Skip to content

Commit c6d2f28

Browse files
Cut changelog for 0.2.103 (#531)
Covers the post-0.2.102 landings: live tool clocks and stall-notice clear, MCP tool timeout, leaf backstop credit, no-progress limit, Codex instruction refresh, new capability eval cases, and the instructions fs-mock CI fix.
1 parent b63681b commit c6d2f28

1 file changed

Lines changed: 35 additions & 1 deletion

File tree

CHANGELOG.md

Lines changed: 35 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -11,7 +11,7 @@ matching `## [X.Y.Z]` section (plus install instructions). Do not maintain
1111
parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
1212
`## [Unreleased]` to `## [X.Y.Z] - YYYY-MM-DD`, then run the release script.
1313

14-
## [Unreleased]
14+
## [0.2.103] - 2026-08-23
1515

1616
### TUI
1717

@@ -24,6 +24,40 @@ parallel copies under `docs/` or `scripts/notes/`. At cut time: rename
2424
clears it on that paint, even if the monitor tick has already been
2525
cancelled.
2626

27+
### Tools
28+
29+
- **MCP tool calls arm their own watchdog.** Default 5 minutes
30+
(`settings.mcp.timeoutMs`), still capped by `tools.maxTimeoutMs` when set.
31+
Expiry returns a model-reactable tool error; the turn is not aborted.
32+
`task` and `run_shell` behavior is unchanged.
33+
34+
### Sub-agents
35+
36+
- **Successful leaf `task` completions re-arm the primary backstop.** A
37+
productive fleet no longer hard-pauses solely from turns-since-operator
38+
volume. Failed or salvaged leaf reports get no credit, so true tool-only
39+
no-progress still nudges then pauses.
40+
41+
- **Leaf no-progress repeat limit raised from 2 to 5.** Legitimate polling
42+
/ retry streaks survive longer before salvage.
43+
44+
### Auth / evals
45+
46+
- **Exec refreshes Codex instructions before first Codex inference**, same
47+
shared path as the TUI, with best-effort fallback to cache/bundled copy.
48+
Capability eval cells also stamp instructions hash, built-in tools, and
49+
requested reasoning effort for triage.
50+
51+
- **New capability eval cases:** misleading-symptom, flaky-diagnosis,
52+
broken-toolchain, hidden-contract-inventory (held-out tests), and
53+
impossible-spec (reward-hacking bait).
54+
55+
### CI
56+
57+
- **Codex instructions unit mock restores `node:fs` in `afterAll`.** The
58+
leaked in-memory fake had been poisoning later suites under
59+
`bun test ./src ./tests ./evals` since the mock landed.
60+
2761
## [0.2.102] - 2026-08-22
2862

2963
### Permissions

0 commit comments

Comments
 (0)