…rade, hardening

Combined salvage of PRs #9661, #9663, #9674, #9677, #9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs #9665 (user msg preservation — duplication risk) and #9675 (dead code).

…rade, hardening

Combined salvage of PRs #9661, #9663, #9674, #9677, #9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs #9665 (user msg preservation — duplication risk) and #9675 (dead code).

…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).

…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).

ContextEngine.should_compress_preflight() is documented as the per-turn
ingest entry for plugin engines, but run_agent.py never calls it. PR
NousResearch#10088 explicitly noted this as dead code when skipping NousResearch#9675:

> NousResearch#9675 (preflight check) — dead code, run_agent.py never calls
> should_compress_preflight()

This breaks plugin context engines that rely on the hook for per-turn
message ingest. hermes-lcm overrides should_compress_preflight() to
persist messages each turn into its DAG store, but with the hook never
called, the lossless message store stays empty until compress() fires
at the threshold (typically ~96K tokens). Reproducible:

  $ hermes chat -q "test" -Q
  $ sqlite3 ~/.hermes/lcm.db "SELECT COUNT(*) FROM messages;"
  0

(Verified on hermes-agent v0.11.0 with hermes-lcm v0.7.0.)

Add two calls to should_compress_preflight(messages):

1. Top of the main loop, right after api_call_count is incremented —
   per-turn ingest before each API call.
2. End of run_conversation(), before the on_session_end plugin hook —
   final flush so the last assistant message reaches the engine when
   the turn exited via the no-tool-calls branch and skipped the
   per-turn hook above.

The return value is discarded; compression is still decided by the
later should_compress(_real_tokens) call which uses the provider-
reported token count. Both calls are wrapped in try/except so a
misbehaving plugin engine cannot break the conversation loop.

Default ContextEngine.should_compress_preflight() returns False with no
work, so this is zero overhead for the built-in ContextCompressor and
any engine that does not override the hook.

After this fix:
  $ hermes chat -q "test" -Q
  $ sqlite3 ~/.hermes/lcm.db "SELECT COUNT(*) FROM messages;"
  2

Refs:
- NousResearch#9675 (closed: feat(compressor): implement preflight compression check)
- NousResearch#10088 (merged body: skipped NousResearch#9675 as dead code)
- stephenschoettler/hermes-lcm#68 (LCM author flagged host integration
  issue but could not file upstream because GitHub Issues was off on a
  different fork)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).

…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).

…rade, hardening

Combined salvage of PRs NousResearch#9661, NousResearch#9663, NousResearch#9674, NousResearch#9677, NousResearch#9678 by kshitijk4poor.

- Smart tool output collapse: informative 1-line summaries replace generic placeholder
- Dedup identical tool results via MD5 hash, truncate large tool_call arguments
- Anti-thrashing: skip compression after 2 consecutive <10% savings passes
- Structured action-log summary template with numbered actions and Active State
- Hardening: max_tokens 1.3x cap, multimodal safety, note idempotency, adaptive cooldown

Follow-up fixes applied during salvage:
- web_extract: reads 'urls' (list) not 'url' (original PR bug)
- Multimodal list content guards in dedup and prune passes
- Kept 'Relevant Files' section in template (original PR removed it)

Skipped PRs NousResearch#9665 (user msg preservation — duplication risk) and NousResearch#9675 (dead code).