[Merge after #12] Thought signatures for gemini need to be passed back to support multistep by hjc-puro · Pull Request #13 · NousResearch/hermes-agent

hjc-puro · 2025-11-20T02:15:05Z

No description provided.

taeyun16 · 2026-03-30T16:13:08Z

🔍 자동 분석 결과

이슈 요약

투표 미션 시스템 — 일일/주간 미션을 제공하여 사용자 목표 유도 + 보상 지급

기술 구현 고려사항

백엔드 (Phoenix/Ash)

미션 템플릿 테이블: mission_templates (type, conditions_json, reward_amount, reset_cron)
유저 미션 인스턴스: user_missions (user_id, template_id, progress, completed_at, claimed)
이벤트 드리븐 발송: Ash change + 로 투표 시마다 progress UPDATE
Oban Worker: 일일 리셋 (KST 자정) —
보상 지급: 활용 (Streak 시스템과 공유 가능)

iOS (SwiftUI)

미션 진행률 View: (progress bar, reward preview, completion checkmark)
API: → MyMissions query + MissionProgress subscription
알림: 로컬 노티피케이션 (KST早上 9시 미션 리셋 알림)

우선순위 제안

Phase 1: 일일 미션 3종 (첫 투표, 투표 달인, 지역 탐험가)
Phase 2: 주간 미션 + 보상 타겟 시스템
Phase 3: 완료 보상 가챠 티켓 + 스트릭 조합

🔍 Code Review for PR #13

Issues found

1. safe_print.py — broken fallback rendering
When rich is unavailable the function falls back to print(*args) without stripping rich markup, so users see literal [bold red]...[/bold red] tags in plain output.

2. profiling.py::aggregate_profiling_stats — statistically wrong

aggregated["tools"][tool_name]["times"].extend(
    [tool_stats.get("mean_time", 0.0)] * tool_stats.get("call_count", 0)
)

Reconstructing per-call timings by repeating the mean N times yields incorrect min/max/median across workers (e.g. min == max == mean for any worker that contributed). The aggregator should combine summary stats directly.

3. run_agent.py — verbose-log spam & lazy imports

logging.debug(f"Tool call attributes: {dir(tc)}") and tc.model_dump() were left in from debugging — they dump every attribute on every tool call when --verbose.
from profiling import reset_profiler as reset_prof inside run_conversation should be hoisted to the module-level import.

4. batch_runner.py — bare except: and unused import

Two bare except: clauses swallow KeyboardInterrupt / SystemExit. Should catch (json.JSONDecodeError, TypeError, AttributeError).
import re is unused.
content.strip() is called inside the bare-except branch without first checking isinstance(content, str), which can re-raise AttributeError if content is non-string JSON.

5. tools/simple_terminal_tool.py — bare except:
Two bare except: clauses around ssh_context_manager.__exit__() swallow control-flow exceptions. Replaced with except Exception.

Resolution

Fixes pushed in commit 7490cc3a on fork branch DavidUmKongs/hermes-agent@thought-sig.

Verified: all 5 files compile cleanly, and a regression test against aggregate_profiling_stats confirms correct call_count/total_time/min/max after the fix (median is now flagged as approximate via median_time_approximate: true since per-call data is not preserved across worker boundaries).

Per P0-C spike decision (c), formalize aiseo_cli.py sync as a permanent fork-only asset. ADR-001 captures Context (5 BLOCKING gaps in upstream), Decision (don't refactor to delegate), Consequences (re-evaluation triggers + quarterly review), and Status. Adds 3 protection tests guarding the BLOCKING gaps that justify (c): - gap NousResearch#3: rolling backup pruning keeps newest 5 by timestamp - gap NousResearch#7: text-level yaml migration preserves all comments + blanks - gap NousResearch#13: _migrate_profile_config calls os.fsync on tmp file CLAUDE.md, NEXT_STEPS.md and seeds README all link back to ADR-001 + spike doc for traceability and quarterly re-evaluation cadence.

- 5/25·5/26 wvb-daily-brief-kst0700 cron 연속 Gemini HTTP 429 (paid Tier 1 1M TPM 초과). v2.5 신설 시 명시한 "429 risk 약간 상승 감수"가 실제 발생. - config.yaml ms365 preset: "mail,files" → "mail" (5/16 안정 복귀) - SKILL.md Step 5 임시 OFF 마킹, Output Format OneDrive 줄 교체, Verification NousResearch#13 strikethrough - 재활성화 조건: (1) Gemini Tier 2 upgrade, OR (2) SKILL.md trim (21K → 8-10K bytes) 후 토큰 여유 확인 Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Path.expanduser() raises RuntimeError 'Could not determine home directory' for an unresolvable ~user token (e.g. a terminal command referencing ~someuser that doesn't exist, or ~ with HOME unset). Two hot-path call sites caught only (OSError, ValueError), so the RuntimeError escaped and crashed entire agent turns. Observed live 2026-06-20 in the SysDesign evolve cron (API call NousResearch#13, 'Outer loop error') — the whole turn died on a single best-effort path hint. Fixes (both purely advisory/best-effort, must never be fatal): - subdirectory_hints._add_path_candidate: add RuntimeError to the except clause. AGENTS.md hint discovery skips an unresolvable candidate instead of raising. - tool_dispatch_helpers._extract_parallel_scope_path: wrap expanduser() in try/except returning None — caller then falls back to SEQUENTIAL execution (the safe default) instead of crashing the dispatch path. Fleet-wide impact: every profile + cron was vulnerable to a ~baduser token in any tool call. Regression tests added to both modules (reproduce the exact crash via terminal command + path-scoped arg). 56 tests pass, ruff clean. Takes effect on next gateway restart (not hot-reloaded).

…ed skills Review rec NousResearch#13 (data-loss item): these were live in the working tree but uncommitted/unbacked-up — they vanish on any checkout/upgrade. Contents: constitution-injection wiring (cognitive_processor.py, prompt_builder.py), daily_summary_facts.py (deferred-asks), file_operations.py (include_ignored), athena_telegram_proxy, cron/jobs.json, SOUL.md, and Athena's self-authored skills (self-audit-cognitive-autonomy, verify-substrate-before-action) + refs. All Python files syntax-checked clean. Backs up to the `personal` fork only — never pushed to nousresearch upstream. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

Path.expanduser() raises RuntimeError 'Could not determine home directory' for an unresolvable ~user token (e.g. a terminal command referencing ~someuser that doesn't exist, or ~ with HOME unset). Two hot-path call sites caught only (OSError, ValueError), so the RuntimeError escaped and crashed entire agent turns. Observed live 2026-06-20 in the SysDesign evolve cron (API call NousResearch#13, 'Outer loop error') — the whole turn died on a single best-effort path hint. Fixes (both purely advisory/best-effort, must never be fatal): - subdirectory_hints._add_path_candidate: add RuntimeError to the except clause. AGENTS.md hint discovery skips an unresolvable candidate instead of raising. - tool_dispatch_helpers._extract_parallel_scope_path: wrap expanduser() in try/except returning None — caller then falls back to SEQUENTIAL execution (the safe default) instead of crashing the dispatch path. Fleet-wide impact: every profile + cron was vulnerable to a ~baduser token in any tool call. Regression tests added to both modules (reproduce the exact crash via terminal command + path-scoped arg). 56 tests pass, ruff clean. Takes effect on next gateway restart (not hot-reloaded).

When agent.max_tools_per_turn trips and build_api_kwargs withholds tools, an enforcement-prompted local reasoning model (observed live: ornith 35B, lmstudio api_mode) whose STATIC system prompt still commands it to always use tools returns an EMPTY response — no content, no reasoning. The empty-response retry runs 3x on identical input and the turn dies reason=empty_response_exhausted with no final answer delivered: API call NousResearch#13: ... in=28847 out=95 latency=17.6s Empty response (no content or reasoning) after 3 retries. Turn ended: reason=empty_response_exhausted ... api_calls=13/30 On the FIRST budget-tripped completion call of a turn, append a single system-channel note telling the model tools are gone for the rest of the turn and a plain-text final answer is expected. Injected at the build_api_kwargs tools-withhold choke point — the same layer as the tools gate — so it reaches every api_mode with no per-transport duplication. System role, never a synthetic user message mid-loop (AGENTS.md). A per-turn latch (_tool_budget_wrapup_injected), reset alongside _tools_dispatched_this_turn at turn start, fires it exactly once: network retries reusing the same api_messages list and empty-response retries that rebuild it never duplicate the note. No-op when the budget is off (byte-identical, zero new messages). Anthropic takes a single system param and its adapter was last-writer-wins, so a trailing system note would clobber the byte-stable primary prompt. Make convert_messages_to_anthropic ACCUMULATE additional system messages (mirroring the Bedrock adapter's append and Gemini's join); byte-identical for the single-system-message case. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NP1MdSdZXeYWX2meR6fZo2

The rank lookup only contained projected *tip* ids, but FTS5 search hits roots and mid-chain sessions just as often (old messages match). Those resolved to None and silently fell back to the sequential result index — producing the mixed garbage the user saw (4, 80, 3, 4, 262, 6, 7 with duplicate 4s and impossible small numbers). Now session_rank() walks forward along continuation children (latest started_at first, max 20 hops) until it finds an id in the projected rank map — every chain's live tip is in the map — and returns that chain's position. Result for 'search tmux' is now 4, 80, 44, 80, 262, 303, 552: all real positions in # Title Model Tok Created Last Preview ID ── ───────────────────────────────────────────────── ────────── ────────── ────────── ──────── ──────────────────────────────────────── ──────────────────────── 1 — deepseek-… 0/0 2026-07-31 4h 20260731_071846_9fa701 2 — deepseek-… 30k/8k 2026-07-31 4h can you fix copy selected to clipboar… 20260731_065709_d308d6 3 fintwit-daily-fetch · Jul 31 06:07 deepseek-… 28k/412 2026-07-31 5h [IMPORTANT: You are running as a sche… cron_9223b2202bd2_20260731_060057 4 Tmux config and scripts review summary deepseek-… 99k/69k 2026-07-31 4h review .tmux config and scripts in .l… 20260731_055644_5e1c37 5 — deepseek-… 30k/10k 2026-07-31 6h is there a way to open all folders in… 20260731_052757_d3bd37 6 Managing Personal Hermes Skills Organization deepseek-… 188k/54k 2026-07-31 7h I realiezd that I need to keep my per… 20260731_030112_fd9939 7 Search sessions for data collector deepseek-… 64k/1k 2026-07-31 9h sessions search data-collector 20260731_014020_1bec3c 8 Update Plan Skill with Discovery Phase deepseek-… 52k/24k 2026-07-31 9h [User attached file: /home/shiro/.her… 20260731_012537_aa571d 9 Search Hermes browser iMac socat session deepseek-… 117k/44k 2026-07-31 9h search session where we worked on /br… 20260731_004120_67edd7 10 Hermes CLI Sessions and Resume Commands NousResearch#13 deepseek-… 121k/56k 2026-07-31 2s [CONTEXT COMPACTION — REFERENCE ONLY]… 20260731_113524_142504 11 Linking iMac Chrome to opencode browser deepseek-… 277k/20k 2026-07-31 7h I want to give opencode the same brow… 20260731_000128_f44ad5 12 Improving Hermes Agent Codebase Access Beyond RAG deepseek-… 72k/18k 2026-07-31 11h Looking to improve hermes agent, trie… 20260730_234830_1a2cdd 13 Add oh-my-openagent docs to qmd collection deepseek-… 108k/40k 2026-07-30 11h can you add oh my openagents docs in … 20260730_232541_0a40bc 14 fintwit-daily-fetch · Jul 30 06:04 deepseek-… 28k/358 2026-07-30 1d [IMPORTANT: You are running as a sche… cron_9223b2202bd2_20260730_060042 15 Monid setup requires API key deepseek-… 35k/6k 2026-07-29 1d set up https://monid.ai/SKILL.md 20260729_230009_5f48fd 16 Quick storage cleanup with approval deepseek-… 45k/15k 2026-07-29 1d find quick way to free storage space,… 20260729_215349_39b397 17 Phone SSH Fix and Termux Venv Setup deepseek-… 95k/39k 2026-07-29 1d help me fix my phone tmux launcher. s… 20260729_205759_70586e 18 Config file fixes for opencode deepseek-… 71k/26k 2026-07-29 1d can you fix @file:opencode.json and @… 20260729_181012_522948 19 Setup FaceSwap Google Colab Notebook deepseek-… 55k/6k 2026-07-29 1d @file:README.md I want to test this t… 20260729_163129_b94fe9 20 Fix Discover Offline Cloudflare Repo Error deepseek-… 41k/5k 2026-07-29 1d trying to update my apps in Updates -… 20260729_065842_ee8947, with the root/tip pair of one chain correctly sharing its slot (80). Rank window raised 500 -> 2000 so the whole store is covered (569 projected rows today). Tests: 95 pass.

…In/Out) The dedup only dropped ancestors that were directly in the FTS5 result set. When intermediate generations were missing (search matched only e.g. NousResearch#7 and NousResearch#13), the walk stopped and both children were listed twice. Now every result is grouped by its deepest compression ancestor (backward walk across compression edges only — branch children stop the walk so distinct conversations are never collapsed) and the newest descendant wins, so one conversation = one row regardless of which generations matched. Uses a backward-only walk instead of get_compression_lineage, whose forward walk assumes a linear chain and fragments on divergent children (seen in the 'Script unique pwd' chain). Also renames the token column header from 'Tok' to 'Tok(In/Out)' (input/output) per request. Tests: 95 pass.

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

cemendes · 2026-08-09T03:27:22Z

Remediation for review findings

New branch fix/pr13-remediation at exact SHA 1a334481c515ae59c7902581adc07c9a79fc4972:

Common ancestor restored: rebased onto origin/main at e92a0da4a0. Merge-base confirmed.
EOF whitespace fixed: trailing blank line at test_kanban_blocked_sticky.py:484 removed. git diff --check now passes (exit 0).
Same two files: hermes_cli/kanban_db.py and tests/hermes_cli/test_kanban_blocked_sticky.py — content identical to reviewed head cb3451cf6 except for the EOF fix.

Awaiting new exact-SHA CI evidence.

hjc-puro added 7 commits November 15, 2025 00:01

add tracking for cluster failurse

31c7333

add simple terminal

f813959

rate limits

1614c15

don't log exit code !=0 as terminal failure

ab7293b

add linewise profiling

349e37d

add profiling

e06a15b

support gemini models

a219e17

hjc-puro changed the title ~~Thought signatures for gemini need to be passed back to support multistep~~ [Merge after 12] Thought signatures for gemini need to be passed back to support multistep Nov 20, 2025

hjc-puro changed the title ~~[Merge after 12] Thought signatures for gemini need to be passed back to support multistep~~ [Merge after #12] Thought signatures for gemini need to be passed back to support multistep Nov 20, 2025

teknium1 closed this Feb 2, 2026

SHL0MS mentioned this pull request Mar 30, 2026

[UX] Respect NO_COLOR env var and TERM=dumb — zero accessibility checks in codebase #4066

Closed

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

Merge pull request 'fix(delegate): restore credential pool lease rele…

c7a6ee2

…ase in finally block' (NousResearch#13) from fix/delegate-credential-lease-leak into main

kshitijk4poor mentioned this pull request Apr 17, 2026

feat: promote kimi-k2.5 to first position in all model suggestion lists #11745

Merged

blaketyndall-eng mentioned this pull request May 14, 2026

W2-T06: TTL-cache _load_gateway_config + hoist call in message handler blaketyndall-eng/hermes-agent#7

Merged

hernandez42 mentioned this pull request May 24, 2026

🔒 Security Audit: 14 confirmed issues in hermes-agent #31715

Closed

BowmanStephen mentioned this pull request May 25, 2026

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

Open

This was referenced May 25, 2026

fix(tests): mock copilot token resolution in auth_remove_copilot test #31953

Closed

fix(agent): ignore stale local base_url for direct providers #30034

Open

fix(tests): monkeypatch copilot token resolution to survive borrowed-credential pruning #31975

Open

OmarB97 mentioned this pull request May 31, 2026

fix(agent): allow bounded repeated stall retries #35642

Closed

momomojo mentioned this pull request Jul 17, 2026

Enhance async tool execution and error handling in Hermes agent for A… #19

Merged

Bruiserhq mentioned this pull request Aug 12, 2026

ACP adapter: turns ending in text_response are silently dropped on Buzz-connected sessions (never published to relay) #84227

Open

Navigation Menu

Uh oh!

[Merge after #12] Thought signatures for gemini need to be passed back to support multistep - #13

[Merge after #12] Thought signatures for gemini need to be passed back to support multistep#13
hjc-puro wants to merge 7 commits into
mainfrom
thought-sig

hjc-puro commented Nov 20, 2025

Uh oh!

taeyun16 commented Mar 30, 2026

Uh oh!

DavidUmKongs commented Apr 27, 2026

Uh oh!

cemendes commented Aug 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

Uh oh!

Conversation

hjc-puro commented Nov 20, 2025

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!

DavidUmKongs commented Apr 27, 2026

🔍 Code Review for PR #13

Issues found

Resolution

Uh oh!

cemendes commented Aug 9, 2026

Remediation for review findings

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants