The helper docstring restated the inline why-comments (exact probe string,
backend $HOME tilde fallback, which statuses prove absence). Keep only the
contract callers need: three return values and that anything not proven
absent is "unavailable" and must fail closed. Inline comments stay.
_probe_regular_file returns "bad_size" only after `[ -f ]` succeeded and
`wc -c` output was unparseable, so a regular file IS present on the
execution target. Treat it as "exists" (the precise existing-binary
refusal) instead of the generic "unavailable" retry message; both refuse,
but the retry hint is wrong for a file that is known to exist. Mirrors
how the other _probe_regular_file callers treat bad_size.
Idea and the original report/Docker reproduction come from #122663, the
first submitted fix for this bug.
Fixes#122662
Co-authored-by: liuzikaii <2319582736@qq.com>
_check_binary_document_write stat'ed the controller host only, so a binary
(.sqlite/.pdf/...) that existed solely in the task's execution target
(Docker/SSH/... namespace) was treated as a new file and destroyed by a
plain-text write_file/patch that even reported verified:true (#122662).
The existence decision now goes through one tri-state helper
(_target_regular_file_state: exists/absent/unavailable) that asks the LIVE
file-ops layer - the same backend the write executes on. Host-backed envs
keep today's Path.is_file semantics (OSError -> proceed); other backends are
probed via ShellFileOperations._probe_regular_file, whose missing/not_regular
answers prove absence while every other status fails closed with a retry
message. Locality comes from the live environment object
(_file_ops_uses_host_paths), never env_type strings or class-name hint tables
(VercelSandboxEnvironment is unclassified there). The PDF and generic binary
refusal messages are unchanged verbatim; opaque-document and SQLite-sidecar
unconditional refusals are untouched.
_stale_overwrite_blocker keeps its host-only probe on purpose: remote reads
never record a full_write_baseline (version stability requires host metadata),
so converting it would refuse every remote overwrite of a file the task fully
read. Tracked as follow-up.
(cherry picked from commit 8b5eb26d57d7ef7d1d975de0ef5e1e8abd9a513d)
Follow-up to the ssh path fix salvaged from #121686.
- Read the ssh anchor raw (session record, override, TERMINAL_CWD): the
shared workspace-root helper expands ~ on the Hermes host, so
TERMINAL_CWD='~/proj' still resolved into the container home.
- Resolve ~ to the remote home the SSH environment detects at connect,
bringing the environment up through the file tools' own creator
(_get_file_ops, same cwd and cache) when none is live. A failed bring-up
is remembered per container for 30s so one call's several resolutions
don't each retry; a live environment is always used first. SSHEnvironment
now records whether the home was detected, and a guessed /home/<user>
(echo $HOME failed) is not used. Results are
absolute and stable from the first call (read tracking and staleness
checks key on them), and '..' normalizes to the real target: relative
traversal like ../../../etc/x from ~ was refused on main and slipped past
the sensitive-path guard on the PR head.
- If the remote home cannot be detected, an ssh ~-path that climbs above ~
cannot be classified; the write guard refuses it.
- ~user passes through for the remote shell instead of becoming ~/~user.
- coerce_ssh_remote_cwd maps paths under the host subprocess home onto ~/,
except when that home is the OS user's real home.
- The outside-workspace warning compares in the remote namespace (it fired
on every correct relative write when the anchor was ~).
- The backend type is looked up once per resolution again (the PR head did three per local path).
The test trim dropped the headline case: text written over an EXISTING .db. Sidecar paths return at is_sqlite_sidecar before the overwrite branch, so a mutant that drops has_binary_extension from the overwrite condition survived every remaining test. Add a third parametrize case that targets the held-open WAL database file itself and asserts the binary refusal with bytes unchanged.
Same commit: the _check_binary_document_write docstring now names SQLite sidecars (always) and every BINARY_EXTENSIONS suffix (on overwrite), and the suffix is computed once at the top instead of three times.
_has_extension_in and is_sqlite_sidecar each re-implemented the same
rfind('.') / -1 / .lower() suffix extraction, and the write guard used a
third spelling (filepath[filepath.rfind("."):]) for its refusal messages
next to os.path.splitext in the overwrite branch. Add _lower_suffix()
("" when no dot) and route both predicates through it; the write guard
now uses os.path.splitext for every message's displayed extension.
