This repository was archived by the owner on May 18, 2026. It is now read-only.
chore: prepare hermes-agent fork for v0.6.0 overlay handoff - #1
Merged
Merged
Conversation
Phase 3a of Fox in the Box's v0.6.0 upstream-separation migration (fox-in-the-box-ai/fox-in-the-box#155, sub-issue NousResearch#170). Fork-side PR. Companion monorepo PR follows once this merges. Three coordinated changes that together hand off all hermes-agent Fox customizations to the sibling fox-overlay package: 1. Revert 5 modified files to upstream content (merge-base e5dad4a). The Fox edits relocate to fox_overlay/agent_plugins/fox_overlay_plugin/ monkey_patches/ in the monorepo PR, where each becomes a runtime monkey-patch with inspect.signature() self-check. - agent/auxiliary_client.py (-20 lines: provider=auto fallback) - cron/jobs.py (-11 lines: failure_history field) - cron/scheduler.py (-58 lines: rolling-5 failure history) - hermes_cli/runtime_provider.py (-1 line: target_model Bedrock fix) - tools/cronjob_tools.py (-8 lines: diagnostics surfacing) 2. Remove plugins/memory/mem0_oss/ (-1218 lines, 3 files). Relocated to packages/fox-overlay/fox_overlay/agent_plugins/mem0_oss/ in the monorepo PR; loaded via upstream's hermes_agent.plugins entry-point group with the missing `kind: backend` manifest field added. 3. Pause .github/workflows/notify-monorepo.yml repository_dispatch. Symmetric with the webui fork's already-paused workflow (also from migration prep) and with the monorepo's already-paused sync-submodules.yml. Without this, the merge push would dispatch a `hermes-agent-updated` event the monorepo silently drops — fighting itself. workflow_dispatch remains for manual invocation. Net effect on this fork once both PRs land: a virgin upstream tree (modulo this still-active workflow file) ready to be re-pointed at NousResearch/hermes-agent directly in Phase 8. Out of scope: - Any monkey-patch implementation — that's the monorepo PR - Restoration of upstream tests (none touched in this fork) - Other workflows (release.yml, lint.yml etc.) — keep as-is Per `feedback_cross_repo_cohesion.md`: this PR merges first; monorepo PR opens with the new submodule SHA.
…ributions Reviewer #2 (DevOps) on PR #1 flagged that `contributor-check.yml` fails because `roadhero@gmail.com` (Fox in the Box maintainer) is unmapped in scripts/release.py AUTHOR_MAP. The check walks $MERGE_BASE..HEAD authors and rejects unknown emails. This fork (fox-in-the-box-ai/hermes-agent) accepted direct pushes from roadhero pre-PR-#1; the migration cohesion convention requires PRs going forward, so the AUTHOR_MAP needs the entry. Keeps the fork's first-ever PR all-green and avoids setting the "red CI is OK here" cultural anchor on Phase 3.
This was referenced May 17, 2026
roadhero
added a commit
to fox-in-the-box-ai/fox-in-the-box
that referenced
this pull request
May 17, 2026
… (Phase 3b) (#221) * feat(fox-overlay): agent plugin + monkey-patches + relocated mem0_oss Phase 3b (monorepo side) of the v0.6.0 upstream-separation migration (epic #155, sub-issue #171). Companion to merged fork PRs fox-in-the-box-ai/hermes-agent#1 (revert Fox edits + remove mem0_oss + pause notify-monorepo) and #2 (bootstrap shim). ## What this does Populates packages/fox-overlay/fox_overlay/agent_plugins/ with the relocated Fox customizations so the gateway boots with Fox bug fixes restored on top of virgin upstream: 1. Three monkey-patches against upstream hermes-agent functions, applied via `inspect.getsource` + textual substitution at register-time (anchor-based; raises AssertionError at startup if upstream refactors the target — caught by Phase 9 nightly watch). - hermes_cli.runtime_provider.resolve_runtime_provider: target_model used for Bedrock api_mode routing - agent.auxiliary_client._resolve_task_provider_model: provider=auto resolves to explicit provider when cfg_model set - agent.auxiliary_client._get_auxiliary_task_config: auxiliary.default falls back when no task-specific config - cron.jobs.create_job + .mark_job_run: rolling failure_history - cron.scheduler.run_job: enriched FAILED output (diagnostics + traceback + session-log path) - cron.scheduler.tick: structured multi-line failure delivery - tools.cronjob_tools._format_job: surface last_error + failure_history + consecutive_failures when present 2. mem0_oss memory provider relocated wholesale from forks/hermes-agent/plugins/memory/mem0_oss/ (byte-identical restored via git show; SHA256 verified). plugin.yaml gains `kind: backend` for upstream v0.13.0's plugin manifest validation. 