wiki: publish pending edits (2026-10-20)

This commit is contained in:
Tony0410
2026-10-20 18:38:11 +08:00
parent 667ab7f41d
commit fbd4f297eb
46 changed files with 1012 additions and 197 deletions
+19 -2
View File
@@ -3,8 +3,8 @@ title: Agent Memory
type: concept
status: active
created: 2026-07-22
updated: 2026-07-22
verified_on: 2026-07-22
updated: 2026-10-20
verified_on: 2026-10-20
confidence: high
tags: [concept, memory]
sources: [raw/configs/hermes-config-sanitized.txt]
@@ -20,6 +20,23 @@ How an agent retains facts/context beyond a single conversation turn or session
- **Verified settings:** memory enabled `true`; message char limit `5000`; user-profile char limit `3000`; flush every `6` turns; `auto_sleep: true`; default scope `session`; vector type `int8`; shared surface `data/shared/mnemosyne.db`.
- **Correction (2026-07-22, per Anthony):** this page previously named Hindsight as a predecessor backend on this box. That was a mix-up with a different Hermes instance — Hindsight has nothing to do with this box. See [[comparisons/memory-backends]] for the full correction.
> **Full `memory:` block re-read 2026-10-20 — the summary above is accurate but incomplete.**
> Live values: `memory_enabled: true`, `provider: mnemosyne`, `memory_char_limit: 5000`,
> `user_profile_enabled: true`, `user_char_limit: 3000`, `flush_min_turns: 6`,
> `nudge_interval: 10`. Mnemosyne sub-block: `auto_sleep: true`, `default_scope: session`,
> `vector_type: int8`, `sleep_threshold: 50`, `profile_isolation: true`,
> `shared_surface_path: data/shared/mnemosyne.db`, `shared_surface_read: true`,
> `sync_roles: [user]`, `ignore_patterns: []`.
>
> **New since 2026-07-22 and worth knowing:**
> - `reflect.disabled_for_cron: true`, `reflect.max_calls_per_session: 3`
> - **`skip_contexts: [cron, flush, subagent, background, skill_loop]`** — Mnemosyne capture is
> deliberately skipped in cron, flush, subagent, background and skill-loop contexts. That is
> exactly why cron runs like this one produce no Mnemosyne working memories: the skip is
> configured, not a malfunction.
> - `profile_isolation: true` — memories do not leak across Hermes profiles, which matters now that
> the `ops` profile has been deleted (see [[runbooks/multiplexer-setup]]).
## Model
- Per-turn conversation content accumulates up to the char limit, then flushes to the Mnemosyne store every 6 turns (rather than every turn) — a batching tradeoff between write frequency and losing very recent context if the process dies mid-batch.
- `scope: session` means recall defaults to the current session rather than a global cross-session pool, unless explicitly widened.
+22 -2
View File
@@ -3,8 +3,8 @@ title: Model Routing
type: concept
status: active
created: 2026-07-22
updated: 2026-07-22
verified_on: 2026-07-22
updated: 2026-10-20
verified_on: 2026-10-20
confidence: high
tags: [routing, model, inference]
sources: [raw/configs/hermes-config-sanitized.txt, current-profile.md]
@@ -20,6 +20,26 @@ How Hermes selects providers, applies combos, and enforces context limits.
- Default model: `freellmapi-auto`
- OmniRoute used for auxiliary roles: vision, skills_hub, approval, mcp, title_generation, web_extract, compression
> **Live `model:` block re-read 2026-10-20 — the default has moved.**
> `model.default` is now **`meituan/longcat-2.5-preview:free`** (provider `nous`), not
> `freellmapi-auto`. The block also carries `base_url`, `persist_switch_by_default` and `api_mode`.
