mirror of
https://github.com/NousResearch/hermes-agent.git
synced 2026-07-06 12:52:11 +00:00
fix(agent): release refreshed compression locks on every exit path (#54465)
This commit is contained in:
Rod Boev
kshitij
parent
53ef954841
commit
f2ace45286
2 changed files with 314 additions and 284 deletions
|
|
@ -579,314 +579,316 @@ def compress_context(
|
|||
_release_lock() # compression aborted — no rotation will happen
|
||||
return messages, _existing_sp
|
||||
|
||||
summary_error = getattr(agent.context_compressor, "_last_summary_error", None)
|
||||
if summary_error:
|
||||
if getattr(agent, "_last_compression_summary_warning", None) != summary_error:
|
||||
agent._last_compression_summary_warning = summary_error
|
||||
agent._emit_warning(
|
||||
f"⚠ Compression summary failed: {summary_error}. "
|
||||
"Inserted a fallback context marker."
|
||||
)
|
||||
else:
|
||||
# No hard failure — but did the configured aux model error out
|
||||
# and get recovered by retrying on main? Surface that so users
|
||||
# know their auxiliary.compression.model setting is broken even
|
||||
# though compression succeeded.
|
||||
_aux_fail_model = getattr(agent.context_compressor, "_last_aux_model_failure_model", None)
|
||||
_aux_fail_err = getattr(agent.context_compressor, "_last_aux_model_failure_error", None)
|
||||
if _aux_fail_model:
|
||||
# Dedup on (model, error) so we don't spam on every compaction
|
||||
_aux_key = (_aux_fail_model, _aux_fail_err)
|
||||
if getattr(agent, "_last_aux_fallback_warning_key", None) != _aux_key:
|
||||
agent._last_aux_fallback_warning_key = _aux_key
|
||||
try:
|
||||
summary_error = getattr(agent.context_compressor, "_last_summary_error", None)
|
||||
if summary_error:
|
||||
if getattr(agent, "_last_compression_summary_warning", None) != summary_error:
|
||||
agent._last_compression_summary_warning = summary_error
|
||||
agent._emit_warning(
|
||||
f"ℹ Configured compression model '{_aux_fail_model}' failed "
|
||||
f"({_aux_fail_err or 'unknown error'}). Recovered using main model — "
|
||||
"check auxiliary.compression.model in config.yaml."
|
||||
f"⚠ Compression summary failed: {summary_error}. "
|
||||
"Inserted a fallback context marker."
|
||||
)
|
||||
|
||||
todo_snapshot = agent._todo_store.format_for_injection()
|
||||
if todo_snapshot:
|
||||
compressed.append({"role": "user", "content": todo_snapshot})
|
||||
|
||||
agent._invalidate_system_prompt()
|
||||
new_system_prompt = agent._build_system_prompt(system_message)
|
||||
agent._cached_system_prompt = new_system_prompt
|
||||
|
||||
if agent._session_db:
|
||||
try:
|
||||
# Trigger memory extraction on the current session before the
|
||||
# transcript is rewritten (runs in BOTH modes — the logical
|
||||
# conversation's pre-compaction turns are about to be summarized
|
||||
# away regardless of whether the id rotates).
|
||||
agent.commit_memory_session(messages)
|
||||
|
||||
if in_place:
