Skip to content

Commit 6fc921a

Browse files
AbirAbbasclaude
andcommitted
chore(skills): sync embedded agentfield-use mirror on branch
Branch-local drift: this PR merged main (incl. #827's skill edit) while carrying its own mirror copy; sync-embedded-skills.sh realigns them so the skillkit drift tests pass regardless of #828. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
1 parent 87621d0 commit 6fc921a

1 file changed

Lines changed: 14 additions & 1 deletion

File tree

  • control-plane/internal/skillkit/skill_data/agentfield-use

control-plane/internal/skillkit/skill_data/agentfield-use/SKILL.md

Lines changed: 14 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -154,7 +154,9 @@ curl -s -X POST http://localhost:8080/api/v1/execute/swe-planner.plan \
154154

155155
Async dispatch is cheap: fire all independent calls up front, then poll them
156156
together. Do NOT serialize multi-agent work — the whole point of the control
157-
plane is managing many agents at once. What to know:
157+
plane is managing many agents at once. When a batch of independent jobs arrives
158+
(ten PRs to review, five repos to scan), the default is to dispatch the whole
159+
batch now and poll as a group — not one-at-a-time. What to know:
158160

159161
- Concurrent calls to the **same reasoner** are safe when the agent is (e.g.
160162
pr-af isolates concurrent reviews per PR). If an agent's docs don't say it's
@@ -173,6 +175,17 @@ launching more heavy runs — if `active_executions >= recommended_max_concurren
173175
finish or await in-flight work first rather than starting more, and tell the
174176
user you're throttling to avoid overloading the machine.
175177

178+
**Canary after reconfiguration, then fan out.** The one exception to
179+
fire-everything-up-front: you just changed a node's runtime config (provider,
180+
model, bin path — `af secrets set` + restart). A misconfigured harness can fail
181+
*silently* — the run reports `succeeded` with empty results in seconds, and an
182+
agent that posts externally (GitHub reviews, Slack, tickets) will publish that
183+
garbage under the user's identity, once per dispatched call. So after any
184+
config change: send ONE representative call, confirm it did real work (nonzero
185+
cost/duration, plausible output — not just `succeeded`), then fan out the rest
186+
at full width. This is a gate on the first call after a config change, not a
187+
reason to serialize steady-state work.
188+
176189
## 4. Get the result
177190

178191
**What's in flight right now** — no IDs needed (also answers "how many agents

0 commit comments

Comments
 (0)