remove the unreachable persistent task-checkpoint writer

[Soph]

https://entire.io/gh/entireio/cli/trails/1096

persistent.go carried a full task-checkpoint writer — tasks/<tool-use-id>/ with per-task checkpoint.json, incremental NNN-<id>.json files, and a redacted subagent transcript — behind if opts.IsTask && opts.ToolUseID != "".

Nothing ever reached it. No producer in the tree sets IsTask: condensation builds one WriteOptions per session, and neither attach nor agentimport sets it. The api/checkpoint write union has no task variant at all, so a task write could not even be expressed. Measured on this repo's entire/checkpoints/v1: 0 of the 150 newest commits contain any tasks/ path.

Deleting the three functions alone would have left the scaffolding that made this look like working storage, so the always-empty parts go with them:

• the IsTask guard in applySessionWrite, and its taskMetadataPath return value
• buildCommitMessage's taskMetadataPath parameter — so a committed checkpoint no longer advertises an Entire-Metadata-Task trailer it never populated
• incrementalCheckpointData / taskCheckpointData
• the WriteOptions fields only that writer read: AgentID, CheckpointUUID, SubagentTranscriptPath, and the four Incremental* fields

IsTask/ToolUseID stay: they feed stored metadata that live readers consume (explain, checkpoint list), and the ephemeral store genuinely sets IsTaskCheckpoint on the pending path. They now carry a comment saying no producer sets them and pointing at #2058, so the seam isn't mistaken for storage again.

Untouched: the shadow branch's live tasks/<tool-use-id>/ writer in ephemeral.go, and the trailer formatter it uses.

Coverage

TestWriteCommitted_SubagentTranscript_JSONLFallback exercised only the dead path and goes with it. Its live twin TestWriteTemporaryTask_SubagentTranscript_RedactsSecrets covers the same invalid-JSONL redaction fallback on the ephemeral store, so the behaviour stays covered.

Doc and comment fixes

Things this made visible:

• manual_commit_condensation.go claimed re-reading subagent transcripts "finds nothing on this path once the agent has cleaned the transcripts up." Agents do not clean them up — Claude Code keeps 180 agent-*.jsonl (121 older than two days), Codex 164 rollouts (all older). Cost is the real and sufficient reason for subagentsDir="", so the false sentence is replaced by the measured fact.
• ref-checkpoint-backend.md listed pre-push OPF among the lifecycle paths that follow the primary, contradicting its own limitations section 200 lines later; and described a checkpoint ref's tree as containing tasks/….
• sessions-and-checkpoints.md now states outright that a committed checkpoint holds one record per session and none per subagent.

Verification

mise run fmt && mise run lint && mise run test:ci — 0 lint issues, no failures, 56 vogon + 4 roger-roger canary tests pass.

Refs #2058. This is the cleanup half of that issue; the storage design (payload location, OPF gate) is still open there.

🤖 Generated with Claude Code

---

[!NOTE] Cursor Bugbot is generating a summary for commit 8dae672. Configure here.

[Copilot]

Pull request overview

Removes the unreachable “persistent task-checkpoint writer” code path from the committed-checkpoint stores (git-branch + git-refs), and updates surrounding docs/comments/tests so committed checkpoints are described accurately as session-only records (no durable per-subagent/task subtree yet).

Changes:

• Deleted the persistent tasks/<tool-use-id>/... write path and its associated commit trailer plumbing from cmd/entire/cli/checkpoint/persistent.go (shared by both backends).
• Simplified the persistent write contract by removing now-unused task-writer-only fields from api/checkpoint.WriteOptions, and removed the dead test that only exercised that path.
• Updated architecture docs and in-code commentary to reflect current durable-storage behavior and the rationale for skipping subagent transcript re-reads during condensation.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
docs/architecture/sessions-and-checkpoints.md	Clarifies that committed checkpoints store one record per session and do not persist tasks/ detail.
docs/architecture/ref-checkpoint-backend.md	Aligns ref-backend description with actual committed subtree layout and OPF limitations.
cmd/entire/cli/strategy/manual_commit_condensation.go	Fixes/updates condensation rationale comment (cost vs. “agents cleaned up transcripts”).
cmd/entire/cli/checkpoint/refs_store.go	Removes now-dead persistent task writer plumbing from the git-refs store write path.
cmd/entire/cli/checkpoint/persistent.go	Removes the unreachable persistent task writer and related commit-message/trailer scaffolding.
cmd/entire/cli/checkpoint/checkpoint_test.go	Drops the test that only covered the removed persistent task writer path.
api/checkpoint/metadata.go	Removes task-writer-only fields from WriteOptions and documents IsTask/ToolUseID as a currently-unused seam.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Soph]

Closing — superseded by #2032, which I didn't know about when I opened this.

#2032's last ~9 commits (5cdde2e3a onward) build the durable subagent record: session.TaskRecords as the mid-turn pointer ledger, WriteOptions.Tasks []TaskPayload materialized into tasks/<tool-use-id>/{task.json, agent-<id>.jsonl} inside the session checkpoint. Along the way it deletes the same unreachable writer, the same WriteOptions fields, and the same TestWriteCommitted_SubagentTranscript_JSONLFallback test this PR did — and goes further, also dropping IsTask/ToolUseID, which I kept here as a seam.

Two of this PR's doc edits would also become wrong once #2032 lands: it states in sessions-and-checkpoints.md and ref-checkpoint-backend.md that a committed checkpoint has no tasks/ subtree, which is exactly what #2032 changes.

What survives, split out:

• docs(checkpoints): OPF does not follow the checkpoint primary #2083 — the one piece feat(claude-code): capture background subagent work durably (SubagentStop + task records) #2032 doesn't touch: ref-checkpoint-backend.md listing pre-push OPF among the paths that follow the checkpoint primary, contradicting its own limitations section.
• The false comment in manual_commit_condensation.go ("finds nothing on this path once the agent has cleaned the transcripts up" — agents don't clean them up) belongs in feat(claude-code): capture background subagent work durably (SubagentStop + task records) #2032, which already edits that file. Raising it there rather than conflicting over it.