fix(compression): adopt live continuation tip at flush across multi-hop chains

[teknium1]

Summary

A turn writing against a session already closed by context compression died with session_persistence_failed and the misleading "this is often a full disk" dialog — even though the store was healthy and a live continuation existed. Fixes #82001.

The depth-1 recovery API (find_live_compression_child) could not resolve lineages with ≥2 compression hops (root(compressed) → mid(compressed) → tip(live)): the direct child is itself compression-ended, so the lookup returned nothing, adoption failed closed, and the turn died. Reproduced independently three times on this issue (two-hop production trace, minimal root → mid → tip probes, and a three-hop deep-research workload).

Changes

• Agent flush (run_agent.py _flush_messages_to_session_db_unlocked): on CompressionSessionClosedError, resolve tip = db.get_compression_tip(old_id) — the canonical bounded transitive walk (excludes _branched_from/_delegate_from/source='tool' children, prefers live over stale closed siblings like ws_orphan_reap). Adopt only when tip != old_id and the tip row is live (ended_at IS NULL), then retry the flush exactly once (adoption budget) — a second closed-parent write fails closed, no loop. get_compression_tip returns the input id when no continuation exists, so tip == old_id is treated as no-adopt/fail-closed.
• Gateway transcript flush (gateway/session.py append_to_transcript): the reroute now uses the same tip + liveness contract instead of the depth-1 lookup, so multi-hop lineages resolve on the gateway write side too.
• Turn-start recovery preflight (agent/conversation_compression.py _adopt_live_compression_child): the last depth-1 consumer in this family. It resolved a unique live direct child, loaded its transcript, and revalidated before mutating the agent — semantics that map cleanly onto tip resolution, so it now resolves via get_compression_tip with the same liveness check and post-load revalidation (confirmed tip == adopted tip). Fail-closed behavior for no-continuation / closed-tip lineages is preserved; recover_rotated_compression_session's orphan-reopen fallback is untouched.
• Diagnostics: classify_persistence_error gains a compression_closed bucket (added to PERSISTENCE_ERROR_CAUSES; matched by exception type and by the "closed by compression" phrase for RPC-wrapped strings). The turn-completion explanation for this cause names compression rotation and tells the client to refresh/pick up the new session id — it never mentions disk. Real disk failures keep the disk guidance. The per-turn _compression_adoption_failed diagnostic flag is reset at turn start.

Fail-closed semantics are strictly preserved: no live tip → no adoption, no message content or ordering is touched (this is session-identity plumbing only), and the adopted retry replays the exact same unpersisted batch.

Tests

tests/run_agent/test_compression_closed_adoption.py (new, real temp SQLite via hermes_state, no live model):

• depth-1 unique live child adoption at flush (message lands in the child; closed parent stays immutable)
• multi-hop chain root(compressed) → mid(compressed) → tip(live) adopts the tip — the regression the depth-1 API missed
• no continuation → fail closed, flush returns False, cause = compression_closed
• stale-closed tip (ws_orphan_reap) → fail closed, never adopt a closed session
• adoption budget: tip lookup runs exactly once; a second closed write after adoption fails closed (no loop)
• wording guards: compression-closed explanation never contains "disk"; the disk cause keeps "full disk" guidance

tests/gateway/test_session.py: multi-hop chain reroute to the live tip, fail-closed on a stale-closed tip, and the existing backlog-migration mock moved to the canonical get_compression_tip API.

tests/run_agent/test_compression_closed_adoption.py ........ 8 passed
tests/gateway/test_session.py (TestGatewaySessionDbRecovery) 6 passed
Regression: tests/gateway/test_session.py, tests/run_agent/test_turn_completion_explainer.py,
tests/state/test_compression_lineage_guard.py, tests/agent/test_compression_* ,
tests/run_agent/test_cross_process_turn_lease.py, tests/test_hermes_state.py,
tests/cron/test_scheduler.py, tests/run_agent/test_413_compression.py,
tests/run_agent/test_860_dedup.py, test_81641 → 539 passed, 1 pre-existing
failure (tests/test_hermes_state.py FTS5 projection test, fails on clean origin/main too)

Relationship to existing work

• Supersedes fix(agent): adopt live compression child on flush #79763 (depth-1 flush adoption) — @KCAYAAI's fail-closed framing is kept as the core contract here; the resolver is upgraded to the full-chain canonical API and the gateway + preflight call sites are covered too.
• Design, working-tree diff, and test design by @Al3xand3r1987; independent root → mid → tip probes by @yuzilongleif-collab; canonical get_compression_tip pointer and landing plan by @ayushnangia (credited via Co-authored-by).
• Complements fix(sessions): recover orphaned compression parents without continuations (#80337) #80487 (orphan reopen, no-continuation case) — untouched.
• Part 2 from the issue (returning the adopted session id through the stream / WebUI session-index convergence) remains a separate client-facing change and is intentionally not in this PR.

Infographic

[github-actions]

૮ >ﻌ< ა ci review

_{running on 3d6cedb — fix(compression): adopt live continuation tip at flush acros}

---

_{waiting for more jobs to start…}

⚠️ Warnings

OSV vulnerability scan · View job

5 known vulnerabilities found in pinned dependencies.

• CVE-2026-67213 in package-lock.json
• CVE-2026-67213 in website/package-lock.json
• CVE-2026-70608 in package-lock.json
• CVE-2026-56876 in package-lock.json
• CVE-2026-70606 in package-lock.json

How to fix:

Review the findings in the Security tab. Update the affected dependencies if a patched version is available.