|
||||
# ── In-place compaction: keep the same session_id ──────────
|
||||
# No end_session, no new row, no parent_session_id, no title
|
||||
# renumber, no contextvar/env/logging re-sync. The session's
|
||||
# id, title, cwd, /goal, and gateway routing all stay put.
|
||||
#
|
||||
# Durable, NON-DESTRUCTIVE replace: soft-archive the
|
||||
# pre-compaction turns (active=0, kept on disk + FTS-searchable +
|
||||
# recoverable) and insert `compressed` as the new live (active=1)
|
||||
# set, atomically. `compressed` already carries the surviving
|
||||
# tail (current-turn messages the compressor kept via
|
||||
# protect_last_n), so we DON'T pre-flush here — a flush would
|
||||
# INSERT current-turn rows that archive_and_compact would then
|
||||
# archive alongside the rest (harmless but wasted writes). The
|
||||
# live-context load filters active=1, so a resume reloads ONLY
|
||||
# the compacted set; the original turns remain under the SAME id
|
||||
# for search/recovery (Teknium review — keep one durable id
|
||||
# WITHOUT destroying history, unlike a hard replace_messages).
|
||||
# See #38763.
|
||||
agent._session_db.archive_and_compact(agent.session_id, compressed)
|
||||
# Reset the flush identity set so the next turn's appends are
|
||||
# diffed against the COMPACTED transcript: the compacted dicts
|
||||
# are passed as conversation_history next turn and skipped by
|
||||
# identity, so only genuinely new turn messages get appended
|
||||
# (no dup of the summary, no resurrection of dropped turns).
|
||||
agent._flushed_db_message_ids = set()
|
||||
# Rotation-independent signal: the conversation was compacted in
|
||||
# place (id unchanged). The gateway reads this (NOT an id-change
|
||||
# diff) to re-baseline transcript handling.
|
||||
compacted_in_place = True
|
||||
else:
|
||||
# ── Rotation (legacy): end this session, fork a continuation ─
|
||||
# Flush any un-persisted current-turn messages to the OLD
|
||||
# session before ending it, so they survive in the preserved
|
||||
# parent transcript (#47202). (In-place skips this — see above.)
|
||||
try:
|
||||
agent._flush_messages_to_session_db(messages)
|
||||
except Exception:
|
||||
pass # best-effort — don't block compression on a flush error
|
||||
# Propagate title to the new session with auto-numbering
|
||||
old_title = agent._session_db.get_session_title(agent.session_id)
|
||||
agent._session_db.end_session(agent.session_id, "compression")
|
||||
old_session_id = agent.session_id
|
||||
agent.session_id = f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:6]}"
|
||||
# Ordering contract: the agent thread updates the contextvar here;
|
||||
# the gateway propagates to SessionEntry after run_in_executor returns.
|
||||
try:
|
||||
from gateway.session_context import set_current_session_id
|
||||
|
||||
set_current_session_id(agent.session_id)
|
||||
except Exception:
|
||||
os.environ["HERMES_SESSION_ID"] = agent.session_id
|
||||
# The gateway/tools session context (ContextVar + env) and the
|
||||
# logging session context are SEPARATE mechanisms. The call above
|
||||
# moves the former; the ``[session_id]`` tag on log lines comes
|
||||
# from ``hermes_logging._session_context`` (set once per turn in
|
||||
# conversation_loop.py). Without this, post-rotation log lines in
|
||||
# the same turn keep the STALE old id while the message/DB/gateway
|
||||
# state carry the new one — breaking log correlation exactly at the
|
||||
# compaction boundary (see #34089). Guarded separately so a logging
|
||||
# failure can never regress the routing update above.
|
||||
try:
|
||||
from hermes_logging import set_session_context
|
||||
|
||||
set_session_context(agent.session_id)
|
||||
except Exception:
|
||||
pass
|
||||
agent._session_db_created = False
|
||||
try:
|
||||
agent._session_db.create_session(
|
||||
session_id=agent.session_id,
|
||||
source=agent.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"),
|
||||
model=agent.model,
|
||||
model_config=agent._session_init_model_config,
|
||||
parent_session_id=old_session_id,
|
||||
else:
|
||||
# No hard failure — but did the configured aux model error out
|
||||
# and get recovered by retrying on main? Surface that so users
|
||||
# know their auxiliary.compression.model setting is broken even
|
||||
# though compression succeeded.
|
||||
_aux_fail_model = getattr(agent.context_compressor, "_last_aux_model_failure_model", None)
|
||||
_aux_fail_err = getattr(agent.context_compressor, "_last_aux_model_failure_error", None)
|
||||
if _aux_fail_model:
|
||||
# Dedup on (model, error) so we don't spam on every compaction
|
||||
_aux_key = (_aux_fail_model, _aux_fail_err)
|
||||
if getattr(agent, "_last_aux_fallback_warning_key", None) != _aux_key:
|
||||
agent._last_aux_fallback_warning_key = _aux_key
|
||||
agent._emit_warning(
|
||||
f"ℹ Configured compression model '{_aux_fail_model}' failed "
|
||||
f"({_aux_fail_err or 'unknown error'}). Recovered using main model — "
|
||||
"check auxiliary.compression.model in config.yaml."
