incognito-private-mode
Private Mode, honestly. A Hermes plugin that adds privacy controls you can
check: provider privacy flags, memory stripping, tool vetoes, and
consent-gated history removal, plus a status command that tells you exactly
what is and is not covered. No blanket "Incognito mode" claim; every
protection is a coverage line with a stated scope.
Pick your profile (30 seconds)
hermes plugins enable incognito-private-mode
hermes incognito-private-mode preset complete # or: provider-private
hermes incognito-private-mode setup # optional: history removal (asks consent)
Both presets write the same provider-side settings; they differ in what
Hermes keeps on your machine. Each cell states what the plugin enforces and
where that stops.
preset complete | preset provider-private | |
|---|---|---|
| Main-chat requests (enforced by middleware) | only routes rated ok, ok-if-flagged (flags injected and checked) or an account setting you confirmed in setup; ⛔ training, ⚠️ unverified, unknown and local/loopback endpoints are blocked | same |
| Background calls (titles, compression, vision, summaries) | classified and disclosed, not enforced: no flags, strip or blocks are applied to them | same |
| Memory blocks in main-chat requests | stripped (request blocked if removal cannot be proven) | stripped (same) |
| Persona (SOUL.md) in main-chat requests | stripped, best-effort (status/audit say when it could not be located) | same |
| Model can save memories / learn skills | no: memory, skill_manage vetoed | yes: saved locally by Hermes; memory blocks are stripped from later main-chat requests while the snapshot's file version is among the last 6 versions this Hermes process observed (history is process-wide, shared across chats, and only records versions seen at memory-strip requests); otherwise — or if the file changed before the chat's first private request — the request is blocked (fail closed) |
Past-chat search (session_search) | vetoed | allowed: results from past chats go into the conversation and to the provider |
| Hermes chat history | removed on close for CLI chats (default scope), only after setup consent, only for chats started after consent; earlier compressed parts are kept (see What stays) | kept |
| Plaintext http:// to remote hosts | blocked | blocked |
| Outbound audit trail | local, content-free | local, content-free |
Strict blocking means a local model (Ollama, LM Studio, a LiteLLM proxy) is
blocked under both presets: a loopback endpoint may forward anywhere, so it is
rated ⚠️. Set allow_unverified: true to use one (it is then allowed with a
warning). Neither preset touches what external memory providers receive (see
What stays).
Everything is still an individual setting; presets just write the bundle.
Tweak any setting afterwards (see Settings below); /private status tells you
which profile you're running (complete, provider-private, or custom).
Install
hermes plugins install <source-url-or-path> # or drop this directory into ~/.hermes/plugins/
hermes plugins enable incognito-private-mode
hermes incognito-private-mode setup # guided wizard (TTY only), records cleanup consent
Requires Hermes >= 0.22.5. Python: whatever Hermes runs (3.11/3.14 tested).
No extra dependencies; Hermes-provided SDK types only. AGPL-3.0.
What it does
| Coverage line | What happens | Enforcement |
|---|---|---|
| 1. Main chat requests | Curated route table classifies the endpoint host (authoritative) + provider id. ⛔ training routes blocked outright; ⚠️ unverified/unknown routes allowed-with-warning by default (strict mode, used by both presets, blocks them); ok-if-flagged routes get extra_body.provider.data_collection=deny, zdr=true injected copy-on-write. | enforced |
| 2. Background calls (titles, compression, vision) | Classification and counting only. Aux calls are not intercepted; status says so explicitly. | disclosed, not enforced |
| 3. On this machine | Consent-gated history removal: sessions are hashed (SHA-256, s:-prefixed) into a provenance ledger; after you close a chat on a surface in apply_to_platforms (default: CLI only), an opportunistic sweep deletes that chat's rows, messages, FTS entries and dump files through Hermes' own fenced public API. Compression parents are not deleted (see What stays). | enforced after setup consent; unproven CLI close paths (/new, resume switch, branch) never auto-delete — only quitting the CLI is proven |
Plus:
- Memory strip (when
use_memoryis false): built-in memory blocks
(MEMORY (your personal notes)/USER PROFILE (who the user is)) and
<memory-context>…</memory-context>spans are removed from outgoing
main-chat requests (copy-on-write; exact text reconstruction, so
multi-paragraph entries, even ones containing the§separator, are
removed whole; when exact reconstruction is impossible the request is
blocked, not sent). Background calls are not stripped. - Tool vetoes:
memory(unlesssave_memory),skill_manage(unless
learn_skills),session_search(unlesskeep_history), and an external
memory provider's tool only when it is named after the provider id.
A veto returns a clear message; an internal error in the check also vetoes — SDK-shaped responses for the three known transports, and for any other transport an allowlist-scrubbed request (only the model field survives; the turn fails downstream with zero user content sent). Residual: a third-party middleware downstream of this plugin could reconstruct content from the middleware chain's original_request context — unsupported transports are therefore best-effort protected and this is disclosed rather than claimed as absolute. - Log redaction: inbound-message, conversation-turn and tool-call preview
log lines carry[redacted]instead of your text while mode is notoff;
the redactor is linear-time (no ReDoS). Lines written before the plugin
loaded are not touched. - Plaintext-remote block: main-chat
http://to anything non-loopback is
blocked (cleartext egress) unless you setallow_plaintext_remote. - Persona strip (opt-in
strip_persona, on in both presets): your SOUL.md
text is located verbatim in main-chat requests and replaced; when it cannot
be located, the audit record and status say so instead of claiming
protection. - Outbound audit log: a local, content-free record of every main-chat
provider request while private mode is on (host, verdict, blocked, persona
outcome) at
audit.jsonl;/private statussummarises the last hour. Never message
content, 0600, size-capped. - API error translation: provider "no endpoint matching data policy" /
billing errors become one plain-English sentence instead of a retry loop. /privatechat command:status · keep · on · off · cleanup · deep-clean.
Status tells the truth
$ hermes incognito-private-mode status
Private Mode: ON — coverage, never a blanket Incognito:
Profile: custom
1. Main chat requests: ⚠️ unverified retention; allowed with a warning
2. Background calls (titles, compression, vision): not enforced — ⚠️ unknown (auto routing)
3. On this machine: history cleanup not enabled (run `hermes incognito-private-mode setup`)
...
Outbound audit (last hour): 4 calls (ok-if-flagged×1, warn×3; 0 blocked); hosts: api.deepseek.com, ...
What stays (never removed by this plugin): final replies kept for retries (by Hermes design),
CLI input history (~/.hermes/.hermes_history), paste files, interrupt_debug.log, deleted bytes in
free DB pages until a deep clean, auxiliary calls without our flags, ...
(Illustrative output with default settings, which match neither preset.
Under either preset line 1 reads "unverified; blocked" for the same route.)
What stays
The plugin never claims these are gone. /private status repeats the list.
| What stays | Why | How to remove it |
|---|---|---|
| Final replies of gateway-handled chats, kept for retries | Hermes design (core state_meta admission results) | Not removable by this plugin; it never deletes core receipt rows |
CLI input history (~/.hermes/.hermes_history): typed and pasted text | Hermes' shared prompt history across all chats | Edit or delete the file yourself; the plugin never erases shared files |
Paste files (~/.hermes/pastes/) | The CLI saves large pastes as files | Delete them yourself |
interrupt_debug.log in your Hermes home | Hermes debug log | Delete it yourself |
| Deleted bytes in free SQLite pages | SQLite marks deleted pages free; it does not overwrite them | /private deep-clean prints the steps: quit Desktop/CLI, hermes gateway stop, hermes sessions optimize (never --force), restart. SSD remanence remains after that |
| Earlier parts of a compressed private chat (compression parents) | Hermes ends them with end_reason='compression', which is not a user close, so the plugin never deletes them; the close notice reports how many remain | Delete them yourself with Hermes' own session tools if you want; the plugin never deletes compression chains |
| Background/auxiliary calls (titles, compression, vision, summaries) | Hermes sends them outside the main-chat middleware; the plugin only classifies and counts them, so no flags, strip or blocks apply | Point auxiliary.<task> at a route you trust; status line 2 shows its rating |
| Chats on TUI, Desktop and messaging surfaces | Cleanup scope defaults to apply_to_platforms: [cli] because those interactive rows are not e2e-verified; messaging cleanup is not built | Add tui / desktop to apply_to_platforms (opt-in); messaging: none |
Chats started before consent, while mode was off, kept with /private keep, started with keep_history: true, or ended by an automatic reason or crash | By design: only consented, armed, user-closed chats are deleted | Delete them yourself with Hermes' own session tools |
| Everything an external memory provider received | A configured memory.provider sees every message; status warns when one is set | Remove the provider from config; its stored data is governed by that provider |
Past chats returned by session_search (wire name chat_history_lookup on Anthropic OAuth) | With keep_history: true (provider-private) the tool is allowed, and its results are part of the conversation sent to the provider | Set keep_history: false to veto it |
Max-iterations summary call, codex_app_server runtime, safe-mode sessions, gateway shutdown snapshots/spools, save_trajectories, spawn-ledger.json for -q, logs from before the plugin loaded, backups, terminal scrollback, messages on messaging platforms | Outside the plugin's hooks | Per item, manually; see /private status |
| MoA advisor fan-out + aggregation | With provider: moa, EVERY turn's full main-chat request (memory and persona included) is sent UNSTRIPPED to each MoA advisor before private mode runs; private mode then blocks the aggregation. Advisors are not classified or counted. Do not use MoA in private chats. | honest disclosure — avoid MoA in private mode |
Disclosures
- In-process
state.dbwrites: deletion runs inside the Hermes process
through Hermes' fenced public API (delete_session(sessions_dir=None, expected_delete_ids=…, exclude_active_write_guards=True)); core checkpoints
the store itself (the plugin opens no extra connection for that — catalog
review). No CLI delete, no--force, no
subprocesses. - Process-wide log filter: attached in
register()to the
gateway.run,agent.turn_contextandhermes.tool_callsloggers and
their handlers; it redacts every record of those formats while mode is not
offand is detached on unload. - Background cleanup thread: sweeps run in a daemon thread
(ipm-sweep) started from the session-start hook (rate-limited) and the
session-finalize hook, at most 20 sessions per batch. - Local plugin state files in the plugin data dir (0700):
ledger.json,
confirmations.json,meta.json, a lock file andaudit.jsonl, all 0600.
Session ids are stored hashed; the audit log holds no message content.
Safety design (the short version)
- Fail closed on deletes. Deletion only proceeds for entries the ledger
can fully prove (armed lineage + closed with an explicit user end-reason,
fresh-ledger re-check under the lock). Any unprovable or unreadable target
refuses the whole operation. - Requests: the rewrite step fails open, the final check fails closed. If
the request-rewrite middleware errors, the request continues to the final
check; if the final check errors, the turn is blocked via the synthetic
shape where the transport allows it. - Hermes' own fence is the only delete path: never the CLI delete, never
--force, never compression-chain deletion. No subprocesses anywhere. - Open rows are sacred: sessions that never closed, or closed via an
automatic reason (ws_disconnect, idle timeout, …), are never deleted. - Pre-setup chats survive: cleanup applies only to sessions born after
you recorded consent ("arming generation"); mode-off chats are excluded
permanently, even after re-enabling. keepis honored: an excluded chat stays excluded across generations
and profile switches; surviving lineage is never falsely marked deleted.- Settings/consent stay plugin-side: the ledger, confirmations and consent
live in 0600, flock-serialized files; the plugin's own messages show only
hashed ids and verdict labels/counts.
Settings
mode (private|off) · apply_to_platforms (cleanup scope, default cli)
· remove_history_on_close · keep_history · use_memory · save_memory
· learn_skills · allow_unverified (default true; false = strict) ·
allow_plaintext_remote (default false) · strip_persona (default
false) · audit_log (default true) · block_training ·
deep_clean_reminder. Per-chat overrides via /private on|off|keep: off
turns protections off for that chat; on and keep force private even when
global mode is off; keep also excludes the chat from cleanup.
Honest limitations
- Background/auxiliary calls are classified and disclosed, not intercepted
(statusline 2). - Blocked-route banner is per api_mode: chat_completions/anthropic_messages/
codex_responses get a synthetic reply; other modes fail the turn with the
provider error (disclosed). - Deleted bytes remain in free SQLite pages until a manual deep clean
(deep-cleanprints the exact steps; it does not run them); SSD remanence
is physics. - Compressed private chats are only partly removed: compression parents stay
(see What stays). - Interactive TUI surfaces were exercised indirectly (one-shot CLI paths were
the e2e workhorse); the TUI uses the same hooks and finalize path, and
cleanup defaults to CLI-only until TUI/Desktop rows pass (opt in via
apply_to_platforms).
Development
source ../harness/env.sh # isolated HOME harness
cd tests && env -u PYTHONPATH "$IPM_PY" -m pytest -q unit/ contract/ # 349 tests at HEAD
hermes plugins validate <dir> # catalog validation
License: AGPL-3.0.
README from the reviewed commit 8430658; it updates when the author re-pins.

Tranquil-Flow