Rl capabilities && File Operator Tools by teknium1 · Pull Request #15 · NousResearch/hermes-agent

teknium1 · 2026-02-05T11:50:35Z

Initial RL training toolset and complete file operations toolset (to make it operate much more like a coding agent)

- Updated `.env.example` to include Tinker and WandB API keys for reinforcement learning training. - Enhanced `model_tools.py` to clarify configuration options and streamline the RL training process. - Expanded `README.md` with detailed instructions for setting up RL training using Tinker and WandB. - Modified `hermes_cli` files to integrate RL training tools and ensure proper configuration checks. - Improved `rl_training_tool.py` to reflect changes in training parameters and configuration management.

- Added the tinker-atropos submodule for enhanced RL training capabilities. - Updated model_tools.py to reorder RL function definitions and improve descriptions. - Modified rl_cli.py to include checks for the tinker-atropos setup and provide user guidance. - Adjusted toolsets.py and __init__.py to reflect changes in RL function availability. - Enhanced rl_training_tool.py to manage training processes directly without a separate API server.

- Modified `model_tools.py` to update default model IDs and add new RL function `rl_test_inference`. - Enhanced `README.md` with installation instructions for submodules and updated API key usage. - Improved `rl_cli.py` to load configuration from `~/.hermes/config.yaml` and set terminal working directory for RL tools. - Updated `run_agent.py` to handle empty string arguments as empty objects for better JSON validation. - Refined installation scripts to ensure submodules are cloned and installed correctly, enhancing setup experience.

… streaming - Added unique run ID generation for WandB tracking during test inference. - Enabled WandB usage for test tracking and updated command-line arguments accordingly. - Implemented real-time output streaming for process execution, improving log visibility and debugging. - Enhanced error handling to display last few lines of stderr for better troubleshooting.

- Introduced file manipulation capabilities in `model_tools.py`, including functions for reading, writing, patching, and searching files. - Added a new `file` toolset in `toolsets.py` and updated distributions to include file tools. - Enhanced `setup-hermes.sh` and `install.sh` scripts to check for and optionally install `ripgrep` for faster file searching. - Implemented a new `file_operations.py` module to encapsulate file operations using shell commands. - Updated `doctor.py` and `install.ps1` to check for `ripgrep` and provide installation guidance if not found. - Added fuzzy matching and patch parsing capabilities to improve file manipulation accuracy and flexibility.

taeyun16 · 2026-03-30T16:13:14Z

🔍 자동 분석 결과

이슈 요약

리더보드 시스템 — 주간/월간/지역별 순위로 경쟁 유도 + 보상 제공

기술 구현 고려사항

백엔드 (Phoenix/Ash)

리더보드 스냅샷: weekly_leaderboards (period_start, period_end, category)
순위 계산: (Oban, 매주 월요일 00:00 KST)
- WeeklyVoteKing: 주간 투표 수 top 100
- MonthlyCollector: 월간 신규 고양이 수집 top 100
- StreakChampion: 연속 투표 최장 기간 top 100
- RegionalRank: PostGIS ST_DWithin로 지역 필터링
순위 변동 추적: leaderboard_changes (user_id, change_amount, direction)
보상 지급: period结束时 자동 + 배지 발송

iOS (SwiftUI)

리더보드 탭: (segmented: 주간/월간/지역)
순위 Cell: (rank, avatar, name, score, change delta)
내 순위 고정: sticky bottom bar — "내 순위: 15위 / 234명"
변경 알림: 순위 진입/변동 시 토스트 + 햅틱

우선순위 제안

Phase 1: 주간 투표왕 (가장 단순 — existing vote_count 활용)
Phase 2: 지역별 순위 (PostGIS regional 쿼리 의존)
Phase 3: 월간 수집왕 (고양이 도감 완료 후)
Phase 4: 연속 기록왕 (Streak 시스템 통합)

관련existing 코드

— 아직 없으면 Streak worker 패턴 참고
Regional ranking: PostGIS queries

🤖 HanNyang Auto Resolver

* fix: cross-platform terminal/browser support and build fixes - Use platform-aware default shell (zsh on macOS, bash on Linux/Windows) instead of hardcoding /bin/zsh everywhere - Mark Playwright packages as external in Vite to fix chromium-bidi bundle errors during dev - Copy pty-helper.py to dist/server/assets/ on build so production server can find it - Fix browser launch guard to check `context` instead of `browser` (which is always null with launchPersistentContext) - Exclude unused diagnostics.ts stub from route tree - Update README with Playwright install step and Extra section Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> * docs: add Ubuntu/Debian prerequisites for Tauri desktop build Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Previously hermes-api.ts helper functions (hermesGet, hermesPost, hermesPatch, hermesDeleteReq, streamChat) sent no Authorization header, causing 401 'Invalid API key' errors when the gateway has API_SERVER_KEY configured. Also exports BEARER_TOKEN from gateway-capabilities.ts so hermes-api.ts can import and reuse it for auth headers. Co-authored-by: mgnyc11 <mgnyc11@users.noreply.github.com>

zhangtobybot-a11y · 2026-04-11T11:11:20Z

已合并到 PR #45

zhangtobybot-a11y · 2026-04-11T11:11:24Z

已合并到 PR #45

…ameter + schema property' (NousResearch#15) from fix/terminal-notify-on-complete into main

Rl capabilities && File Operator Tools

Fixes 12 remaining MEDIUM issues from the deep audit (19 total, 7 fixed in Round 12): design_agent: - NousResearch#15: add asyncio.wait_for(300s) around LLM API call to prevent infinite hangs - NousResearch#17: replace 2x hardcoded 'claude-opus-4-8' with shared DEFAULT_MODEL constant qa_agent / validate_agent: - NousResearch#20,NousResearch#22,NousResearch#23: already fixed in Round 12 (verified — dynamic timeout/threshold values used) memory.py: - NousResearch#24: frontmatter parser uses regex r'^---$' instead of str.split('---',2), preventing false splits on content containing '---' (SQL, markdown tables) - NousResearch#25: parse and preserve 'description' field from frontmatter in metadata, fixing write→load roundtrip data loss profiles.py: - NousResearch#26: ProfileConfig now frozen=True (immutable dataclass per coding standards) deploy_agent: - NousResearch#31: replace 2x sync subprocess.run with asyncio.create_subprocess_exec - fix 5x .decode() → .decode('utf-8', errors='replace') for Windows CJK safety - remove unused import subprocess db.py: - NousResearch#27: add class docstring explaining RLock + _unlocked pattern - NousResearch#28: FK constraints already in DDL (verified PRAGMA foreign_keys=ON active) - NousResearch#29: add _ensure_connection() with PRAGMA integrity_check(1) + auto-reconnect on 4 critical methods (create_task, get_task, claim_task, submit_result) - extract _create_connection() static method for reuse by reconnect Tests: 79 passed, 0 failed

…s.md

hutchutchutch · 2026-06-24T22:08:15Z

Refactor progress on PR #15 file operations, fifth slice: extracted read/delete/move behavior into tools/file_operations_read.py and moved shared result dataclasses into tools/file_operations_results.py while preserving tools.file_operations exports and ShellFileOperations wrapper methods. Commit: a20c28974 (refactor(file-ops): extract read and file actions). Validation: py_compile for touched modules; focused file-operation/parser/live suite 241 passed; standard gateway batches 265 passed and 115 passed; git diff --check clean. Current line counts: tools/file_operations.py 879, read 221, results 82, search 380, lint 339, support 164.

hutchutchutch · 2026-06-24T22:14:00Z

Refactor progress on PR #15 file operations, sixth slice: extracted write/patch mutation behavior into tools/file_operations_write.py and moved the abstract FileOperations interface into tools/file_operations_interface.py. Commit: 80da36609 (refactor(file-ops): extract write and patch actions). Validation: py_compile for touched modules; focused mutation/read/parser/LSP suite 284 passed; standard gateway batches 265 passed and 115 passed; git diff --check clean. The original hotspot is now under target: tools/file_operations.py 476 lines; interface 75; write 222; read 221; results 82; search 380; lint 339; support 164.

hutchutchutch · 2026-06-24T22:23:10Z

Progress update for PRD #15:

Implemented the next file-tool refactor slice in e6683b654 (refactor(file-tools): extract tracker state).

What changed:

Added tools/file_tools_state.py for read dedup state, read timestamp/staleness helpers, read tracker cap enforcement, and patch failure tracking.
Kept tools.file_tools as the compatibility adapter: existing private names like _read_tracker, _read_tracker_lock, _patch_failure_tracker, _cap_read_tracker_data, _check_file_staleness, reset_file_dedup, and notify_other_tool_call remain importable from tools.file_tools.
Preserved legacy monkeypatch behavior for _READ_HISTORY_CAP, _DEDUP_CAP, and _READ_TIMESTAMPS_CAP by syncing patched adapter constants into the state module before cap enforcement.

Line counts after this slice:

tools/file_tools.py: 1435 lines, down from 1671 before this slice.
tools/file_tools_state.py: 249 lines.

Validation:

python -m py_compile tools/file_tools.py tools/file_tools_state.py
focused tracker/read/staleness suite with TMPDIR rooted under /Users/home/.hermes/tmp-pytest-hermes-agent-issue15 -> 89 passed
tests/tools/test_gateway_cwd_contract.py with the same non-sensitive temp root -> 4 passed
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

Notes:

A combined path/cwd batch hit existing test-state leakage where one test left live cwd cache state for a reused task id; tests/tools/test_gateway_cwd_contract.py passes in isolation.
tests/tools/test_file_tools.py still has the existing macOS /tmp vs /private/tmp expectation mismatch on three mocked write/patch assertions; this predates and is unrelated to the tracker-state extraction.

hutchutchutch · 2026-06-24T22:29:02Z

Progress update for PRD #15:

Implemented another file-tool refactor slice in 6c6ead02f (refactor(file-tools): extract path safety helpers).

What changed:

Added tools/file_tools_paths.py for cwd resolution, TERMINAL_CWD sentinel handling, live cwd lookup, path-resolution warnings, blocked-device checks, sensitive path checks, Hermes config path resolution, and cross-profile/container mirror guards.
Kept tools.file_tools as the compatibility adapter for existing private imports and monkeypatches, including _resolve_path_for_task, _path_resolution_warning, _is_blocked_device, _is_blocked_device_path, _check_sensitive_path, _check_cross_profile_path, and _hermes_config_resolved/_hermes_config_resolved_loaded.
Preserved live file-ops cache lookup through a small adapter so path helpers do not own the file-ops cache.

Line counts after this slice:

tools/file_tools.py: 1139 lines, down from 1435 after the previous slice and 1671 before today.
tools/file_tools_paths.py: 313 lines.
tools/file_tools_state.py: 249 lines.

Validation:

python -m py_compile tools/file_tools.py tools/file_tools_paths.py tools/file_tools_state.py
focused tracker/read/staleness suite with TMPDIR rooted under /Users/home/.hermes/tmp-pytest-hermes-agent-issue15 -> 89 passed
focused Hermes config sensitive-path compatibility class -> 5 passed
tests/tools/test_gateway_cwd_contract.py with the non-sensitive temp root -> 4 passed
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

Notes:

The combined path/cwd batch still hits the existing order-dependent live cwd cache leakage for a reused task id; the affected cwd contract file passes in isolation.
This slice keeps the upstream-facing tool function names and private test patch points intact to reduce nightly merge risk.

hutchutchutch · 2026-06-24T22:32:42Z

Progress update for PRD #15:

Implemented the third file-tools slice in 37c3bdd13 (refactor(file-tools): extract env cache helpers).

What changed:

Added tools/file_tools_env.py for terminal-environment-backed ShellFileOperations creation and cache population.
Kept tools.file_tools as the compatibility adapter for _get_file_ops, _file_ops_lock, _file_ops_cache, and clear_file_ops_cache so existing tests and monkeypatches continue to target the same module.
The new helper receives the existing cache, lock, ShellFileOperations factory, and logger as arguments; it does not own the public cache state.

Line counts after this slice:

tools/file_tools.py: 1031 lines, down from 1139 after the path slice and 1671 before today.
tools/file_tools_env.py: 129 lines.
tools/file_tools_paths.py: 313 lines.
tools/file_tools_state.py: 249 lines.

Validation:

python -m py_compile tools/file_tools.py tools/file_tools_env.py tools/file_tools_paths.py tools/file_tools_state.py
focused env/cache/read/staleness suite with TMPDIR rooted under /Users/home/.hermes/tmp-pytest-hermes-agent-issue15 -> 80 passed
focused cwd/container suite -> 32 passed
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

This continues the low-merge-risk pattern: move implementation behind small modules while preserving the old tools.file_tools import and patch surface.

hutchutchutch · 2026-06-24T22:40:06Z

Progress update for PRD #15:

Implemented the next file-tools slice in 9a3e19b4d (refactor(file-tools): extract read handler).

