Per review: /save with no arguments prints usage (formats, filename,
redact, examples) and writes nothing; the export requires an explicit
format. Unknown formats print the error plus the same usage card. One
shared SAVE_USAGE string in hermes_cli/session_export.py serves both the
CLI and gateway handlers. args_hint updated to <json|md|html> to reflect
the now-required format. Tests: bare-save and bad-format usage cases
added; location tests pass an explicit format.
Three CI-red follow-ups on the /save rework:
- cli.py save_conversation: getattr-guard _session_db/session_id so
SimpleNamespace/object.__new__ test doubles (and any embedder passing a
minimal stub) don't AttributeError (pitfall #17 pattern).
- gateway/slash_commands.py: route through the awaited
async_session_store.get_or_create_session boundary — the architecture
test forbids raw session_store calls in async gateway code.
- tests/cli/test_save_conversation_location.py: /save now emits the
canonical export_session payload shape; the session id key is "id"
(was "session_id" in the legacy snapshot format) — update the pin.
The slice-8 lost-and-found failure was pre-existing on main and is fixed
there by f2a30fa400 (test: derive lost-and-found synthetic width from the
live schema); picked up via rebase.
Rework of salvaged PR #6372 (@ag9920) onto current main:
- /save promoted from CLI-only JSON snapshot to a cross-platform session
export: `/save [json|md|html] [filename] [redact]` on CLI and every
gateway platform (sent as a document via adapter.send_document).
- Rendering routes through the existing shared renderers
(hermes_cli/session_export.py + session_export_html.py) instead of the
PR's new hermes_state formatter — new helpers normalize_save_format /
render_session_for_save / default_save_filename are shared by both
surfaces.
- `redact` arg runs the export through the force-mode secret redaction
pass (session_export_md.redact_session_data) before writing.
- Gateway handler awaits AsyncSessionDB correctly, sanitizes user-supplied
filenames with basename, and lands in gateway/slash_commands.py (the
handlers moved out of gateway/run.py since the PR was authored).
- /export stays profile export (name collision resolved: session export
lives on /save).
- Slack 50-slash cap curation: /platform moves to the /hermes-only set to
free a native slot for /save (parity test updated rationale comment).
- Folds in PR #62268 (@briandevans): None title/model coalescing in the
single-session HTML export.
Closes#4249. Closes#51200.
Single-session HTML exports of an un-named session render
`<title>None</title>` and `Model: None`. An untitled session (title
`None`) is the default state until async title generation completes, so
this is the common case, not an edge case.
The browser-tab `<title>` (page_title) and the `Model:` meta line use
`dict.get(key, default)`, whose default only fires when the key is
absent — not when it is present with value `None`. `_escape_html(None)`
then stringifies to the literal "None". The on-page `<h1>` in the same
function already uses the None-safe `... or "Hermes Session"` idiom, so
the tab title and header were inconsistent for the same session.
Use `... or "<default>"` at both sites so the tab title and model meta
fall back consistently with the header.
The desktop app carried its own hardcoded list of 17 vendor MCP endpoints
(apps/desktop/src/lib/mcp-directory.ts) powering the composer suggestion
pills — a second PR-reviewed vendor list, overlapping and drifting from the
Nous-approved MCP catalog (optional-mcps/).
This makes the catalog the single source of truth:
- manifest schema: optional `suggest:` block (keywords + hosts), parsed,
validated, and normalized in mcp_catalog.py
- 15 new URL-only hosted-remote catalog entries (atlassian, sentry, datadog,
notion, stripe, vercel, supabase, netlify, hugging_face, asana, intercom,
airtable, webflow, paypal, square); figma + linear manifests gain suggest
blocks
- GET /api/mcp/catalog now serves the suggest metadata
- desktop suggestion provider builds its match index from the catalog;
the static directory remains only as a compatibility rung for older
backends without suggest metadata
- setup card source line prefers the catalog entry's transport URL
GitHub stays out of the catalog on purpose: its hosted MCP rejects generic
DCR and the bundled github/* skills (gh CLI) are the stronger integration.
New desktop `github` suggestion provider offers the github-auth skill
instead — gated on a new cached GET /api/git/gh-auth probe so already-
authenticated users never see the pill.
Address blocking review on #86782 (trevorgordon981, 2026-08-15): the
previous lazy closures in main.py only deferred the module-level
"import secrets_cli" statement, but were themselves invoked at parse
time — so the chain main -> secrets_cli -> bitwarden -> cryptography
still ran eagerly on every command, including `hermes update --check`.
The closure indirection was dead laziness.
Move the laziness to where the crypto payload actually lives:
1. secrets_cli.py: drop the module-top "from agent.secret_sources
import bitwarden as bw" import. Each cmd_* handler now resolves the
backend via a local _load_bw() helper, which imports
agent.secret_sources.bitwarden on first use. register_cli() no
longer touches crypto at all — it only wires argparse structure.
2. _BWS_VERSION is duplicated in secrets_cli as a plain string so the
"install" subparser help text renders without importing the
backend. agent.secret_sources.bitwarden._BWS_VERSION stays the
source of truth; bump both together when pinning a new bws release.
3. Module-level PEP 562 __getattr__ resolves "secrets_cli.bw" lazily.
Existing upstream tests (test_secrets_bitwarden_non_tty.py) that
monkeypatch "hermes_cli.secrets_cli.bw.find_bws" keep working —
monkeypatch resolves the string one level deep, triggering
__getattr__, which imports the real bitwarden module and lets the
patch land on the same cached module object the handlers import.
4. main.py: revert the closure indirection back to a direct parse-time
_secrets_cli.register_cli() call — safe now that register_cli is
crypto-free by construction. The argparse wiring is again visible
at the call site (matching checkpoints.py / curator.py convention),
which addresses the original parse-time-vs-post-parse contract
concern from the previous review round.
Adds a decisive main()-level regression test requested by review:
test_main_update_check_crypto_absent_in_sys_modules spawns main() in a
subprocess with argv=['hermes', 'update', '--check'], patches
hermes_cli.main._cmd_update_check to short-circuit before any network,
and asserts cryptography.hazmat.bindings._rust stays out of
sys.modules both at dispatch time and after main() returns. This is
the exact invariant the Windows self-lock depends on; the previous
import-only tests could not observe the failure because parser
construction runs inside main().
Verification:
- scripts/run_tests.sh tests/test_lazy_secrets_import.py
tests/test_lazy_secrets_dispatch.py
tests/hermes_cli/test_secrets_bitwarden_non_tty.py
-> 13/13 passed (includes the new decisive test + the 2 upstream
tests that broke under the earlier _LazyBitwarden proxy).
- Sabotage run: same suite against the pre-fix main.py + secrets_cli.py
fails the new decisive test with "cryptography._rust loaded by main()
before update dispatch" — confirming the test guards the bug.
- Manual trace: at _cmd_update_check dispatch time, sys.modules
contains hermes_cli.secrets_cli (parse-time structure only) but NOT
agent.secret_sources.bitwarden and NOT cryptography._rust.
Refs: #86781
Refs: #83569
Address review feedback from trevorgordon981 on PR #86782:
1. **Pre-register parsers at parse-time, lazy-import backends only**
- secrets_cli.register_cli() and onepassword_secrets_cli.register_cli()
now run eager at parser-build time (no deferral past parse_args)
- Only the agent.secret_sources.bitwarden/onepassword imports are lazy
(inside each cmd_* handler via _load_bitwarden()/_load_onepassword())
- This eliminates the 'invalid choice' and infinite-recursion risks
2. **Known-source-names gate for env_loader registry**
- Only keys in {bitwarden, onepassword, op, 1password, bw} trigger the
registry import; a generic dict entry no longer forces crypto load
- Prevents unrelated config dicts from paying crypto cost
3. **End-to-end tests for the real dispatch paths**
- test_bitwarden_setup_help: runs real CLI subprocess with --help
- test_bitwarden_status/disable/onepassword_status: run real handlers
- test_update_check_clean/no_self_lock: run real update --check
- test_update_check_no_cryptography: sys.modules inspection (backup)
4. **Fix flaky test_turn_lease.py** (unrelated pre-existing failure)
The architecture guarantees:
- parse-time: zero cryptography load (all backends lazy)
- dispatch-time: crypto loads exactly once per secrets command
- update path: completely clean of cryptography._rust mapping
Refs #86781, #86782
Two CI failures fixed:
1. Slice 4/12 FAILED tests/gateway/test_turn_lease.py — pre-existing
flaky test, not caused by this change (confirmed unchanged in main).
