fix(pformat): tag each argument with its slot number instead of counting empties

crates/tinyagents-harness/src/tool_calling/dialect/pformat.rs

@@ -1,13 +1,16 @@
-//! The positional dialect: `<tool_call>read_file[src/main.rs]</tool_call>`.
+//! The slot-indexed dialect: `<tool_call>read_file[0|src/main.rs]</tool_call>`.
 //!
 //! Roughly an 80% token saving over the JSON form on the call side, and more
 //! than that on the catalogue side, since a signature replaces a schema. See
 //! [`crate::tool_calling::pformat`] for the grammar itself.
 //!
 //! The interesting property is that it degrades rather than fails: a body that
-//! is not a well-formed positional call falls through to the JSON parser per
-//! tag, so a model that mixes the two forms in one response — or ignores the
-//! protocol entirely and emits JSON — is still understood.
+//! is not a well-formed p-format call falls through to the JSON parser per tag,
+//! so a model that mixes the two forms in one response — or ignores the protocol
+//! entirely and emits JSON — is still understood. That fallback is also what
+//! makes the parser's strictness affordable: a call with a miscounted or
+//! non-numeric index is refused here and retried as JSON, rather than being
+//! bound to whichever parameters it happens to line up with.
 
 use std::sync::Arc;
 
@@ -58,19 +61,25 @@ impl PFormatDialect {
         let mut instructions = String::new();
         instructions.push_str("## Tool Use Protocol\n\n");
         instructions.push_str(
-            "Tool calls use **P-Format** (Parameter-Format): compact, positional, \
+            "Tool calls use **P-Format** (Parameter-Format): compact, slot-indexed, \
              pipe-delimited syntax wrapped in `<tool_call>` tags. ~80% cheaper on tokens \
              than JSON.\n\n",
         );
         instructions
-            .push_str("```\n<tool_call>\nget_weather[London|metric]\n</tool_call>\n```\n\n");
+            .push_str("```\n<tool_call>\nget_weather[0|London|1|metric]\n</tool_call>\n```\n\n");
         instructions.push_str(
             "**Rules:**\n\
-             - Form: `name[arg1|arg2|...|argN]`. Arguments are positional and must match the \
-               order shown in each tool's `Call as:` signature in the `## Tools` section above \
-               (alphabetical by parameter name).\n\
-             - Empty calls: `name[]` for zero-arg tools.\n\
-             - Empty argument: `name[||value]` is three positional values, the first two empty.\n\
+             - Form: `name[index|value|index|value|...]`. Each value is preceded by the slot \
+               number it fills, taken from that tool's `Call as:` signature in the `## Tools` \
+               section above.\n\
+             - **Send only the arguments you mean to send.** To pass just the third slot, \
+               write `name[2|value]` — there are no empty slots to count.\n\
+             - The signature shows each slot as `index|<name>`, e.g. \
+               `search[0|<query>|1|<limit>]`. `<name>` is a placeholder: replace it with the \
+               value, and do not send the name itself.\n\
+             - Empty calls: `name[]` for zero-arg tools, or for a call sending no arguments.\n\
+             - A call whose indices are missing, non-numeric, or not in the signature is \
+               **rejected** — it will not run. Copy the numbers from the signature.\n\
              - Escapes inside argument values: `\\|` → `|`, `\\]` → `]`, `\\\\` → `\\`.\n\
              - You may emit multiple `<tool_call>` blocks in a single response. Each tag holds \
                exactly one call.\n\

crates/tinyagents-harness/src/tool_calling/dialect/test.rs

@@ -65,12 +65,12 @@ fn xml_dialect_embeds_the_full_schema_catalogue() {
 }
 
 #[test]
-fn pformat_dialect_parses_a_positional_call() {
+fn pformat_dialect_parses_an_indexed_call() {
     let registry = build_registry([("get_weather", weather_schema().parameters)]);
     let dialect = PFormatDialect::new(registry);
 
     let (_text, calls) = dialect.parse_response(&response(
-        "<tool_call>get_weather[London|metric]</tool_call>",
+        "<tool_call>get_weather[0|London|1|metric]</tool_call>",
     ));
 
     assert_eq!(calls.len(), 1);
@@ -85,7 +85,7 @@ fn pformat_dialect_falls_back_to_json_per_tag() {
     let dialect = PFormatDialect::new(registry);
 
     let (_text, calls) = dialect.parse_response(&response(
-        "<tool_call>get_weather[London|metric]</tool_call>\n\
+        "<tool_call>get_weather[0|London|1|metric]</tool_call>\n\
          <tool_call>{\"name\": \"other_tool\", \"arguments\": {\"x\": 1}}</tool_call>",
     ));
 
@@ -104,7 +104,7 @@ fn pformat_dialect_leaves_the_catalogue_to_the_prompt() {
     // (`get_weather` and `Call as:` still appear — as the syntax example and as
     // a pointer at the `## Tools` section that owns the real listing.)
     assert!(!instructions.contains("Look up the weather"));
-    assert!(!instructions.contains("get_weather[location|unit]"));
+    assert!(!instructions.contains("get_weather[0|<location>|1|<unit>]"));
     assert!(!PFormatDialect::new(PFormatRegistry::new()).embeds_tool_catalogue());
 }
 
@@ -532,15 +532,15 @@ fn catalogue_signature_matches_what_the_parser_reconstructs() {
     let rendered = render_pformat_catalogue(&tools);
 
     assert!(rendered.starts_with(CATALOGUE_HEADING));
-    assert!(rendered.contains("Call as: `get_weather[location|unit]`"));
+    assert!(rendered.contains("Call as: `get_weather[0|<location>|1|<unit>]`"));
 
     // The catalogue order is the order the parser assigns, not a coincidence.
     let dialect = PFormatDialect::new(build_registry([(
         "get_weather",
         weather_schema().parameters,
     )]));
     let (_text, calls) = dialect.parse_response(&response(
-        "<tool_call>get_weather[London|metric]</tool_call>",
+        "<tool_call>get_weather[0|London|1|metric]</tool_call>",
     ));
     assert_eq!(calls[0].arguments["location"], "London");
     assert_eq!(calls[0].arguments["unit"], "metric");

crates/tinyagents-harness/src/tool_calling/parse_test.rs

@@ -416,7 +416,7 @@ fn a_pformat_tag_does_not_suppress_a_sibling_glm_tag() {
     );
 
     let response = concat!(
-        "<tool_call>echo[hello]</tool_call>\n",
+        "<tool_call>echo[0|hello]</tool_call>\n",
         "<tool_call>shell/command>ls -la</tool_call>"
     );
     let (_narrative, calls) = parse_tool_calls_with_pformat(response, &reg);
@@ -446,7 +446,7 @@ fn a_pformat_tag_does_not_suppress_a_sibling_fenced_json_tag() {
     );
 
     let response = concat!(
-        "<tool_call>echo[hello]</tool_call>\n",
+        "<tool_call>echo[0|hello]</tool_call>\n",
         "<tool_call>\n```json\n{\"name\": \"shell\", \"arguments\": {\"command\": \"ls\"}}\n```\n</tool_call>"
     );
     let (_narrative, calls) = parse_tool_calls_with_pformat(response, &reg);
@@ -478,7 +478,7 @@ fn a_json_body_is_not_double_counted_by_the_glm_fallback() {
     );
 
     let response = concat!(
-        "<tool_call>echo[hello]</tool_call>\n",
+        "<tool_call>echo[0|hello]</tool_call>\n",
         "<tool_call>{\"name\": \"shell\", \"arguments\": {\"command\": \"cat a/b>c\"}}</tool_call>"
     );
     let (_narrative, calls) = parse_tool_calls_with_pformat(response, &reg);
@@ -510,7 +510,7 @@ fn a_tagged_body_still_honours_argument_key_aliases() {
     );
 
     let response = concat!(
-        "<tool_call>echo[hello]</tool_call>\n",
+        "<tool_call>echo[0|hello]</tool_call>\n",
         "<tool_call>{\"name\": \"shell\", \"args\": {\"command\": \"ls\"}}</tool_call>"
     );
     let (_narrative, calls) = parse_tool_calls_with_pformat(response, &reg);