|
||||
)
|
||||
except Exception as _cs_err:
|
||||
# The child row could not be created (e.g. FK constraint,
|
||||
# contended write). Previously the outer handler simply
|
||||
# warned and let the agent continue on the NEW id — which
|
||||
# has no row in state.db, producing an orphan: the parent
|
||||
# is ended, the child is never indexed, and every
|
||||
# subsequent message is attributed to a session that
|
||||
# doesn't exist (#33906/#33907). Roll the live id back to
|
||||
# the parent so the conversation stays attached to a real,
|
||||
# indexed session instead of a phantom.
|
||||
logger.warning(
|
||||
"Compression child session create failed (%s) — "
|
||||
"rolling back to parent session %s to avoid an orphan.",
|
||||
_cs_err, old_session_id,
|
||||
)
|
||||
agent.session_id = old_session_id
|
||||
|
||||
todo_snapshot = agent._todo_store.format_for_injection()
|
||||
if todo_snapshot:
|
||||
compressed.append({"role": "user", "content": todo_snapshot})
|
||||
|
||||
agent._invalidate_system_prompt()
|
||||
new_system_prompt = agent._build_system_prompt(system_message)
|
||||
agent._cached_system_prompt = new_system_prompt
|
||||
|
||||
if agent._session_db:
|
||||
try:
|
||||
# Trigger memory extraction on the current session before the
|
||||
# transcript is rewritten (runs in BOTH modes — the logical
|
||||
# conversation's pre-compaction turns are about to be summarized
|
||||
# away regardless of whether the id rotates).
|
||||
agent.commit_memory_session(messages)
|
||||
|
||||
if in_place:
|
||||
# ── In-place compaction: keep the same session_id ──────────
|
||||
# No end_session, no new row, no parent_session_id, no title
|
||||
# renumber, no contextvar/env/logging re-sync. The session's
|
||||
# id, title, cwd, /goal, and gateway routing all stay put.
|
||||
#
|
||||
# Durable, NON-DESTRUCTIVE replace: soft-archive the
|
||||
# pre-compaction turns (active=0, kept on disk + FTS-searchable +
|
||||
# recoverable) and insert `compressed` as the new live (active=1)
|
||||
# set, atomically. `compressed` already carries the surviving
|
||||
# tail (current-turn messages the compressor kept via
|
||||
# protect_last_n), so we DON'T pre-flush here — a flush would
|
||||
# INSERT current-turn rows that archive_and_compact would then
|
||||
# archive alongside the rest (harmless but wasted writes). The
|
||||
# live-context load filters active=1, so a resume reloads ONLY
|
||||
# the compacted set; the original turns remain under the SAME id
|
||||
# for search/recovery (Teknium review — keep one durable id
|
||||
# WITHOUT destroying history, unlike a hard replace_messages).
|
||||
# See #38763.
|
||||
agent._session_db.archive_and_compact(agent.session_id, compressed)
|
||||
# Reset the flush identity set so the next turn's appends are
|
||||
# diffed against the COMPACTED transcript: the compacted dicts
|
||||
# are passed as conversation_history next turn and skipped by
|
||||
# identity, so only genuinely new turn messages get appended
|
||||
# (no dup of the summary, no resurrection of dropped turns).
|
||||
agent._flushed_db_message_ids = set()
|
||||
# Rotation-independent signal: the conversation was compacted in
|
||||
# place (id unchanged). The gateway reads this (NOT an id-change
|
||||
# diff) to re-baseline transcript handling.
|
||||
compacted_in_place = True
|
||||
else:
|
||||
# ── Rotation (legacy): end this session, fork a continuation ─
|
||||
# Flush any un-persisted current-turn messages to the OLD
|
||||
# session before ending it, so they survive in the preserved
|
||||
# parent transcript (#47202). (In-place skips this — see above.)
|
||||
try:
|
||||
agent._flush_messages_to_session_db(messages)
|
||||
except Exception:
|
||||
pass # best-effort — don't block compression on a flush error
|
||||
# Propagate title to the new session with auto-numbering
|
||||
old_title = agent._session_db.get_session_title(agent.session_id)
|
||||
agent._session_db.end_session(agent.session_id, "compression")
|
||||
old_session_id = agent.session_id
|
||||
agent.session_id = f"{datetime.now().strftime('%Y%m%d_%H%M%S')}_{uuid.uuid4().hex[:6]}"
|
||||
# Ordering contract: the agent thread updates the contextvar here;
|
||||
# the gateway propagates to SessionEntry after run_in_executor returns.
|
||||
try:
|
||||
from gateway.session_context import set_current_session_id
|
||||
|
||||
set_current_session_id(agent.session_id)
|
||||
except Exception:
|
||||
os.environ["HERMES_SESSION_ID"] = agent.session_id
|
||||
# The gateway/tools session context (ContextVar + env) and the
|
||||
# logging session context are SEPARATE mechanisms. The call above
|
||||
# moves the former; the ``[session_id]`` tag on log lines comes
|
||||
# from ``hermes_logging._session_context`` (set once per turn in
|
||||
# conversation_loop.py). Without this, post-rotation log lines in
|
||||
# the same turn keep the STALE old id while the message/DB/gateway
|
||||
# state carry the new one — breaking log correlation exactly at the
|
||||
# compaction boundary (see #34089). Guarded separately so a logging
|
||||
# failure can never regress the routing update above.
|
||||
try:
|
||||
from hermes_logging import set_session_context
|
||||
|
||||
set_session_context(agent.session_id)
|
||||
except Exception:
|
||||
pass
|
||||
# Re-open the parent: it was ended above, but we're
|
||||
# continuing on it, so it must not stay closed.
|
||||
agent._session_db_created = False
|
||||
try:
|
||||
agent._session_db.reopen_session(old_session_id)
|
||||
except Exception:
|
||||
pass
|
||||
old_session_id = None # no rotation happened
|
||||
# The parent row already exists in state.db, so mark the
|
||||
# session as created — _ensure_db_session would otherwise
|
||||
# retry a (harmless INSERT OR IGNORE) create next turn.
|
||||
agent._session_db.create_session(
|
||||
session_id=agent.session_id,
|
||||
source=agent.platform or os.environ.get("HERMES_SESSION_SOURCE", "cli"),
|
||||
model=agent.model,
|
||||
model_config=agent._session_init_model_config,
|
||||
parent_session_id=old_session_id,
|
||||
)
|
||||
except Exception as _cs_err:
|
||||
# The child row could not be created (e.g. FK constraint,
|
||||
# contended write). Previously the outer handler simply
|
||||
# warned and let the agent continue on the NEW id — which
|
||||
# has no row in state.db, producing an orphan: the parent
|
||||
# is ended, the child is never indexed, and every
|
||||
# subsequent message is attributed to a session that
|
||||
# doesn't exist (#33906/#33907). Roll the live id back to
|
||||
# the parent so the conversation stays attached to a real,
|
||||
# indexed session instead of a phantom.
|
||||
logger.warning(
|
||||
"Compression child session create failed (%s) — "
|
||||
"rolling back to parent session %s to avoid an orphan.",
|
||||
_cs_err, old_session_id,
|
||||
)
|
||||
agent.session_id = old_session_id
|
||||
try:
|
||||
from gateway.session_context import set_current_session_id
|
||||
set_current_session_id(agent.session_id)
|
||||
except Exception:
|
||||
os.environ["HERMES_SESSION_ID"] = agent.session_id
|
||||
try:
|
||||
from hermes_logging import set_session_context
|
||||
set_session_context(agent.session_id)
|
||||
except Exception:
|
||||
pass
|
||||
# Re-open the parent: it was ended above, but we're
|
||||
# continuing on it, so it must not stay closed.
|
||||
try:
|
||||
agent._session_db.reopen_session(old_session_id)
|
||||
except Exception:
|
||||
pass
|
||||
old_session_id = None # no rotation happened
|
||||
# The parent row already exists in state.db, so mark the
|
||||
# session as created — _ensure_db_session would otherwise
|
||||
# retry a (harmless INSERT OR IGNORE) create next turn.
|
||||
agent._session_db_created = True
|
||||
raise
|
||||
agent._session_db_created = True
|
||||
raise
|
||||
agent._session_db_created = True
|
||||
# Carry a persistent /goal onto the continuation session.
|
||||
# Compression mints a fresh child id; load_goal does a flat
|
||||
# per-session lookup with no parent walk, so without this an
|
||||
# active goal silently dies at the boundary (#33618).
|
||||
try:
|
||||
from hermes_cli.goals import migrate_goal_to_session
|
||||
migrate_goal_to_session(old_session_id, agent.session_id, reason="compression")
|
||||
except Exception as _goal_err:
|
||||
logger.debug("Could not migrate goal on compression: %s", _goal_err)
|
||||
# Auto-number the title for the continuation session
|
||||
if old_title:
|
||||
# Carry a persistent /goal onto the continuation session.
|
||||
# Compression mints a fresh child id; load_goal does a flat
|
||||
# per-session lookup with no parent walk, so without this an
|
||||
# active goal silently dies at the boundary (#33618).
|
||||
try:
|
||||
new_title = agent._session_db.get_next_title_in_lineage(old_title)
|
||||
agent._session_db.set_session_title(agent.session_id, new_title)
|
||||
except (ValueError, Exception) as e:
|
||||
logger.debug("Could not propagate title on compression: %s", e)
|
||||
from hermes_cli.goals import migrate_goal_to_session
|
||||
migrate_goal_to_session(old_session_id, agent.session_id, reason="compression")
|
||||
except Exception as _goal_err:
|
||||
logger.debug("Could not migrate goal on compression: %s", _goal_err)
|
||||
# Auto-number the title for the continuation session
|
||||
if old_title:
|
||||
try:
|
||||
new_title = agent._session_db.get_next_title_in_lineage(old_title)
|
||||
agent._session_db.set_session_title(agent.session_id, new_title)
|
||||
except (ValueError, Exception) as e:
|
||||
logger.debug("Could not propagate title on compression: %s", e)
|
||||
|
||||
# Shared post-write steps (both modes target agent.session_id, which
|
||||
# in-place keeps and rotation has already reassigned to the new id):
|
||||
# refresh the stored system prompt and reset the flush cursor so the
|
||||
# next turn re-bases its append diff.
|
||||
agent._session_db.update_system_prompt(agent.session_id, new_system_prompt)
|
||||
agent._last_flushed_db_idx = 0
|
||||
except Exception as e:
|
||||
# If the rotation rolled back to the parent (orphan-avoidance
|
||||
# above), agent.session_id is the still-indexed parent and
|
||||
# old_session_id was cleared — so this is recovery, not an
|
||||
# un-indexed orphan. Otherwise an earlier step failed before the
|
||||
# child was created and the warning's original meaning holds.
|
||||
if locals().get("old_session_id") is None and not in_place:
|
||||
logger.warning(
|
||||
"Compression rotation aborted and rolled back to the "
|
||||
"parent session (%s): %s", agent.session_id or "?", e,
|
||||
)
|
||||
else:
|
||||
logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e)
|
||||
# Shared post-write steps (both modes target agent.session_id, which
|
||||
# in-place keeps and rotation has already reassigned to the new id):
|
||||
# refresh the stored system prompt and reset the flush cursor so the
|
||||
# next turn re-bases its append diff.
|
||||
agent._session_db.update_system_prompt(agent.session_id, new_system_prompt)
|
||||
agent._last_flushed_db_idx = 0
|
||||
except Exception as e:
|
||||
# If the rotation rolled back to the parent (orphan-avoidance
|
||||
# above), agent.session_id is the still-indexed parent and
|
||||
# old_session_id was cleared — so this is recovery, not an
|
||||
# un-indexed orphan. Otherwise an earlier step failed before the
|
||||
# child was created and the warning's original meaning holds.
|
||||
if locals().get("old_session_id") is None and not in_place:
|
||||
logger.warning(
|
||||
"Compression rotation aborted and rolled back to the "
|
||||
"parent session (%s): %s", agent.session_id or "?", e,
|
||||
)
|
||||
else:
|
||||
logger.warning("Session DB compression split failed — new session will NOT be indexed: %s", e)
|
||||
|
||||
# Compaction-boundary bookkeeping, computed once. `old_session_id` is only
|
||||
# bound in the rotation branch; in-place leaves it unset. `_boundary_parent`
|
||||
# is the id the boundary notifications attribute the prior state to: the old
|
||||
# id on rotation, the (unchanged) current id in-place.
|
||||
_old_sid = locals().get("old_session_id")
|
||||
_is_boundary = bool(_old_sid) or in_place
|
||||
_boundary_parent = _old_sid or agent.session_id or ""
|
||||
# Compaction-boundary bookkeeping, computed once. `old_session_id` is only
|
||||
# bound in the rotation branch; in-place leaves it unset. `_boundary_parent`
|
||||
# is the id the boundary notifications attribute the prior state to: the old
|
||||
# id on rotation, the (unchanged) current id in-place.
|
||||
_old_sid = locals().get("old_session_id")
|
||||
_is_boundary = bool(_old_sid) or in_place
|
||||
_boundary_parent = _old_sid or agent.session_id or ""
