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
#963 shipped only its first row — the nested-run refusal (PR #1080). The rest of the host-worker guard is still convention, not a tested contract, so a worker that writes outside its allowed dirs or asks for an unbounded run is caught by nothing but review. This is the remainder of #963.
Add write-dir enforcement on every write path:
apply AGENTDESCENT_SURFACE_WRITE_DIRS in apply, in show --diff output and in the materialised tree, not only on the MCP surface;
reject the write before it touches disk, with a test that a path outside the allowed dirs is refused.
Add host-worker budget defaults:
max_calls, max_usd (or max_tokens) and max_concurrency, plus a per-Ticket cap the caller cannot silently exceed;
a run that reaches the cap stops with stop_reason=budget (or the existing equivalent) instead of crashing.
Add the child-env scrubbing test:
an offline test that the child env carries no host session markers and no host MCP config.
Acceptance: apply outside the allowed write dirs is rejected before disk is touched; a capped run stops with stop_reason=budget rather than crashing; the child env has no host session markers and no host MCP config; all three behaviours covered by offline tests, no new sandbox and no auth layer; docs/changelog.
Remainder of #963. Not assigned; on the bench until it is picked up.
Claimed for AgentDescent Dev-1, row 1 only: write-dir enforcement on every write path.
Branch: agent/1082-write-dir-enforcement. The PR will say Partial: write-dir enforcement for #1082 and will not close this issue: rows 2 (host-worker budget caps) and 3 (child-env scrubbing) stay here and get their own carriers after row 1 lands.
Row 1 scope: AGENTDESCENT_SURFACE_WRITE_DIRS applied in apply, in show --diff output and in the materialised tree, not only on the MCP surface; a write outside the allowed dirs is refused before it touches disk, pinned by a test.
Claiming one slice of this remainder, not the whole issue.
Branch agent/1082-operator-run-cap off main @ 14ba9529: an operator per-run call ceiling, AGENTDESCENT_MAX_CALLS, clamped (min only) into runstore.launch / runstore.execute so a caller cannot raise it. Calls only — a process-wide AGENTDESCENT_MAX_USD would make every un-priced run a SpecError through require_budget_pair, so the USD question stays open. Not in this slice: max_concurrency, the MCP tool arguments, doctor, write-dir and child-env work (the last two rows are already shipped, PRs #1087 and #1088).
Row status, so nobody re-opens the shipped ones the way I just did:
Row 1 (write-dir enforcement on every write path) — shipped, PR #1087; recorded in this thread at 2026-11-20T05:12:58Z. I opened #1188 for it from the acceptance text alone and closed it the same hour with that evidence.
Row 3 (child-env scrubbing test) — shipped, PR #1088 (23979b73), tests/test_host_env_scrub.py at f5017f93 pins the spawn boundary (three launch rows) and every SESSION_MARKERS entry. I opened #1189 for it and closed it the same way.
Row 2 (host-worker budget defaults) — the only live part: AGENTDESCENT_MAX_CALLS shipped as the operator ceiling in #1181 (Partial: #1082), and the MCP surface arguments are in review as #1186. Still open here: max_concurrency, a per-Ticket cap, and the USD question. Not sliced yet on purpose — it edits the same mcp.py helper #1186 is changing, and a second PR on that helper would be the real conflict.
So the remaining acceptance for this issue is row 2 only.
Activity
HugoXK commentedon Nov 14, 2026
Claimed for AgentDescent Dev-1, row 1 only: write-dir enforcement on every write path.
Branch:
agent/1082-write-dir-enforcement. The PR will sayPartial: write-dir enforcement for #1082and will not close this issue: rows 2 (host-worker budget caps) and 3 (child-env scrubbing) stay here and get their own carriers after row 1 lands.Row 1 scope:
AGENTDESCENT_SURFACE_WRITE_DIRSapplied inapply, inshow --diffoutput and in the materialised tree, not only on the MCP surface; a write outside the allowed dirs is refused before it touches disk, pinned by a test.fix(runstore): scrub the host session out of a detached run's env (#1082
feat(cli): refuse a nested run with a named reason (#963) (#1080)
feat(cli): apply stays inside the worker's write dirs (#1082) (#1087)
fix(runstore): share one host-session denylist and re-scrub the merge…
fix(runstore): keep the host config dirs a detached run reads (#1082)
fix(runstore): keep XDG_CONFIG_HOME for the detached run too (#1082)
fix(runstore): keep the host session out of a detached run's env (#1082…
docs(hosts): state the kept config roots as an exception to the env s…
HugoXK commentedon Nov 20, 2026
Claiming one slice of this remainder, not the whole issue.
Branch
agent/1082-operator-run-capoffmain@14ba9529: an operator per-run call ceiling,AGENTDESCENT_MAX_CALLS, clamped (min only) intorunstore.launch/runstore.executeso a caller cannot raise it. Calls only — a process-wideAGENTDESCENT_MAX_USDwould make every un-priced run aSpecErrorthroughrequire_budget_pair, so the USD question stays open. Not in this slice:max_concurrency, the MCP tool arguments,doctor, write-dir and child-env work (the last two rows are already shipped, PRs #1087 and #1088).fix(runstore): keep the USD stop silent about the call ceiling; label…
fix(runstore): only name the ceiling when the applied cap is the ceil…
test(runstore): cover both hand-set request cases and the empty-strin…
feat(runstore): operator call ceiling `AGENTDESCENT_MAX_CALLS` (Partial
show --diffchecks only the destination, so it advertises a writeapplyrefuses on the backup path #1188HugoXK commentedon Nov 20, 2026
Row status, so nobody re-opens the shipped ones the way I just did:
#1087; recorded in this thread at2026-11-20T05:12:58Z. I opened#1188for it from the acceptance text alone and closed it the same hour with that evidence.#1088(23979b73),tests/test_host_env_scrub.pyatf5017f93pins the spawn boundary (threelaunchrows) and everySESSION_MARKERSentry. I opened#1189for it and closed it the same way.AGENTDESCENT_MAX_CALLSshipped as the operator ceiling in#1181(Partial: #1082), and the MCP surface arguments are in review as#1186. Still open here:max_concurrency, a per-Ticket cap, and the USD question. Not sliced yet on purpose — it edits the samemcp.pyhelper#1186is changing, and a second PR on that helper would be the real conflict.So the remaining acceptance for this issue is row 2 only.
feat(mcp): cap a run from the MCP surface -- `plan` and `start` take …