fix(acp): publish current context and compression depth

[stefanpieter]

What does this PR do?

Make ACP context telemetry represent the current effective model window and durably expose completed compression depth.

ACP usage_update.used now prefers provider-reported last_prompt_tokens and uses the existing request estimator only before real model usage is available. Each completed worker-thread model step schedules a fresh update, including non-negative _meta.hermes.compressionCount.

Compression count round-trips through session storage, survives restore and compression-driven internal rotation, and is cleared durably by an explicit same-session ACP /reset.

Restored sessions replay existing history and then await the current usage/compression update before either session/load or session/resume returns, so a fresh client is current at the response boundary.

This complements #70267: that PR changes PromptResponse.usage.inputTokens; this PR fixes the authoritative ACP usage_update stream, compression metadata, and durable reset/rotation behavior.

Related Issue

Related to #70204 and #70267.

Changes made

• acp_adapter/server.py: publish current-window usage/compression depth after model steps and during load/resume restore; persist cleared count on ACP /reset
• agent/context_compressor.py: load, persist, rotate, and explicitly reset compression depth
• hermes_state_common.py and hermes_state.py: add the authoritative schema column plus bounded session accessors after current main moved SCHEMA_SQL into the common module
• regressions cover current-window usage, estimate fallback, mid-turn scheduling, metadata, load/resume response boundaries, rotation, explicit reset, and durable state
• contributor attribution remains mapped by the repository helper

Review remediation

• replaced deferred resume telemetry scheduling with an awaited update after history replay, matching the existing load response-boundary contract
• added an exact resume regression that failed on the prior head because no UsageUpdate was visible when resume_session() returned
• reconciled current-main schema/test splits without resurrecting obsolete inline schema or removed test blocks

Current exact-head verification

• head: 1e0784da1333b06370567fa77ca59fed86da4843
• scripts/run_tests.sh tests/acp/test_server.py tests/agent/test_context_engine_host_contract.py tests/test_hermes_state.py -q — 180 passed
• Ruff, compileall, git diff --check, and added-line security scan — passed
• independent exact-SHA review — PASS
• branch contains reviewed base f3cda0ceb18d8ba7465a6d223098ef0e56c8fee1; live main later advanced only in tools/file_operations.py and its test, with no dependency overlap and a clean git merge-tree proof
• upstream CI is awaiting external-fork workflow approval: https://github.com/NousResearch/hermes-agent/actions/runs/30619238988

Type of change

• Bug fix (non-breaking)
• Tests

Checklist

• Contributing guide reviewed
• Conventional commits
• Focused regressions added
• Tested on macOS
• Full repository suite not claimed; affected hermetic suites are reported above

[stefanpieter]

CI is awaiting the repository-admin approval required for workflows from external forks (run 30442121432). The submitter approval endpoint returns HTTP 403 (Must have admin rights to Repository).

Exact head eb05e60e3e31984efa5f1627ded239278bbef667 passed locally:

• affected ACP/state suites: 589 passed
• full tests/acp/: 332 passed
• Ruff, compileall, and the 853-file Windows-footgun scan
• independent exact-tree review: passed with no security or logic findings

[stefanpieter]

Post-submit live validation found and fixed a restore-only delivery gap: session/load had deferred usage_update, leaving a fresh client stale until its first prompt. The updated head awaits the restored update before returning.

Fresh evidence on the updated exact tree:

• 590 affected/state tests passed
• 333 full ACP tests passed
• Ruff, compileall, and 853-file Windows footgun scan passed
• real ACP wire probe emitted used=182079, size=272000, compressionCount=2 at message 305 before the load response at message 307
• independent final review: passed, no findings

[stefanpieter]

@kshitijk4poor this is ready for maintainer action on exact head f0015386044fbc3191858aea4701faba4f0a3eef.

• independent exact-head review: PASS, no blockers
• unmodified fork CI: All required checks pass (8 Python slices, E2E, Ruff/ty, Windows, attribution, OSV, supply chain): https://github.com/stefanpieter/hermes-agent-upstream-fork/actions/runs/30450355455
• upstream workflow awaiting admin approval: https://github.com/NousResearch/hermes-agent/actions/runs/30450097470
• latest main added one unrelated cron-test file after the reviewed base; no path overlap and the merge tree is clean

Could you please approve the external-fork workflow and review/merge if satisfied? The PR body has the exact base/head, patch hash, local evidence, and live ACP wire proof.

[teknium1]

Thanks for tracing the ACP usage stream through provider usage, restore delivery, and durable compression state. The current-main premise is real: acp_adapter/server.py:857-867 still prefers the rough estimate and load_session defers its usage update.

Problems

• acp_adapter/server.py:1431 makes session/load publish synchronously, but resume_session has the same in-call replay contract at acp_adapter/server.py:1453-1455 and still defers its usage update at acp_adapter/server.py:1466. A resumed client can therefore retain the stale-telemetry behavior this change fixes for loads.

Suggested changes

• Apply the awaited usage publication to resume_session and add a response-boundary regression alongside the new load test.
• Current main moved SCHEMA_SQL to hermes_state_common.py:135 in 21c7ae8563; preserve the declarative column addition there during salvage.

Automated hermes-sweeper review.

[teknium1]

resume_session has the same before-response replay contract at lines 1453-1455, but it still defers _schedule_usage_update(state) at line 1466. Please await _send_usage_update(state) there too and add a response-boundary regression; otherwise resumed clients can still miss the restored telemetry this change guarantees for session/load.

[stefanpieter]

Resolved the current review blocker on 1e0784da1333b06370567fa77ca59fed86da4843.

• session/resume now replays history and awaits the restored usage/compression update before returning, matching session/load
• added the exact response-boundary regression that failed on the prior head
• forward-ported durable compression-count state to current main's authoritative hermes_state_common.py schema without resurrecting removed schema/test blocks

Verification:

• scripts/run_tests.sh tests/acp/test_server.py tests/agent/test_context_engine_host_contract.py tests/test_hermes_state.py -q — 180 passed
• Ruff, compileall, diff check, added-line security scan — passed
• independent exact-SHA review — PASS
• remote branch and PR head both match the SHA above

The exact-head upstream CI run is awaiting maintainer approval for external-fork workflows: https://github.com/NousResearch/hermes-agent/actions/runs/30619238988

@digironin @teknium1 please re-review when available. Formal RequestReviewsByLogin is unavailable to this fork author; the historical gate label and unresolved review thread remain reviewer-controlled.