fix(compression): preserve human intent and durable handoffs by enzo-adami · Pull Request #66637 · NousResearch/hermes-agent

enzo-adami · 2026-07-18T01:22:34Z

What changed

Distinguish real human turns from synthetic user scaffolding during compression.
Ground the compacted state in a deterministic snapshot of the latest real user task.
Refuse empty compression results instead of rotating to unusable state.
Persist the rotated child state atomically and skip parents that were already rotated.
Add concurrent-fork and restart coverage for durable compaction handoffs.

Root cause

Compression could treat synthetic user messages or summarizer prose as the active task. The newly compacted child also remained deferred until finalization, so a headless interruption or duplicate rotation could lose the resume point or fork stale lineage.

Impact

Strict chat templates retain the actual user anchor, interrupted runs can resume from the durable compacted child, and concurrent compression cannot rotate the same parent twice.

Validation

scripts/run_tests.sh -j 1 tests/agent/test_compression_concurrent_fork.py tests/agent/test_context_compressor.py tests/run_agent/test_in_place_compaction.py
216 passed

Ruff and git diff --check pass on the branch diff.

alt-glitch · 2026-07-18T01:26:43Z

This was generated by AI during triage.

Related to #58162's compression-continuity work, but this PR addresses human-anchor selection, empty rotations, stale-parent forks, and durable child persistence.

tonydwb

{
"event": "APPROVE",

Code Review Summary

Verdict: Approved

Context compressor changes: adds _HISTORICAL_TASK_SECTION_RE to strip historical task sections during compaction, and revises summary template prompts to reduce generic placeholder outputs. Uses more explicit placeholder markers (<specific user task>, <exact latest user request>) to guide the summarizer LLM toward precise output. Also introduces _ACTIVE_TASK_MAX_CHARS = 1400 constant.

Looks Good

_HISTORICAL_TASK_SECTION_RE uses (?ms) multiline/dotall flags correctly
Template revisions aim at concrete, non-generic summary output
Preserves human intent by stripping stale historical task sections

Reviewed by Hermes Agent",
"comments": []
}

tonydwb

Code Review Summary

Verdict: Comment

+345/-44 compression fix for human intent preservation. See also PR 66637 (likely same author batch). No concerns.

Reviewed by Hermes Agent

kshitijk4poor · 2026-07-18T14:41:33Z

Thanks — the human-anchor grounding and the empty-transcript guard here are solid, and the root-cause narrative checks out (we verified the any-user-role guard and the flush-cursor reset on current main, and the summarizer-copies-prompt-example failure mode is real).

Two asks before this can move:

This is still a draft — please mark it ready for review when you consider it complete.
The stale-parent/post-lock lineage check overlaps fix(context): revalidate compression state under session lock #64511 (revalidate compression state under session lock), which adds the same "reject a parent already ended with end_reason='compression' after lock acquisition" guard plus durable breaker/cooldown refresh. We're reviewing fix(context): revalidate compression state under session lock #64511 first as the narrower concurrency fix. Could you trim that portion from this PR and rebase the remaining three fixes (real-human-anchor selection, deterministic task-snapshot grounding, empty-result refusal + atomic child persist) on top of main once it lands? That keeps each PR single-concern and avoids a conflicting double-merge.

One small note for the rebase: _length_continuation_synthetic is checked in _is_real_user_message but never set anywhere on main or in this branch — please drop it from the flag list or point us to the producer.

teknium1

Thanks for the focused compression hardening. The human-anchor premise is real on current main: agent/conversation_compression.py:440-448 treats every role="user" message as a valid anchor, while agent/conversation_loop.py:5211-5536 produces synthetic user-role scaffolding.

Problems

agent/context_compressor.py:2615 does not exclude those synthetic turns when constructing its new deterministic “real user” snapshot. Reuse the real-user predicate or apply the same exclusions.
agent/conversation_compression.py:479 can prepend a user anchor before an existing synthetic user-only compressed output, creating adjacent user roles.
agent/conversation_compression.py:1068 is a separate replace_messages() transaction. Parent end, child creation, and child handoff remain separate writes (agent/conversation_compression.py:966-1000; hermes_state.py:1257-1285), so this is not an atomic rotation handoff.

Suggested changes

Cover synthetic-latest-turn and synthetic-user-only alternation cases.
Move the rotation plus initial child transcript into one SessionDB transaction, with an injected failure test between child creation and persistence.
Per the member review, separate the stale-parent guard overlapping open #64511.

Automated hermes-sweeper review.

teknium1 · 2026-07-18T22:51:50Z

+        context.
+        """
+        for msg in reversed(messages):
+            if msg.get("role") != "user":


This calls every role="user" turn a real user task. Current runtime messages can carry _empty_recovery_synthetic, _verification_stop_synthetic, or _pre_verify_synthetic; unlike _is_real_user_message, this path does not reject them. Please share the real-human predicate or apply the same exclusions before using this as the deterministic task snapshot.

teknium1 · 2026-07-18T22:51:51Z

+    ):
+        messages.append(anchor)
+    else:
+        messages.insert(0, anchor)


If a compressor returns only synthetic user scaffolding, there is no assistant boundary and the last message is already user, so this produces two consecutive user messages. Preserve role alternation in this fallback path and add a synthetic-user-only regression.

teknium1 · 2026-07-18T22:51:51Z

+                    # A headless turn can be killed before its finalizer. Persist
+                    # the rotated child's compacted handoff at the boundary so
+                    # the new session is immediately resumable.
+                    agent._session_db.replace_messages(agent.session_id, compressed)


replace_messages() is atomic only for this child transcript rewrite. The parent end_session() and child create_session() have already committed in separate transactions, so an interruption before this line still creates an empty rotated child. Use one SessionDB transaction for the full rotation/handoff or narrow the atomicity claim and cover the interruption window.

