Files
agent-tools/README.md
T
unkin-agent 387653a3c0
ci/woodpecker/pr/test Pipeline was successful
ci/woodpecker/pr/pre-commit Pipeline was successful
ci/woodpecker/pr/build Pipeline was successful
Distrust origin/<branch> when prune's fetch fails
A stale remote-tracking ref survives a failed fetch and the next successful
--prune deletes it, so it cannot prove a branch's commits survive upstream.
Record whether the pruning fetch succeeded and gate the origin/<branch>
existence and containment proofs on it; a failed fetch removes the worktree and
keeps the branch. Local-object proofs and the merged head SHA are unaffected.
2026-09-10 00:18:54 +10:00

267 lines
11 KiB
Markdown

# agent-tools
Small Gitea-automation CLIs, shipped together in one RPM (`agent-tools`). They
act as an agent user (**`unkin-agent`** by default) by minting a scoped Gitea
token from Vault, so automated PRs, comments and pushes are attributed to the
agent — not to whoever happens to run the command. Set `AGENT_LOGIN` to act as a
different agent identity.
- **`agentpr`** — create pull requests and post PR comments as the agent user.
- **`watchpr`** — poll one or more PRs and exit when one changes in a way worth
acting on.
- **`agentws`** — manage per-branch git worktrees for `unkin-agent`, cloning
into Ben's source checkout and isolating agent work under the XDG cache.
- **`agentvault`** — run deterministic Vault flows in one invocation, so agents
never plumb secret material through a shell.
## How it gets a token
On first use each tool performs a Vault AppRole login (`role_id` only, no
`secret_id`), then reads `gitea/creds/$AGENT_LOGIN` — or `GITEA_CREDS_PATH` when
set — to obtain a short-lived Gitea token, cached in-process for the run. With
neither variable set that is `gitea/creds/unkin-agent`, as before.
Everything is configured by environment variables, all with defaults:
| Variable | Default | Purpose |
|---|---|---|
| `VAULT_ADDR` | `https://vault.service.consul:8200` | Vault/OpenBao address |
| `AGENT_APPROLE_ROLE_ID` | built-in default | AppRole role_id (overridable) |
| `GITEA_URL` | `https://git.unkin.net` | Gitea base URL |
| `AGENT_LOGIN` | `unkin-agent` | agent identity: selects `gitea/creds/<login>`, and the login whose comments `watchpr` ignores |
| `GITEA_CREDS_PATH` | `gitea/creds/$AGENT_LOGIN` | Vault path minting the Gitea token (wins over `AGENT_LOGIN`) |
| `AGENTWS_SRC_ROOT` | `~/src/prodenv` | source-of-truth checkout root (`agentws`) |
| `AGENTWS_ROOT` | `~/.cache/agentws` | worktree root (`agentws`) |
| `AGENTWS_OWNER` | `unkin` | Gitea org that owns the repos (`agentws`) |
| `AUTHENTIK_URL` | `https://identity.k8s.syd1.au.unkin.net` | Authentik base URL (`agentvault`) |
## agentpr
```bash
# Verify identity (prints the agent login, unkin-agent by default)
agentpr whoami
# Open a PR
agentpr pr create --repo unkin/argocd-apps \
--base main --head benvin/my-change \
--title "Add woodpecker SA" --body "Adds the ServiceAccount ..."
# prints: #<number> <html_url>
# Comment on a PR
agentpr pr comment --repo unkin/argocd-apps --pr 42 --body "Rebased, CI green."
agentpr --version
agentpr --help
```
Non-zero exit on any API error.
## watchpr
Poll PRs and exit (reporting what changed) when a tracked PR **merges/closes**,
gets a **new comment from someone other than the agent**, its **CI fails**
(failure/error), or it **loses mergeability** (a conflict appears). Benign
transitions — CI `pending``success`, the agent's own comments — are ignored.
```bash
# Watch until something meaningful happens (default interval 60s)
watchpr unkin/argocd-apps#42
# Multiple PRs, custom interval; refs accept #N or :N
watchpr --interval 30s unkin/argocd-apps#42 unkin/terraform-vault:98
# --interval takes a duration (30s, 2m, 1h30m) or a bare number of seconds
watchpr --interval 30 unkin/argocd-apps#42
# One-shot: print current state and exit 0 (great for scripts)
watchpr --once unkin/argocd-apps#42
watchpr --once --json unkin/argocd-apps#42
```
On a meaningful change `watchpr` prints the reason and the PR's current state,
then exits 0. Use `--json` for machine-readable output.
## agentws
`agentws` gives an agent an isolated git worktree per branch without disturbing
Ben's shared checkouts. Repos are cloned into the **source root**
(`~/src/prodenv/<repo>`) so branches created here are visible in the main
checkout too; the worktrees themselves live under the **worktree root**
(`~/.cache/agentws/<repo>__<branch>`).
```bash
# Clone unkin/argocd-apps into ~/src/prodenv if missing, then add a worktree for
# a new branch off the remote default branch. Prints the worktree path.
agentws new argocd-apps --branch benvin/my-change
# Branch off a specific base instead of the remote default
agentws new argocd-apps --branch benvin/hotfix --from release-1.2
# List managed worktrees (repo, branch, path)
agentws list
# Remove a worktree (by path or branch); refreshes the source repo afterwards
agentws rm benvin/my-change
agentws rm ~/.cache/agentws/argocd-apps__benvin-my-change --delete-branch
# Classify every managed worktree; dry run unless --yes is given
agentws prune
agentws prune --yes
agentws prune --yes --keep-branches
# Remove every managed worktree and prune each source repo
agentws clean
# Print a fresh Gitea token for the agent login
agentws token
```
### prune
`agentws prune` decides, per worktree, whether its work is safely upstream:
| Signal (first match wins) | Verdict |
|---|---|
| uncommitted or untracked changes | keep |
| branch has an open PR | keep |
| tip contained in `origin/<default>` | remove worktree + local branch |
| every commit patch-equivalent to one in `origin/<default>`'s history | remove worktree + local branch |
| PR merged **and** HEAD contained in the PR's head commit (or in a verified `origin/<branch>`) | remove worktree + local branch |
| PR closed **and** HEAD contained in a verified `origin/<branch>` | remove worktree + local branch |
| anything else | remove worktree, keep the branch |
A branch is deleted only where git proves its commits survive elsewhere. PR
state alone never authorises that: a merged or closed PR whose branch picked up
commits since keeps its branch, because those commits exist nowhere but here.
The delete runs `git branch -d` first so git's own unmerged check is a backstop,
falling back to `-D` only for a proven branch — squash merges keep the guard
tripping even once the work has landed.
Patch equivalence comes from `git cherry`, which these squash-merging repos need
because a merged branch's commits carry different SHAs upstream. It proves the
patches reached the default branch's history at some point — a later revert
still counts — not that they stand at its tip.
`origin/<branch>` counts as evidence only when this run's `git fetch --prune`
succeeded. A tracking ref left over from an earlier fetch may name a branch that
is already gone upstream and is itself due for deletion, so a failed fetch
downgrades those verdicts to `remove` and keeps the branch. Proofs that read
only local objects — containment in `origin/<default>`, patch equivalence, and
containment in a merged PR's head SHA — stand on their own.
Gitea PR state only adds to the git answer: when it cannot be reached, prune
says so and never deletes a branch it could not prove, and a PR listing that
hits the pagination cap is reported rather than read as "no PR". Matching a
branch to its PR uses `head.label`, since Gitea rewrites `head.ref` to
`refs/pull/<n>/head` once the branch is deleted on merge.
`--keep-branches` removes worktrees only, and its verdicts print as `remove`.
### Auth / credential-helper design
Gitea tokens minted from Vault are short-lived (~1h), so `agentws` never
persists one in a remote URL or in git config. Instead it wires itself as an
**ephemeral git credential helper**:
- `agentws token` prints a fresh token to stdout (handy for scripts).
- `agentws credential get` speaks the git credential protocol on stdin and, for
the configured Gitea host only, emits `username=$AGENT_LOGIN` +
`password=<fresh token>`.
`agentws new` sets this up per worktree without touching the shared checkout: it
enables `extensions.worktreeConfig` on the repo once, then writes
`user.name` / `user.email` and `credential.helper = !<agentws> credential` to
the **per-worktree** config. Clone/fetch use the same helper via a transient
`-c credential.helper=...`; the shared `origin` URL is left clean. On worktree
removal `agentws` fetches in `~/src/prodenv/<repo>` so its default branch stays
current.
## agentvault
Deterministic Vault flows, each a single self-contained invocation: the tool
reads and writes the secrets itself, and prints only identifiers.
### seed-outpost
Copy an Authentik outpost's token into Vault KV-v2. `agentvault` reads the
Authentik API token from `kv/service/authentik/agent-api-token`, resolves the
named outpost's `token_identifier`, fetches its key via
`/api/v3/core/tokens/<identifier>/view_key/` and writes it to the destination
KV path. The token value is never printed or logged.
```bash
agentvault seed-outpost \
--outpost k8s-outpost \
--dest-path kubernetes/namespace/authentik/default/outpost-token
```
```
outpost: k8s-outpost
token_identifier: ak-outpost-k8s-outpost
dest: kv/kubernetes/namespace/authentik/default/outpost-token
version: 3
```
Re-running is safe: it writes a new KV version. Flags: `--outpost` and
`--dest-path` are required; `--dest-key` (default `token`), `--kv-mount`
(default `kv`), `--token-path` (default `service/authentik/agent-api-token`) and
`--authentik-url` override the rest.
Errors name the failing stage: AppRole login, KV read denied (policy not
applied), outpost not found (terraform not applied), `view_key` failure, or KV
write denied.
### seed-oauth
Make a Vault KV-v2 path hold a complete oauth2-proxy credential set. It is a
read-modify-write: `client_id` is set from the flag, `client_secret` and
`cookie_secret` are generated (32 bytes from `crypto/rand`) only when missing,
every other key on the path is written back untouched, and nothing is written
at all when the secret is already correct. `cookie_secret` is base64url so it
decodes to exactly the 32 bytes oauth2-proxy requires.
```bash
agentvault seed-oauth \
--path kubernetes/namespace/repospawner/default/oauth-credentials \
--client-id 4f1c…
```
```
path: kv/kubernetes/namespace/repospawner/default/oauth-credentials
keys: client_id, client_secret, cookie_secret
client_id: created
client_secret: kept
cookie_secret: created
version: 4
```
That is the common case: the provider's `client_secret` already lives on the
path, so only the missing keys are added. A run with nothing to do prints
`version: unchanged` and issues no write.
Flags: `--path` and `--client-id` are required; `--kv-mount` (default `kv`) and
`--rotate` override the rest. `--rotate` regenerates both secrets — only use it
when the IdP provider's secret is being rotated alongside, since a rotated
`client_secret` no longer matches the provider.
Errors name the failing stage: AppRole login, KV read denied, or KV write
denied. Only key names, actions and the KV version are printed.
## Build & package
```bash
make build # -> dist/agentpr, dist/watchpr, dist/agentws, dist/agentvault
make test # go test -race ./...
make rpm # build + package dist/agent-tools-<version>-1.x86_64.rpm
```
Release is tag-driven (`v*`) via Woodpecker: builds the RPM, `PUT`s it to the
artifactapi `rpm-internal` yum repo, and cuts a Gitea release with
cross-compiled binaries attached.
### Version bump
```bash
make patch # or: make minor / make major — tags vX.Y.Z and pushes the tag
```