feat(gateway): config-driven per-chat channel_context injection

[Lazymonter]

Problem

The inbound pipeline already accepts adapter-provided event.channel_context, but operators cannot attach deterministic context to a specific chat through configuration. Dynamic orchestrators therefore have to patch gateway/run.py and re-apply that patch after every update.

This keeps the inline/file config shape proposed by @liuhao1024 in #43728 and provides a current-main implementation that incorporates the review feedback there around canonical keys, cache identity, runtime config resolution, and integration coverage.

Solution

Add gateway.channel_context_map with inline and JSON-file forms:

gateway:
  channel_context_map:
    "telegram:-1002285219667": "This chat is bound to the release workspace."

gateway:
  channel_context_map: chat-context.json

• Keys use one canonical <platform>:<chat_id> form, preserving native colons and never falling back to an unqualified ID.
• Relative paths resolve from the active profile's HERMES_HOME; ~ and ${ENV_VAR} are supported through the existing runtime config path.
• Adapter/backfill context remains first. Configured context is appended under [Configured chat context] without mutating the inbound event.
• File-backed maps are checked on each inbound message and parsed only when their path/signature changes. Cache identity includes path, device, inode, size, mtime, and ctime, with bounded LRU state.
• Missing, malformed, oversized, non-regular, or concurrently changing files fail closed for that message; stale bindings are not reused. The docs specify atomic replacement for external writers.
• Validation limits files to 1 MiB, 4096 entries, 512-byte UTF-8 keys, and 2 KiB UTF-8 context values. Warning deduplication is also bounded and does not log map keys or values.

Testing

• Added direct resolver coverage for canonical keys, inline/file config, profile isolation, cache invalidation, atomic replacement, TOCTOU changes, size/type limits, parser/path failures, and bounded caches.
• Added runtime integration coverage for merge order, event immutability, profile-scoped config loading, and existing sender/context behavior.
• scripts/run_tests.sh tests/gateway/test_channel_context_map.py tests/gateway/test_session.py tests/gateway/test_shared_group_sender_prefix.py tests/gateway/test_context_ref_expansion_runtime.py tests/gateway/relay/test_channel_context_consume.py -q — 145 passed.
• Ruff and git diff --check pass.

Closes #43650

[teknium1]

Thanks for incorporating the #43728 review feedback. Current main still only prepends adapter-provided event.channel_context in gateway/run.py:10566-10570, so the runtime-configured/file-backed behavior remains needed.

The PR constructs a platform-qualified key in gateway/channel_context.py:67-79, validates bounded inline/file maps, and uses a path plus device/inode/size/mtime/ctime signature before cache reuse (gateway/channel_context.py:287-405). The integration resolves under the existing profile-aware runtime-config path and applies in the shared preprocessing method, covering both normal and queued inbound paths (gateway/run.py:2426-2441, gateway/run.py:10517-10896, current main locations).

No blocking correctness issue was identified in the PR diff. The existing channel_overrides.system_prompt mechanism is a static system-prompt alternative (gateway/config.py:409-440), whereas this PR supplies the distinct dynamic inbound-context map described in #43650.

This is an automated hermes-sweeper review.

[GottZ]

This was generated by AI during triage.

Summary

Two open PRs address #43650 through inline and file-backed per-chat context injection. #43728 adds the initial loader and inbound integration, while #65148 retains that configuration shape and adds canonical platform-qualified keys, profile-aware resolution, robust cache invalidation, bounded validation, non-mutating context composition, integration coverage, and documentation.

Related pull requests

• feat(gateway): config-driven per-chat channel_context injection #43728 duplicate — (+329/-0) — n/a: Implements inline/file-backed context maps, but its lookup uses raw source.chat_id while its documented examples use platform-qualified keys, and its cache identity stores only mtime. Despite the COMMENTED keep_open review on feat(gateway): config-driven per-chat channel_context injection #43728, feat(gateway): config-driven per-chat channel_context injection #65148's diff directly supersedes this implementation by addressing the review's canonical-key, path/signature-cache, runtime-config, and integration-test findings.
• feat(gateway): config-driven per-chat channel_context injection #65148 best fix — (+921/-5) — n/a: The recorded best fix for [Feature]: Config-driven per-chat channel_context injection for inbound messages #43650 implements canonical <platform>:<chat_id> lookup, profile-scoped inline/file configuration, bounded fail-closed parsing, path/device/inode/size/mtime/ctime cache identity, non-mutating context composition, tests for adapter-built sources and routed profiles, and user documentation. Consistent with the automated keep-open verdict, its diff preserves the salvageable feature while resolving the concrete defects identified in feat(gateway): config-driven per-chat channel_context injection #43728.

Duplicates

#43728 and #65148 implement the same configuration-driven per-chat inbound-context feature for #43650; #65148 is the superseding implementation that incorporates the documented review feedback on #43728.

Suggested consolidation

Keep #65148 open with a salvage path: preserve its canonical-key resolver, profile-aware runtime integration, robust bounded file cache, non-mutating merge behavior, regression coverage, and documentation while the author rebases or splits out a mergeable cut. Close #43728 as a duplicate of #65148 because the latter supersedes its implementation and addresses the COMMENTED keep_open review's specific defects.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I43650(["issue #43650 (open)"])
    subgraph Dup43728 ["PRs duplicating each other"]
        P43728["PR #43728 (open)"]
        P65148["PR #65148 (open)"]
    end
    P65148 -->|best fix| I43650
    class I43650 open
    class P43728 open
    class P65148 open
    class P65148 best
    class P65148 target
    click I43650 "https://github.com/NousResearch/hermes-agent/issues/43650"
    click P43728 "https://github.com/NousResearch/hermes-agent/pull/43728"
    click P65148 "https://github.com/NousResearch/hermes-agent/pull/65148"

Loading

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 2 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 50 kB of PR diffs, 8 kB of issue/PR text, 3 kB of discussion (3 comments), 3 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.