@thegpvc/gp-foundry 0.2.10 → 0.2.12
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/actions/agent-context/README.md +60 -9
- package/actions/agent-context/action.yml +22 -1
- package/actions/agent-context/dist/index.cjs +79 -45
- package/actions/config-loader/dist/index.cjs +1 -1
- package/actions/run-agent/README.md +9 -0
- package/actions/run-agent/action.yml +84 -0
- package/actions/sanitize-untrusted-input/README.md +24 -5
- package/actions/sanitize-untrusted-input/dist/index.cjs +30 -28
- package/actions/setup-agent/README.md +1 -0
- package/actions/setup-agent/action.yml +13 -1
- package/dist/cli/index.cjs +120 -111
- package/package.json +1 -1
- package/schema/config.schema.json +7 -1
- package/skill/reference/node-types.md +28 -2
- package/skill/templates/foundry.config.yaml +1 -0
- package/skill/templates/scope.yaml +2 -1
|
@@ -1,6 +1,7 @@
|
|
|
1
1
|
# agent-context
|
|
2
2
|
|
|
3
|
-
Fetches GitHub issue/PR data via the Octokit REST API
|
|
3
|
+
Fetches GitHub issue/PR data via the Octokit REST API (plus one GraphQL query,
|
|
4
|
+
for review threads) and formats it into a
|
|
4
5
|
labeled plaintext context file for an agent prompt. Writes the file to a temp
|
|
5
6
|
path and exposes it via the `context-file` output.
|
|
6
7
|
|
|
@@ -14,10 +15,12 @@ This is a Node20 JavaScript action authored in TypeScript
|
|
|
14
15
|
| `type` | yes | — | Context type: `issue`, `pr-diff`, `pr-review`, or `pr-full`. |
|
|
15
16
|
| `number` | yes | — | Issue or PR number. |
|
|
16
17
|
| `token` | yes | — | GitHub token used for API calls. |
|
|
18
|
+
| `checks-token` | no | `""` | Token used only to read check-run conclusions on the PR head. Pass `GITHUB_TOKEN` (job permission `checks: read`); empty skips the section. See below. |
|
|
17
19
|
| `include-diff` | no | `"true"` | Include the full unified diff for PR types. Any value other than `"false"` is treated as true. |
|
|
18
20
|
| `triggering-comment` | no | `""` | Body of the comment that triggered the workflow; appended as a final section when non-empty. |
|
|
19
21
|
| `base-branch` | no | `""` | Optional base branch noted in the log for the comparison. The PR diff is always taken relative to the PR's own base ref; this input is surfaced for logging and downstream steps, not to re-target the diff. |
|
|
20
22
|
| `trusted-authors` | no | `""` | Comma-separated logins whose posts are the harness's own output; framed as the agent's work item rather than as untrusted input. See below. |
|
|
23
|
+
| `trusted-framing` | no | `"work-item"` | How a `trusted-authors` post is framed: `work-item` (act on it) or `baseline` (the agent's own previous output, to compare against). Any other value fails the step. See below. |
|
|
21
24
|
| `sanitize` | no | `"true"` | Neutralize prompt-injection in untrusted text before it lands in the context file. See below. |
|
|
22
25
|
|
|
23
26
|
## Outputs
|
|
@@ -29,13 +32,46 @@ This is a Node20 JavaScript action authored in TypeScript
|
|
|
29
32
|
## Context types
|
|
30
33
|
|
|
31
34
|
- **`issue`** — issue header (number, title, labels, body) plus all issue comments.
|
|
32
|
-
- **`pr-diff`** — PR header (title, file/line stats, body)
|
|
33
|
-
-
|
|
35
|
+
- **`pr-diff`** — PR header (title, head SHA, file/line stats, body), the
|
|
36
|
+
changed-file list, check runs (if `checks-token` is set), and the diff.
|
|
37
|
+
- **`pr-review`** — everything in `pr-diff`, plus issue comments, reviews, and
|
|
38
|
+
review threads, ahead of the diff.
|
|
34
39
|
- **`pr-full`** — identical to `pr-review`.
|
|
35
40
|
|
|
36
|
-
Empty sections are omitted.
|
|
37
|
-
|
|
38
|
-
|
|
41
|
+
Empty sections are omitted. What the PR sections carry:
|
|
42
|
+
|
|
43
|
+
- **Changed files** — one line per file: status, path (and the previous path of
|
|
44
|
+
a rename), `+additions -deletions`.
|
|
45
|
+
- **Check runs on head** — `name: conclusion` (or the status, while a run is not
|
|
46
|
+
completed), sorted by name, the latest attempt per check name only. Read with
|
|
47
|
+
`checks-token`, not `token`: the Checks permission has no fine-grained-PAT
|
|
48
|
+
equivalent, so the harness token cannot see check runs and the workflow's
|
|
49
|
+
`GITHUB_TOKEN` can. Without `checks-token` there is no section; if the fetch
|
|
50
|
+
fails, the section is a one-line note that conclusions were unavailable.
|
|
51
|
+
- **Reviews** — each with the commit it was submitted against, marked
|
|
52
|
+
`(current head)` or `(not the current head)`, so a review of this head can be
|
|
53
|
+
told from a review of an older one.
|
|
54
|
+
- **Review threads** — grouped per thread, each led by its node id, `RESOLVED` /
|
|
55
|
+
`UNRESOLVED` (and `OUTDATED`), and `path:line`, then its comments, then the
|
|
56
|
+
reactions on the first comment with who left them. Threads exist only in the
|
|
57
|
+
GraphQL API; if that query fails, the section falls back to the flat REST
|
|
58
|
+
inline comments (`INLINE REVIEW COMMENTS`) with a note saying so.
|
|
59
|
+
|
|
60
|
+
None of the best-effort fetches (threads, check runs) can fail the step; the
|
|
61
|
+
underlying API error is logged as a workflow warning.
|
|
62
|
+
|
|
63
|
+
### The diff
|
|
64
|
+
|
|
65
|
+
The diff is capped at ~100KB (100,000 bytes); larger diffs are truncated with a
|
|
66
|
+
`[Diff truncated at 100KB. Use Read tool to examine full files.]` marker.
|
|
67
|
+
|
|
68
|
+
GitHub refuses to serve a diff that is too large (HTTP 406, or a 422
|
|
69
|
+
`too_large`). The step does not fail on that: the diff section is assembled
|
|
70
|
+
instead from the per-file patches of the changed-file list, in order, up to
|
|
71
|
+
200KB, under a line stating that the diff was too large for the API and how many
|
|
72
|
+
files are shown in full, truncated, and omitted (a file is omitted when the
|
|
73
|
+
budget has run out or when GitHub supplies no patch for it — binaries, very
|
|
74
|
+
large files). Any other error fetching the diff still fails the step.
|
|
39
75
|
|
|
40
76
|
## Untrusted text
|
|
41
77
|
|
|
@@ -52,8 +88,14 @@ collision-proof fence inside an `UNTRUSTED USER INPUT` banner. Per-body fencing
|
|
|
52
88
|
over distinguishable.
|
|
53
89
|
|
|
54
90
|
Titles are scrubbed in place instead of bannered — a banner per title would
|
|
55
|
-
drown the structure it labels
|
|
56
|
-
|
|
91
|
+
drown the structure it labels. File paths and check-run names get the same
|
|
92
|
+
one-line scrub: both are chosen by the PR author.
|
|
93
|
+
|
|
94
|
+
The diff's text is left verbatim — no marker redaction, no secret masking; it
|
|
95
|
+
has to stay a faithful diff — under a header noting that text inside a diff is
|
|
96
|
+
data under review, never instructions. It is wrapped in a collision-proof fence
|
|
97
|
+
like every other body, so a diff that contains a markdown fence or a line like
|
|
98
|
+
`=== REVIEWS ===` cannot pass itself off as a section of the context.
|
|
57
99
|
|
|
58
100
|
Neutralized markers and masked secrets are reported as workflow warnings, so an
|
|
59
101
|
injection attempt shows up in the run log. Set `sanitize: "false"` only if the
|
|
@@ -71,6 +113,14 @@ control-stripped, secret-masked and fenced, but the banner says they are the wor
|
|
|
71
113
|
item, and injection-marker redaction is skipped so ordinary review prose survives
|
|
72
114
|
intact. Everyone else in the same file stays untrusted.