What changed:

Added tools/file_tools_read.py for the read_file_tool implementation: pagination normalization, device guard, extractable document handling, binary/read safety guards, dedup/stub-loop handling, max-character guard, redaction, large-file hinting, and file-state read registration.
Kept tools.file_tools.read_file_tool as the registered compatibility wrapper.
The wrapper injects existing adapter functions/state (_get_file_ops, _get_max_read_chars, _resolve_path_for_task, _cap_read_tracker_data, _read_tracker, _read_tracker_lock), so existing tests and monkeypatches keep targeting tools.file_tools.

Line counts after this slice:

tools/file_tools.py: 788 lines, down from 1031 before this slice and 1671 before today.
tools/file_tools_read.py: 246 lines.
Existing extracted helpers remain under target: env 129, paths 313, state 249.

Validation:

python -m py_compile tools/file_tools.py tools/file_tools_read.py tools/file_tools_env.py tools/file_tools_paths.py tools/file_tools_state.py
focused read guards / read loop / read_extract / file_state registry suite -> 98 passed
tests/tools/test_file_staleness.py -> 11 passed
tests/agent/test_file_safety_credentials.py -> 20 passed
tests/tools/test_file_tools.py::TestReadFileHandler -> 4 passed
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

Note:

One intentionally broad combined read/staleness batch still hit the known order-dependent live cwd cache leakage; the affected staleness test passes in isolation and the focused stale suite passes fully.

hutchutchutch · 2026-06-24T22:43:12Z

Progress update for PRD #15:

Implemented the next file-tools slice in 6138eaa1a (refactor(file-tools): extract search handler).

What changed:

Added tools/file_tools_search.py for the search_files implementation: pagination normalization, repeated-search loop guard, ShellFileOperations search dispatch, match redaction, warning/hint formatting, and structured error handling.
Kept tools.file_tools.search_tool as the compatibility wrapper.
The wrapper injects existing adapter functions/state (_get_file_ops, normalize_search_pagination, _read_tracker, _read_tracker_lock), so existing tests and monkeypatches keep targeting tools.file_tools.

Line counts after this slice:

tools/file_tools.py: 741 lines, down from 788 after the read slice and 1671 before today.
tools/file_tools_search.py: 95 lines.

Validation:

python -m py_compile tools/file_tools.py tools/file_tools_search.py
tests/tools/test_file_tools.py::TestSearchHandler + tests/tools/test_read_loop_detection.py::TestSearchLoopDetection -> 13 passed
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

Next remaining PR #15 seam is the write/patch mutation handler block; extracting that should bring tools/file_tools.py close to or under the 500-line target.

hutchutchutch · 2026-06-24T22:51:28Z

Progress update for PRD #15:

Implemented the mutation/schema slice in 50e42d119 (refactor(file-tools): extract mutation handlers).

What changed:

Added tools/file_tools_mutation.py for write_file_tool and patch_tool implementation: sensitive/cross-profile guards, stale warnings, file_state locks, read timestamp refresh, files_modified/resolved_path reporting, V4A traversal guard, and patch failure escalation hints.
Added tools/file_tools_schemas.py for static file tool schemas.
Kept tools.file_tools.write_file_tool, tools.file_tools.patch_tool, _handle_write_file, and _handle_patch as compatibility/registry wrappers.
Restored compatibility re-exports for _detect_line_ending and _normalize_line_endings from tools.file_operations after the earlier support-helper extraction.

Line counts after this slice:

tools/file_tools.py: 455 lines (under the ~500 target; down from 1671 before today).
tools/file_operations.py: 478 lines (still under target).
New helper modules remain small: mutation 235, schemas 103, search 95, read 246, env 129, paths 313, state 249.

Validation:

python -m py_compile tools/file_operations.py tools/file_tools.py tools/file_tools_mutation.py tools/file_tools_schemas.py and related helper modules
tests/tools/test_line_ending_preservation.py -> 12 passed
tests/tools/test_file_staleness.py -> 11 passed
tests/tools/test_patch_failure_tracking.py -> 5 passed
mutation handler unit subset excluding known macOS /tmp normalization assertions -> 22 passed
mutation state/cross-profile/cwd combined batch -> 72 passed, with the known order-dependent live-cwd cache case failing when combined but passing in isolation
Standard gateway batch A -> 265 passed
Standard gateway batch B -> 115 passed
git diff --check / staged diff check clean

Known existing caveats kept separate:

tests/tools/test_file_tools.py still has three macOS /tmp vs /private/tmp mocked-call expectation failures; production behavior was not changed for this refactor.
A broad combined cwd/staleness batch can still hit the existing live cwd cache order leak; the focused staleness suite passes.