3. pyproject.toml entry-points wire fox_overlay + mem0_oss into upstream's hermes_agent.plugins group (discovery mechanism documented in plan §3 Agent 2). 4. Dockerfile inserts `COPY packages/fox-overlay /app/fox-overlay` + `pip install -e /app/fox-overlay` between the upstream pip install (line 141) and the Phase 2 static-overlay ENV block (line 143). Layer-cache friendly: static-asset edits don't invalidate the (slow) upstream install layer. 5. Submodule pin bumped to e301105ec (fork main after PR#1 + PR#2 merged). ## Validation gates - Container rebuilds clean (FITB_VERSION=v0.6.0-p3-shim-qa, 1.45 GB) - Container starts, /health 200, supervisord all RUNNING - Gateway startup log (/data/data/hermes/logs/agent.log) shows all 8 `[fox-overlay] patched ...` lines confirming each substitution - All 8 runtime sentinel attributes True after gateway boot - Zero AssertionError / Traceback / ERROR in any log ## Engineer-discovered gap (handled by fork PR #2) Upstream's PluginManager.discover_and_load() is invoked only from hermes_cli commands and tui_gateway, never from gateway/run.py. The plan assumed entry-points auto-fire at gateway boot; they don't. Resolved via the ~23-line bootstrap shim in fork PR #2 (plan deviation flagged + Dennis-approved). Net agent patch surface goes from plan's optimistic 0 files to 1 file / ~23 lines — still within Phase 8's patch-series model. ## _helpers.py textual-substitution pattern `substitute_function` reads upstream source via inspect.getsource, applies (old, new) substitutions sequentially with EXACTLY-ONCE anchor check, recompiles in a copy of the upstream module's globals, and rebinds the attribute. Anchor drift → AssertionError at register time. Idempotent via per-target sentinel. Notably does NOT call textwrap.dedent — dedent normalizes whitespace-only blank lines to bare \\n, breaking anchors that match verbatim file content. Engineer-discovered during integration smoke (cron.jobs.mark_job_run anchor with 16-space blank line failed until dedent removed); fix documented in _helpers.py docstring. ## Out of scope (deferred per architect brief) - Bootstrap.py changes — Phase 4 owns those (webui-side shim) - Webui-side patches — Phase 4-7 - Submodule re-point to virgin upstream — Phase 8 (extracts the fork-PR-#2 shim to packages/fox-overlay/patches/agent/001-*.patch at that time) Closes #171. * chore(fox-overlay): revert mem0_oss relocation (Phase 3 rollback) Reverts the mem0_oss-relocation portion of this PR per Reviewer #2 and Reviewer #3's HIGH finding during Phase 3b review. Architectural mismatch: upstream's memory-provider discovery is a separate directory-scan path (plugins/memory/__init__.py) with its own FakePluginContext exposing register_memory_provider — incompatible with the standard PluginContext that entry-point-discovered plugins receive. Relocated mem0_oss could not register a MemoryProvider at gateway runtime via either the upstream PluginManager OR the gateway bootstrap shim. ## File changes - removed packages/fox-overlay/fox_overlay/agent_plugins/mem0_oss/ (3 files, ~1218 lines) - reverted pyproject.toml: drop mem0ai + qdrant-client dependencies (mem0_oss is back as a bundled fork plugin; upstream's own discovery path installs its plugin.yaml deps if needed) - reverted pyproject.toml entry-points: drop mem0_oss entry; keep fox_overlay entry for the 3 monkey-patches - reverted agent_plugins/__init__.py: register() no longer imports or invokes mem0_oss; docstring updated with deferral rationale - reverted MANIFEST.toml: drop mem0_oss entry; add deferral comment - bumped hermes-agent submodule pin from e301105e to 0a278fc1 (fork main after PR #3 merge which restores plugins/memory/mem0_oss/ byte-identical to pre-PR-#1 baseline) - added packages/fox-overlay/tests/test_agent_overlay.py — 9 pytest cases verifying each of the 8 monkey-patch sentinels + idempotency of register() ## Net Phase 3 scope (after this revert) The 3 Fox monkey-patches (Bedrock target_model, auxiliary.default fallback, cron failure diagnostics) ship as planned via the fox-overlay entry-point + gateway bootstrap shim. mem0_oss continues as a bundled fork plugin, working as it did in v0.5.4 — zero user- visible regression. ## Future work (out of scope for v0.6.0) A dedicated future phase handles proper memory-provider relocation, requiring either: (a) An upstream PR to NousResearch/hermes-agent adding register_memory_provider to the standard PluginContext, OR (b) A Fox-side fake-context construction in the overlay that mirrors plugins/memory/__init__.py's discovery path. Refs Reviewer #2 + Reviewer #3 findings on PR #221; rollback decision by Dennis 2026-05-17 (Option A of three presented).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 3a of Fox in the Box's v0.6.0 upstream-separation migration (fox-in-the-box-ai/fox-in-the-box epic NousResearch#155, sub-issue NousResearch#170). Fork-side PR. Companion monorepo PR follows once this merges.
What this does
Hands off all hermes-agent Fox customizations from the fork to the sibling
fox-overlaypackage via three coordinated changes plus a CI-attribution fix. Net effect once both PRs land: the fork is a virgin upstream tree (modulo the now-paused workflow), ready to be re-pointed atNousResearch/hermes-agentdirectly in Phase 8.File changes
Reverted to upstream merge-base content (5 files, byte-perfect to
e5dad4ac5):agent/auxiliary_client.pycron/jobs.pycron/scheduler.pyhermes_cli/runtime_provider.pytools/cronjob_tools.pyThe 5 edits relocate to
packages/fox-overlay/fox_overlay/agent_plugins/fox_overlay_plugin/monkey_patches/in the monorepo PR (NousResearch#171), where each becomes a runtime monkey-patch withinspect.signature()self-check.Deleted (3 files):
plugins/memory/mem0_oss/{__init__.py, README.md, plugin.yaml}— relocated topackages/fox-overlay/fox_overlay/agent_plugins/mem0_oss/in monorepo PR.kind: backendmanifest field added during relocation. Loaded via upstream'shermes_agent.pluginsentry-point group.Modified (2 files):
.github/workflows/notify-monorepo.yml— pausedrepository_dispatchpush trigger..github/workflows/notify-monorepo.ymlas-is"). Justified: symmetric with the webui fork's already-paused workflow and with the monorepo's already-pausedsync-submodules.yml. Without this, the merge push would dispatch ahermes-agent-updatedevent the monorepo silently drops — fighting itself.workflow_dispatchremains active —scripts/release.py— added"roadhero@gmail.com": "roadhero"toAUTHOR_MAP(commit 2). Resolves thecontributor-check.ymlCI red on this PR.Revert mechanics
Reverts use
git checkout e5dad4ac5 -- <path>wheree5dad4ac5= merge-base between forkmainandupstream/main. This restores byte-perfect upstream content from the fork point. Pulls NO upstream evolution — Phase 8 (NousResearch#200) handles the full upstream-tag re-point separately. Verified bit-exact viashasum -a 256parity across all 5 files.Out of scope
5-hat process
Reviewer findings actioned
contributor-check.ymlred —roadhero@gmail.comnot in AUTHOR_MAPscripts/release.pyworkflow_dispatchescape hatch warningauxiliary_clientrevert removes bug-fix-shaped editsCross-repo cohesion
Per
feedback_cross_repo_cohesion.md: this PR merges first; monorepo PR opens with the new submodule SHA. The fork between this merge and monorepo-PR merge runs with NO Fox customizations active on the agent side — that's expected; the monorepo's currently-pinned submodule SHA still points at44a6808f(pre-merge fork main with Fox edits intact), so deployed:stablecontainer is unaffected until the monorepo PR bumps the pin.Authorship
Both commits authored by
roadhero(Dennis Vorobyov) per Fox migration policy.