|
|
73
115
|
|
|
116
|
+
"Work item" is the right framing for a Fixer. It is the wrong one for a
|
|
117
|
+
re-reviewer, whose earlier review it would present as something to act on: the
|
|
118
|
+
agent would be told to obey its past self. `trusted-framing: baseline` frames the
|
|
119
|
+
same posts as the agent's own previous output — a baseline to compare the current
|
|
120
|
+
state against, not instructions to follow and not a conclusion to defer to, since
|
|
121
|
+
the code may have changed since. Only the banner text differs; the scrubbing,
|
|
122
|
+
the fence, and the note that quoted third-party text is still data are the same.
|
|
123
|
+
|
|
74
124
|
## Example
|
|
75
125
|
|
|
76
126
|
```yaml
|
|
@@ -79,7 +129,8 @@ intact. Everyone else in the same file stays untrusted.
|
|
|
79
129
|
with:
|
|
80
130
|
type: pr-review
|
|
81
131
|
number: ${{ github.event.pull_request.number }}
|
|
82
|
-
token: ${{ secrets.
|
|
132
|
+
token: ${{ secrets.AGENT_PAT }}
|
|
133
|
+
checks-token: ${{ secrets.GITHUB_TOKEN }}
|
|
83
134
|
triggering-comment: ${{ github.event.comment.body }}
|
|
84
135
|
|
|
85
136
|
- name: Use the context
|
|
@@ -11,6 +11,15 @@ inputs:
|
|
|
11
11
|
token:
|
|
12
12
|
description: GitHub token for API calls
|
|
13
13
|
required: true
|
|
14
|
+
checks-token:
|
|
15
|
+
description: >-
|
|
16
|
+
Optional token used ONLY to read check-run conclusions on the PR head
|
|
17
|
+
(latest attempt per check name). Separate from `token` because the Checks
|
|
18
|
+
permission has no fine-grained-PAT equivalent: the harness PAT cannot read
|
|
19
|
+
check runs, so pass GITHUB_TOKEN here (the job needs `checks: read`) while
|
|
20
|
+
`token` stays the harness token. Empty skips the checks section entirely;
|
|
21
|
+
a failed fetch renders a one-line "unavailable" note instead of failing.
|
|
22
|
+
default: ""
|
|
14
23
|
include-diff:
|
|
15
24
|
description: Whether to include the full diff (for PR types)
|
|
16
25
|
default: "true"
|
|
@@ -25,13 +34,25 @@ inputs:
|
|
|
25
34
|
Fixer's checklist IS the Reviewer's review, and banning it from acting on
|
|
26
35
|
it is incoherent. Leave empty to treat every body as untrusted.
|
|
27
36
|
default: ""
|
|
37
|
+
trusted-framing:
|
|
38
|
+
description: >-
|
|
39
|
+
How a `trusted-authors` post is framed for the agent reading it.
|
|
40
|
+
"work-item" (default): the post is the agent's work item, to act on --
|
|
41
|
+
right for a Fixer handed the Reviewer's review. "baseline": the post is
|
|
42
|
+
the agent's own previous output, a baseline to compare the current state
|
|
43
|
+
against -- not instructions to follow and not a conclusion to defer to,
|
|
44
|
+
since the code may have changed since. Use it for a re-reviewer, which
|
|
45
|
+
"work-item" would tell to obey its past self. Scrubbing and fencing are
|
|
46
|
+
identical either way. Any other value fails the step.
|
|
47
|
+
default: "work-item"
|
|
28
48
|
sanitize:
|
|
29
49
|
description: >-
|
|
30
50
|
Neutralize prompt-injection in attacker-controlled text (issue/PR bodies,
|
|
31
51
|
comments, review bodies, titles) before it is written to the context file:
|
|
32
52
|
strips control/bidi sequences, caps length, neutralizes role-override
|
|
33
53
|
markers, masks secret-looking strings, and wraps each body in a
|
|
34
|
-
collision-proof fence inside an UNTRUSTED banner.
|
|
54
|
+
collision-proof fence inside an UNTRUSTED banner. The diff is fenced too,
|
|
55
|
+
but its text is left verbatim. On by default — the
|
|
35
56
|
agent reading this file holds a write token and a shell. Set "false" only
|
|
36
57
|
if the consumer sanitizes some other way.
|
|
37
58
|
default: "true"
|