Fix: repair an incomplete JSONL tail before resuming session appends, so the next entry doesn't get lost too (#928)

[kaluli123123]

Summary

A crash mid-write can leave a session's .jsonl file's final record truncated. This is #928, one of the "related reports" under the still-open tracker #1380 (Make persisted session and configuration state crash-safe). Reproduced fresh on current main (06860844e) with a new failing-first regression test — not just read from the source.

Root cause

parseEntriesFromBuffer (and friends) already silently skip any line that fails JSON.parse, including a truncated trailing fragment — so in-memory state after resume looks fine. But the raw bytes of the fragment are never removed from disk, and the hot append path in _persist() calls appendFileSync() directly onto the existing file:

} else {
	mkdirSync(dirname(this.sessionFile), { recursive: true });
	appendFileSync(this.sessionFile, `${JSON.stringify(entry)}
`);
	...
}

The next real entry lands right after the fragment, producing one combined malformed line that also fails to parse — so the new entry is silently lost too, and this repeats on every future crash/resume cycle at that same file. Confirmed with a direct repro: write a valid header + message + a truncated trailing record (no closing brace, as a crash mid-fwrite would leave it), open the session, append a new message, reload — the new message is gone.

Fix

All in packages/coding-agent/src/core/session-manager.ts:

• The buffer/stream parsers now also track whether the file's final non-blank line existed but failed to parse, reporting it separately from ordinary mid-file skips. Only the trailing case is a safe, unambiguous repair target — an earlier malformed line is followed by content that already round-tripped, so it isn't attributable to an in-flight write and stays silently skipped exactly as before (out of scope, same as the row-validation done for Validate session entries and make append rollback re-entrancy-safe #709, per the original report's own note).
• New loadEntriesFromFileWithTail / loadEntriesFromFileAsyncWithTail expose { entries, incompleteTail }; the existing loadEntriesFromFile / loadEntriesFromFileAsync keep their original FileEntry[] signature as thin wrappers, so the one other call site (context-tree.ts) is unaffected.
• SessionManager.setSessionFile (used by both the sync SessionManager.open and, via a new preloadedIncompleteTail parameter, the async SessionManager.openAsync) now calls a new _repairIncompleteTail() after loading: it quarantines the fragment's raw bytes to a <file>.incomplete-tail-<ts>-<id>.jsonl sidecar for forensic recovery, logs a warning, and immediately rewrites the session file through the existing atomic temp-file+rename path (_rewriteFile) — so the corrupt bytes are gone from disk before any subsequent append can concatenate onto them, not just before the next write happens to occur.

Acceptance criteria from #928 — all covered by new tests

• Resume distinguishes a complete final record from an incomplete tail.
• Invalid tail bytes are quarantined before append (and truncated immediately, not deferred).
• A subsequent entry round-trips.
• Earlier valid records are preserved.
• Recovery is observable (a log line via getLogger("coding-agent.session-manager"), plus the quarantine sidecar).

Tests

Added 6 tests to packages/coding-agent/test/session-manager/file-operations.test.ts, all verified to fail against pre-fix main and pass after:

• loadEntriesFromFileWithTail/-AsyncWithTail report a truncated final record without dropping earlier entries.
• No tail is reported when the last record is valid but simply unterminated (no trailing newline) — not a corruption case.
• No tail is reported for a malformed line that isn't the last one (unchanged legacy skip behavior, intentionally out of scope).
• SessionManager.open and SessionManager.openAsync both: quarantine the fragment, keep earlier records, physically truncate the on-disk file immediately, and a subsequently appended entry round-trips through a reload instead of being lost.

Also ran the full session-manager-adjacent suite for regressions: test/session-manager/, session-manager-flush.test.ts, session-manager-git-state.test.ts, sdk-session-manager.test.ts, context-tree.test.ts — 161 + 14 passed, 0 failed. npx tsgo -p tsconfig.json --noEmit and the root npm run check (biome, tsgo, installer render, browser smoke) both pass. Added a packages/coding-agent/.changes/ fragment per the changelog-fragment CI check.

Patch

Branch: https://github.com/kaluli123123/prime-agent/tree/fix/session-jsonl-incomplete-tail-repair

Diff: main...kaluli123123:prime-agent:fix/session-jsonl-incomplete-tail-repair

Happy to open a PR from this branch if a maintainer wants to invite implementation, per CONTRIBUTING.md.