Repository navigation
Context window rollover still fails after #5205 closed as completed: checkpoint.md absent recurs on 1.0.105 #5312
Description
Filing this as a new issue because #5205 — which carries the original report, an independent reproduction, and my 1.0.105 reproduction — was closed as completed on 2026-10-30 with no closing comment, no linked PR or commit, and no release containing a fix. The latest release is still 1.0.105 (2026-10-28), and no changelog entry in any version mentions checkpoint.md or the rollover refusal. Only the author or maintainers can reopen #5205, so the post-closure evidence also lives here. (Reopen request with the same evidence: #5205 (comment); my original reproduction: #5205 (comment))
Environment: Copilot CLI 1.0.105 (current latest release), Windows 11.
Failure: "Failed to start a fresh context window: checkpoint.md is absent for this agent; write it before requesting rollover"
Recurrences since the #5205 closure, both on 1.0.105:
- 2026-10-30 ~18:44 UTC — resuming a parked long-running session: the rollover was refused with the exact error above. The session had zero compaction events and its checkpoints/ folder held only the stub index.md, so no checkpoint existed to satisfy the gate.
- 2026-10-31T03:05:38Z — a delegated agent launch died at startup with the same refusal delivered as a session.error event ("Execution failed: Failed to start a fresh context window: checkpoint.md is absent for this agent; write it before requesting rollover"), aborting the parent run's shutdown records mid-sequence.
Pattern (unchanged across all reports): in a session where no compaction has run, no checkpoint is ever registered — automatic compaction does not create one — and any fresh-context rollover in that session is then refused. The agent cannot create the checkpoint itself, and no documented mechanism lets it; the only workaround is a manual /compact by the user before the rollover point.
Expected: either rollover does not require a checkpoint the system never created, or the checkpoint is created by the automatic path the documentation describes. If a fix has shipped somewhere not visible in the releases or on #5205, a pointer to the PR or release would settle both issues.
Activity
dewers99 commentedon Oct 31, 2026
Third recurrence since the #5205 closure — this one ~15 minutes after this issue was filed. Copilot CLI 1.0.105, Windows 11, session 1a3bd087 (long-running orchestrator session). Notably, a manual /compact had completed in this session at 2026-10-31T11:20:35Z and registered a checkpoint (checkpoints/001-wo-11-r4-pause-wrap-up.md, 7,649 bytes, plus index.md). At 2026-10-31T12:36:38.807Z the session still logged session.error: "Execution failed: Failed to start a fresh context window: checkpoint.md is absent for this agent; write it before requesting rollover." So the refusal is not explained by the absence of any compaction in the session — a registered checkpoint existed and the rollover was refused anyway.
dewers99 commentedon Nov 1, 2026
Recurrence -- 2026-10-31, CLI 1.0.105 on Windows 11. Resuming a multi-agent dispatcher run, the launch of a stage subagent (auditor) failed before its first turn with the same error: "Failed to start a fresh context window: checkpoint.md is absent for this agent; write it before requesting rollover". A second launch attempt under a different session ID produced an output file but no completion, leaving the run's state claiming the stage is running when it is not. The parent session stayed alive and coherent throughout -- it banked its records and reported the conflict itself -- so the failure is specifically the subagent's fresh-window start with no checkpoint present: automatic compaction never registered one, and the rollover refuses without it.
dewers99 commentedon Nov 1, 2026
Follow-up and correction, 2026-11-01 — I found the root cause of the checkpoint.md absence in my own setup, and it narrows this report.
My multi-agent workflow runs custom agents whose definition files declare a
tools:allowlist. Those lists omitted the context-management tools (session_artifacts, get_context_remaining, session_history, new_context). The CLI setting contextManagementTools was on, but it only allows the tools — an agent running under a restrictivetools:list never receives them, so my agents could never create a checkpoint, no matter how often their instructions told them to. Every checkpoint that existed during the affected runs was created by me typing /compact by hand.Controlled check: an agent definition with no
tools:list receives all four tools. After adding the four tool names to my agents'tools:lists, fresh sessions and reopened existing sessions list all four, and the previously stuck session can checkpoint itself. Fixed on my side, in my agent definitions.So I withdraw the implication that the CLI failed to create checkpoint.md. What remains, and what I am still reporting: when checkpoint.md is absent — for whatever reason, including a misconfiguration like mine — a context-window rollover does not fail cleanly. On 2026-10-31 (~21:52 ET) a long-running session died mid-work during rollover, with no error surfaced and nothing recorded; the in-flight work survived only because an external record existed. A visible refusal — "cannot roll over: no checkpoint" — would have changed the outcome.
If that behavior is by design, I am happy to close this. I am leaving it open for now because the silent death is the part I believe is worth a look.
copilot-cli-bot commentedon Nov 11, 2026
This appears to be resolved as of GitHub Copilot CLI v1.0.109. Please update to that release or a newer stable version. We're closing this as completed. If it still happens, please comment with your version and reproduction steps.