_SQLITE_EXTENSIONS was `{.db,.sqlite,.sqlite3,.db3} & BINARY_EXTENSIONS`,
which silently dropped .db3 (not a binary extension anywhere in the
codebase, so `x.db3-wal` was never a sidecar). Spell the set as the three
members that actually take effect and assert the subset invariant instead
of hiding it behind an intersection. Behaviour is unchanged.
The refactor that moved sidecar detection into binary_extensions dropped the
contributor's unconditional refusal for the db family and made every binary
extension a plain overwrite guard. That is right for the MAIN .db/.sqlite
(text fixtures named *.db exist), but a -wal/-shm/-journal path is never a
legitimate text target: a checkpointed database has no sidecar on disk, so
write_file("kanban.db-wal", text) silently dropped a garbage WAL next to a
live database.
Add is_sqlite_sidecar(path) and refuse such paths in
_check_binary_document_write regardless of is_file(). Scope the marker
stripping in _has_extension_in to SQLite suffixes so report.docx-wal no
longer counts as an opaque document. Hoist the duplicated is_pdf_path call.
Parametrize the write_file WAL test over sidecar existing/absent.
Move the .db-wal/-shm/-journal detection from a write-guard-local regex
into tools/binary_extensions._has_extension_in so has_binary_extension
(read guard AND write guard) agree: a sidecar counts as its database's
extension. read_file now refuses sidecars instead of returning lossy
text, which also gives write_file the no-baseline overwrite refusal.
Behaviour change vs the contributor's commit: creating a NEW .db /
.sqlite file via write_file stays allowed (main allows it and text
fixtures named *.db exist) — only overwriting an existing binary is
refused. The generic has_binary_extension overwrite branch is kept and
merged with the PDF branch because patch has no full-read baseline
check: on main, patch on an existing .db with a matching old_string
rewrites the header in place.
Tests folded into the existing guard test classes (one write_file, one
patch refusal on a real WAL sidecar); the PR's 12-test file and its
private-regex assertions are dropped.
The text tools' read->modify->write round-trip re-encodes binary bytes
lossily: the terminal transport decodes stdout with errors=replace, so
a patch on a SQLite database reads mojibake and writes it back, silently
destroying the file (observed in the wild: a kanban board DB lost 550k
bytes / 26 days of history this way on 2026-09-15).
read_file already refuses to display binary files (extension + byte
sniffing), but write_file/patch had no write-side equivalent, and
patch_replace reads via plain cat so the read-side guard never fires
for it.
Guard rules (extends _check_binary_document_write, the existing
binary-document write guard):
- .db/.sqlite/.sqlite3 and their -wal/-shm/-journal sidecars: ALWAYS
refused (even new-file creation) — plain text can never be a valid
database, mirroring the opaque-document rule. The sidecars need a
basename regex because os.path.splitext('x.db-wal') yields '.db-wal',
which is not in BINARY_EXTENSIONS.
- Other BINARY_EXTENSIONS (.png, .zip, ...): refused when OVERWRITING an
existing file, mirroring the existing .pdf-overwrite rule — the model
can only have read mojibake, so writing back destroys it. Creating a
NEW file with such an extension stays allowed.
Tests: new test_binary_db_write_guard.py covers the regex (main names,
sidecars, non-db names), the guard function, and the write_file/patch
tool paths against real SQLite files (integrity_check still ok, bytes
untouched). 29 new+existing guard tests pass; file-tools/patch suites
green (196 tests).
When the CLI approval callback raises, when no callback is registered on the
thread while prompt_toolkit owns the terminal, or when the input() read is
interrupted, prompt_dangerous_approval returned "deny" and the command gate
rendered "BLOCKED: User denied this command" — attributing a refusal to a
user who was never asked (#22992). #112308 fixed the gateway half of the
class (withdrawn prompts -> outcome "cancelled" with a cause); this closes
the CLI residual on the same shape.
- tools/approval_prompt.py: those three paths return an Unanswered("cancelled")
sentinel carrying the cause; MCP elicitation consent maps it to "cancel".
- tools/approval.py: the CLI gate renders "BLOCKED: <noun> was not approved: the
approval prompt could not be delivered or was not answered (<cause>)" with
outcome "cancelled" — still fail-closed, "Silence is not consent".
- tools/file_tools_write_guards.py: the protected-instruction write gate
reports the undelivered prompt instead of "was denied by the user".
- Shared metrics: "cancelled" is a counted approval outcome (contract + v2
schema) instead of falling into "unknown".
- Docs: hook `choice="cancelled"` now covers the CLI causes.
Fixes#22992
When a gateway approval wait ends without anyone answering — the parent's
delegate_task finishing and tearing the child down, a /stop, or the turn's
notifier being unregistered at turn end — the tool result said
"BLOCKED: Command denied by user" (outcome="denied", user_summary "You denied
this command"). The user never saw or answered the prompt, so the parent agent
went on reasoning about a refusal that never happened (#112026, #22992).
The action stays fail-closed (the command does not run, the model still gets
the NOT-consented stop text), but the attribution is now truthful:
- tools/approval_gateway_wait.py: `_cancel_cause()` reads the existing
per-thread interrupt-cause channel (`get_interrupt_reason()`, a trusted fixed
category — no string matching) for the interrupted state and marks a
notifier-unregister wake (event set, result None) as "the turn ended before
the prompt was answered". Both the direct and the coalesced-follower wait
return `cancelled=<cause>`; the post_approval_response hook fires
choice="cancelled" instead of "deny"/"timeout".
- tools/approval.py: a cancelled decision renders
"BLOCKED: Command approval was withdrawn before the user answered (<cause>)."
with outcome="cancelled" and its own user_summary; an explicit /deny is
untouched.
- tools/delegate_tool_child_run.py: `_signal_child_stop` publishes a fixed
tool_reason ("parent delegation ended"; the late-child mirror forwards the
parent's own category) so a child's pending approval can tell teardown from a
user /stop — previously it rode the default "explicit stop requested".
- tools/file_tools_write_guards.py / tools/approval_prompt.py: the protected
instruction-file gate and MCP elicitation consume the same key instead of
reporting "denied by the user" / "decline".
Co-authored-by: zccyman <16263913+zccyman@users.noreply.github.com>
Co-authored-by: KoNit-K <124019182+KoNit-K@users.noreply.github.com>
Claude Code v2.1.234 (Aug 17, 2026) hardened its pre-approval file
accesses to reject Windows NT-namespace (\??\) paths against the NTLM
credential-leak vector. Port the same guard into Hermes file safety:
- agent/file_safety.py: is_nt_namespace_path() / get_nt_namespace_error()
raw-string check (never resolves — resolving IS the leak trigger).
Wired as the first check in get_read_block_error() and the write
denial classifier.
- tools/file_tools.py: raw-string guard at read_file_tool entry and in
_check_sensitive_path (covers write_file_tool + patch_tool), before
the task-base join can anchor the prefix under a POSIX base dir.
- Blocks \??\, \\.\, \\?\UNC\, \\?\GLOBALROOT. Extended-length
local drive paths (\\?\C:\...) and plain UNC shares stay allowed.
- tests/agent/test_nt_namespace_guard.py: 10 blocked forms, 11 allowed
forms, no-resolve proof, tool-layer chokepoint coverage.
- docs: protected-paths table in user-guide/security.md
The stale-overwrite refusal made write_file permanently unusable for any
existing file it could not show in one read_file page: every >2000-line
(or >100K-char) page was recorded as partial, no full baseline ever
existed, and the refusal told the model to "re-read the whole file", which
the tool cannot do. Track the line ranges each task pages through per path
at one mtime; contiguous pages from line 1 to total_lines are a full read
(a new mtime between pages restarts the coverage). The same gap hit two
siblings: the extracted-document branch (.ipynb, text-authorable) returned
before any read bookkeeping, so an existing notebook could never be
overwritten; and reset_file_dedup dropped every baseline on compaction
while keeping read_timestamps, so every write after compaction was refused
even for files unchanged on disk. Baselines now survive compaction exactly
like the dedup mtime map does — only while the recorded mtime still matches.
Refusal texts no longer embed the pre-PR "Warning: … Consider re-reading"
copy inside "Refusing to overwrite", and every refusal names a recovery the
model can perform: read the remaining pages, or use patch.
Require an explicit full-file baseline before replacing existing host-visible files with write_file, and fail closed when that baseline is stale. This prevents stale conversation context from clobbering manual or external edits.\n\nRefs #65604
The file-safety salvage carried five tests and the execute_code salvage four;
fold them into the two behaviour contracts per fix: (a) the named-profile scope
exempts the root's direct files and the write lands with no prompt; a child env
sees the ACTIVE profile's home per turn; (b) negatives hold: a checkout's
.hermes/config.yaml stays gated fail-closed, a lookalike profiles/ tree exempts
nothing; a dedicated process with no override is untouched. Both red on base.
Also compress the _hermes_exempt_homes docstring (WHY only; cite #110630).
Under a named profile (``hermes -p <name>``, i.e. ``HERMES_HOME=<root>/profiles/<name>``)
the protected-instruction gate exempted only the profile dir, so the ROOT's DIRECT
files (LEDGER.md / MEMORY.md / USER.md / SOUL.md / AGENTS.md / DECISIONS.md ...) fell
through to the ``.hermes`` component rule and were treated as project-local
``<repo>/.hermes/config.yaml``. That gate always-asks and fails closed without a human
channel, so every write there was refused headless — it blocked #54 (LEDGER.md edit
could not land). Subdirectories (scripts/, cron/, data/) were unaffected, which is why
#55 sailed through and #54 did not.
``_hermes_exempt_homes()`` now returns the active profile home plus the Hermes ROOT
when that home really is a named profile (``named_profile_home`` validates: ``.hermes``
name, home markers, tombstone, or the resolved default root), so a coincidental
``profiles/`` directory never exempts its parent. ``_get_real_hermes_home()`` keeps its
per-profile semantics — the #107327 multiplex regression test pins that — and the
exemption consumes the new helper instead.
Refs: https://gitlab.yx.netease.com/agent-projects/platforms/agent-mentor/-/issues/60
The per-call home/config getters fell back to `_expand_tilde("~/.hermes...")`
when the primary resolver raised. `_expand_tilde` follows the subprocess-HOME
contract, which under host `auto` mode can be the real/default user home rather
than the active multiplex `HERMES_HOME`. So on the exception path the guards
recreated the very cross-profile authority bug the happy path fixed: beta's
`config.yaml` was compared against the default/root config (hard-block fails
open), and the protected-instruction exemption resolved against the wrong home.
Re-derive both fallbacks from the same `get_hermes_home()` key the happy path
uses (`Path(home)/config.yaml`, `realpath(home)`), and substitute no unrelated
home if the active security path cannot be established — a `None` fails closed
at the protected-instruction consumer (exemption skipped, gate runs).
Adds opposite-side regressions: forcing the primary config resolver to raise
keeps beta's own config refused (and does not spuriously protect alpha's under
beta's scope); forcing the primary home resolver to raise keeps beta's
instruction-file exemption resolved against beta.
Addresses the fallback-authority review on #107335 (thanks @andrexibiza).
(cherry picked from commit 119d88b46f745ba081f12adf3f6ebba457d95d68)
In a multiplexed gateway (`gateway.multiplex_profiles: true`) each profile turn
scopes `HERMES_HOME` through a per-turn contextvar. But
`tools/file_tools_write_guards.py` memoised the resolved home and config path in
process-global module state, filled once by whichever profile ran first. Both
the protected agent-instruction approval gate (`_get_real_hermes_home` →
exemption for a profile's own home) and the `config.yaml` hard-block
(`_get_hermes_config_resolved`) therefore became order-dependent: a later
profile's own `workspace/AGENTS.md` was gated against a *sibling* profile's home,
and — worse — its own `config.yaml` stopped matching the block, so a
prompt-injected agent could rewrite the very file the block exists to protect
(reproduced end-to-end in #107327).
Resolve both values per call instead. `get_hermes_home()` / `get_config_path()`
are contextvar-scoped, so the getters now track the active profile; the guard
already pays a `realpath` per call, so the extra cost is negligible. The two
module slots are kept purely as a test-override surface (set the slot + its
`_loaded` flag to pin a value); production leaves them unset and resolves live,
which also removes the cross-test poisoning the process memo could cause.
Adds regression coverage: both getters track the active profile after a prior
profile's scope, and the `config.yaml` hard-block fires for beta's own config
even after an alpha turn ran first.
(cherry picked from commit 36b257391da497ac31e6c440727dc57ecd584e71)
For each issue anchor present in BASE 63279301bc non-test .py and absent on HEAD, the BASE comment/docstring block was re-attached at the HEAD location of the code it explained (matched by the distinctive code line / enclosing def). Sentences already covered by an existing HEAD comment were deduped; the issue number always survives. Insert-only: no code lines changed.