…ded for version bumps) (NousResearch#15) * [agento] fix: green-gate Gate 3 — auto-PASS dependabot PRs (no CR needed for version bumps) dependabot[bot] PRs are automated dependency bumps. Requiring a CodeRabbit review blocks them indefinitely when CR hits rate limits. Add PR_GATE3_AUTHOR check: if author == dependabot[bot], Gate 3 auto-PASSes. Human-authored PRs continue to require CR APPROVED. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [agento] fix: set LATEST_CR=APPROVED for dependabot Gate 3 auto-PASS (Gate 5 fast-path) BugBot caught that LATEST_CR was unset when dependabot PRs auto-PASS Gate 3, making Gate 5's CR-approved non-blocking fast-path unreachable for those PRs. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * [agento] fix: skeptic-cron Gate 3 — dependabot PRs auto-PASS (no CR needed) Mirrors green-gate Gate 3 dependabot exemption so dependabot version bumps can be auto-merged by skeptic-cron without requiring a formal CR review. Adds author field to PR JSON payload so dependabot detection works in the loop. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

Rl capabilities && File Operator Tools

…de changes) Ran a read-only dry-run against a real exported skill library (76 SKILL.md, 25 candidates, nested category dirs + CJK) in an isolated HERMES_HOME. Pipeline ran cleanly and safely (no mutation); LLM decisions were sensible (2 consolidations proposed, rest kept standalone). Two real gaps fixtures could never expose: - dry-run under-reports consolidation proposals (counts/arrays stay 0 while prose proposes merges) — asymmetric vs split/deprecate (KNOWN_ISSUES NousResearch#15) - keyword-retention guard is near-inert on nested/CJK skills: _load_skill_keywords resolves only flat paths -> name-only fallback (KNOWN_ISSUES NousResearch#16) Guards were not exercised (model used the YAML channel; 25 skill_view, 0 skill_manage). No code changed. Co-Authored-By: Claude <noreply@anthropic.com>

…n report (NousResearch#15 / P0) In dry-run nothing is removed, so the removal-based classification left consolidated/pruned empty even when the model proposed merges — while splits/deprecations already surfaced their YAML proposals. A dry-run preview therefore under-reported what a real run would do (REPORT.md printed 'consolidated: 0' above a prose body proposing merges). Surface the YAML-block consolidation/pruning proposals into run.json counts/arrays + REPORT.md, tagged source='model (proposed, dry-run)' and behind a DRY-RUN banner. The fold runs AFTER the cron-rewrite block so a dry-run never mutates cron/jobs.json; real-run classification is unchanged (guarded by a test). Co-Authored-By: Claude <noreply@anthropic.com>

…h-pass real-data re-verification Co-Authored-By: Claude <noreply@anthropic.com>

…o code changes) Ran a read-only dry-run against a larger, deliberately uncleaned backup (130 usage records vs 123 files, path-prefixed keys, all-active-despite-archive, external symlink). Result: no errors/exceptions handling the dirty data; LLM emitted a well-formed empty structured block (keep all 27); guard not exercised (YAML channel). NousResearch#15 fix makes the reported 0 trustworthy (matches empty block) vs the 8th-pass ambiguous 0. One untested dimension noted: the active-vs-archived usage mismatch's effect on the deterministic prune is not covered by dry-run (prune is skipped). No code changed. Co-Authored-By: Claude <noreply@anthropic.com>

Anthropic bills any >200K-input request on a subscription (OAuth) account to the EXTRA-USAGE budget, not plan limits — even without the context-1m beta header (1M context is GA on Claude 4.6+). A 339K-token desktop session got permanent 'You are out of extra usage' 400s while a small probe on the same token returned 200. Fix: _apply_subscription_context_cap() clamps the resolved context to 200K when the token is an OAuth/setup token (sk-ant-oat*, eyJ*, cc-*). Metered API keys (sk-ant-api*) are NOT capped. Escape hatch: anthropic.long_context. Tests: test_critical26_subscription_context_capped_at_plan_lane, test_critical26_long_context_escape_hatch (both pass) See CRITICAL NousResearch#15 in CRITICAL-ISSUES.md for full RCA.

Findings #1 and NousResearch#2, together, because NousResearch#2 makes #1 bypassable: a gate in the model-tools path is worth nothing while `registry.dispatch()` reaches handlers without consulting it. `resolve()` now has a production caller. A non-AUTO tool cannot execute without an execution capability, and a capability cannot exist without a decision — not "approved" as a boolean somebody might set early, but a token minted at consent, bound to the exact call, and destroyed by being used. Four properties, each with a test that fails without it. Bound to the tool and an argument fingerprint, so approving one call cannot execute a different one and mutating arguments between consent and execution invalidates the token. One-use, so a retry loop cannot turn one approval into several executions. Short-lived, because consent goes stale. And `consume()` raises rather than returning a boolean — a caller that forgets to check a boolean executes anyway, and this is the one check where forgetting must not be survivable. Fail-closed throughout. A broken permission lookup refuses; `requires_capability` returns True on any exception; an unreadable trust list trusts nothing; minting without a `tool_call_id` is refused outright, because a capability that could match any call is not a capability. Every refusal test asserts the handler ran **zero** times — a gate that refuses after the side effect is not a gate. **The default is `observe`, and that is a gap rather than a preference.** `get_tier()` returns ALWAYS_APPROVAL for unregistered tools — deliberately, so an unknown tool is never assumed safe — and most tools here were never registered. Switching straight to `enforce` would refuse nearly every call in the product, which is an outage, not a safety improvement. `observe` audits every call that *would* be refused, with its tool name, so the registration backlog is measurable against real traffic and the switch can be flipped once the audit goes quiet. This mirrors HERMES_APPROVAL_INTEGRITY_MODE, which exists for the same reason; both must reach `enforce` for the tier system to mean anything, and neither is there yet. Enforce mode is fully tested. Also corrected, per the review: the baseline-failure classification, which I had called environmental wholesale. Three of them are real defects at baseline — `atomic_config_write` NameErrors, a SQLite journal-mode assertion, and a jobs asset-contract failure — and the evidence now says so per failure. And the rollback procedure, which claimed one merge revert covered a many-commit range; it now gives a non-destructive runtime pin plus three repository options, with the history-rewriting one marked as needing explicit approval. Regression check: tests/tools 8532 passed / 19 failed — the failing set byte-identical to baseline, so the gate broke nothing. Still not merged to main. Remaining findings: NousResearch#6 undo wiring, NousResearch#10 Now composition, NousResearch#11 native resume, NousResearch#12 New Chat, NousResearch#13 free-form clarify, NousResearch#14 approval acknowledgement, NousResearch#15 reconnect, NousResearch#16 readiness coupling, NousResearch#17 sensor delivery. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Nu2Qaq5Y7EScuooGz8co34

…, coupling **NousResearch#14 — a card could resolve on a decision the gateway never recorded.** `approval.respond` answers with how many approvals it resolved, and zero is a *successful RPC that decided nothing* — the request expired, or was answered elsewhere. The client read the envelope and closed the card, so the owner believed they had answered while the agent stayed blocked. Now only a `resolved >= 1` resolves it, and only an explicit `status: "ok"` resolves a clarify. **NousResearch#13 — free-form clarify answers went out as new prompts.** The native path had a steering step that was a no-op returning true, after which the text was sent via `prompt.submit`: the question stayed open and the answer arrived as an unrelated message. Replaced by one `answerClarifyFreeText` that addresses the request by id natively and walks the menu on the terminal — one method, because splitting it into "steer, then send" is what let the two halves disagree. **NousResearch#11 — resume threw the transcript away.** `session.resume` returns `messages`; `open()` read only the identity fields. A refresh reattached to a running session and showed an empty feed, so execution continued against a conversation the owner could no longer see. History is now delivered through an `onHistory` hook that fires *before* `subscribe()` — ordering is the other half of the fix, since applying the past after live deltas have landed interleaves them — and merges through the existing hydration path rather than a second one. **NousResearch#12 — New Chat did nothing natively.** Both controls rotated a PTY identity that native mode does not use. The session stayed open, its durable id stayed in storage, and the next prompt continued the old conversation under a heading that said it was new. `startNew()` closes the session, clears the profile-scoped durable id before the rebuild so the effect cannot read a stale one, drops the live id immediately, and opens fresh without resuming. **NousResearch#16 — a missing TUI disabled the transport that exists to avoid the TUI.** The chat-readiness probe asks whether the *TUI* can start, and its verdict also gated native session creation. Each transport is now gated on its own prerequisite, and the probe does not run at all under native chat — the banner it raises describes a dependency native chat does not have. Gates: typecheck 0, eslint 0, 74 files / 706 frontend tests, build green. Remaining from the review: NousResearch#6 undo not wired to production, NousResearch#10 Now not consuming the ranking modules, NousResearch#15 reconnect, NousResearch#17 sensor delivery. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Nu2Qaq5Y7EScuooGz8co34

