dongarandClaude Opus 5.5
4904fa0b63
A model call that passed request_timeout_s failed the whole run, though the model was still writing. The chat route, which every local model takes, now streams its replies and puts them back together into the response the server would have sent, so the call record and parsing are unchanged. request_timeout_s becomes the longest the model may go silent. Measured on the 4090 through LiteLLM: an abandoned plain request kept the server busy (a tiny request after it waited 51 s); an abandoned stream freed it (0.3 s). A plain retry would have waited behind the abandoned call, so there is none. The passthrough route, hosted Anthropic models only, stays unstreamed: its cost comes only in a header that streaming loses. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01J5Utyfzw7gF7UhjRDEtjTk
129 lines
6.3 KiB
YAML
129 lines
6.3 KiB
YAML
# Example configuration for nisse-dev. Copy to nisse-dev.yaml (or point
|
|
# NISSE_CONFIG at your copy) and edit.
|
|
#
|
|
# - Structure lives here. Secrets do not: write them as ${VARIABLE} references,
|
|
# resolved from the environment (or .env, see .env.example) at startup.
|
|
# - Keep this file outside the workspace clone. Startup refuses otherwise,
|
|
# because the agent commits and pushes everything in the clone.
|
|
# - Unknown keys are errors, so a misspelt knob fails at startup instead of
|
|
# silently doing nothing.
|
|
|
|
state_db: ~/nisse-demo-workspace/state.db
|
|
# loop: new # the new conversational loop (M2.1); default old (control.py)
|
|
|
|
# D8: one entry per project. A second project is a second entry (the old loop
|
|
# runs exactly one). Changed by PR like the rest of this file.
|
|
projects:
|
|
demo:
|
|
repo: owner/demo-repo
|
|
base_branch: main
|
|
workspace: /root/work/demo # the clone, on the build host
|
|
build_host: # L13: every command runs here, over SSH
|
|
address: build-demo.example
|
|
key: ~/.ssh/nisse-dev # the agent's key for the build host
|
|
# user: root
|
|
# port: 22
|
|
push_key: ~/.ssh/push-demo-repo # a deploy key with write access (D4)
|
|
# labels: [discuss] # added to every issue's own (D1)
|
|
# secrets: # by name; values only by reference (D7)
|
|
# DATABASE_URL: {value: "${DEMO_DATABASE_URL}", visible: commands}
|
|
# DEPLOY_TOKEN: {value: "${DEMO_DEPLOY_TOKEN}", visible: harness}
|
|
# egress: [pypi.org] # beyond the default; recorded, not yet enforced
|
|
# trust: # the trust dial (L3, D5)
|
|
# merge: approved # I merge PRs an operator approved; default operator.
|
|
# # Needs protection: approvals only operators can give,
|
|
# # stale approvals dismissed, me on the merge whitelist.
|
|
|
|
llm: # L9: LiteLLM is the only LLM destination
|
|
base_url: http://litellm.example:4000
|
|
token: ${NISSE_LLM_TOKEN}
|
|
# Ceiling on each response. The model's own max_output_tokens in LiteLLM
|
|
# caps it further, which is where to hold a local model to less.
|
|
max_output_tokens: 16384
|
|
# How long the model may go silent before its request is abandoned. Local
|
|
# models' replies are streamed, so a slow one that keeps writing is never
|
|
# cut off; this bounds reading a long prompt, or a hung model.
|
|
request_timeout_s: 600
|
|
|
|
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-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
|
|
|
|
# Capability profiles (L9), from benchmarks: what one role is told about
|
|
# another (never a model's name). Optional; a model without one is briefed as
|
|
# "no profile recorded".
|
|
profiles:
|
|
qwen3.8-27b:
|
|
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-5-5:
|
|
tools: reliable
|
|
strengths: quick, well-scoped changes and short answers
|
|
# 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
|
|
weaknesses: cost; use it for what needs it
|
|
|
|
matrix: # L5
|
|
homeserver: https://matrix.example
|
|
user_id: "@nisse-dev:example"
|
|
token: ${NISSE_MATRIX_TOKEN}
|
|
room_id: "!roomid:example"
|
|
operators: ["@you:example"]
|
|
|
|
gitea: # D4, D5
|
|
url: https://gitea.example
|
|
token: ${NISSE_GITEA_TOKEN}
|
|
operators: [your-gitea-login]
|
|
|
|
# The knowledge vault (architecture L10), read through note tools. Optional:
|
|
# without it, the agent has no note tools. Only the folders listed exist for
|
|
# the agent: rw ones it writes (one writer per folder), ro ones it only reads.
|
|
vault:
|
|
root: /var/lib/nisse/vault
|
|
folders:
|
|
memory/nisse-dev: rw # its own memory
|
|
notes: ro # the operator's material
|
|
# ignore: [".*", "_plug"] # the editor's own files, never notes (default)
|
|
|
|
# What each knob does, and how to choose a value: docs/Tuning-Guide.md.
|
|
# These are the defaults; the section can be left out entirely. The file is
|
|
# read at startup, so a change takes effect on the next restart.
|
|
tuning:
|
|
driver:
|
|
max_turns: 40 # model calls per driver run before it stops as capped
|
|
failure_limit: 2 # consecutive failed commands before escalating (L9)
|
|
circle_repeats: 3 # a run repeating itself this often ends, escalated (L2); 0: never
|
|
circle_similar: 0.9 # how alike its repeats must be
|
|
preflight: # the toolchain step (D3)
|
|
max_turns: 20 # model calls one toolchain step may make
|
|
command_timeout_s: 900 # one command in it: installs take a while
|
|
expert:
|
|
consult_budget: 3 # ask_expert calls per milestone (L9)
|
|
delegation: # hard caps on one delegation (L9)
|
|
max_turns: 30
|
|
max_tokens: 200000
|
|
wall_clock_s: 900
|
|
tools:
|
|
command_timeout_s: 120 # one run_command before it is killed
|
|
max_output_chars: 4000 # tool output kept for the model, head and tail
|
|
context: # fitting each request into the model's window
|
|
keep_recent_turns: 6 # turns never cut or dropped
|
|
safety_margin: 0.1 # fraction of the window held back
|
|
loop: # the new loop (loop: new)
|
|
bug_grace_s: 600 # a `bug` restatement proceeds after this long unanswered
|
|
general_idle_s: 86400 # a conversation closes after this long without a message
|
|
memory:
|
|
max_turns: 10 # model calls the memory pass at an item's close may make
|
|
build_vm: # an unreachable build VM (D3, #137)
|
|
retry_s: 300 # how often it is tried again
|
|
settle_s: 180 # once it answers, how long until work goes on
|