| name | api-discovery |
| description | Resolves the on-disk location of a Maven artifact's source code, so it can be inspected directly instead of running `unzip` against JARs in the Gradle cache. Use when inspecting a library's API or implementation — definitions of public types, method signatures, KDoc, internal helpers, etc.
|
API discovery
Before reading library source code, run the discover script in
.agents/scripts/api-discovery/. It returns a path you can hand
straight to normal search and file-reading tools such as rg, sed,
or the active agent's file viewer.
Do not run find ~/.gradle/caches or unzip against cache JARs.
Each unzip decompresses the archive afresh — slow and token-heavy.
How to call it
From the consumer repository root:
.agents/scripts/api-discovery/discover <query>
Where <query> is one of:
| Form | Example | Notes |
|---|
group:artifact:version | io.spine:spine-base:2.0.0-SNAPSHOT.390 | Most explicit |
group:artifact | io.spine:spine-base | Version inferred from buildSrc |
artifact | spine-base | Spine-only; group inferred from buildSrc |
The script writes the absolute resolved path to stdout, and any
freshness/diagnostic warnings to stderr. Always read stderr — a
silent stdout means clean resolution; a noisy stderr means caveats
the user should know about.
Exit codes
| Code | Meaning | What you do |
|---|
0 | Path on stdout is usable. | Search or read files under that path directly. If stderr is non-empty, surface the warning to the user before relying on the path. |
1 | Unresolvable (no sibling AND no JAR). | Report the failure. Do not fall back to unzip ~/.gradle/caches/.... |
10 | Cache directory not initialized. | Run the bootstrap flow below. |
Bootstrap flow (exit 10)
On the first run in a fresh workstation the per-workstation cache
directory does not yet exist. The script exits 10 and names the
path it would create. Ask the user:
The shared cache directory <workspace-root>/.agents/caches/api-discovery/
does not exist yet. How would you like to proceed?
- Approve — create the directory at the default path.
- Alternative root — pick a different parent for the shared
.agents/ directory (e.g., ~/SpineWorkspace, /srv/spine).
- Non-cached — skip the extraction cache. Sibling-first
discovery still works for Spine artifacts; non-Spine deps will
not be served by
api-discovery in this repo.
Then act on the user's reply:
Check that memory at session start. If it exists, skip cache-touching
paths entirely.
Workflow
- Always call
discover before reading library source.
- Use the returned path with search or file-reading tools directly. Do not
cd into the directory — that adds path-prefix noise to tool calls
and makes line citations harder to read.
- If stderr contains
STALE: ..., the sibling on disk does not match
the version declared in buildSrc. Surface the warning AND offer
to refresh — see Refreshing a stale sibling below.
- If the script exits
1, report the failure with its stderr
message and stop. Do not try unzip as a workaround.
Refreshing a stale sibling
The user keeps siblings cloned locally as the source of truth and
sometimes works across several siblings at once with a feature branch
checked out in each. So a STALE line has two possible meanings, and
they require different handling:
- Sibling is behind
master/main. A git pull --ff-only will
bring it up to date.
- Sibling is on a feature branch. This is intentional — the user
is staging changes across multiple subprojects. The local code is
the right code; do not pull.
You cannot tell which case applies without inspecting the sibling. The
companion script update-sibling handles both safely: it pulls only
on the default branch with a clean tree and a tracked upstream, and
exits 0 without touching anything when on a feature branch.
Procedure
When you see a STALE: ... line from discover:
-
Surface the warning to the user.
-
Ask, in one short prompt:
The sibling at <path> is stale. Want me to try updating it?
I'll only git pull --ff-only if it's on master/main with
a clean working tree; if you have a feature branch checked out,
I'll leave it as-is.
-
If the user agrees, run:
.agents/scripts/api-discovery/update-sibling <sibling-path-or-name>
<sibling-path-or-name> is either the absolute path shown by
discover (preferred — unambiguous) or just the sibling repo name
(resolved under <workspace-root>).
-
Read stdout to decide what to do next — it is a single stable
token, not free-form English:
pulled — HEAD advanced. Re-run discover so the STALE warning
clears (or, more rarely, reports a different discrepancy).up-to-date — sibling was already at upstream tip. The STALE
warning is informational — the declared buildSrc version and
the sibling's versionToPublish simply disagree. Proceed
without re-running discover.skipped-branch — sibling is on a feature branch and was left
untouched. Use the local code as-is; proceed without re-running.
stderr always carries the human-readable diagnostics; surface it
to the user, but do not parse it to drive control flow.
-
If the user declines, proceed without pulling. Do not ask again
for the same sibling in the same session unless the user revisits.
update-sibling exit codes
Exit 0 is split into three outcomes by the stdout token — read
that, not the stderr text.
| Code | stdout | Meaning | What you do |
|---|
0 | pulled | HEAD advanced to upstream tip. | Re-run discover so the STALE warning clears. |
0 | up-to-date | Already at upstream tip; nothing to do. | Proceed; surface the STALE warning to the user as informational. |
0 | skipped-branch | On a non-default branch; left untouched. | Use the local code as-is; proceed without re-running. |
1 | (empty) | Sibling not on disk. | Report the error. |
2 | (empty) | Not a git repository. | Report the error; do not retry. |
3 | (empty) | Detached HEAD — refused. | Tell the user; do not retry. |
4 | (empty) | Working tree dirty — refused. | Tell the user; do not retry. |
5 | (empty) | No upstream tracking on default branch — refused. | Tell the user. |
6 | (empty) | git pull --ff-only failed (divergence, network, etc.). | Surface the git error verbatim. |
64 | (empty) | Usage error (no/too many arguments). | Fix the invocation; do not retry blindly. |
Failure paths produce empty stdout so the agent can never misread
an error message as a result token.
"Don't ask me again"
If the user says something like "stop offering" or "skip the prompt
this session", remember that for the rest of the conversation and do
not prompt on subsequent STALE warnings — just surface the warning
and move on. This is per-session state; do not write it to
auto-memory.
Anti-patterns
Stop doing these — they are exactly what this skill exists to replace:
find ~/.gradle/caches/modules-2/files-2.1/ -name '*-sources.jar'unzip -l <jar> to list classesunzip -p <jar> path/in/jar to read a file- Any chain of
unzip + grep against a Gradle-cache JAR
If you find yourself wanting to do those, run discover instead.
Examples
Spine artifact, fresh sibling on disk:
$ .agents/scripts/api-discovery/discover io.spine:spine-base
/Users/<you>/Projects/Spine/base-libraries/base
$ echo $?
0
Follow-up searches then look like:
rg --files /Users/<you>/Projects/Spine/base-libraries/base.rg -n 'class Identifier' /Users/<you>/Projects/Spine/base-libraries/base.
Spine artifact, stale sibling:
$ .agents/scripts/api-discovery/discover io.spine.tools:validation-java
api-discovery: STALE: validation-java declared 2.0.0-SNAPSHOT.433 in Validation.kt but sibling publishes 2.0.0-SNAPSHOT.440
api-discovery: sources at /Users/<you>/Projects/Spine/validation/java may differ from the published artifact
/Users/<you>/Projects/Spine/validation/java
Surface the STALE line, then offer to refresh — see Refreshing a
stale sibling. After the user agrees and the pull succeeds, re-run
discover and the warning clears.
Stale sibling, refresh on master:
$ .agents/scripts/api-discovery/update-sibling /Users/<you>/Projects/Spine/validation
Updating abc1234..def5678
Fast-forward
...
api-discovery: /Users/<you>/Projects/Spine/validation pulled 'master': abc1234... -> def5678...
pulled
$ echo $?
0
Stdout is pulled — re-run discover to clear the STALE warning.
Stale sibling, already at upstream tip:
$ .agents/scripts/api-discovery/update-sibling /Users/<you>/Projects/Spine/validation
Already up to date.
api-discovery: /Users/<you>/Projects/Spine/validation already up-to-date on 'master' (def5678...)
up-to-date
$ echo $?
0
Stdout is up-to-date — the sibling is fresh; the STALE warning
reflects a declared-version vs. versionToPublish discrepancy that
git pull cannot resolve. Surface it to the user as informational.
Stale sibling, feature branch (no-op):
$ .agents/scripts/api-discovery/update-sibling /Users/<you>/Projects/Spine/validation
api-discovery: /Users/<you>/Projects/Spine/validation is on 'feature/new-rule' (not master/main); using local code as-is
skipped-branch
$ echo $?
0
Stdout is skipped-branch — feature branch is intentional local
state. Use the code as-is.
Non-Spine artifact, first use (extraction):
$ .agents/scripts/api-discovery/discover com.google.guava:guava:33.5.0-jre
api-discovery: extracted com.google.guava:guava:33.5.0-jre -> .../guava/33.5.0-jre
/Users/<you>/Projects/Spine/.agents/caches/api-discovery/sources/com.google.guava/guava/33.5.0-jre
Second call returns the same path with no stderr (already cached).
Unresolvable:
$ .agents/scripts/api-discovery/discover io.spine:does-not-exist:9.9.9
api-discovery: io.spine:does-not-exist:9.9.9 is not in the local Gradle cache
api-discovery: run './gradlew dependencies' (or rebuild) to fetch it, then retry
$ echo $?
1
Report the failure verbatim; do not try unzip as a workaround.
Related
- Implementation reference:
.agents/scripts/api-discovery/README.md.
- Sibling refresh on STALE:
.agents/scripts/api-discovery/update-sibling.
- Manual cache pruning:
.agents/scripts/api-discovery/clean-cache.
