Files
orchestra/AGENTS.md
T
2026-07-28 16:30:02 +04:00

118 lines
6.8 KiB
Markdown

# Orchestra
A Go implementation of `orchestra-spec (1).md` — an unattended multi-agent
task orchestrator that leases coding tasks to CLI harnesses (Claude Code,
Codex, opencode) running inside `herdr`-managed panes, rotates them across
context-window limits, and hands off work via a git-anchored continuity
protocol.
Layout: `internal/{domain,store,provider,registry,router,herdr,orchestrator,
continuity,federation,delivery,authz,operations,admin}` + `cmd/orchestra/main.go`.
## Ground truth over documentation
This repo has a documented history of code that *looks* wired but isn't —
packages with tests that pass in isolation while the live call path silently
no-ops (bare `continue` on error, discarded return values). See `AUDIT.md`
for the full audit and `progress.md` for a running log. **Before trusting a
claim in progress.md that something "works" or "is fixed," check the actual
call site** — the file is written by past sessions of this same assistant and
has previously overstated completion.
The single most reliable way to verify herdr-adapter code is right: don't
read `internal/herdr/adapter.go` and assume the method names are real. Ping
the live herdr instance and check.
## herdr protocol — verified against a live instance, 2026-07-27
- herdr speaks JSON-RPC over a raw TCP (or unix-socket) connection — **not
HTTP**. `internal/herdr/herdr.go`'s `Client.Call` is the only correct way
to talk to it; a bare `curl` to the port returns nothing.
- Request shape: `{"id":"<n>","method":"<name>","params":<object>}`. Herdr's
Rust JSON-RPC decoder requires `params` to be present and rejects a bare
`null` — always send `{}` for parameterless calls (the client does this
automatically).
- Full real method list is committed at `deploy/herdr-schema.json`, captured
live from `192.168.1.105:9245` (the `workpc` herdr) since no local `herdr`
CLI is available in this sandbox — the schema was reconstructed by sending
an unknown method name and reading the `unknown variant ... expected one
of ...` error, then probing each method of interest with `params:{}` /
`params:{pane_id:"nonexistent"}` to read Rust serde's `missing field
<x>` errors for its param shape.
- **Confirmed invented (do not use, they don't exist):** `pane.release`,
`pane.kill`, `pane.rotation_signal`, `pane.status`. If you see these
anywhere, it's a bug, not a valid call.
- **Real replacements:** `pane.close({pane_id})` for kill;
`pane.release_agent({pane_id, source, agent})` for release (structurally
different — does *not* return a `handoff_ref`, see below). No replacement
exists for `rotation_signal` — herdr has no concept of Orchestra rotation.
- **Architectural point that's easy to get wrong:** herdr never produces a
handoff. The agent writes the handoff artifact (§6.1 of the spec); herdr's
role in "release" is only to drop its own claim on the pane/agent binding.
Any adapter code that expects herdr to hand back a `handoff_ref` is wrong
by construction, independent of whether the method name is right.
- Protocol version is returned as a **JSON number** (`17`), not a string,
even though `config.jsonc` declares `"protocol": "17"` as a string.
`CheckProtocol`'s raw-bytes fallback happens to make this compare correctly
today — don't "clean up" that code without checking this note first, or it
might start doing a real numeric-vs-string comparison and break.
## Deployment topology (as of 2026-07-27)
- Runs as `orchestra.service` on **homesrv** (this machine's own systemd —
`journalctl -u orchestra.service` for logs; `sudo` isn't available in this
sandbox environment, but plain `journalctl` without sudo works here).
- Config: `.orchestra-config/orchestra.env` (env vars) +
`.orchestra-config/config.jsonc` (projects/machines/herdrs registry).
`ORCHESTRA_CONFIG=/etc/orchestra/config.jsonc` — the deployed copy, not the
repo's `deploy/config.example.jsonc`.
- Two machines in the registry: `homesrv` (192.168.1.104) and `workpc`
(192.168.1.105), each nominally running 3 herdrs (claude/codex/opencode).
In practice **homesrv has no local herdr running** (connection refused on
9245) — only workpc's herdr is live and reachable. `main.go` only logs
herdr connection *failures* at startup, never successes, so "no log line"
for a herdr does not mean it's down — check reachability directly.
- There was a real, live, stuck task as of 2026-07-27: workspace `wA`, task
id `06FT6CKD9Y98AZRX6X8K3QXFZG`, opencode harness, pane `wA:p1`,
`agent_status: "blocked"`. Likely stuck because rotation/release could
never reach it (B2/B5). Check whether it's still stuck before assuming
fixes here have taken effect operationally — code fixes don't retroactively
unstick an already-orphaned pane; that needs a manual kill/restart once the
release path is trustworthy.
## The federation fork — read before touching anything cross-machine
Two incompatible designs coexist. **Design A** ("drive the remote socket",
currently deployed via `clients/herdr-bridge.go`) has homesrv call
`worktree.create`/`agent.start` etc. directly on workpc's herdr over TCP as
if it were local — meaning anchor validation (`git rev-parse HEAD`) run by
the coordinator executes on the *wrong machine* relative to the actual
checkout. **Design B** ("workers pull tasks", `/v1/federation/*`) is fully
built server-side but has zero clients — no worker binary exists. Decision
(AUDIT.md, 2026-07-27): keep Design A through Phase 5, commit to Design B in
Phase 6, with two guardrails landed immediately (refuse to rotate/cleanup a
lease held by a non-local herdr rather than validate against the wrong
checkout). Don't build on top of Design A's cross-machine calls without
reading that section first.
## Working conventions
- **workpc worker deployment target:** copy the built worker binary to
`workpc:~/orchestra-deploy/orchestra-worker` (that is,
`/home/kami/orchestra-deploy/orchestra-worker`), not directly to
`/usr/local/bin`. The workpc deployment process installs from this staging
path. Verify the remote checksum and Go build revision before restart.
- `go build ./...`, `go vet ./...`, and `go test ./...` must all pass — `go
vet` was broken for a while (duplicate JSON struct tags) and nobody
noticed because only `build`/`test` were being checked. Always run all
three.
- Silent `continue`-on-error is the recurring bug pattern in this codebase
(adapter lookups, rotation, expiry). When touching `internal/orchestrator`
or `internal/herdr`, prefer a recorded/observable failure
(`MonitorHealth` fields) over a bare `continue` — that's literally what
turned B1/B2 invisible for as long as they were.
- Don't invoke destructive herdr calls (`pane.close`, `pane.release_agent`)
against a real pane from an investigative/audit session without asking
first — there is live operator state on the other end (see the stuck-task
note above).