fix: harden retrieval routing and runtime hints

[acidember]

Summary

Supersedes #29141 with the same fix stack rebased onto the newer upstream main at 6a6766fb8.

This PR preserves two local Hermes fixes:

• Hardens retrieval/source-of-truth routing guidance so current/system/file/git/math facts are grounded through tools instead of memory.
• Wires retrieval route hints into the run-agent API payload only, avoiding persistence into session history.
• Stabilizes adjacent Kanban regression tests after upstream drift.

Fresh rebase note

Upstream moved again after #29141 was opened, causing a narrow conflict in hermes_cli/kanban_db.py around additive index creation. The conflict was resolved by keeping the newer upstream migration block that creates all additive task indexes after their columns exist, including idx_tasks_session_id.

Verification

Targeted verification after the fresh rebase:

python -m py_compile agent/conversation_loop.py agent/prompt_builder.py hermes_cli/kanban_db.py run_agent.py tests/agent/test_prompt_builder.py tests/run_agent/test_run_agent.py tests/hermes_cli/test_kanban_core_functionality.py
PASS

python -m pytest tests/agent/test_prompt_builder.py -q -o addopts= --tb=short
147 passed

python -m pytest tests/hermes_cli/test_kanban_core_functionality.py::test_default_spawn_auto_loads_kanban_worker_skill tests/hermes_cli/test_kanban_core_functionality.py::test_default_spawn_appends_per_task_skills tests/hermes_cli/test_kanban_core_functionality.py::test_default_spawn_dedupes_kanban_worker_from_task_skills tests/hermes_cli/test_kanban_core_functionality.py::test_gateway_dispatcher_disables_corrupt_board_without_traceback -q -o addopts= --tb=short
4 passed

python -m pytest tests/run_agent/test_run_agent.py -q -o addopts= --tb=short
345 passed

[teknium1]

Thanks for the contribution. I verified that the underlying gap still exists on current main, but I would not salvage this older branch as-is.

Problems

• The route classifier is too broad: agent/prompt_builder.py:304 sends any prompt containing review, task, decision, report, or ticket to shared_work whenever that surface is available. That can misroute normal PR/code-review or task-execution prompts away from the live-system/session grounding this PR is trying to enforce.
• The runtime shared_work mapping in run_agent.py:2179 checks for fabric_recall, fabric_search, and fabric_pending, but current main has no fabric_* tool definitions (git grep -n "fabric_recall\|fabric_search\|fabric_pending" origin/main -- . returned no matches), so that branch is not grounded in a current tool surface.
• This exact branch is mechanically stale: gh pr view 29143 --json mergeStateStatus reports DIRTY, and same-author #32206 is an open refreshed stack containing these commits plus later related work.

Suggested changes

• Narrow shared_work matching to explicit Fabric/shared-work phrasing, or remove it until the real tool names exist.
• Add regression cases for ambiguous prompts like review this PR, show this task status, and what did we decide last time.

Automated hermes-sweeper review.

[teknium1]

This marker set is too broad for a routing primitive. Words like review, task, decision, report, and ticket appear in ordinary PR-review and task-execution prompts; when shared_work is available this would steer those prompts away from the live-system/session-search sources the PR is meant to enforce.

[teknium1]

I could not find fabric_recall, fabric_search, or fabric_pending tool definitions on current main. Unless those names are introduced in the same stack, this shared_work surface is ungrounded and the branch is effectively dead in the current runtime.

[GottZ]

This was generated by AI during triage.

Summary

Two PRs address the retrieval-routing gap. #29141 and its fresher rebase #29143 both add source-of-truth classification and API-only route hints, but the shared-work classifier is broader than the reported grounding need and references Fabric tool surfaces absent from current main.

Related pull requests

• #29141 [closed] duplicate — (+596/-18) — duplicate: This earlier version adds retrieval guidance, route classification, API-only hint injection, and related tests, but it remains relevant only as the predecessor of #29143; it was closed after a contributor identified #29143 as the same fix rebased onto newer main at 6a6766f.
• #29143 related — (+592/-12) — superseded: This rebase carries essentially the same retrieval-routing implementation as #29141, including the overbroad shared_work keywords and mappings for fabric_* tools that the contributor review found absent from current main. Despite the keep_open verdict recorded on the #29143 review, the review text says not to salvage this DIRTY branch as-is and identifies same-author #32206 as the refreshed stack containing these commits plus later related work.

Duplicates

#29141 and #29143 are substantively the same fix stack; #29141 was closed as a duplicate of #29143, and #29143 is in turn superseded by the refreshed same-author stack in #32206.

Suggested consolidation

Close #29143 as duplicate of #32206, making the chain explicit: #29141 → #29143 → #32206. This departs from the visible keep_open verdict on #29143 because the accompanying contributor review documents that the branch is DIRTY, should not be salvaged as-is, contains an overbroad shared_work classifier and ungrounded fabric_* mappings, and has already been refreshed in #32206; #29141 can remain closed as the older duplicate.

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
    subgraph Dup29141 ["PRs duplicating each other"]
        P29141["PR #29141 (closed)"]
        P29143["PR #29143 (open)"]
    end
    class P29141 closed
    class P29143 open
    class P29143 target
    click P29141 "https://github.com/NousResearch/hermes-agent/pull/29141"
    click P29143 "https://github.com/NousResearch/hermes-agent/pull/29143"

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 0 issues in this complex. Each diff was read against this issue; Assessment working set: 63 kB of PR diffs, 3 kB of issue/PR text, 2 kB of discussion (3 comments), 0 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.