@uzysjung/agent-harness 26.119.0 → 26.120.0
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/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@uzysjung/agent-harness",
|
|
3
|
-
"version": "26.
|
|
3
|
+
"version": "26.120.0",
|
|
4
4
|
"description": "Curate vetted AI-coding skills & plugins by your tech stack — install only what you need, across Claude Code, Codex, OpenCode & Antigravity",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"publishConfig": {
|
|
@@ -143,6 +143,13 @@ after their job ended. The clutter compounds per delegation.
|
|
|
143
143
|
still-open agent is a declared decision, not a leak.
|
|
144
144
|
- **Sweep at checkpoints**: at phase end and during [[compaction-handoff]], list running agents
|
|
145
145
|
and stop every finished one before moving on.
|
|
146
|
+
- **Collect the result as a file, not as a return message.** A long final message can be dropped
|
|
147
|
+
in transit, and the failure is silent — it reads as "the worker produced nothing", not "the
|
|
148
|
+
transport lost it", so you retry the work instead of the delivery. Name a scratchpad path in the
|
|
149
|
+
spawn prompt, have the worker write there (an early partial, overwritten as it refines), then
|
|
150
|
+
read the file. This is a **spawn-time** decision: retrofitting costs a re-request per worker, and
|
|
151
|
+
one that has already ended may not reproduce what it had. Inline returns are for short answers
|
|
152
|
+
you can afford to lose.
|
|
146
153
|
|
|
147
154
|
## V&V separation
|
|
148
155
|
|
|
@@ -102,6 +102,12 @@ one gets the artifact + goal + audience + the *same* rubric, and **must not see
|
|
|
102
102
|
output** — independence is the precondition that makes aggregation add information. Anchoring on a
|
|
103
103
|
peer collapses the panel toward one effective vote.
|
|
104
104
|
|
|
105
|
+
**Have each persona write its findings to a file, not return them inline** — a panel report is
|
|
106
|
+
exactly the long payload that gets dropped in transit, and the loss is silent. Losing one persona's
|
|
107
|
+
report does not merely cost that report: it quietly shrinks the panel, which is the one variable
|
|
108
|
+
the whole method depends on (see "Nine Judges, Two Effective Votes" above). Name the path when you
|
|
109
|
+
spawn, not after — see [[model-orchestration]] "Worker lifecycle" for the general rule.
|
|
110
|
+
|
|
105
111
|
Prefer pinning the persona sub-agents to a cheaper tier (Sonnet) — see the cost-tier note. But this
|
|
106
112
|
degrades gracefully: if the harness can't pin sub-agents to a specific model, just run the panel on
|
|
107
113
|
the default sub-agent model and note in the step-6 coverage caveat that the panel ran at the
|