the three plain lines and "wrong if" sit behind Show details in the footer: the card answers "send it or not" first, and "why" on request.

• X restates the concrete thing in your own words (usually the assistant line you are answering).
• Y states the relation in one human sentence: why it works or fails. At most one abstraction.
• Z gives one other place the same relation shows up. If it does not look alike on the surface, the card carries a separate line saying how it maps; if it cannot, this beat scores zero.
• W is a checkable action for today: engineers get a file or command, everyone else gets something they can see ("refresh and look whether the badge is still red").
• Wrong if: what result would prove the card wrong.

Three plain lines on every card:

• Now: what is happening and where it is stuck, in symptoms you can see.
• Why: the result this step is for, not the action.
• After: what you will see or get, specific enough to judge "is this what I want".

One missing line and the card is not shown. Wild mode is included: the line Tab inserts must be readable at the current level, not a stack of unexplained terms.

• It generates 3 directions internally: advance, use an existing tool, reframe, verify, clarify, constrain, close.
• It scores them for value, novelty, worth_extra_step, and action type. The relation beat is judged by the host, not scored by the model: a surface-alike example needs no mapping line, and a surface-unlike example with no written mapping is not shown at all.
• Level: the card reads the user's own messages and adapts — everyday language and one action at a time for non-engineers, shorthand and judgement for engineers. Override it with level in the plugin config.
• The default mode answers a specific line of the assistant's last message (reply_to, quoted verbatim), but picks the answer worth sending, not the smoothest one.
• It shows one card, with X/Y/Z/W and its falsifier.
• If nothing beats the obvious “continue”, it shows nothing. A quiet card is a feature.
• Tab inserts only W. You can edit it. ACE never sends for you.

Composer

ACE also works inside the input box:

• While you type, a ghost completion appears at the caret; Tab accepts it.
• ` (backquote) or F1 toggles the ghost. F1 also prints one diagnostic line per press, newest at the bottom.
• Alt+1 … Alt+4 picks one of the alternative card rows; Escape dismisses the card or the settings panel.
• The loaded build is exposed as <html data-ace-version>; quote it in a bug report.

Demo

The card above is the intended loop:

1. The assistant asks, “run the full suite?”
2. The obvious next line is “run it”.
3. The card: X “run the full suite?” → Y the retry swallows the flake, so a full run only spreads the odds → Z upload retries and CI flaky reruns are the same mechanism → W “remove the retry and run this one case once”.
4. “Wrong if”: one clean run means the retry was not masking anything.
5. Tab inserts only W; you decide whether to send.

Quick start

1. Install as above; restart DSH.
2. Open a conversation. The composer row shows ✦ ACE; click it, or wait for the next completed turn if automatic ACE is on.
3. Press Tab on the card to fill the draft.

If you are migrating from dsh-composer-predict, read MIGRATION.md first. Do not load both packages at once.

A missing card is not a failure. It means nothing beat the obvious "continue" for that turn. Silent turns are logged as no-card, and the score threshold can be swept offline with npm run sweep instead of being tuned on live chats.

Three cases

These are examples, not a controlled benchmark. They show the kind of added value ACE is aiming for.

1. A test that may be hiding the bug

• Obvious: “Run the full suite.”
• ACE Move: “Remove the retry and run this one test once; compare repeated runs with and without retries; a higher failure rate would support the masking hypothesis.”
• What it adds: it tests the mechanism instead of increasing the sample size.
• Risk: the test may fail for an unrelated reason.

2. A cheaper implementation path

• Obvious: “Write a small script to batch this.”
• ACE Move: “Use the existing CLI’s --json output and pipe it; no new script to maintain.”
• What it adds: a smaller surface area and a faster path to the same outcome.
• Risk: the CLI may not expose the field you need.

3. A task that is drifting

• Obvious: “Keep polishing the current card.”
• ACE Move: “The card is not the bottleneck; the install step still fails on a fresh profile. Fix that before the visual polish.”
• What it adds: it names the real constraint and changes the order of work.
• Risk: the install failure may be environment-specific.

How it works

1. The browser sends the recent transcript and the session cwd to the host route.
2. The host builds a bounded, read-only evidence pack: git status, diff stat, last commit and a bounded diff, plus the project's top-level tree and its dependency/script names, so a structural claim can land on a real module.
3. A deterministic situation block is built from the last user goal, the assistant’s last offer, and explicit constraints. There is no separate situation-model call.
4. One generation call returns 3 line-delimited JSON candidates plus a selection line.
5. The host scores and selects exactly one move. Fixed slot order is gone; verify/constrain candidates are penalized so “check again” cannot win by default. The winner then goes through one bounded review call: X/Y/Z, the relation, the analogue and the falsifier are checked against the evidence, and a card that cannot support them is rejected rather than silently downgraded.
6. If the best score is below the threshold, the plugin returns no card.
7. Tab fills the draft. The card’s why/evidence/risk are not sent with the message.

The parser scans balanced JSON objects and skips a truncated or malformed line while keeping valid earlier candidates. A bad line does not discard the whole response.

Settings and cost

• Modes: safe, standard, brainstorm, wild (labelled ACE) are equal-status modes, not a volume dial. Each changes the prompt policy and the output budget, the card shows which mode produced it, and the settings panel switches it. No mode collapses into safe.
• Cache: the rules and the voice samples form a byte-identical head, the conversation is appended instead of re-windowed every turn, and only the volatile tail (mode, situation, evidence) changes, so the provider's automatic prefix cache can hit on most of the prompt.
• Pause: stops automatic suggestions; ✦ ACE still works on demand.
• Text size: one slider in the settings panel drives --dsh-fs-w / --dsh-fs-meta / --dsh-fs-x. The value is stored in chrome.storage.local when the container exposes the extension API (localStorage otherwise) and set on document.documentElement.style, so changing the size never needs a rebuild.
• Model: uses the active DSH agent default, or the plugin’s configured route.
• Privacy: the recent transcript and the local evidence pack go to the selected model. Nothing is written back to the DSH session. captureSamples: false disables prediction samples only; ledger: false separately disables the local conversation ledger. Existing files are not deleted by these switches. Rejected-card feedback and usage/error diagnostics are separate local records; this is not a zero-retention mode.
• Ledger: each session keeps an append-only record of what the predictor saw; the prompt is rebuilt from it, so the prefix only grows, and past ledgerMaxEntries the oldest part is written to an archive file and replaced by a deterministic digest. ledger: false turns it off.
• Samples: the last 50 expert predictions (candidate moves, the raw answer, the last user line) stay in diagnostics/dsh-ace-samples.jsonl; npm run sweep re-scores them offline at every threshold, with no model calls.
• Accounting: every prediction appends one row to diagnostics/dsh-ace-usage.jsonl with the session, turn index, token counts, cache hits and a priced cost using the configurable pricing block.
• Cost: normally two model calls per ACE card: candidate generation and one bounded reasoning review. Recovery may add a generation retry. Generation caps are safe 1600, standard 2600, brainstorm 3200, wild 3600 tokens, also bounded by expertMaxTokens. Review and recovery each cap output at 2400 tokens. These are output limits, not total usage estimates.
• Waiting and review: elapsed seconds do not identify the backend stage. A review timeout retains the candidate with a visible unreviewed warning; content rejection returns no card. Default model-stage budgets are 25s generation + up to 12s recovery + 12s review. History and evidence work add time; the browser cancels an expert request after 58s. These limits are not a latency promise.
• Trial feedback: use the trial and latency report to distinguish installation, waiting, and card-quality problems. A restart loads the installed package; it does not download a new release.
• Automatic ACE: generates and reviews a card after each completed turn. Pause it to use only the ✦ button.

Development

npm test

Measure real provider token usage against a running DSH instance. macOS and Linux:

COMPOSER_PREDICT_BASE=http://127.0.0.1:13770 node scripts/token-tiers.mjs

Windows (PowerShell):

$env:COMPOSER_PREDICT_BASE='http://127.0.0.1:13770'
node scripts/token-tiers.mjs

The suite covers prompt/parser behavior, JSONL bad-line salvage, selection scoring, no-card behavior, route envelopes, and the single-card UI interactions.

Compatibility is declared in package.json.dsh.compatibility. Treat a version not listed there as unverified; do not assume a package-manager success is a runtime acceptance.

Uninstall

Remove the bundle and dependency from the profile, restart DSH, and delete the package directory. The old composer-predict route and settings key are kept for migration; removing the new package does not delete your local settings.

Author

A DSH plugin by micker.

License

MIT. Contributions use DCO sign-off; no CLA. See CONTRIBUTING.md and SECURITY.md.