# Agent settings.
#
# Every key is optional: what you leave out keeps its built-in default
# (see internal/config/config.go). Nothing here requires recompiling —
# `go run .` picks the file up at startup.
#
# Use another file with:  AGENT_CONFIG=./fast.yaml go run .

# Which server is behind the URL. Left out, it means `dmr`: every agent.yaml
# written before this key existed keeps working unchanged. `llamacpp` speaks
# the same OpenAI-compatible protocol, but with its own port, its own way of
# refusing tool calls (no --jinja) and its own words in error messages — the
# provider is what turns those into one useful line. See agent.llamacpp.yaml.
# AGENT_PROVIDER overrides this key for one run.
provider: dmr

# The chat model, in the provider's own naming (DMR: `ai/…`, `hf.co/…`). It
# MUST be able to emit tool_calls, otherwise the agent never runs a single
# command. AGENT_MODEL overrides it.
model: huggingface.co/jetbrains/mellum2-12b-a2.5b-instruct-gguf-q4_k_m:Q4_K_M

# Endpoint. `baseUrl` is tried first; if it does not answer, `fallback` is used
# — handy from a container or a sandbox, where the host is reachable as
# host.docker.internal. Both are LEFT OUT here on purpose: each provider has its
# defaults, and an explicit value is always respected — so with DMR's URL
# written in this file, `AGENT_PROVIDER=llamacpp ./bob agent.yaml` still talked
# to port 12434. Leaving them out keeps the file provider-neutral. These are the
# values the `dmr` provider fills in:
#   baseUrl: http://localhost:12434/engines/v1
#   fallback: http://host.docker.internal:12434/engines/v1
# AGENT_BASE_URL — or, for dmr, the older DMR_BASE_URL — overrides both without
# probing. `fallback: ""` (an explicit empty string) disables the fallback.

# The NAME of the environment variable holding the API key — never the key
# itself, so this file can be committed and shown on screen. DMR ignores the
# key, so nothing is needed here; llama-server checks one only when started
# with --api-key (default variable: LLAMA_API_KEY).
# apiKeyEnv: LLAMA_API_KEY

# How many tokens the server actually serves. The model does not know this:
# it is what the operator started the server with (`llama-server -c 32768`,
# `docker model configure`). Shown at start-up next to its origin, because a
# number without an origin never gets corrected. 0 = ask the server
# (llama-server tells its n_ctx on /props), else "unknown".
contextWindow: 0

# Max number of characters a tool returns to the model. Context safeguard:
# beyond that, the output is truncated (beginning + end kept).
maxOutput: 16000

# Max number of model ↔ tools round trips for a single question. Raise it for
# tasks that need long chains of commands.
maxTurns: 40

# The built-in file tools: read_file, write_file, edit_file. This is the switch
# this part exists for. `true`: the model edits files through tools it can SEE
# in its tool list. `false`: the agent is part 09 again — bash and read_skill —
# and edits files through the `edit` CLI if it is on the PATH. Same binary,
# same prompts, two set-ups: measured side by side, that is the comparison
# (in part 07, `read_skill` as a tool was loaded 3/3 times where a catalogue
# in the prompt plus `cat` managed 1/11). Paths are relative to the current
# directory and are not confined to it.
editTools: true

# Where the `read_skill` tool looks for its markdown procedures. The path is
# relative to the CURRENT DIRECTORY, so run the agent from the directory that
# holds skills/ — not from src/. When the directory has no *.md file, the tool
# is not declared at all.
skillsDir: skills

# How many lines of a command's output are echoed to the terminal. "Show me
# that file" means show it: running `cat` is not displaying it. 0 disables it.
previewLines: 20

# The system prompt: what the agent is, and what it is allowed to do.
# This is the knob to play with — it is the shortest path to changing behaviour.
system: |
  Your name is Bob.
  You are a coding agent working in a terminal.
  You have a "bash" tool to run shell commands.
  Use it to explore files, run tests, inspect the repository, etc.
  Chain several commands if needed, then answer clearly in English.

  A request often mixes things you answer from yourself ("say hello") with
  things only a command can answer ("list the files"). Handle every part, in
  the order asked, and run a command for each part that needs one.
  Never state the contents of a file, the output of a command, or the state of
  the repository unless a command in THIS answer returned it. What you did not
  read, you do not know: run the command instead of recalling it.

  SKILLS
  You have a second tool, `read_skill`. Its description lists the procedures
  available for this project — one per kind of task.

  Any request to DO something to a Go project is a skill, not a shell command
  you invent. Match the request against that list, call `read_skill` FIRST,
  before any bash command, and then follow what it says step by step.

  FILE EDITING
  You have three tools for files: `read_file`, `edit_file` and `write_file`.
  They are how a file gets read and changed here: each change is exact,
  checked before it is written, and comes back as a diff with line numbers.
  bash is for running things — building, testing, listing, searching.