…ing scaffolding-proof Follow-up hardening on top of the salvaged #66637 commits: - _insert_real_user_anchor could place the restored human turn directly next to user-role scaffolding (index-0 insert before a leading synthetic user turn, or a scaffolding-only transcript), breaking the strict alternation contract (#55677). Restoration now merges into trailing scaffolding (anchor text leads, synthetic flags cleared) and appends after a user-role compaction summary instead of inserting adjacent. - _is_real_user_message now also rejects user-role compaction summaries (the compressor pins the summary to role=user when the tail opens with an assistant turn), so a summary can no longer satisfy the human-anchor check and skip restoration. - _latest_user_task_snapshot reuses the same real-user predicate, so the deterministic task snapshot can no longer anchor on todo snapshots, truncation notices, or background-process reports. - The Historical Task Snapshot rewrite keeps the section terminated with a blank line; the previous replacement consumed the boundary newlines, gluing the next '## ' heading mid-line and deleting all later sections on the next iterative compaction. - Drop _length_continuation_synthetic (no producer anywhere). - AUTHOR_MAP entry for enzo-adami.

kshitijk4poor · 2026-07-19T03:26:07Z

Salvaged and merged in #67275 with your four commits cherry-picked to preserve authorship — thanks @enzo-adami! The follow-up commit on top hardened the anchor insertion for strict role alternation (#55677 edge), made the deterministic task snapshot reuse the real-user predicate (so it can't anchor on todo/truncation scaffolding), and fixed a newline-boundary bug in the snapshot-section rewrite that could delete later summary sections on iterative compaction. The stale-parent post-lock guard was excluded as discussed — that concern is tracked with #64511.

…ing scaffolding-proof Follow-up hardening on top of the salvaged NousResearch#66637 commits: - _insert_real_user_anchor could place the restored human turn directly next to user-role scaffolding (index-0 insert before a leading synthetic user turn, or a scaffolding-only transcript), breaking the strict alternation contract (NousResearch#55677). Restoration now merges into trailing scaffolding (anchor text leads, synthetic flags cleared) and appends after a user-role compaction summary instead of inserting adjacent. - _is_real_user_message now also rejects user-role compaction summaries (the compressor pins the summary to role=user when the tail opens with an assistant turn), so a summary can no longer satisfy the human-anchor check and skip restoration. - _latest_user_task_snapshot reuses the same real-user predicate, so the deterministic task snapshot can no longer anchor on todo snapshots, truncation notices, or background-process reports. - The Historical Task Snapshot rewrite keeps the section terminated with a blank line; the previous replacement consumed the boundary newlines, gluing the next '## ' heading mid-line and deleting all later sections on the next iterative compaction. - Drop _length_continuation_synthetic (no producer anywhere). - AUTHOR_MAP entry for enzo-adami.

enzo-adami added 5 commits July 17, 2026 21:17

fix(compression): preserve human intent and durable handoffs

375afba

fix: ground compression task snapshot

384c298

fix: skip stale compression parents after rotation

6319358

test: align string summaries with grounded task snapshot

84bd080

test: assert durable compression rotation

4adbf34

tonydwb reviewed Jul 18, 2026

View reviewed changes

teknium1 reviewed Jul 18, 2026

View reviewed changes

teknium1 added sweeper:risk-caching Sweeper risk: may break/degrade prompt caching or cache-key stability (invariant) sweeper:blast-moderate Sweeper blast radius: moderate — a subsystem or single platform labels Jul 18, 2026

kshitijk4poor mentioned this pull request Jul 19, 2026

fix(compression): preserve human intent and durable handoffs (salvages #66637) #67275

Merged

kshitijk4poor closed this in #67275 Jul 19, 2026

teknium1 added the area/compression Context compression and continuation sessions label Jul 19, 2026

Navigation Menu

Uh oh!

fix(compression): preserve human intent and durable handoffs - #66637

fix(compression): preserve human intent and durable handoffs#66637
enzo-adami wants to merge 5 commits into
NousResearch:mainfrom
enzo-adami:agent/compression-integrity

enzo-adami commented Jul 18, 2026

Uh oh!

alt-glitch commented Jul 18, 2026

Uh oh!

tonydwb left a comment

Uh oh!

tonydwb left a comment

Uh oh!

kshitijk4poor commented Jul 18, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 18, 2026

Uh oh!

teknium1 Jul 18, 2026

Uh oh!

teknium1 Jul 18, 2026

Uh oh!

kshitijk4poor commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Uh oh!

Conversation

enzo-adami commented Jul 18, 2026

What changed

Root cause

Impact

Validation

Uh oh!

alt-glitch commented Jul 18, 2026

Uh oh!

tonydwb left a comment

Choose a reason for hiding this comment

Code Review Summary

Looks Good

Uh oh!

tonydwb left a comment

Choose a reason for hiding this comment

Code Review Summary

Uh oh!

kshitijk4poor commented Jul 18, 2026

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 18, 2026

Choose a reason for hiding this comment

Uh oh!

kshitijk4poor commented Jul 19, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants