Merge branch 'main' into feat/injection-guard-dataset · ThinkFlowLab/system1-agents@dcc7392 · GitHub
Skip to content

Commit dcc7392

Browse files
hsliuustc0106hsliuustc0106
authored
Merge branch 'main' into feat/injection-guard-dataset
2 parents daa64f6 + adcc5d1 commit dcc7392

38 files changed

Lines changed: 2398 additions & 54 deletions

‎.agents/skills/add-agent-recipe/SKILL.md‎

Lines changed: 7 additions & 2 deletions

‎.agents/skills/self-review/SKILL.md‎

Lines changed: 21 additions & 14 deletions
Original file line numberDiff line numberDiff line change
@@ -72,28 +72,35 @@ demo is not a task-success rate. Mark missing metrics unmeasured, and remove or
7272
qualify unsupported claims rather than manufacturing a comparison.
7373

7474
Use [CONTRIBUTING.md's video guide](../../../CONTRIBUTING.md#agent-video-demos)
75-
to prepare, record and attach an agent demo. For agent behavior changes, prefer a
76-
short real video (about 20–45 seconds); a terminal recording works for text agents
77-
and rails. Screenshots can support the clip or explain a recording gap. Show the
75+
to classify the PR and prepare, record and attach its demo. Important PRs require
76+
a video of the application/task, System1-Agents decision-model agent and actual
77+
System1-Omni inference in the same run. Check all three parts against the linked
78+
trace; a terminal recording works for text agents and rails. Screenshots and logs
79+
support the clip. Use PR #35's recording linked in the guide as the example and
80+
choose a relevant README application candidate from the guide. Show the
7881
relevant input, action sequence, and result, with failures or human intervention
79-
visible. Label cuts, replay speed,
80-
and elapsed timing honestly; link a fuller trace when a clip omits context. Do not
81-
stage screens or present a replay as a live run. A replay must identify the source
82+
visible. Label cuts, replay speed, and elapsed timing honestly; link a fuller
83+
trace when a clip omits context. Do not stage screens or present a replay as a
84+
live run. A replay must identify the source
8285
run/commit, workload, and speed; a historical or upstream model demo is not proof
8386
that the current agent integration works.
8487

8588
Choose figures that answer the review question: workflow screenshots, a short
8689
action timeline, or task-success comparisons backed by the run records. There is
87-
no asset quota. Nonvisual, docs-only, and test-only changes may use test output or
88-
a concise explanation instead of video. If a relevant run cannot be made within
89-
the available authorization, resources, or budget, report the gap and its impact;
90-
do not turn a missing run into a pass or require a production-scale demonstration.
90+
no asset quota. Only the guide's minor docs/formatting/test-only exemption permits
91+
`N/A` with a reason; a nonvisual task still needs a terminal video when the PR is
92+
important. If a required run cannot be made within the available authorization,
93+
resources, or budget, report the gap and its impact; keep the important PR draft
94+
until the video is supplied or a maintainer accepts
95+
the documented exception. Do not turn a missing run into a pass or require a
96+
production-scale demonstration.
9197

9298
## PR demo/evidence section
9399

94100
Prepare a **Demo / evidence** section for the PR containing what applies:
95101

96-
- Task and observable result, or `N/A` with a concrete reason.
102+
- Required application + agents + Omni video and trace, or the documented exemption/blocker.
103+
- Task and observable result, or `N/A` for an exempt change with a concrete reason.
97104
- Reproduction command/fixture, configuration, and baseline/head commits.
98105
- Measured comparison and raw result links, including failures and limitations;
99106
distinguish personally run checks, author-reported results, and observed CI.
@@ -111,9 +118,9 @@ performance claims. Link durable, reviewer-accessible artifacts rather than loca
111118
paths. If rights or safe disclosure are unresolved, omit the asset and state why.
112119

113120
Check System1-Omni's current model, modality and hardware support as described in
114-
the video guide. Try a supported serving path when the branch has a compatible
115-
client; otherwise record the missing integration or configuration. Pin both
116-
repositories and verify the actual worker/frontend from run evidence. Keep
121+
the video guide. Use a supported serving path for the required video when the
122+
branch has a compatible client; otherwise record the missing integration or
123+
configuration. Pin both repositories and verify the actual worker/frontend from run evidence. Keep
117124
agent-task evidence separate from serving/kernel measurements. Do not add an
118125
unrequested backend integration or run outside the authorized resources/budget
119126
merely to produce a demo.

‎.claude/skills/build-s1a-agent/SKILL.md‎

Lines changed: 6 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -84,9 +84,13 @@ ask the user before starting one.
8484
For a reusable use case, add a recipe using [the recipe template](../../../recipes/TEMPLATE.md) and
8585
[add-agent-recipe skill](../../../.agents/skills/add-agent-recipe/SKILL.md), then update the recipe index.
8686
Follow [CONTRIBUTING.md's video guide](../../../CONTRIBUTING.md#agent-video-demos)
87-
for a short recording of the actual agent's input, selected actions and independently checked result.
87+
for the required video of the application/task, the System1-Agents decision-model agent and System1-Omni
88+
inference in the same run. Show actual input, selected/executed actions, engine identity and checked result.
89+
Choose an application from the guide's README candidates that exercises the new agent's actual behavior.
8890
Use a terminal recording for a text agent or rail; label scripted checks and recorded replays.
89-
Identify the inference engine and check the supported System1-Omni path or record its integration gap.
91+
Use a supported System1-Omni path; record missing support or integration as a review gap. Follow PR #35's
92+
worked example linked in the guide. Keep an important PR draft until its video is supplied or a maintainer
93+
accepts the documented exception.
9094
Prepare the PR's **Demo / evidence** section with the clip, reproduction command and linked run artifacts.
9195
Recording, model execution and uploading each require the authorization/resources already established for
9296
the task. If any are unavailable, report the gap and hand off the prepared instructions.

‎.env.example‎

Lines changed: 1 addition & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -9,6 +9,7 @@ MODEL_NAME=google/gemini-2.5-flash
99
# Jev through OpenRouter's /api/alpha/decisions with OPENROUTER_API_KEY: TYPESAFE_API_URL selects that backend.
1010
# TYPESAFE_API_URL=https://openrouter.ai/api/alpha/decisions
1111
# TYPESAFE_MODEL=typesafe/jev-1.13
12+
# S1A_DECISION_TIMEOUT_S=5 # seconds per decision; raise it (e.g. 30) for a slower local System One server
1213
# Any OpenAI-compatible endpoint instead of OpenRouter:
1314
# OPENAI_API_KEY=
1415
# OPENAI_BASE_URL=https://api.openai.com/v1

‎.github/PULL_REQUEST_TEMPLATE.md‎

Lines changed: 10 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -23,11 +23,19 @@
2323
## Demo / evidence
2424

2525
<!-- Follow CONTRIBUTING.md#agent-video-demos and .agents/skills/self-review/SKILL.md:
26-
task input → actions → observed result; identify the actual inference engine;
26+
Important PRs require a video of application/task + System1-Agents decision-model agent
27+
and System1-Omni inference in the same run, with input → actions → observed result.
28+
Choose a relevant README use case; use PR #35's recording as the example. Identify the actual inference engine;
2729
reproduction command/fixture and exact baseline/head commits; measured comparisons
2830
and raw results; useful demo/figure links with source revision, replay speed, and
2931
limitations. Check asset rights and redaction. Label illustrations separately from
30-
actual runs. Video is optional; use N/A with a reason where evidence does not apply. -->
32+
actual runs. Use N/A only for the guide's minor-change exemption. A missing required
33+
video is a review gap: name its blocker/follow-up and keep the PR draft until it is
34+
supplied or a maintainer accepts the documented exception. -->
35+
36+
- Video link, or exemption/blocker with follow-up:
37+
- Application/task → agent actions → independently checked result:
38+
- System1-Agents / System1-Omni commits, checkpoint and worker/frontend provenance:
3139

3240
For an agent-assisted self-review, use the [self-review skill](https://github.com/ThinkFlowLab/system1-agents/blob/main/.agents/skills/self-review/SKILL.md).
3341
For recording and upload instructions, use the [agent video guide](https://github.com/ThinkFlowLab/system1-agents/blob/main/CONTRIBUTING.md#agent-video-demos).

‎.github/workflows/test.yml‎

Lines changed: 9 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -14,15 +14,22 @@ concurrency:
1414
jobs:
1515
core:
1616
# The install the README promises: openjiuwen alone, no game or report extras, no keys.
17+
# Keep these names in sync with main's required status checks.
18+
name: ${{ matrix.check_name }}
1719
runs-on: ${{ matrix.os }}
1820
timeout-minutes: 15
1921
strategy:
2022
matrix:
21-
os: [ubuntu-latest]
22-
python: ["3.11", "3.13"]
2323
include:
24+
- os: ubuntu-latest
25+
python: "3.11"
26+
check_name: core (3.11)
27+
- os: ubuntu-latest
28+
python: "3.13"
29+
check_name: core (3.13)
2430
- os: windows-latest
2531
python: "3.11"
32+
check_name: core (windows, 3.11)
2633
steps:
2734
- uses: actions/checkout@v7
2835
- uses: astral-sh/setup-uv@v10.2.0

‎CHANGELOG.md‎

Lines changed: 15 additions & 0 deletions

0 commit comments

Comments
 (0)