>
> **The five `custom_providers` and their live base URLs:**
>
> | Name | Base URL | Host |
> |---|---|---|
> | `litellm` | `http://litellm:4000/v1` | LiteLLM proxy (m93p CT500) |
> | `guacano` | `http://ollama:11000/v1` | Guanaco on CT205 |
> | `Omniroute:20128` | `http://omniroute:20128/v1` | OmniRoute on m93p CT101 |
> | `command-code` | `https://api.commandcode.ai/provider/v1` | remote |
> | `omni-101` | `http://100.78.63.20:20128` | OmniRoute node `omniroute-1` (direct IP, not a DNS name) |
>
> Note the **two distinct OmniRoute endpoints**: `Omniroute:20128` (DNS, m93p CT101) and
> `omni-101` (raw IP `100.78.63.20:20128`). `omni-101` is the pin on effectively every
> model-calling cron job, and it has been the single point of failure in the 10-14, 10-16/17 and
> 10-19 incidents — its `/api/health` reads `200 {"status":"ok"}` throughout, so **reachability
> probes do not detect that failure class**; only job-level error text does.
## Context Enforcement
- Three layers: model native → combo `data.config.context_length` → Hermes client cap
- Effective context is the minimum of the three
+10 -2
View File
@@ -3,8 +3,8 @@ title: Persistent Shells
type: concept
status: active
created: 2026-07-22
updated: 2026-07-22
verified_on: 2026-07-22
updated: 2026-10-20
verified_on: 2026-10-20
confidence: high
tags: [concept, terminal]
sources: [raw/configs/hermes-config-sanitized.txt]
@@ -23,6 +23,14 @@ A background shell session that survives across separate Hermes turns/requests,
- **Home mode:** auto
- **Container CPU:** 1
> **Full `terminal:` block re-read 2026-10-20 — the above still holds, plus more.**
> Live values: `backend: local`, `persistent_shell: true`, `timeout: 180`, `home_mode: auto`,
> `container_cpu: 1`, `container_memory: 5120`, `container_disk: 51200`,
> `container_persistent: true`, `lifetime_seconds: 300`,
> `docker_mount_cwd_to_workspace: false`, `docker_run_as_host_user: true`.
> The container-* keys are inert while `backend: local` — the shell is not containerised on this
> box, so `container_memory`/`container_disk` describe a path that isn't taken.
## Why it matters here
Enables multi-step CLI workflows (e.g. a long `git` operation, a build, a watch process) to be resumed across turns without Hermes re-establishing context each time. The tradeoff is a shell can be left in a bad state (stuck process, wrong directory) that silently affects the *next* turn — see [[runbooks/troubleshoot-tool-loop]] for the related tool-loop guardrails that catch some of this.
+16 -2
View File
@@ -3,8 +3,8 @@ title: Tool Calling
type: concept
status: active
created: 2026-07-22
updated: 2026-07-22
verified_on: 2026-07-22
updated: 2026-10-20
verified_on: 2026-10-20
confidence: high
tags: [concept, tool-calling]
sources: [raw/configs/hermes-config-sanitized.txt]
@@ -35,6 +35,20 @@ hard_stop_after:
## Why it matters here
This is the mechanism behind [[runbooks/troubleshoot-tool-loop]] — that runbook's "agent loops on one tool" symptom is exactly what `warn_after`/`hard_stop_after` are meant to catch. With hard-stop disabled, a human (or the messaging integrations) is the actual backstop today.
> **Re-verified 2026-10-20 — the `tool_loop_guardrails:` block is byte-for-byte unchanged.**
> Live values in `~/.hermes/config.yaml`: `warnings_enabled: true`, `hard_stop_enabled: false`,
> `warn_after: {exact_failure: 2, same_tool_failure: 3, idempotent_no_progress: 2}`,
> `hard_stop_after: {exact_failure: 5, same_tool_failure: 8, idempotent_no_progress: 5}`.
> 90 days of no drift on this guardrail.
>
> **The 2026-10-18 incident is a *different* loop and the guardrails did not catch it.**
> The degenerate-stream case (see
> [[incidents/2026-10-18-degenerate-stream-canned-substitute]]) was a **model output** repetition
> loop — the *model* repeating itself in the response stream, not a tool call repeating. Hermes's
> degenerate-stream guard cut it and delivered a canned non-answer, but `tool_loop_guardrails`
> (which watches *tool* calls) never fired. Two separate loop-detection mechanisms exist; do not
> assume one covers the other.
## Related
- [[concepts/persistent-shells]]
- [[runbooks/troubleshoot-tool-loop]]