Tool-call parsing splits coding agents into tiers on some models, not on native tool-call models

The failure on qwen2.5-coder starts in parsing. Served through Ollama, all three sizes of that model write their tool calls as plain text inside the reply ^[9]. The text is a usable `{"name": "edit_file", ...}` block. An agent that listens only on the native tool channel sees no call. It treats the reply as a final answer, and the task ends with nothing done ^[9]. According to the post, pi, Hermes and opencode made no tool call in any of the 30 runs on those models ^[9]. goose made some calls but completed none of the tasks ^[10]. The exception is goose's toolshim. It hands the model's text to a second, small model whose job is to rewrite it as a proper tool call ^[11]. Running a second model to read the first one's output is a lot of machinery. On qwen2.5-coder 32B it won anyway, scoring 28 against Polyglot's 24 ^[11]. Under the pre-set tiers, 28 is reliable and 24 is works sometimes ^[3]. On gpt-oss, which emits calls in its own trained format, the same shim dropped goose from 18 to 5 [c12, c13]. That moved goose from works sometimes to fails ^[2]. Polyglot does the job with a parser ^[14]. A parser handles the formats it was written for. Outside the 7B, Polyglot missed 13 of 150 runs, and the author blames the parser for up to 5 of them: a model stuttering an empty tag before the real one, or running past a broken closing tag into an invented next turn ^[20]. Those formats, and the others found in reruns, are fixed in 0.13.2, according to the author ^[20]. The tier cut-offs were set before the first run ^[3]. With 30 runs, 28 successes is consistent with a true rate anywhere from 79% to 98% ^[4]. "So a 30 next to a 28 is a tie, not a win," the author wrote ^[5]. Polyglot's own builds show the noise. Three pre-release builds of 0.13.2 scored 179, 173 and 172 out of 210, all in the same tiers ^[21]. That is a spread of seven runs, or about 85% against 82% ^[4]. The author attributes the spread to the models, not the code ^[21]. Where models emit calls through the native channel, the agents converge. Every agent was reliable on Qwen3.8-27B, four of six on qwen3-coder and three on Devstral ^[7]. "If you run one of these models, pick your agent on other grounds," the author wrote ^[8]. Prompt size is one. Measured from Ollama's request log, Polyglot opens a task with about 1.2k tokens of fixed prompt and pi with 1.2-1.6k, against 3.9-4.7k for Hermes and 5.5-7.6k for goose and opencode ^[16]. Ollama's prompt cache is reused better by pi than by Polyglot, a gap the author called "something for me to fix, not to advertise" [c17, c18]. Prompt size also explains the correction to the earlier post. Ollama defaults to a 4,096-token context and silently cuts anything longer ^[15]. The opening prompts of goose and opencode overshoot that by roughly 1.4k to 3.5k tokens before any work starts ^[5]. In the first runs, goose, Hermes and opencode were working with most of their instructions missing ^[15]. Guarding against this is the best engineering in the post. Every model now runs at a 32k context, and the harness refuses to start below 16k ^[19]. After each agent's runs it scans Ollama's log and marks the results invalid if it finds a truncated prompt ^[19]. The author builds Polyglot, the only agent that does not fail on any of the seven models [c1, c6]. Harness, raw results and failed-run transcripts are on GitHub ^[1]. They replace a first version that used three runs per task and a private harness ^[22]. No rerun by anyone else is reported in the post.