**NousResearch#15 — messages could sit "sending" for the rest of the session.** The hook opened once and, on failure, reported an error and stopped. Held messages waited on a `ready` that was never coming. Reconnect is now bounded with backoff and has its own visible state: `reconnecting` is distinct from `connecting` because the owner needs to know a previous attempt failed, and `gaveUp` is distinct from both because it names the point at which nothing further happens without them. When it gives up, held messages are marked failed — an optimistic row that never resolves is a message the owner believes was delivered, and telling the truth eventually is the entire purpose of the optimistic row. The retry budget resets on a successful open, so a later drop gets its own full allowance rather than inheriting an exhausted one. It retries the *open*, never the send: a fresh session is not a reason to re-run a prompt whose fate is unknown. **NousResearch#10 — the ranking modules were asserted and unenforced.** `nowOrder.ts` and `outcomeRanking.ts` had tests while `NowPage.tsx` imported neither and rendered its old composition, which is the same shape of gap as a tier that classifies and a dispatcher that never asks. Now projects real hub state through `nowOutcomes.ts` into the ranker and renders the top three: one recommendation marked *in words* rather than by colour, its reason underneath, each item's work state spelled out, and a native keyboard-operable button to replace the recommendation with a way back to Imperator's. The projection is where invented data would have come from, so it is the narrow place to look: every candidate derives from something the hub reported, there is no placeholder or "get started" row, and an empty hub renders no section at all. A halted guardrail outranks everything because nothing else on the page can proceed while it holds. Tests render the *real* `NowView` against a real payload. Testing the modules in isolation is exactly what let the page ignore them, and after two rounds of my own fixtures agreeing with my own mistakes, a test that builds its own version of the subject is not evidence. Gates: typecheck 0, eslint 0, 74 files / 714 frontend tests, build green. Remaining from the review: NousResearch#6 undo not wired to production, NousResearch#17 sensor capture paths not built. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Nu2Qaq5Y7EScuooGz8co34

… — Codex PR NousResearch#30 review (NousResearch#31) Two chatgpt-codex-connector[bot] review comments on PR NousResearch#30 that still applied to current main (the other 13 were already addressed in the merged PR): 1. client.py NousResearch#13 — Cancel timed-out MCP calls: _run_sync now cancels the concurrent.futures.Future on timeout before re-raising, so a stalled MCP endpoint doesn't leave a pending HTTP session running on the shared background loop after the caller has given up / opened the breaker. 2. __init__.py NousResearch#15 — Parse the host before allowing keyless local mode: is_available() now urlparse's the URL and requires an exact loopback host (localhost/127.0.0.1/::1) instead of a substring match, so a URL like 'https://localhost@example.com/mcp' (host=example.com) is no longer trusted as local keyless mode. Adds regression tests: test_memgw_client_timeout.py (cancel-on-timeout fails against pre-fix client) + TestKeylessLocalModeHostParsing (3/4 fail pre-fix). Deferred (already fixed in merged PR NousResearch#30): MCP dep import check, MCP dep pin, loop lock, isError handling, stale-prefetch generation, sync/delegation thread tracking + shutdown join, session-switch prefetch invalidation, no join on turn path, default provider unset. Deferred (complex, needs interface change): NousResearch#3/NousResearch#4 per-turn user_id refresh for shared gateway sessions — sync_turn/prefetch only receive session_id, not user_id, so threading identity through requires a MemoryProvider interface change; NousResearch#1 in-tree placement is an architectural call (AGENTS.md says existing in-tree providers stay).

teknium1 added 6 commits February 3, 2026 23:41

initial RL training tools and loop

f018999

teknium1 merged commit 8dd3831 into main Feb 5, 2026

SHL0MS mentioned this pull request Mar 30, 2026

[UX] Tool approval UX: "always" label is misleading, timeout is not configurable #4067

Closed

h4x3rotab pushed a commit to Clawdi-AI/hermes-agent that referenced this pull request Apr 10, 2026

Merge PR NousResearch#15: Final header cleanup — subtractive pass

58fc261

malaiwah pushed a commit to malaiwah/hermes-agent that referenced this pull request Apr 11, 2026

Merge pull request 'fix(terminal): add missing notify_on_complete par…

7e557c0

…ameter + schema property' (NousResearch#15) from fix/terminal-notify-on-complete into main

kshitijk4poor mentioned this pull request Apr 15, 2026

feat: add hermes-blender skill for 3D modeling and rendering #10191

Closed

njiangk mentioned this pull request Apr 21, 2026

fix(whatsapp): mitigate libsignal protobufjs CVE in bridge deps #13259

Closed

3 tasks

angelburgosrosado pushed a commit to angelburgosrosado/hermes-agent that referenced this pull request Apr 27, 2026