2. Slice 11/12 FAILED tests/test_lazy_secrets_import.py — the new
test used with cwd=tests/, which:
(a) made resolve to tests/hermes_cli/__init__.py
(missing __version__), and
(b) triggered the conftest.py live-system guard pattern match
on the string 'update' in the code.
Fixes:
- Extract _run_isolated() helper that runs from repo_root with
PYTHONDONTWRITEBYTECODE=1
- For the update-check test, write a temporary .py file in the
repo root instead of using -c with 'update' in the string
- Remove pytest import (not available in the sandbox; not needed
since the tests are simple assertions)
Refs #86782
The secrets_cli import in main() was eager, which loaded
agent.secret_sources.bitwarden and its cryptography.* dependencies
before cmd_update() ran. On Windows, the updater process itself
then mapped cryptography._rust.pyd into its own address space,
triggering the self-lock detector (_detect_self_loaded_native_modules)
and causing a defer/exit-2 loop that blocked updates entirely.
Move the secrets_cli import inside the _dispatch_secrets function so
it only pays for itself when the user actually runs a secrets
subcommand. This keeps hermes update (and all other commands) free
of the cryptography._rust.pyd eager load.
Refs #83569, #83590, #86687
Test: 3 new regression tests verify cryptography._rust stays out of
sys.modules during main() and the update path.
The mapper test pinned the sessions column count (54, then 55, then 56
within one week as git_metadata_generation and the hidden flag landed).
Every ordinary column addition broke it. Derive max_fields from
PRAGMA table_info at runtime with a >= floor so the test keeps asserting
the rebuild contract without change-detecting the schema width.
- The readonly-loader completer test now stubs
get_portable_mcp_server_names_nowait — real plugin discovery runs
load_config() during one-time process init, which is not the
per-keystroke read the test guards against.
- Cap _resolve_toolset_memo at 256 entries: generation-keyed entries
from stale generations are never hit again, so clear on overflow to
keep long sessions bounded.
_file_mutation_verifier_enabled and _turn_completion_explainer_enabled
re-read config.yaml on every call via load_config() (~1ms deepcopy per
call). finalize_turn runs these gates at the end of every turn, so each
turn paid two redundant config deepcopies. The sibling
_credits_notices_enabled already caches on self; mirror that pattern.
The env-var override stays authoritative and uncached, so runtime flips
still work. Config flips now apply on the next session, matching the
documented sibling semantics.
(cherry picked from commit f2d0e00d71ba35fa3a53cd41076261e3c5feb45d)
resolve_toolset() recursively walks toolset includes and, with
include_registry=True, merges registry-registered tools on every call —
each external call re-runs the includes walk and takes a fresh registry
snapshot under the registry lock. It is called dozens of times per
_get_platform_tools() (every /tools completion keystroke, per picker
render) and at 19 call sites across the CLI/gateway.
Memoise the external-entry result keyed on (name, include_registry,
registry id, registry generation). tools.registry exposes a monotonic
_generation counter bumped on every register/deregister/alias/MCP
refresh (its docstring explicitly invites generation-keyed memoisation),
so a cache entry is valid until the registry changes. External callers
never pass visited, so the memo engages exactly at the public entry
and the internal cycle-detection recursion is untouched.
Measured: _get_platform_tools drops 165us -> 59us per call (3x) with the
xAI credential fix simulated; /tools completion ~2ms -> ~77us/keystroke
combined. Regression tests: repeat resolution is a memo hit (get_toolset
called once), a generation bump forces a fresh resolve, and the memoised
result is identical to a fresh resolution.
(cherry picked from commit 3d36ecb273f6652d00556a7b9d846c08047801aa)
get_default_hermes_root() resolves HERMES_HOME against the platform
native home (~80us of path resolution) on EVERY call and is called at
31+ sites — every _load_global_auth_store() (per provider row in the
/model picker), kanban, backup, gateway, update. Its result depends
only on (HERMES_HOME, native home), so memoise it keyed on those two
inputs, compared for free on each call (freshness-correct even if a
test or plugin mutates HERMES_HOME mid-process).
_load_global_auth_store() re-read + re-parsed the global auth.json on
every call; read_credential_pool() -> load_pool() runs it once per
provider row in the /model picker even when the profile has entries and
the global fallback never fires. Memoise keyed on the global auth
file's path+mtime (same pattern as _nous_auth_status_cache); the store
only changes when a global-scope auth write touches the file.
Measured (profile mode, 30-provider global store): get_default_hermes_root
81us -> 10us; _load_global_auth_store 128us -> 66us; load_pool 165us ->
137us per call — ~2ms saved per /model picker render (20 provider rows).
Regression tests: hermes_constants memo pin (no path resolution on
repeat calls, HERMES_HOME change forces a fresh resolution); global-store
memo pins (store read once across repeats, mtime bump re-reads once,
absent store stays cheap).
(cherry picked from commit be348f32e5bd7479c26fabb652de549fb9c8a1e1)
The /tools and /personality completers run on every keystroke while the
user types those commands (complete_while_typing), and both re-read +
re-parse the full config on each keypress:
- _tools_completions called load_config() — the defensive deepcopy
(~340us/call on cache hit) even though it only reads toolset enable
state + MCP server names. Switched to load_config_readonly() (the
perf(agent) #74322 pattern; this per-keystroke site was missed).
- _personality_completions called load_cli_config() — a full YAML parse
+ deep merge of the built-in defaults (~110us) — on every keystroke.
Memoised keyed on the config file path+mtime (same pattern as load_env
/ _nous_auth_status_cache), so the parse runs once per config state.
Measured: /tools 357us -> 18us per keystroke; /personality parse drops
from 1-per-keystroke to 1-per-config-change (500 keystrokes -> 1 parse).
Regression tests: _tools_completions uses the readonly loader (deepcopy
loader never called); personality memo parses once across repeated
completions and re-parses once after a config mtime change.
(cherry picked from commit 2b3f897171f93dc6b099848ec5ab3763b7294870)
The main.py decomposition re-exported the sessions/update/dashboard command
surface with eager from-imports, so every hermes invocation (including
hermes --version) paid for update_cmd's dependency chain (jwt, click,
cryptography). Resolve the re-exports through the existing PEP 562 module
__getattr__ (same pattern as _PROVIDER_MODELS) so each module loads on
first actual use. Internal call sites go through a _self() helper because
bare-name lookups do not trigger __getattr__; _self() imports sys locally
since update tests patch hermes_cli.main.sys. The
_warn_stale_dashboard_processes back-compat alias moves into the lazy
surface, and the sessions argparse dispatch defers sessions_cmd to call
time. Monkeypatching hermes_cli.main.<name> keeps working: a patch sets a
real module attribute, which shadows __getattr__.
Measured (Windows 11, Python 3.11, median of 7 warm runs):
import hermes_cli.main 253ms -> 196ms (-22%).
(cherry picked from commit cad1083b71635f98815b698d69a18f7f58e15517)
Read the durable display transcript when creating a branch instead of copying the compacted model projection. Hydrate the Desktop branch boundary from persisted history, avoid stale whole-chat counts, and seed the new tile from the backend snapshot. Add regression coverage for compacted histories, visible-message counts, selected prefixes, and hydration races.
(cherry picked from commit c3d2d759ae104395adde215b2e293ccf8e895684)
CLI-origin sessions have no session_key; the Desktop history-truncation
path called replace_messages(session["session_key"], ...) with None,
whose reinsert violated the messages.session_id FK -> "FOREIGN KEY
constraint failed" -> "Restore failed" on resume. Key the persist off
the durable session id instead.
Extracted from PR #81904 (the scope=compacted API half was superseded by
include_compacted, #86595). The PR's companion change defaulting
session_key to the session id at insert time is deliberately NOT taken:
main treats a non-NULL session_key as "this is a gateway session"
(list_gateway_sessions, orphan gateway-session repair), so the default
would misclassify every CLI session.
(extracted from PR #81904, commit cef9b9b27d)
When the 1h provider_models_cache.json TTL lapses, the model picker
serially fetches /v1/models for each authenticated provider. With 10+
providers this stacks to 15-30s of blocking before the picker renders.
Add a parallel prefetch step before the serial picker build loops:
- _collect_authed_provider_slugs(): lightweight credential pre-scan
that mirrors sections 1/2/2b without fetching model lists
- _prefetch_provider_models_parallel(): ThreadPoolExecutor-based
concurrent fetch of stale/missing cache entries (max 8 workers)
- update_provider_cache_entry(): thread-safe single-entry cache writer
with threading.Lock to prevent concurrent write races
Guardrails:
- Skipped when <=3 authed providers (overhead not worth it)
- Skipped when refresh=True (serial path force-refreshes)
- Exception-isolated (falls back to serial path on any failure)
- No behavioral change (same model lists, same picker output)
Closes#80413
(cherry picked from commit 89dddd6cb5d53d73278e0518c375fb5b878e5c6b)
_create_request_anthropic_client() built a fresh anthropic.Anthropic
client (and httpx pool) on every single LLM call, and
_close_request_anthropic_client() always fully closed it right after
- unlike the OpenAI-wire path, which caches and reuses one warm
client across sequential calls via a single-slot cache keyed on the
effective client kwargs.
Add the same single-slot cache to the Anthropic-wire path: keyed on
credentials, base URL/Bedrock region, per-model timeout, and the
1M-beta flag; in_use guards concurrent calls from sharing one pool's
close/abort lifecycle; poisoned marks a cross-thread-aborted slot so
the owner-thread close discards it; reuse only on request_complete /
stream_request_complete (the same _REQUEST_CLIENT_REUSE_REASONS the
OpenAI path already uses). Wires a teardown hook into
release_clients()/close() mirroring _close_cached_request_openai_client.
Fixes #HPA-02
(cherry picked from commit 37f90df15593e6ded0390f827f8e5604bf0acc86)
Every search route in _search_messages_impl (FTS, CJK bigram, trigram,
LIKE fallback, rebuild-gap supplement) selected m.content, then the
result tail popped it unread. On DBs with multi-MB tool rows, each
search read and materialized up to `limit` full rows only to discard
them. Snippets come from snippet()/substr() in SQL and the context
window is re-fetched by id, so no code path ever read the column.
Drop the column from all six SELECT lists. Returned dicts are
unchanged: content was never part of the public result (the pop ran
before return), and tests/test_hermes_state.py already documents that
contract.
(cherry picked from commit d0c3af167e7dd4eb18e1bea29ba107a94911ea24)
Hermes Console registers `checkpoints prune`, `clear` and `clear-legacy` as
mutating, so it takes a console-level confirmation before dispatching any of
them. `_apply_confirmed_defaults` then exists to keep the CLI layer from
asking a second time — its docstring says so — but it only force-defaults
`clear` and `clear-legacy`. `prune` was left out, even though `cmd_prune`
gates its orphan preview on the identical `not args.force` shape.
`_capture_output` redirects stdout and stderr but never stdin, so the
unskipped `_confirm()` call hits `input()` with no terminal behind it:
`EOFError` propagates into `_confirm`, which returns False, and `cmd_prune`
prints "Aborted." and returns 1. The console turns that non-zero exit into a
ConsoleCommandError, so `checkpoints prune` fails outright for any user who
has at least one orphan checkpoint project — after that user already
confirmed. When the server does happen to inherit a foreground terminal, the
same call instead blocks a console worker thread and eats the operator's
keystrokes.
Forcing the flag is the documented behavior here rather than a weakening of
the recent orphan-allowlist hardening. `orphan_allowlist` binds a deletion to
the identities shown in the preview, guarding the window where a workdir
disappears while the command waits on `input()`. Under the console there is
no preview and no wait, which is exactly the `--force` case the comment on
`cmd_prune` describes as "no restriction".
Migrates 8 of 12 native dialog call sites in the kanban dashboard plugin
to the SDK's ConfirmDialog primitive (added in PR #50550):
- moveTask, moveSelected, applyBulk, deleteTask, deleteSelected,
archiveBoard, removeAttachment, doPatch
The 4 remaining carve-outs (window.prompt for completion summary,
window.alert for missing summary, cli_hint clipboard fallback) are
documented inline — the host's ConfirmDialog hardcodes onClick → unmount,
preventing the keep-open-across-validation behavior the completion-summary
form needs. Followup: upstream a `disabled` prop to ConfirmDialog and
rebuild the completion body using host Dialog components.
New architecture:
- useKanbanDialogs(t) — Promise-based dialog state machine at
KanbanPage scope. request({kind, ...}) returns {confirmed, summary?}.
- KanbanDialog component — renders ConfirmDialog from SDK for kind=confirm.
- performMoveTask(taskId, newStatus, count, summary) — extracted shared
dispatch path for single + bulk moves (optimistic UI + PATCH/POST +
error recovery).
- requestDialog prop threading — KanbanPage → BoardSwitcher,
TaskDrawer → TaskDetail → doPatch/AttachmentsSection. Every call
site has a defensive fallback to window.confirm if the prop is
missing (verified by test_dashboard_done_actions_prompt_for_completion_summary
counting the cancel guards + destructive:true markers in the bundle).
New host i18n keys (web/src/i18n/en.ts + types.ts):
- kanban.confirmDoneMany / confirmArchiveMany / confirmBlockedMany
- kanban.trash.confirmTitle / confirmManyTitle
Tests:
- Replaced bundle-string-only completion-summary test with behavioral
coverage: bundle cancel-guard count + destructive marker count, plus
backend tests that confirm cancel preserves old status and confirm
dispatches the expected PATCH/DELETE body.
- Removed the SDK_CONTRACT_VERSION snapshot test from
web/src/plugins/registry.test.ts (forbidden by AGENTS.md
"Don't write change-detector tests"; the two remaining tests in that
file already cover the new SDK surface behaviorally).
Closes#50547 (consumers of #50550).
Cross-vendor re-review: Gemini 3.5 Flash + GPT-OSS 120B (both SHOULD-FIX,
no remaining BLOCKERs after these fixes).
The salvaged #76716 adds git_metadata_generation to sessions (54 -> 55
columns). Update the synthetic-rebuild test's pinned widths and row
builders to the new current layout.
Creating a project whose resolved primary path already belongs to a
non-archived project now raises a clear ValueError naming the existing
project (create_project) — duplicated projects each seeded an identical
copy of the repo subtree, multiplying the duplicate-lane bug per copy.
The agent-facing project_create tool is idempotent instead: it re-activates
the existing project rather than erroring. allow_duplicate_path=True keeps
deliberate duplicates possible. Also updates the legacy non-git lane-id
expectation to the branch-style id introduced for #53329.
_build_repos emptied each lane's sessions array for the overview (hydrate=False)
payload BEFORE _sort_lanes ran. _lane_sort_key derives a lane's activity from
max(_session_time(s) for s in group["sessions"]), so with the rows already gone
every non-trunk lane scored activity=0.0 and the sort key
(is_trunk, is_kanban, -activity, label) collapsed to alphabetical-by-label. The
documented intent — branches and linked worktrees sort by most-recent activity,
then label — was silently defeated on the projects.tree RPC that feeds the
desktop sidebar overview, while the drill-in path (hydrate=True) kept the rows
and sorted correctly. Any repo with two or more non-trunk lanes showed a
different order in the overview than when opened.
Move the session-clearing to after _sort_lanes/_disambiguate_labels so the sort
reads real recency. Lane counts are still captured before clearing, so
sessionCount and the slim overview payload are unchanged — only the order is
fixed, and the overview now matches the drill-in.
_place_by_heuristic used the raw path as the lane key for non-git
project folders, while the desktop overlay independently computed
::branch::main for the same session (since git_branch was null).
The ID mismatch caused duplicate lanes — one from the backend with
the folder name, one from the overlay labeled 'main'.
Use _branch_lane_id(path, DEFAULT_BRANCH_LABEL) so the backend's
lane key matches the overlay's expected ::branch::main scheme,
eliminating the duplicate lane.
A gateway active-session lease is acquired against the root HERMES_HOME,
but release_active_session()/transfer_active_session() re-resolved the
registry path from the *current* HERMES_HOME. Under native multiplex a
routed turn runs agent cleanup inside _profile_runtime_scope, so the
release looked under the named profile while the root entry stayed
alive — after max_concurrent_sessions routed turns every new session was
rejected with 'Hermes is at the active session limit' (#85431).
Pin state/lock paths on the lease at acquisition time and prefer them on
release and transfer. Fixes#85431.
Older nemo-relay bindings reject metadata= on scope.pop, which aborted
turn finalization and left scopes open. Filter kwargs to what the live
binding accepts so close paths can complete.
generate_base_drafts hardens every base draft to a transparent PNG with
_harden_transparency. When the provider returned a non-PNG file (webp,
jpg, or gif), the hardened PNG is saved under a new path and the original
draft is left in cache/images. Nothing prunes that directory outside the
gateway housekeeping loop, so a CLI or desktop draft round leaks one
original per non-PNG draft.
Remove the original after a successful hardening when the output path
differs from the input. PNG inputs (including mixed-case suffixes like
.PNG) are hardened in place so a case-insensitive filesystem cannot
treat with_suffix(".png") as a different file and unlink the output.