You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
fix(cli): --json evolve and --json show drop stop_detail, so the machine output cannot say which cap stopped the run #1190
Implementation: AgentDescent (ours) — draft PR #1191 exists; it was opened outside our release order and is being traced, do not pick up.
The run's own machine output cannot say which cap stopped it. status.json and status --json can; the JSON of the command that actually stopped cannot.
What is true today (read at f5017f93)
--json evolve builds its payload at cli.py:2763-2770 and never reads st.stop_detail, so it prints "stop_reason": "budget" with no key naming the limit.
--json show builds at cli.py:3118-3119 (stop_reason, error) and also drops it.
stop_detail appears in cli.py only at :83 (comment), :185 and :186 — nothing on the JSON side.
Repro (offline, counting-agent spec, all three from --json evolve)
limit
--json evolve
status.json
spec max_calls: 6
{"stop_reason": "budget"}, no key
max_calls: 6 call(s) made, cap 6
AGENTDESCENT_MAX_CALLS=2
same
max_calls (AGENTDESCENT_MAX_CALLS): 2 call(s) made, cap 2
--max-usd 0.01 --usd-per-call 0.5
same
max_usd: $0.0000 spent, cap $0.0100
#1181 sharpened the cost: the operator label is now the only thing distinguishing the ceiling from the caller's own flags, and it does not reach this surface.
Acceptance
stop_detail appears in both payloads only when it is set, so an uncapped run's payload is byte-identical to today (additive, no schema break);
one test row per builder asserting the key's absence without a stop and its presence with the ceiling's label; the absence row is what stops a blanket None from being emitted;
no change to stop_reason, to the human line, or to status.json;
Partial: #1166 (its acceptance named status.json + status only — this is the gap it left, not drift).
Evidence the PR must carry
red-first: the new row raises KeyError: 'stop_detail' on the PR's base, with the failing id and the exact command;
green at the tip, same command, plus the touched-file suite;
the per-file diff.
Sequencing: the red test and the 8-line fix can be prepared now; the PR opens when a queue slot frees (the Lead says when), so the review seat never holds three AD PRs at once.
Activity