  - Read before you write: call `read_file` on the file (numbered=true when
    you need line numbers). You cannot target text you have not seen; never
    rely on what you think you remember about a file.
  - To change an existing file, call `edit_file` with one or more {old, new}
    pairs. `old` is copied from the file character for character — same
    spaces, same indentation, same line breaks — and appears exactly once:
    add the surrounding lines until it is unique. Several pairs are applied
    together, against the original file. An empty `new` deletes the text.
  - Call `write_file` only to create a file, or to rewrite one entirely and
    on purpose. On an existing file it replaces everything, including what
    you did not intend to touch.
  - Read the diff the tool returns: it says exactly what changed and on which
    line. If `edit_file` refuses — text not found, ambiguous, overlapping
    edits — read the file again and fix `old`. Do not fall back to
    `write_file` to force the change through.
  - After editing code, run the narrowest check with bash: the formatter, the
    compiler, or the test covering that file.

  RULES
  - Keep everything the file already does, unless the user asked to remove it.
  - Touch only the files the request is about. Do not add tests, files or
    features that were not asked for.
  - Never run a git command unless the user says git, commit or push.
  - Never move, rename or delete a file unless the user asked for it.
  - Then answer in English, in a few lines.
  - If you don't know how to use a <cli>, run `<cli> --help` (or `<cli> help`)
    to understand the options, then run the command.

  BACKGROUND JOBS
  Never let a command block the answer. Anything that serves, watches or runs
  long goes to the background, with BOTH streams redirected and its pid kept:

    nohup <command> > /tmp/<job>.log 2>&1 & echo $! > /tmp/<job>.pid

  Redirecting only stdout still blocks until the process exits. Read the
  `bg-jobs` skill before you wait on, inspect or stop such a job — each has a
  rule you cannot guess. Stop every job you started before you finish, and say
  which ones you left running.

# Generation settings (OpenAI API keys). Low temperature for a coding agent:
# we want precise and reproducible answers, not creativity.
sampling:
  temperature: 0.0
  # Une seule commande par tour. Sans ça, le modèle peut en demander
  # plusieurs d'un coup : Genkit les exécute EN PARALLÈLE, dans le même
  # répertoire, et on ne sait plus quelle sortie appartient à quelle
  # commande. Pour le détecteur de boucle, c'est pire : deux appels du même
  # tour s'enregistrent dans un ordre indéterminé, et « la même action deux
  # fois de suite » perd son sens. Mettre `true` pour retrouver le
  # comportement par défaut de l'API.
  parallel_tool_calls: false
  top_p: 0.9
  max_tokens: 4096

watchdogTimeout: 30s

# Context compression, carried over from part 08 (its CONTEXT_WINDOW.md has the
# reasoning). OFF by default: with `enabled: false` the agent behaves exactly
# as before. Why it exists: nothing in the agent ever shortens the history —
# measured with the fake engine of part 03, the messages sent to the model grew
# 2 → 5 → 7 → 9 over four requests, and every `bash` output (up to maxOutput
# characters) and every skill read stays until /quit. On a local model the
# window is fixed at load time and the prefill runs on the laptop, so a long
# session first shows up as a false "[watchdog: connection hang]", then as an
# error from the server.
#
# The window it measures against is the top-level `contextWindow` above — ONE
# key for the banner and for the trigger, so the two cannot disagree. When
# that key is 0, the value the provider's probe learned from the server is
# used (llama-server tells its n_ctx on /props; DMR tells nothing). When both
# are unknown, only `maxMessages` can trigger, and the agent says so at
# start-up.
context:
  enabled: false

  # Compress when the history reaches this share of the window. 75 leaves a
  # quarter for the next question, the outputs of its turns and the answer —
  # a single `bash` output can be maxOutput characters, about 4-5k tokens.
  threshold: 75

  # Fallback on the message count, for when the window is unknown or the
  # estimate is off. One command costs 2 messages (call + response), so 80 is
  # roughly 30 commands of history.
  maxMessages: 80

  # Question turns kept raw at the end. The recent turns are where the model
  # works; summarising them makes it re-run what it just did. Three covers
  # "do X", "now fix it", "and test it" — the usual shape of a demo.
  keepLastTurns: 3

  # max_tokens of the summary request. Seven sections of one line per item
  # fit in far less; the cap stops a runaway model from filling the window it
  # was asked to empty.
  summaryMaxTokens: 1200

  # Replaces the built-in summary prompt (08-context-compression/CONTEXT_WINDOW.md
  # § 4.1) when set. Empty = built-in.
  prompt: ""

  # One line after each compression: what was replaced, what was kept,
  # tokens before → after. Worth showing on a screen, noise in a log.
  showStats: true

displayCommands: true