deepclause-sdk 0.0.88 → 0.1.0

package/dist/system/assets/docs/HARNESS_EXAMPLES.md

@@ -0,0 +1,172 @@
+# Idiomatic harness examples
+
+Copy-paste shapes for REPL harnesses. Keep each skill small and compose them.
+(The snippets are tested against the bundled SWI-Prolog.)
+
+Most skills should look like §2–§3: a short sequence of `task/N`/`prompt/N` calls
+with thin guards between them. §4–§7 (facts, pure logic, judgments, CLP) are the
+exception — reach for them only when the task is genuinely data- or
+logic-shaped, not as the default style. DML is the orchestration layer over model
+calls, not the logic library (`DML_AUTHORING.md` §0). For file work, shell out
+(`exec(bash(...))`) instead of reading files into DML (§2).
+
+Every skill carries a `doc(Name/Arity, Text)` annotation and the harness a
+`module_doc(Text)` line (`DML_AUTHORING.md` §4). `module_doc/1` is the harness's
+**skill description** — what it does and when to use it; it is the first line of
+the manifest, which the agent reads at session start and whenever it changes. A
+file with only `module_doc/1` is still a valid skill (an advertised capability
+with an empty body). Keep both terse.
+
+```prolog
+module_doc("Example harness: a few small skills used as patterns. Use when you need a copy-paste starting point.").
+```
+
+## 1. Wrap a workspace script as a tool
+
+Materialize the program as a workspace file (run a DML goal with
+`make_directory_path("scripts"), write_file("scripts/count_words.py", "…")`),
+link it with `script_file/2`, and define a predicate for it. `tool/2` is only an
+annotation that exposes an existing predicate to `task/N`; keep it a one-line
+delegation so the predicate is also callable from `execute` / `/goal`:
+
+```prolog
+script_file(count_words, ".dc-repl/scripts/count_words.py").
+
+% The capability: callable from `execute` / `/goal` and from `task/N`.
+count_words(File, Count) :-
+    script_file(count_words, Script),
+    format(atom(Cmd), "python3 '~w' '~w'", [Script, File]),
+    exec(bash(command: Cmd), Result),
+    get_dict(stdout, Result, Count).
+
+% The annotation exposing it to the nested agent (delegation only).
+tool(count_words(File, Count), "Count words in a workspace file.") :-
+    count_words(File, Count).
+```
+
+Untrusted data goes through files/stdin, never interpolated into the shell
+command. `script_file/2` is just a fact (the exporter bundles the file); the
+write lands on disk immediately, so `bash` can run the script in the same turn.
+
+## 2. A skill that calls the model
+
+`task/N` runs a bounded agentic turn. Name each output variable in the
+description so the model fills it:
+
+```prolog
+doc(classify/2, "Route a support message. classify(+Message, -Team). Example: /goal classify(\"I was charged twice\", T).").
+
+classify(Message, Team) :-
+    task("Route this support message to exactly one of billing, orders, or account. Message: {Message}. Store the choice in Team.", Team).
+
+demo :-
+    classify("I was charged twice for the same order", Team),
+    format(string(Out), "Routed to ~w", [Team]),
+    answer(Out).
+```
+
+## 3. Multi-step pipeline: task -> task -> answer
+
+```prolog
+report(Topic, Result) :-
+    task("List three risks of using an LLM for arithmetic. Store them in Risks.", Risks),
+    format(string(Desc), "Rank these risks and explain the top one: ~w", [Risks]),
+    task(Desc, Top),
+    format(string(Result), "Top risk: ~w", [Top]).
+```
+
+Compose the steps and let the model do the thinking; keep only thin guards
+between them — an if-then-else for a branch, `once/1` to commit the first
+answer, or a simple check that fails a step:
+
+```prolog
+summarise(Doc, Summary) :-
+    task("Extract the document's main claim. Store it in Claim.", Claim),
+    (   Claim == ""
+    ->  Summary = "no claim found"
+    ;   task("Summarise this claim in one sentence: {Claim}", Summary)
+    ).
+```
+
+## 4. Data as facts (data = code)
+
+Small/medium data belongs in the module and is queried exactly like code:
+
+```prolog
+parent(alice, bob).
+parent(alice, carol).
+parent(bob, dave).
+parent(carol, eve).
+
+doc(ancestor/2, "Transitive parent. ancestor(+A, -B) enumerates descendants. Example: /goal ancestor(alice, X).").
+
+ancestor(A, B) :- parent(A, B).
+ancestor(A, B) :- parent(A, C), ancestor(C, B).
+```
+
+```text
+/goal ancestor(alice, X).
+X = bob ; X = carol ; X = dave ; X = eve
+```
+
+## 5. Pure logic + a test predicate
+
+```prolog
+doc(tally/2, "Count occurrences of each item. tally(+List, -Counts) yields Key-N pairs. Example: /goal tally([a,b,a,c,a], C).").
+
+tally(List, Counts) :-
+    list_to_set(List, Items),
+    findall(Item-N, (member(Item, Items), aggregate_all(count, member(Item, List), N)), Counts).
+
+test_tally :-
+    tally([a,b,a,c,a], C),
+    C == [a-3, b-1, c-1].
+```
+
+`/goal test_tally.` succeeds when the behavior is correct. When a skill changes,
+update or add its test and run it (`execute` during a turn, or `/goal`).
+
+## 6. Semantic judgments (the model decides, Prolog enforces)
+
+```prolog
+route(Message, Team) :-
+    choose(Message, "Which team should handle this?", [billing, orders, account], Team).
+
+urgent(Message) :-
+    holds(Message, "Is this urgent?").
+```
+
+`choose/4` always returns one of your options; `verify/3` is three-valued
+(`yes`/`no`/`unknown`); `holds/2` is the semidet wrapper.
+
+## 7. Constraints with CLP(FD)
+
+```prolog
+:- use_module(library(clpfd)).
+
+doc(solve/3, "X+Y=Total with X<Y over 0..100. solve(+Total, -X, -Y). Example: /goal solve(10, X, Y).").
+
+solve(Total, X, Y) :-
+    [X, Y] ins 0..100,
+    X + Y #= Total,
+    X #< Y,
+    label([X, Y]).
+```
+
+The `:- use_module(...)` directive must be **top-level** (an inline
+`use_module` inside a body is not run at read time, so `#=`/`ins` are unknown).
+
+## 8. Vision: describe a workspace image
+
+```prolog
+doc(describe_image/2, "Describe an image file. describe_image(+Path, -Text). Example: /goal describe_image(\"photos/fox.png\", T).").
+
+describe_image(Path, Text) :-
+    read_media(Path, Image),         % media(Path, Mime) — no bytes
+    attach_media(Image),             % reference-only; sent by the next task/N
+    task("Describe the attached image in one sentence.", Text).
+```
+
+Images are files; `media(Path, Mime)` is the reference. The bytes are read from
+the workspace at call time, only PNG/JPEG/GIF/WebP are accepted, and they never
+enter clauses or memory. Judgments stay text-only.