Files
archie/config.example.toml
T
sam 101c133239 feat(capture): store unbound webhook events, retention-bounded
Adds POST /webhooks/capture/{source}: an unauthenticated endpoint (mounted
alongside /healthz, bypassing dashboard token auth by design) that persists
any inbound webhook archie has no binding for yet, so it can be inspected
and mapped later (t2db.2/t2db.3).

docs/prds/event-capture-storage.md decides SQLite over NATS JetStream for
storage: NATS is not unconditionally available until 7d5u.2 ships embedded
support, and capture's single-writer/single-reader access pattern doesn't
need JetStream's fan-out/replay machinery anyway.

internal/store/captures.go adds a captured_events table with prune-on-write
(retention age + row-count cap, both configurable). The HTTP handler applies
webhookguard redaction before persisting, and enforces two more independent
disk-bound mechanics before that: a per-request body-size cap and a
per-remote-address rate limit. New [capture] config section with documented
defaults; all fields optional, capture is on by default.

Adversarial review (manual, after the usual fresh-subagent pass hit a
session API limit mid-run) caught and fixed three real defects before this
landed, each with a regression test confirmed to fail against the pre-fix
code:
- the body-size cap silently no-op'd when CaptureMaxBodyBytes was left
  unset, the one bound that limits a single payload's size rather than row
  count or write rate
- the rate limiter was keyed by the request's own source path segment --
  unregistered and attacker-chosen by design for this endpoint, violating
  webhookguard.RateLimiter's documented bounded-key-space assumption and
  letting a sender bypass the limit by rotating the segment; rekeyed to
  remote address
- the retention prune compared time.RFC3339Nano strings, whose trimmed
  variable-width fractional seconds break lexicographic ordering within the
  same second; switched to fixed-width RFC3339

Closes archie-core-t2db.1.
2026-08-22 04:51:25 +10:00

264 lines
12 KiB
TOML
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
#
# archied configuration
#
# Secrets are NOT stored here: each one is a {engine, key}
# reference resolved through a secrets engine at startup.
# The built-in "env" engine reads key as an environment variable name;
# "bws" (Bitwarden Secrets Manager) is compiled in. Custom engines are
# Yaegi plugins loaded from this directory -- reference implementations
# for sops, vault/openbao and age ship in examples/secret-engines/ and
# each needs its CLI on PATH plus engine-specific env (sops: SOPS_FILE;
# vault: VAULT_ADDR, VAULT_TOKEN, optional VAULT_SECRET_PATH;
# age: AGE_FILE, AGE_IDENTITY).
# secret_engine_dir = "~/.config/archie/secret-engines"
# work_dir = "~/.local/share/archie/work" # default
# db_path = "~/.local/share/archie/archie.db" # prefix for independent task and conversation SQLite files
poll_interval = "60s"
label = "archie"
bot_user = "vectis-labs-archie"
diff_cap_lines = 400
# how many operator-triggered retries a parked task gets
# before it moves to dead (global default; set per-repo to
# override). Enforced by the dashboard's retry action.
# max_retries = 3
[forge]
type = "github"
host = "https://github.com"
token = { engine = "env", key = "ARCHIE_GITHUB_TOKEN" }
[dispatch]
# Can be of "assignee", "label", or "either"
# this sets the method archie uses to pick up tasks.
trigger = "assignee"
# Optionally set a reaction to show that archie has picked up a task.
# setting this to "off" disables the pickup reactions.
ack_reaction = "eyes"
# dispatch.labels maps each task lifecycle state to the label
# Archie applies on the issue.
# You can rename labels to match your project's conventions,
# or comment one out to skip that transition.
# The labels are the *source of truth* for the dispatch loop: Archie
# scans for issues bearing queued, then moves them through working →
# waiting → pr → parked (→ dead when max_retries is exhausted).
[dispatch.labels]
queued = "archie:queued"
working = "archie:working"
waiting = "archie:waiting"
pr = "archie:pr"
parked = "archie:parked"
# dead = "archie:dead" # state label for a task that exhausted max_retries
[models] # role -> ai-sdk runtime ref (used by agent workflows)
triage = "openai/gpt-5.4"
planner = "openai/gpt-5.4"
builder = "openai/gpt-5.4"
[providers.openai]
class = "openai"
api_key = { engine = "bws", key = "OPENAI_API_KEY" }
# api_key_env = "OPENAI_API_KEY" # backwards-compatible environment form
[providers.ollama]
class = "ollama"
# [chat]
# `operator` is the display name of user. It is
# passed to the chat agent as runtime metadata, which means it is sent to the
# configured model provider on every turn -- leave it unset to tell the agent
# nothing about who is running it.
# operator = "John Smith"
# `workspace` roots the chat agent's file and shell tools (read, write, edit,
# shell, grep, find). If unser, the tools aren't registered.
# Set it to let the agent read, modify and execute inside that directory
# choose it deliberately
# workspace = "~/.local/share/archie/workspace"
# `max_steps` caps the model/tool round-trips in a single chat turn
# (default 100). This is a task budget, not a safety limit.
# max_steps = 100
# Let read/write/edit/find/grep tools reach absolute paths outside the
# workspace jail. Relative paths still resolve against workspace.
# The shell tool is never confined, so a jail only pushes the agent into
# using shell for everything, losing truncation, line numbering, and
# read-before-write protection. Enable when the agent needs to work
# cross-filesystem (disk ops, system edits, etc.). Grant deliberately.
# unrestricted_filesystem = true
# Render each completed tool call inline in the reply, naming the tool and one
# line of its result, before the answer built from them. Off by default: it is
# there to make a wrong tool choice visible from the chat, at the cost of
# narrating every internal step. Applies to every chat channel -- Telegram
# and the web dashboard both read this one setting.
# show_tool_calls = true
# Optional interactive-chat catalog. /provider groups these references by
# their first path segment, and /model shows only the selected provider's
# entries. When omitted, Archie uses the distinct refs in [models].
# models = [
# "openai/gpt-5.6-sol",
# "openrouter/openai/gpt-5.6-sol",
# ]
# Optional self-hosted update adapter. Both values are argv arrays (never
# shell snippets). The check command must write a releaseupdate.Snapshot JSON
# document to stdout; configure the installer as well to show Approve/Defer.
# [chat.telegram]
# update_check_command = ["/usr/local/bin/archie-update-check"]
# update_install_command = ["/usr/local/bin/archie-update-install"]
# External MCP tools are launched and managed by Archie's MCP client. A
# command of "npx" gets a persistent package cache under work_dir
# (top of this file) automatically, so a daemon restart reuses it instead
# of re-resolving and re-downloading the package from the npm registry
# every time. This does not cover a package's own separate downloads (e.g.
# a Playwright/Puppeteer-based server fetching a browser binary from a CDN
# via its postinstall script) -- consult that package's own env vars for
# those.
# [[tools.mcp_servers]]
# name = "desktop-commander"
# transport = "stdio"
# command = "npx"
# args = ["-y", "@wonderwhy-er/desktop-commander@0.2.46", "--no-onboarding"]
# work_dir = "/path/to/workspace"
[agent]
# How the daemon invokes the agent:
# "inprocess" – run agent logic directly in the daemon (default)
# "subprocess" – spawn `command` as a child, communicating over stdin/stdout
# "nats" – dispatch tasks over NATS (requires [nats]; add [containers] for Docker sandboxing)
mode = "inprocess"
# Binary to launch when mode = "subprocess".
command = "archie-agent"
# Environment variables forwarded to the agent process (subprocess/nats modes).
# env = ["GOCACHE"]
# Workspace code-search index. Paths default under work_dir when unset.
[indexing]
# Code-search index files (default: <work_dir>/indexes).
# index_dir = "/var/lib/archie/indexes"
# Per-index lifecycle metadata DB (default: <work_dir>/workspace-indexes.db).
# db_path = "/var/lib/archie/workspace-indexes.db"
# Web dashboard and API: task board, activity stream, chat, config,
# logs, memory, SSE events. Served as an embedded SPA with a full REST
# and streaming API. Auth is token-based for non-loopback addresses;
# loopback binds (127.0.0.1, ::1, localhost) skip auth automatically.
# Set to "off" (or leave empty) to disable.
[web]
listen = "127.0.0.1:8484"
# Notifications for human-in-the-loop moments
# (approval requests, gate failures, parked tasks).
# Archie POSTs a JSON payload to the webhook
# wire it to n8n, email, Slack, or any HTTP endpoint.
[notify]
# webhook = "https://n8n.example.com/webhook/archie"
[budgets]
max_steps = 60
max_tokens = 400000
wall_clock = "30m"
gate_max_failures = 5
# NATS JetStream for distributed task queue. Remove or leave commented out
# to run in single-process SQLite mode.
# [nats]
# url = "nats://localhost:4222"
# token_env = "NATS_TOKEN" # optional; named variable must be set
# Unbound webhook capture: POST /webhooks/capture/{source} accepts and
# stores any inbound payload archie doesn't yet have a binding for, so it can
# be inspected and mapped later (see docs/prds/event-capture-storage.md).
# Stored in the same SQLite database as tasks, not NATS -- capture has one
# writer and one (human, dashboard) reader, so it doesn't need JetStream's
# fan-out/replay machinery, and SQLite is available in every deployment
# shape unconditionally. All fields default when the section is absent;
# capture is on by default.
# [capture]
# retention = "168h" # 7 days; prune-on-write deletes anything older
# max_events = 5000 # row cap, pruned on every insert regardless of age
# max_body_bytes = 262144 # 256 KiB; larger POSTs are rejected with 413
# rate_per_second = 1.0 # per-remote-address token bucket
# rate_burst = 5
# Docker sandbox execution of archie-agent. Requires [nats] and agent.mode = "nats".
# [containers]
# enabled = true
# image = "archie-agent:latest" # Docker image to run
# max_concurrency = 4 # max simultaneous tasks/containers
# max_uptime = "60m" # container lifetime cap
# volume_ttl = "72h" # max age for persistent repo volumes and Git clone caches
# pull_policy = "missing" # "missing" or "always"
# At least one [[repos]] entry is required before archied will poll
# anything -- it is deliberately left commented rather than shipped active:
# an active example here validates as a normal, working repo (owner/name are
# only checked for non-empty), so archied would silently start polling a
# repository that doesn't exist. archied setup does not write this section
# ([[repos]] is an array of tables, which the text-patching setup uses
# cannot target) -- add at least one entry by hand.
#
# [[repos]]
# owner = "your-github-username"
# name = "todo"
# base = "main"
# ecosystem = "go" # "go", "python", "node", "rust", or "custom"
# # Override the ecosystem defaults when needed:
# # preflight = [["go", "version"]]
# # test_glob = "*_test.go"
# protect = ["_templ.go"]
# gate = [
# ["templ", "generate", "./components/"],
# ["go", "vet", "./..."],
# ["go", "build", "./..."],
# ["golangci-lint", "run"],
# ["go", "test", "./...", "-count=1"],
# ]
# # allow_concurrent = true # let the daemon run multiple tasks for this repo at once instead of
# # FIFO-serializing them; only safe when concurrent worktrees won't collide
# # (e.g. different base branches or packages per task). Still bounded by
# # [containers].max_concurrency globally.
# # persistent_storage = true # reuse a per-repo Docker volume and Git object cache until volume_ttl
# ── Tools ───────────────────────────────────────────────────────────────────
# Limits on what a tool may return. These bound how much of the context window
# one tool call can spend: the workspace file tools truncate themselves, but
# skill bodies and MCP server results otherwise arrive whole.
#
# Zero (or absent) takes the default. A NEGATIVE value disables the limit --
# zero cannot mean "off" because it is indistinguishable from an unset key.
# [tools.tool_policy]
# max_result_chars = 50000 # cap on one tool result; -1 for no cap
# turn_budget_chars = 200000 # cap on all tool results in one chat turn; -1 for no cap
# spill_dir = "" # unset by default, which truncates oversized results
# # inline. When set, a result over the cap is written
# # here in full and the model is handed the path.
# # It MUST be inside chat.workspace: the read tool
# # refuses paths outside it, so a spill anywhere else
# # is a reference the model cannot open -- worse than
# # truncating. Startup warns if it is outside.
# # Task agents work in their own worktree and cannot
# # read it at all; they always truncate.
# The web_fetch tool: retrieve a URL and return readable text.
# [tools.web_fetch]
# enabled = true # set false to withdraw the tool entirely
# timeout = "30s"
# max_bytes = 2000000
# allow_private_networks = false
# # Off by default, and worth leaving off. archied's own dashboard, NATS and
# # the Docker API answer on loopback and private addresses, and the URL to
# # fetch arrives through chat -- an untrusted path, since a forge issue body
# # or an inbound email can be quoted into a turn. The address is checked at
# # dial time rather than on the URL string, so a public name that resolves to
# # a private address is refused too. Turn this on only for a homelab endpoint
# # you intend Archie to reach.