Merge pull request 'Expert and integration roles on claude-haiku-5-5' (#134) from haiku-5-5 into main

docs/Development-Plan.md

@@ -52,8 +52,8 @@ Only `nisse-dev` has real code; `nisse-core` and `nisse-ops` are placeholders.
 - **Local inference:** llama-swap on an RTX 4090 host (320 W cap). The benchmark
   (PR #55) chose `qwen3.8-27b` for local work.
 - **Models by role today:** driver and helper `claude-opus-5-5`, expert
-  `claude-haiku-4-5`. `qwen3.8-27b` has had no role since the executor's went
+  `claude-haiku-5-5`, classifier `qwen3.8-27b` (M3.2); the driver benchmark
-  (M2.8); the driver benchmark measures where it fits.
+  measures where the local model fits beyond that.
 - **Bring-up** runs on this development VM (`scripts/bringup.sh`), against test
   resources: a test repository, the test Matrix room, a budget-capped LiteLLM key;
   configuration in `/root/nisse-e2e/`.

nisse-dev.example.yaml

@@ -48,7 +48,7 @@ roles:                                # L9: role -> LiteLLM model alias
   # Each work item pins the mapping when it opens.
   driver: claude-opus-5-5       # drives the loop: talks, restates, works
   helper: claude-opus-5-5       # takes over what the driver got stuck on
-  expert: claude-haiku-4-5      # one question at a time, no tools: cheap is enough
+  expert: claude-haiku-5-5      # one question at a time, no tools: cheap is enough
   # Reads go-aheads and judges actions the fast path doesn't know (L4); unset,
   # the expert's. It never sees the transcript, so the driver's model is fine.
   classifier: qwen3.8-27b
@@ -61,10 +61,11 @@ profiles:
     tools: reliable (37/40 on the executor benchmark, PR #55)
     strengths: careful multi-step edits that follow a clear brief
     weaknesses: speed (minutes per hard task) and open-ended design
-  claude-haiku-4-5:
+  claude-haiku-5-5:
     tools: reliable
     strengths: quick, well-scoped changes and short answers
-    weaknesses: long multi-file work (15/20 on the harder benchmark tasks)
+    # Measured on its predecessor, claude-haiku-4-5; not benchmarked yet.
+    weaknesses: long multi-file work (15/20 on the harder benchmark tasks, as 4.5)
   claude-opus-5-5:
     tools: reliable
     strengths: design, debugging, and work other models got stuck on

tests/integration/e2e_support.py

@@ -16,7 +16,7 @@ from nisse_dev.gitea import GiteaClient, GiteaError
 from nisse_dev.llm import RoleMap
 
 # The end-to-end tests run every role on a local model, which costs nothing
-# per call; NISSE_E2E_MODEL overrides it (e.g. claude-haiku-4-5 for a quick,
+# per call; NISSE_E2E_MODEL overrides it (e.g. claude-haiku-5-5 for a quick,
 # paid run). The no-thinking alias (LiteLLM: chat_template_kwargs
 # enable_thinking false): with thinking, some answers ran to 6-8k tokens,
 # minutes each, and stretched a test from 3.5 to 20 minutes.

tests/integration/nisse-test.yaml

@@ -36,9 +36,9 @@ llm:
 # not local-model competence (measured separately by test_model_capability),
 # so every role uses a model that holds the tool-call contract.
 roles:
-  driver: claude-haiku-4-5
+  driver: claude-haiku-5-5
-  helper: claude-haiku-4-5
+  helper: claude-haiku-5-5
-  expert: claude-haiku-4-5
+  expert: claude-haiku-5-5
 
 matrix:
   homeserver: https://matrix.dongar.net