Merge pull request NousResearch#15 from NousResearch/rl-capabilities

8552bd9

Rl capabilities && File Operator Tools

blaketyndall-eng mentioned this pull request May 14, 2026

W2-T08: lru_cache get_tool_output_limits to skip per-call config parse blaketyndall-eng/hermes-agent#10

Merged

acoastalfog mentioned this pull request May 23, 2026

Bridge legacy KB dashboard plugin to standalone app #31190

Closed

Renge-Star mentioned this pull request May 24, 2026

[--tui] gateway exits mid-turn with stdin EOF (TUI closed the command pipe) — NOT byterover-related #27282

Closed

BowmanStephen mentioned this pull request May 25, 2026

[feature-adoption] Adopt remaining unused Hermes features (9 unstarted) BowmanStephen/hermes-agent#9

Open

madlouse mentioned this pull request Jun 2, 2026

Feature: Gateway-level hard routing gate for outbound message classification #37271

Open

xzmzm added a commit to xzmzm/hermes-agent that referenced this pull request Jun 3, 2026

fix: reorder patches NousResearch#15/NousResearch#16 in hermes-patche…

87a9100

…s.md

waefrebeorn pushed a commit to waefrebeorn/slermes that referenced this pull request Jul 2, 2026

Merge pull request NousResearch#15 from NousResearch/rl-capabilities

b33d361

Rl capabilities && File Operator Tools

seiichi3141 mentioned this pull request Jul 4, 2026

feat(kai): broadcast CLI — OBS 起動/配信開始停止/クリーン終了の一括制御（M4） HyuCode/kai-agent#3

Merged

falkoro mentioned this pull request Jul 9, 2026

[Bug]: TypeError: Completions.create() got an unexpected keyword argument 'system' in chat_completions.py via _build_kwargs_from_profile #60821

Closed

1 task

feiguang50-hub pushed a commit to feiguang50-hub/hermes-agent that referenced this pull request Jul 19, 2026

docs: mark NousResearch#15/NousResearch#16 P0s resolved + record nint…

64b644a

…h-pass real-data re-verification Co-Authored-By: Claude <noreply@anthropic.com>

ElystrumCore mentioned this pull request Jul 19, 2026

feat(plugins): fail-closed required enforcers + scheduled-run/model-spend governance ElystrumCore/hermes-agent#1

Merged

bbasketballer75 mentioned this pull request Aug 1, 2026

test: isolate the tool-definition caches between tests #75837

Closed

ComBba mentioned this pull request Aug 1, 2026

fix(gateway): expose agent exit metadata to hooks ComBba/hermes-agent#2

Closed

ComBba mentioned this pull request Aug 2, 2026

fix(gateway): report abnormal turn exits, including crashes ComBba/hermes-agent#3

Merged

teknium1 mentioned this pull request Aug 8, 2026

feat(skills): add product-price-monitor skill + price-watch blueprint (salvages #78673) #81906

Merged

ragnos-dev mentioned this pull request Aug 13, 2026

fix(sidecar): restore opt-in inbound-liveness heartbeat file #84996

Closed

choi138 added a commit to choi138/hermes-agent that referenced this pull request Aug 18, 2026

merge: integrate M8 turn tracing (NousResearch#15)

366f376

Navigation Menu

Uh oh!

Rl capabilities && File Operator Tools - #15

Rl capabilities && File Operator Tools#15
teknium1 merged 6 commits into
mainfrom
rl-capabilities

teknium1 commented Feb 5, 2026

Uh oh!

taeyun16 commented Mar 30, 2026

Uh oh!

zhangtobybot-a11y commented Apr 11, 2026

Uh oh!

zhangtobybot-a11y commented Apr 11, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

teknium1 commented Feb 5, 2026

Uh oh!

taeyun16 commented Mar 30, 2026

🔍 자동 분석 결과

이슈 요약

기술 구현 고려사항

백엔드 (Phoenix/Ash)

iOS (SwiftUI)

우선순위 제안

관련existing 코드

Uh oh!

zhangtobybot-a11y commented Apr 11, 2026

Uh oh!

zhangtobybot-a11y commented Apr 11, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

hutchutchutch commented Jun 24, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants