`harness events` serves POST /hooks/{redmine,discourse,gitea} and
GET /healthz on stdlib net/http until SIGINT/SIGTERM (graceful
shutdown). Flow per delivery: verify (401, generic body) -> normalize
(400) -> append-only store with dedup (200 stored/duplicate + id +
action) -> hand stored actionable events to the conductor via the
Dispatcher interface. Conductor.DispatchEvent is the wiring point and
prints what it will do once phase 3 lands the real event-to-turn
dispatch. Body cap 1 MiB (413); audit log carries normalized fields +
digest only, never headers, secrets or payload. Server refuses to start
without at least one resolvable webhook secret.
Signature verification (gitea hex HMAC-SHA256 over the raw body via
constant-time hmac.Equal; redmine/discourse constant-time shared-secret
headers) with one generic ErrUnverified so rejects give attackers no
oracle. Tolerant normalization of the known Redmine/Discourse/Gitea
payload variants into one Event record (canonical subject ids, actor,
title, repo, sha256 payload digest) plus the DESIGN action mapping
(dispatch_turn / respond_turn / pipeline_step / ignore). Store: JSONL
under state dir, 0600, dedup keyed on provider event id (delivery
header, payload-digest fallback), index rebuilt at startup so replays
across restarts still dedup; torn tail lines skipped, not fatal.
[events] listen/state_dir plus per-source secret refs (env:/file:/literal:)
for redmine, discourse and gitea webhooks. Defaults: :4100, state/events,
X-Redmine/X-Discourse-Webhook-Secret headers. Refs validate fail-fast; set
refs must be well-formed or config load names the [events.*] section.
`harness once` runs one bounded iteration (intake -> routing -> bounded turn
-> REPORT) then exits, so callers chain it without a daemon. The bounded
turn loops request -> tool calls -> results up to max_rounds; gate denials
feed back to the model and are counted instead of failing the turn.
--dry-run prints the plan (task, resolved model, tool bounds) and makes zero
LLM calls; --demo runs the [demo] issue ("tell me about yourself" through
LiteLLM -> GLM self-description as the REPORT), the MVP acceptance bar.
Distinct exit codes for config/usage, intake, and llm failures. Ships
harness.toml.example (scope query, tier map, allow-lists); real configs are
gitignored. Loop tests run end-to-end against a scripted fake OpenAI server:
demo turn, dry-run zero-call, tool round-trip, denial counting, round limit,
and error-class mapping.
💘 Generated with Crush
Assisted-by: Crush:glm-5.2
OpenAI-compatible chat client for the LiteLLM proxy (base_url normalization,
Bearer auth, retry/backoff on 429/5xx/transport, usage accounting) - stdlib
http only. Intake lists issues in scope from /issues.json with the task
class read from a configurable custom field. Writeback lands each turn as
REPORT-<vertical>-<task>-<ts>.md plus REPORT-latest.md (atomic rename) with
model/tier/token telemetry. The bash tool ports maki's permission semantics
without tree-sitter: segment-by-segment compound-command checks, deny beats
allow, word-boundary "cmd *" matching, $()/backtick/subshell denied, output
truncation, per-command timeout with process-group cleanup.
💘 Generated with Crush
Assisted-by: Crush:glm-5.2
harness.toml loading via a stdlib-only TOML subset parser (tables, bare
keys, strings/ints/bools, multi-line arrays; anything richer fails loudly).
Secrets are refs only (env:/file:/literal:, bw: reserved) and are redacted
from every error path. Model routing v0: static class -> tier alias ->
concrete model map per the DESIGN model-selection layer; requests carry the
resolved concrete model and unknown classes fail hard so routing stays
auditable. Table-driven tests cover the parser, validation, key refs, and
routing decisions.
💘 Generated with Crush
Assisted-by: Crush:glm-5.2
mrcharles