|
||||
|
||||
# Notify the context engine that a compaction boundary occurred. Plugin
|
||||
# engines (e.g. hermes-lcm) use boundary_reason="compression" to preserve
|
||||
# DAG lineage / checkpoint per-session state across the boundary instead of
|
||||
# re-initializing fresh. See hermes-lcm#68. Built-in ContextCompressor
|
||||
# ignores kwargs. Fires in BOTH modes: rotation passes old→new ids; in-place
|
||||
# passes the SAME id (the boundary is real even though the id didn't move).
|
||||
try:
|
||||
if _is_boundary and hasattr(agent.context_compressor, "on_session_start"):
|
||||
agent.context_compressor.on_session_start(
|
||||
agent.session_id or "",
|
||||
boundary_reason="compression",
|
||||
old_session_id=_boundary_parent,
|
||||
platform=getattr(agent, "platform", None) or "cli",
|
||||
conversation_id=getattr(agent, "_gateway_session_key", None),
|
||||
)
|
||||
except Exception as _ce_err:
|
||||
logger.debug("context engine on_session_start (compression): %s", _ce_err)
|
||||
|
||||
# Notify memory providers of the compaction boundary so provider-cached
|
||||
# per-session state (Hindsight's _document_id, accumulated turn buffers,
|
||||
# counters) refreshes. reset=False because the logical conversation
|
||||
# continues. See #6672. Fires in BOTH modes: in-place uses the same id as
|
||||
# parent (the conversation didn't fork, but the buffer must still be told
|
||||
# the transcript was compacted so it doesn't double-count dropped turns).
|
||||
try:
|
||||
if _is_boundary and agent._memory_manager:
|
||||
agent._memory_manager.on_session_switch(
|
||||
agent.session_id or "",
|
||||
parent_session_id=_boundary_parent,
|
||||
reset=False,
|
||||
reason="compression",
|
||||
)
|
||||
except Exception as _me_err:
|
||||
logger.debug("memory manager on_session_switch (compression): %s", _me_err)
|
||||
|
||||
# Warn on repeated compressions (quality degrades with each pass).
|
||||
# Route through _emit_status (like the other compression warnings above)
|
||||
# so the warning reaches the TUI / Telegram / Discord via status_callback,
|
||||
# not just CLI stdout. _emit_status still _vprints for the CLI, and
|
||||
# storing it on _compression_warning lets replay_compression_warning
|
||||
# re-deliver it once a late-bound gateway status_callback is wired (#36908).
|
||||
_cc = agent.context_compressor.compression_count
|
||||
if _cc >= 2:
|
||||
_cc_msg = (
|
||||
f"{agent.log_prefix}⚠️ Session compressed {_cc} times — "
|
||||
f"accuracy may degrade. Consider /new to start fresh."
|
||||
)
|
||||
agent._compression_warning = _cc_msg
|
||||
agent._emit_status(_cc_msg)
|
||||
|
||||
# Emit session:compress event so hooks (e.g. MemPalace sync) can ingest
|
||||
# the completed old session before its details are lost. In in-place mode
|
||||
# there is no old id (same session); ``in_place=True`` tells hooks the
|
||||
# transcript was compacted on the same id rather than rotated.
|
||||
if getattr(agent, "event_callback", None):
|
||||
# Notify the context engine that a compaction boundary occurred. Plugin
|
||||
# engines (e.g. hermes-lcm) use boundary_reason="compression" to preserve
|
||||
# DAG lineage / checkpoint per-session state across the boundary instead of
|
||||
# re-initializing fresh. See hermes-lcm#68. Built-in ContextCompressor
|
||||
# ignores kwargs. Fires in BOTH modes: rotation passes old→new ids; in-place
|
||||
# passes the SAME id (the boundary is real even though the id didn't move).
|
||||
try:
|
||||
agent.event_callback("session:compress", {
|
||||
"platform": agent.platform or "",
|
||||
"session_id": agent.session_id,
|
||||
"old_session_id": _old_sid or "",
|
||||
"in_place": in_place,
|
||||
"compression_count": agent.context_compressor.compression_count,
|
||||
})
|
||||
except Exception as e:
|
||||
logger.debug("event_callback error on session:compress: %s", e)
|
||||
if _is_boundary and hasattr(agent.context_compressor, "on_session_start"):
|
||||
agent.context_compressor.on_session_start(
|
||||
agent.session_id or "",
|
||||
boundary_reason="compression",
|
||||
old_session_id=_boundary_parent,
|
||||
platform=getattr(agent, "platform", None) or "cli",
|
||||
conversation_id=getattr(agent, "_gateway_session_key", None),
|
||||
)
|
||||
except Exception as _ce_err:
|
||||
logger.debug("context engine on_session_start (compression): %s", _ce_err)
|
||||
|
||||
# Surface the compaction mode to the caller (run_conversation / gateway)
|
||||
# via a rotation-independent flag. The gateway uses this — NOT an
|
||||
# id-change diff — to re-baseline transcript handling (history_offset=0 +
|
||||
# rewrite on the same id) when compaction happened in place. See #38763.
|
||||
agent._last_compaction_in_place = compacted_in_place
|
||||
# Notify memory providers of the compaction boundary so provider-cached
|
||||
# per-session state (Hindsight's _document_id, accumulated turn buffers,
|
||||
# counters) refreshes. reset=False because the logical conversation
|
||||
# continues. See #6672. Fires in BOTH modes: in-place uses the same id as
|
||||
# parent (the conversation didn't fork, but the buffer must still be told
|
||||
# the transcript was compacted so it doesn't double-count dropped turns).
|
||||
try:
|
||||
if _is_boundary and agent._memory_manager:
|
||||
agent._memory_manager.on_session_switch(
|
||||
agent.session_id or "",
|
||||
parent_session_id=_boundary_parent,
|
||||
reset=False,
|
||||
reason="compression",
|
||||
)
|
||||
except Exception as _me_err:
|
||||
logger.debug("memory manager on_session_switch (compression): %s", _me_err)
|
||||
|
||||
# Keep the post-compression rough estimate for diagnostics, but do not
|
||||
# treat it as provider-reported prompt usage. Schema-heavy rough estimates
|
||||
# can remain above threshold even after the next real API request fits.
|
||||
_compressed_est = estimate_request_tokens_rough(
|
||||
compressed,
|
||||
system_prompt=new_system_prompt or "",
|
||||
tools=agent.tools or None,
|
||||
)
|
||||
agent.context_compressor.last_compression_rough_tokens = _compressed_est
|
||||
agent.context_compressor.last_prompt_tokens = -1
|
||||
agent.context_compressor.last_completion_tokens = 0
|
||||
agent.context_compressor.awaiting_real_usage_after_compression = True
|
||||
# Warn on repeated compressions (quality degrades with each pass).
|
||||
# Route through _emit_status (like the other compression warnings above)
|
||||
# so the warning reaches the TUI / Telegram / Discord via status_callback,
|
||||
# not just CLI stdout. _emit_status still _vprints for the CLI, and
|
||||
# storing it on _compression_warning lets replay_compression_warning
|
||||
# re-deliver it once a late-bound gateway status_callback is wired (#36908).
|
||||
_cc = agent.context_compressor.compression_count
|
||||
if _cc >= 2:
|
||||
_cc_msg = (
|
||||
f"{agent.log_prefix}⚠️ Session compressed {_cc} times — "
|
||||
f"accuracy may degrade. Consider /new to start fresh."
|
||||
)
|
||||
agent._compression_warning = _cc_msg
|
||||
agent._emit_status(_cc_msg)
|
||||
|
||||
# Clear the file-read dedup cache. After compression the original
|
||||
# read content is summarised away — if the model re-reads the same
|
||||
# file it needs the full content, not a "file unchanged" stub.
|
||||
try:
|
||||
from tools.file_tools import reset_file_dedup
|
||||
reset_file_dedup(task_id)
|
||||
except Exception:
|
||||
pass
|
||||
# Emit session:compress event so hooks (e.g. MemPalace sync) can ingest
|
||||
# the completed old session before its details are lost. In in-place mode
|
||||
# there is no old id (same session); ``in_place=True`` tells hooks the
|
||||
# transcript was compacted on the same id rather than rotated.
|
||||
if getattr(agent, "event_callback", None):
|
||||
try:
|
||||
agent.event_callback("session:compress", {
|
||||
"platform": agent.platform or "",
|
||||
"session_id": agent.session_id,
|
||||
"old_session_id": _old_sid or "",
|
||||
"in_place": in_place,
|
||||
"compression_count": agent.context_compressor.compression_count,
|
||||
})
|
||||
except Exception as e:
|
||||
logger.debug("event_callback error on session:compress: %s", e)
|
||||
|
||||
logger.info(
|
||||
"context compression done: session=%s messages=%d->%d rough_tokens=~%s awaiting_real_usage=true",
|
||||
agent.session_id or "none", _pre_msg_count, len(compressed),
|
||||
f"{_compressed_est:,}",
|
||||
)
|
||||
# Release the lock on the OLD session_id only AFTER rotation completed
|
||||
# and all post-rotation bookkeeping (memory manager, context engine,
|
||||
# file dedup) ran. A concurrent path that wakes up the moment we
|
||||
# release will see the NEW session_id in state.db / SessionEntry and
|
||||
# acquire on that — no race against our just-finished work.
|
||||
_release_lock()
|
||||
return compressed, new_system_prompt
|
||||
# Surface the compaction mode to the caller (run_conversation / gateway)
|
||||
# via a rotation-independent flag. The gateway uses this — NOT an
|
||||
# id-change diff — to re-baseline transcript handling (history_offset=0 +
|
||||
# rewrite on the same id) when compaction happened in place. See #38763.
|
||||
agent._last_compaction_in_place = compacted_in_place
|
||||
|
||||
# Keep the post-compression rough estimate for diagnostics, but do not
|
||||
# treat it as provider-reported prompt usage. Schema-heavy rough estimates
|
||||
# can remain above threshold even after the next real API request fits.
|
||||
_compressed_est = estimate_request_tokens_rough(
|
||||
compressed,
|
||||
system_prompt=new_system_prompt or "",
|
||||
tools=agent.tools or None,
|
||||
)
|
||||
agent.context_compressor.last_compression_rough_tokens = _compressed_est
|
||||
agent.context_compressor.last_prompt_tokens = -1
|
||||
agent.context_compressor.last_completion_tokens = 0
|
||||
agent.context_compressor.awaiting_real_usage_after_compression = True
|
||||
|
||||
# Clear the file-read dedup cache. After compression the original
|
||||
# read content is summarised away — if the model re-reads the same
|
||||
# file it needs the full content, not a "file unchanged" stub.
|
||||
try:
|
||||
from tools.file_tools import reset_file_dedup
|
||||
reset_file_dedup(task_id)
|
||||
except Exception:
|
||||
pass
|
||||
|
||||
logger.info(
|
||||
"context compression done: session=%s messages=%d->%d rough_tokens=~%s awaiting_real_usage=true",
|
||||
agent.session_id or "none", _pre_msg_count, len(compressed),
|
||||
f"{_compressed_est:,}",
|
||||
)
|
||||
return compressed, new_system_prompt
|
||||
finally:
|
||||
# Release the lock on the OLD session_id only AFTER rotation completed
|
||||
# and all post-rotation bookkeeping (memory manager, context engine,
|
||||
# file dedup) ran. A concurrent path that wakes up the moment we
|
||||
# release will see the NEW session_id in state.db / SessionEntry and
|
||||
# acquire on that — no race against our just-finished work.
|
||||
_release_lock()
|
||||
|
||||
|
||||
def try_shrink_image_parts_in_messages(
|
||||
|
|
|
|||
|
|
@ -225,6 +225,34 @@ def test_lock_refresh_keeps_owner_live_past_initial_ttl(tmp_path: Path, monkeypa
|
|||
assert db.get_compression_lock_holder(parent_sid) is None
|
||||
|
||||
|
||||
def test_post_compress_exception_stops_lock_refresher(tmp_path: Path, monkeypatch) -> None:
|
||||
"""A warning-path exception after compress() returns must still release the lock."""
|
||||
real_try_acquire = SessionDB.try_acquire_compression_lock
|
||||
|
||||
def _short_ttl(self, session_id: str, holder: str, ttl_seconds: float = 300.0) -> bool:
|
||||
return real_try_acquire(self, session_id, holder, ttl_seconds=1.0)
|
||||
|
||||
monkeypatch.setattr(SessionDB, "try_acquire_compression_lock", _short_ttl)
|
||||
|
||||
db = SessionDB(db_path=tmp_path / "state.db")
|
||||
parent_sid = "REFRESH_EXCEPTION_TEST"
|
||||
db.create_session(parent_sid, source="discord")
|
||||
|
||||
agent = _build_agent_with_db(db, parent_sid)
|
||||
agent._compression_lock_ttl_seconds = 1.0
|
||||
agent._compression_lock_refresh_interval = 0.1
|
||||
agent.context_compressor._last_summary_error = "summary failed"
|
||||
agent._emit_warning = lambda *_a, **_k: (_ for _ in ()).throw(RuntimeError("warn boom"))
|
||||
|
||||
messages = [{"role": "user", "content": f"m{i}"} for i in range(20)]
|
||||
|
||||
with pytest.raises(RuntimeError, match="warn boom"):
|
||||
agent._compress_context(messages, "sys", approx_tokens=120_000)
|
||||
|
||||
time.sleep(1.3)
|
||||
assert db.try_acquire_compression_lock(parent_sid, "probe", ttl_seconds=1.0) is True
|
||||
|
||||
|
||||
class _NoLockSubsystemDB:
|
||||
"""Wraps a real SessionDB but simulates a pre-#34351 version skew.
|
||||
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue