The read_file description added in #79781 states that PDF, legacy
Office, OpenDocument, RTF, and EPUB convert via the optional anydoc
converter, unconditionally. Conversion actually depends on the lazy
install succeeding, security.allow_lazy_installs, and the file being
readable from the Hermes host, so the schema overpromises and the model
learns to expect conversion in environments that can never provide it.
The description now says these formats convert when the optional anydoc
converter is available, and that the auto-install applies where
installs are permitted.
read_file's auto-extraction covered only the stdlib trio (.ipynb/.docx/
.xlsx). firecrawl-anydoc (MIT, Rust core, imports as `anydoc`) converts
Word, PowerPoint, Excel — including legacy .doc/.ppt/.xls — OpenDocument,
RTF, EPUB, and PDF to clean Markdown through one shared document model.
Wiring follows the footprint ladder: no new tool, no hard dependency.
- tools/read_extract.py gains an ANYDOC_EXTENSIONS set that is active
only when the converter imports; the stdlib extractors remain
authoritative for their three formats so behavior is identical with
or without the package.
- tools/lazy_deps.py adds tool.doc_extract (firecrawl-anydoc==0.1.6),
installed on first read of such a file with prompt=False so read_file
can never block. Lazy-only for now: the package's first release was
2026-08-04, inside uv's 14-day exclude-newer quarantine, so the
mirrored pyproject extra lands after it clears.
- Any anydoc ConvertError maps to ExtractionError, falling back to the
existing path/binary handling instead of erroring the tool.
Tests: real-binding suite skips cleanly when the wheel is absent
(verified: 15 passed/3 skipped without it, 18 passed with it), plus an
absent-dep contract class that pins the fallback regardless of local
install state.
write_file confirmed only SIZE (wc -c) after writing — never content.
Models compensated by re-reading files immediately after writing them
(154 verify-reads in a 400k-msg production window), and a corrupted
write (truncated pipe, backend FS oddity) could silently pass.
The write path now compares the on-disk sha256 against the intended
content (one shell call). Three outcomes:
- match -> result carries verified: true; the schema tells the model
an explicit contract: do NOT re-read to check the write landed.
- mismatch -> hard error ('The write did not persist correctly'),
mirroring patch_replace's existing post-write verification.
- backend can't hash (no sha256sum) -> flag omitted, write unaffected.
Hashes the shim-adjusted content (after CRLF/BOM preservation) so
Windows-line-ending and BOM round-trips verify correctly; surrogatepass
encoding matches the rest of the codebase's hashing of model text.
Production mining (state.db, 28.5k read_file calls in the recent
window) shows 74.3% of reads truncated — nearly all by the 500-line
default, not the char budget (22,443 line-limit vs 13 char-budget
truncations). That churn produced 12,229 redundant re-reads, and after
a truncated result the most common next move was fleeing to terminal
cat/sed (4,608 times) — the pagination contract was not trusted.
Median truncated file is 2,422 total lines, so a 2000-line default
makes 44% of today's truncated first-reads complete in one call while
the unchanged ~100K-char budget still caps worst-case result size
(same ceiling as before: 500 lines x 2000-char line cap = the same
100K). Schema max was already 2000.
Touchpoints: DEFAULT_READ_LIMIT + both read_file signatures
(file_operations.py), read_file_tool + schema text (file_tools.py),
execute_code sandbox stub docs (code_execution_tool.py), 3 tests
pinning the old default.
Simplify-pass advisory on the #25387 salvage: _read_tracker_lock is one
global lock guarding every task's read/search bookkeeping (15 sites). A
hung stat on a dead network mount inside it would stall all of them.
Check-then-recheck matches the dedup mtime pattern 30 lines below: read
the entry under the lock, stat outside, reacquire only to evict.
Review follow-ups on the #25387 salvage:
1. CRITICAL: a cached miss survived out-of-band file creation (terminal
command, external process) for the full 60s TTL — breaking the common
agent pattern 'check for file -> create it -> read it' (live-repro'd).
Serve-side existence guard: one ~free stat before serving a cached
miss; if the path now exists the entry is evicted and the real read
runs. Also fixes the search-root variant (write under a cached-missing
directory). Both mutation-checked.
2. notify_other_tool_call now clears the task's not_found entries too
(belt: the dispatcher calls it for every non-read tool).
3. Tracking parity: the record sites no longer early-return. On upstream,
error results flow through consecutive-loop detection and dedup
bookkeeping; short-circuiting skipped that and broke
TestDedupInvalidationTaskResolution when preceded by
TestSilentFileMisplacementE2E (bisected: the early return at the
read record site was the trigger). Recording is now
side-effect-identical to upstream; serving from the cache remains the
optimization. Also reuse the already-computed _resolved instead of
resolving a second time.
When read_file or search hits a non-existent path, ShellFileOperations
spawns a subprocess to stat the path and another to walk the parent
directory for "did you mean..." suggestions. A typo'd path retried 13
times (observed in the wild) costs 26 subprocess invocations + 13 ls
walks for a result we already know.
Add a per-task negative-result cache keyed by (op, resolved_path) with
a 60s TTL and a hard cap of 500 entries. On hit, return the cached
error JSON immediately and skip the subprocess + suggestion walk.
The cache is namespaced by operation ("read" vs "search") because the
two callers return different error JSON shapes ("File not found:" vs
"Path not found:"). Eviction:
* TTL (60s) — short, so a path that appears later isn't masked.
* write_file / patch on the same path — _invalidate_dedup_for_path
now also drops the negative-cache entry so a freshly-written file
is read from disk on the next call instead of returning a stale
"not found" stub.
Tests in tests/tools/test_file_tools.py cover:
* read cache hit skips the subprocess on retry
* cache is per-task (no cross-task pollution)
* successful reads do not poison the cache
* search cache hit skips the subprocess on retry
* read and search caches are namespaced (different error shapes)
* write_file invalidates the read negative cache
* TTL expiry evicts stale entries
Three lower-severity core-tool robustness fixes from a targeted audit, each
reproduced live:
1. terminal_tool did not validate non-positive timeouts. 'timeout or default'
silently coerced 0 to the config default (0 can't mean 'no timeout'), and a
negative value is truthy so it flowed into 'deadline = now + timeout' and
fired an immediate '-Ns' timeout. Reject timeout <= 0 with a clear message.
2. fuzzy_find_and_replace accepted a whitespace-only old_string, which matches
trivially (blank line / run of spaces) and mass-replaces under replace_all
or raises an opaque ambiguity error. Reject it alongside the empty check.
3. The '/private/var/' sensitive-path prefix over-blocked ALL macOS temp-file
writes: , /tmp, and /var/folders realpath into /private/var/folders
on macOS (and paths are resolved through symlinks), and /private/var/tmp is
a normal temp dir. Narrowed to the genuinely-sensitive subtrees
(/private/var/db, /private/var/root); /etc and /private/etc stay blocked.
All verified with sabotage-checked regression tests. 85 terminal/fuzzy/file
tests pass; normal timeouts, legit replacements, and /var + /boot + /etc
blocking are unaffected.
Two independent HIGH-severity correctness bugs found in a core-tools audit,
each reproduced live against current main:
1. Read-dedup was never evicted after a write on non-default tasks.
_invalidate_dedup_for_path looked up the read-tracker under the correct
task_id but resolved the path with _resolve_path(filepath) — which
DEFAULTS task_id='default'. The dedup cache is keyed by the task-resolved
absolute path, so for any task whose workspace cwd differs from the process
cwd (every -w worktree / Desktop / ACP session using relative paths) the
computed key never matched and the stale entry was never removed. A
read_file after a write_file/patch could then return the OLD content stub
when mtime coincided. Fix: pass task_id through.
2. A per-command workdir override permanently hijacked the session cwd.
The post-command dual-write unconditionally recorded env.cwd (stamped to
the transient workdir) into the durable session-cwd store, so every later
command that omitted workdir inherited the one-off directory — contradicting
the documented 'Working directory for this command' contract. Fix: skip the
session-cwd record when workdir was explicitly supplied.
Both verified with sabotage-checked regression tests (fail without the fix).
patch_tool resolved V4A header paths against the task workspace for
locking, staleness, and reporting, but handed the original (often
relative) patch text to file_ops.patch_v4a — which re-resolved headers
against the backend env's own cwd. When the two diverge (the git-worktree
cwd bug), a relative header landed in a different directory than
everything the tool locked and reported: a silent wrong-file write.
Rewrite Update/Add/Delete/Move File headers to the resolved absolute
paths before apply, only for host-filesystem backends (container/remote
namespaces keep their own paths). Header patterns mirror patch_parser
(no-space ***Update File: form) and cover Move File: src -> dst.
Salvage of #53176 by @necoweb3, reimplemented onto current main (the
original branch predates the sensitive-path/Move-header extraction and
per-path locking now in patch_tool).
Co-authored-by: necoweb3 <sswdarius@gmail.com>
The per-session record store is now the ONLY cwd mechanism. Deleted:
- env.cwd_owner stamping + prev_owner threading (terminal_tool): the
shared env no longer carries ownership metadata at all
- _resolve_command_cwd's env/prev_owner params: resolution is
workdir > session record > config/override default
- file_tools._live_cwd_if_owned + _get_live_tracking_cwd: path
resolution never consults the shared env's live cwd
- file_tools._last_known_cwd + _remember_last_known_cwd +
_last_known_cwd_for: the #26211 preserved-anchor registry is
subsumed by the session record, which never lived on the env and
therefore cannot be lost to env cleanup. The _get_file_ops
stale-cache rescue now writes the record instead.
- env recreation (both _get_file_ops and terminal_tool) seeds the
fresh env from override > session record > config
Why no transition fallback: the legacy state was process-local and
in-memory exactly like the record store — after a restart both start
empty, and within a running process every legacy write site has been
dual-writing the record since step 1. There is no populated-legacy/
empty-record state to fall back for.
Tests updated to drive the record store instead of the deleted
mechanism; the cross-session isolation suite now asserts the same
behavior contracts (no leak, cd isolation, #26211 persistence)
against the new architecture, plus a new "session C inherits nothing"
case that the old ownership guard could not express.
Flips the read side of the cwd rearchitecture onto the _session_cwd
store introduced in the previous commit.
_authoritative_workspace_root now resolves:
1. the session's own cwd record (get_session_cwd) — per-session by
construction, so one session's cd can never leak into another
session's file resolution, with no ownership heuristics at all
2. registered override (fallback for cleared/never-written records)
3. legacy shared-env live cwd + preserved anchor (transition-only,
for commands that ran before this code loaded)
4. sentinel-free absolute TERMINAL_CWD
delegate_task children get their record seeded from the parent's at
spawn: they keep starting in the parent's directory (current behavior)
but their subsequent cds stay isolated in their own record instead of
bleeding back through the shared env.
The wrong-worktree leak class is now solved structurally on this path —
there is no shared cwd for sessions to inherit. The legacy env-side
tracking (cwd_owner, _live_cwd_if_owned, _last_known_cwd) remains only
as a transition fallback and is deleted in the next step.
Git Bash hands file tools paths like /c/Users/... which Path() on native
Windows treats as relative \\c\\Users\\... under the process cwd. Reuse
local._msys_to_windows_path (extended for /cygdrive and /mnt drive forms)
in _resolve_path_for_task / _resolve_base_dir so read/write/search land on
the real drive. Container/WSL Linux paths are left untouched.
Salvages #50488 (drops unrelated desktop artifact commit); tests adapted
from #46995.
Co-authored-by: Jeff Watts <186512915+lEWFkRAD@users.noreply.github.com>
Co-authored-by: LeonSGP43 <cine.dreamer.one@gmail.com>
When a single line exceeds the entire char budget, its tail is
unreachable via offset pagination (offsets are line-granular). Tell
the model so it doesn't assume it saw the full line.
read_file previously hard-rejected any read whose formatted output exceeded
the ~100K char safety limit, returning an error with zero content. A file
with few but very long lines (logs, wide CSV rows, minified data) sails past
the line-count limit and then trips the char guard, so the model gets nothing
and must guess a smaller limit — wasting a full round-trip.
Now the read is trimmed to the last complete line that fits the budget and
returns the partial content plus truncated_by="bytes" and a next_offset, so
the model paginates forward instead of starting over. A single line larger
than the whole budget is clamped on a code-point boundary (never empty) and
the cursor still advances. Applies at both read paths (normal + extracted
documents).
Adapted from IronClaw's Rust dual line/byte cap to hermes's Python tool-layer
char guard, which is the single uniform chokepoint over the gutter-rendered
content for every backend.
Follow-up to the salvaged toggle commit:
- file_tools.py / code_execution_tool.py: carry docker_network in their
container_config dicts so those environment-creation paths honor the
lockdown instead of silently defaulting back to bridge (the probe/exec
asymmetry class reported on #46358).
- docker.py: cross-process reuse now inspects HostConfig.NetworkMode when
docker_network=false and removes a mismatched (networked) container
before starting a fresh air-gapped one. Fails closed when inspect fails.
Default-network config never churns containers, so operators using
docker_extra_args --network=none are unaffected.
- tests: AST invariant that every container_config site carrying
docker_run_as_host_user also carries docker_network, plus three reuse
guard tests (reject bridge under lockdown / keep matching none /
no inspect when network enabled).
- docs: configuration.md gains terminal.docker_network + env var row.
auxv leaks AT_RANDOM (stack canary seed) + AT_BASE/AT_PHDR load
addresses — an ASLR oracle on par with maps. pagemap exposes
virtual->physical translation. Both slipped through the endswith
tuple alongside the maps family covered by the salvaged commit.
Adds regression coverage for auxv/pagemap and for the per-thread
/proc/<pid>/task/<tid>/<file> alias form (endswith catches both).
Follow-up on #32238, closes#34430.
PR #4609 blocked /proc/*/maps to prevent ASLR layout leakage, but the
endswith("/maps") check does not match /proc/*/smaps or
/proc/*/smaps_rollup — both expose the same virtual-address layout and
bypass the guard. /proc/*/numa_maps carries the same data with NUMA
annotations and is equally bypassed. /proc/*/mem (raw process memory)
is added as defence-in-depth; it requires address knowledge to exploit
but is blocked for consistency.
Extends the endswith tuple in _is_blocked_device_path() to cover all
four variants and adds regression assertions for all new paths to
test_proc_sensitive_pseudo_files_blocked.
Partially addresses #4427.
patch_tool extracts V4A patch paths so _check_sensitive_path can refuse
writes to /etc/*, /boot/*, etc. before they reach the low-level file ops.
The extraction regex had two gaps:
1. `*** Move File: src -> dst` was never extracted (regex only matched
Update/Add/Delete), so a Move targeting /etc/crontab skipped the
pre-check and fell back on the narrower file_operations deny list.
2. The regex required `\\s+` after `***` but patch_parser uses `\\s*`, so
`***Update File: /etc/hosts` (no space) parsed + applied while
skipping the check.
Loosen the leading whitespace to \\s* and add a Move regex that checks
both endpoints. Move endpoints also run through the same '..' traversal
rejection as the other V4A headers (closes the sibling gap on current
main, which gained that traversal guard after this PR was opened).
When security.redact_secrets is on (default), read_file/search_files/cat
applied redact_sensitive_text(code_file=True) to file content, which still
ran prefix masking. An API key in config.yaml (ghp_..., sk-..., xai-..., etc.)
came back as a head/tail mask like `ghp_S1...Pn2T` — a plausible-looking
truncated key. When an agent read that and wrote it back to config, the masked
value replaced the real credential, silently breaking auth (401). Production
evidence: a config.yaml found containing the exact 13-char masked GitHub PAT.
The two community PRs (#35529, #35534) fixed the corruption by NOT redacting
prefixes for config reads — but that exposes the user's real keys to the agent
context, model, and logs (a security regression). This takes the safer route:
keep redacting, but for file content emit a NON-REUSABLE sentinel.
- New `_mask_token_nonreusable`: prefix secrets -> `«redacted:ghp_…»` (vendor
label preserved for debuggability; zero secret bytes; angle-bracket/ellipsis
wrapper is syntactically invalid as a token so it can't be mistaken for or
written back as a usable key).
- New `redact_sensitive_text(file_read=True)` routes prefix matches through it
(implies code_file=True). Default/log/display mode is UNCHANGED — `_mask_token`
still keeps head/tail (fine for logs, never written back).
- Wired the 3 file_tools.py call sites (read_file / search_files / cat) to
file_read=True.
Fixes both the corruption AND avoids the secret-exposure of the un-redact
approach. 6 new tests (sentinel shape, no-leak, not-a-plausible-key, default
mode unchanged, file_read implies code_file, sk- prefix); 88 redact tests pass;
mutation-verified (reverting to the old mask fails the sentinel/leak tests).
Co-authored-by: liuhao1024 <sunsky.lau@gmail.com>
Co-authored-by: adammatski1972 <289282750+adammatski1972@users.noreply.github.com>
Closes#35519. Supersedes #35529, #35534.
The durable _last_known_cwd anchor is keyed by the shared 'default' container,
so a non-owning worktree session could inherit the owning session's cwd through
it — breaking the wrong-worktree-routing fix (test_file_tools_cwd_resolution::
test_resolution_routes_to_resolving_sessions_worktree).
Reorder _authoritative_workspace_root so the session-specific registered cwd
override (keyed by raw session id) is checked BEFORE the shared-container
_last_known_cwd fallback. A non-owning session now resolves into its own
registered worktree; the durable anchor only fills in when there's no
session-specific override (the #26211 single-session case). Adds a regression
test covering the owner-mirrors-then-other-session-resolves interaction.
Belt-and-suspenders on top of the cherry-picked cwd-preservation fix:
- Proactively mirror every live terminal cwd into _last_known_cwd on each
successful read, so the durable anchor survives even when the cleanup
thread pops both _file_ops_cache and _active_environments before
_get_file_ops' stale-cache save branch can fire.
- Fall back to _last_known_cwd in _authoritative_workspace_root. write_file_tool
resolves the path (via _resolve_path_for_task) BEFORE _get_file_ops rebuilds
the env, so restoring only the rebuilt env's cwd was insufficient — the
resolution that decides where the file lands runs first. This closes that gap.
The local env's persisted _cwd_file can't serve this role: it's keyed by a
random per-session uuid and deleted on cleanup (the same cleanup that triggers
the bug). The in-memory _last_known_cwd registry is the durable anchor instead.
Adds a real-IO E2E regression (TestSilentFileMisplacementE2E) exercising the
actual write_file_tool path after env cleanup.
Root cause: when the terminal environment (`_active_environments` entry) is
cleaned up and re-created during a long conversation, the new environment
always starts with the default config CWD (typically `~/.hermes/hermes-agent`)
instead of preserving the user's last-known working directory. Subsequent
relative-path writes (`write_file`, `execute_code`, shell commands) silently
land in the default CWD, making files appear to be "created but absent."
Fix: add `_last_known_cwd` dict that preserves the old environment's CWD
before the stale cache entry is invalidated. When a new environment is
created for the same task_id, we check `_last_known_cwd` first and use the
preserved CWD instead of the config default.
Changes:
- tools/file_tools.py: add `_last_known_cwd` dict, save CWD before stale
cache invalidation, restore CWD on env recreation
- tests/tools/test_file_tools.py: add `TestLastKnownCwd` with 2 tests
verifying CWD preservation and fallback behavior
Fixes#26211
File tools (read_file, write_file, patch, list_directory, etc.) used
os.path.expanduser() which reads the gateway process HOME env var.
In Docker/systemd/s6 deployments where the gateway HOME differs from
interactive sessions, tilde expanded to the wrong directory.
Add _expand_tilde() helper that delegates to get_subprocess_home() when
available, falling back to os.path.expanduser(). Replace all 9
expanduser() call sites in file_tools.py with _expand_tilde().
The read_file device guard now walks symlink hops before the file operation
layer, but that hop walk still interpreted relative paths against the Python
process cwd. In sessions where TERMINAL_CWD points at the task workspace, a
relative workspace symlink to a blocked alias such as /dev/../dev/stdin could
therefore miss the intermediate device target before later task-cwd resolution.
Anchor relative device checks to the task base before symlink-hop inspection so
the pre-I/O guard sees the same workspace path that read_file would otherwise
read. Absolute device paths and the existing final realpath fallback remain
unchanged.
Refs #10141
Refs #29158
Security-hardening fix for the read_file device guard, not a new sandbox
boundary. The guard already rejects direct device paths and upstream now
has a resolved-path pass for workspace symlinks to blocked devices, but
its concrete-path helper still compared the expanded path before
normalization. That leaves residual alias cases where the dangerous path
is visible before final terminal-specific resolution, for example:
1. /dev/../dev/zero and /dev/./urandom should match the blocked-device
list as concrete paths, not only after final realpath;
2. /dev/stdin-style aliases can disappear once realpath follows them
to /proc/self/fd/0 and then to a tty path;
3. a user symlink to /dev/../dev/stdin exposes the dangerous
intermediate target before final resolution, but not necessarily
after it.
Normalize expanded paths before matching and inspect each symlink hop
before falling back to realpath. This preserves the existing /proc fd and
/proc pseudo-file guards while enforcing the intended security invariant:
model-supplied read paths must not reach blocking or infinite device
streams through spelling, normalization, or symlink-hop tricks.
Classification: security hardening / residual bypass fix for the
read_file device blocklist. This is defensive code at the file-tool
boundary, but it fixes a concrete denial-of-service class tracked as
security in #10141 and #29158.
Tests:
- normalized /dev/../dev/zero and /dev/./urandom aliases
- symlink to /dev/../dev/stdin blocked before realpath
- existing symlink-to-device and regular-symlink guards still pass
Fixes#10141Fixes#29158
* feat(search_files): path-grouped lossless densification of content matches
Content-mode search_files results repeat the {path,line,content} JSON keys
and the full path string for every match. Group consecutive same-path matches
under one path header with indented '<line>: <content>' rows — lossless (every
path/line/content byte preserved), self-describing (matches_format key), and
readable by the model with no decode step.
57.8% mean token reduction on real search_files content outputs (422-output
corpus), fires on 97% of them. Gated at >=5 matches; below that the verbose
array is left untouched. Default to_dict(densify=False) is unchanged, so no
other caller is affected.
ripgrep emits matches path-ordered, so consecutive grouping never reorders
results.
* test: accept densify kwarg in _FakeSearchResult.to_dict
The search loop-detection tests stub SearchResult with a fake whose
to_dict() must mirror the real signature now that it takes densify=.
* test(search_files): edge-case losslessness battery for densification
Adversarial single-line content (colons, indentation, unicode/emoji, empty,
trailing whitespace, quotes+commas), paths with spaces, and an explicit
one-line-per-match invariant documenting the ripgrep contract the format
relies on (0/6775 real match contents contained a newline).
_configured_terminal_cwd and _registered_task_cwd_override carried a
byte-identical sentinel + expanduser + isabs validation tail. Extract it
into _sentinel_free_abs_cwd(raw) so the relative/sentinel rejection rule
lives in one place. Behaviour unchanged (the str() coercion the override
path relied on is preserved in the helper).
The session-cwd fix inserted a registered task/session cwd override step
between the live-cwd and $TERMINAL_CWD fallbacks, but three docstrings still
described the old two-step order — _resolve_base_dir's numbered list was
outright wrong. Update _authoritative_workspace_root, _resolve_base_dir, and
_path_resolution_warning to reflect the actual four-step resolution order.
No behaviour change.
The raw-key-first-then-collapsed override lookup was hand-rolled in three
places with subtly different spellings: terminal_tool's command setup, and
both file_tools._registered_task_cwd_override and _get_file_ops. Since that
exact raw-vs-collapsed invariant is what the session-cwd fix depends on,
keeping three copies invites the drift that caused the original bug.
Add terminal_tool.resolve_task_overrides(task_id) as the single source and
route all three sites through it. Behaviour is unchanged (verified
byte-equivalent across raw/collapsed/isolation/None/subagent inputs).
Completes the worktree-misroute fix from #35399, which made misroutes
visible (resolved_path) but did not prevent them: its divergence warning
only fired once a terminal command had populated the live cwd registry.
A fresh worktree session (registry still empty) with a stale TERMINAL_CWD='.'
got neither a worktree anchor nor a warning, so a relative write_file/patch
silently landed in the MAIN checkout.
Two changes in tools/file_tools.py:
- Treat sentinel TERMINAL_CWD values ('', '.', './', 'auto', 'cwd') and any
relative value as UNSET rather than a literal anchor. Previously '.' was
joined onto the process cwd, silently routing edits to wherever the process
happened to be (the main repo, in a worktree session). The gateway already
sanitizes the same set at import time; the file-tool layer now matches.
- New _authoritative_workspace_root(): prefers the live terminal cwd, else a
sentinel-free absolute TERMINAL_CWD (the worktree path cli.py/main.py set
for -w). _resolve_base_dir() and _path_resolution_warning() both use it, so
a worktree session resolves into — and warns about escaping — the worktree
from the very first write, before any cd has run.
Validation: 11 new/parametrized tests (sentinel handling, empty-registry
anchoring, early divergence warning, live-cwd precedence). 32/32 pass under
scripts/run_tests.sh. Live E2E: relative write in an empty-registry worktree
session lands in the worktree, main untouched.
* fix(file-safety): extend sandbox-mirror guard to cover inner-container path (#32049)
Brian's shape-based guard (#32213) catches paths that still carry the
full sandboxes/<backend>/<task>/home/.hermes/… prefix on the host side.
The inner-container case is not covered: when file tools execute inside
Docker the bind-mount strips that prefix, so the guard receives plain
/root/.hermes/… and passes through. The root:root ownership on the
divergent SOUL.md in #32049 confirms this is the primary failure mode.
Add a ContextVar (_CONTAINER_HERMES_MIRROR) set by DockerEnvironment
when persistent=True. classify_container_mirror_target / get_container_
mirror_warning detect any write whose resolved path falls under that
prefix, using the same warning format and cross_profile=True bypass
contract as the existing guards. Chain the new guard in
_check_cross_profile_path after the two existing detectors.
* fix(file-safety): derive Docker mirror guard from task
---------
Co-authored-by: Ben <ben@nousresearch.com>
#32049 reports that under terminal.backend: docker, write_file / patch
calls to authoritative profile state (SOUL.md, memories, etc.) land on
the sandbox-local mirror at
``<HERMES_HOME>/profiles/<name>/sandboxes/<backend>/<task>/home/.hermes/...``
— a path the host Hermes process never reads. The tool reports success,
the user sees no behavior change, and on disk two divergent copies of
SOUL.md (or any other profile file) accumulate.
The existing classify_cross_profile_target guard does not catch this:
its parts[2] check sees "sandboxes" and returns None, and the path is
in-profile from the inner-mirror perspective so even a fixed version
would not fire.
Add a parallel sandbox-mirror classifier in agent/file_safety:
* classify_sandbox_mirror_target() detects the
``…/sandboxes/<backend>/<task>/home/.hermes/…`` shape via path parts.
Detection is path-shape only — backend-agnostic, does not require
the file to exist, and works regardless of which HERMES_HOME resolves.
* get_sandbox_mirror_warning() returns a model-facing warning that
names the mirror root and the inner authoritative path the agent
likely meant.
Wire both detectors through tools/file_tools._check_cross_profile_path
so the existing write_file and v4a patch call sites pick up the new
guard with no API change. The bypass kwarg (``cross_profile=True``)
remains shared between the two guards — same "I know what I'm doing"
escape valve after explicit user direction.
This is the defense-in-depth piece of the proposal in #32049 ("any
…/sandboxes/<backend>/…/home/…hermes/… path as sandbox-mirror"). It
catches the host-side speculation case where the agent writes a literal
sandbox-mirror path. The inner-container case (where the bind mount
strips the ``sandboxes/`` prefix from the agent's path view) is out of
scope for this surgical change — that requires either a dispatch-layer
host-side check before the container handoff, or the host-side
``profile_state`` / ``soul`` tool the issue also proposes.
Soft guard, NOT a security boundary — matches the existing
classify_cross_profile_target contract.
Co-authored-by: briandevans <252620095+briandevans@users.noreply.github.com>
Co-authored-by: Ben Barclay <ben@nousresearch.com>
Relative paths in write_file/patch could resolve against the agent PROCESS cwd
instead of the terminal's working directory. In a git-worktree session with a
stale TERMINAL_CWD='.' (a relative base), early edits silently landed in the
MAIN checkout, verified there, and reported success — while the agent inspected
the worktree and saw nothing, misreading it as the patch tool no-op'ing.
- _resolve_base_dir(): resolution base is now ALWAYS absolute. A relative
TERMINAL_CWD is anchored to the process cwd once, deterministically, instead
of being left to resolve()-time cwd. Live terminal cwd stays authoritative.
- write_file/patch pass the resolved absolute path to the shell FileOps layer
so the tool layer and shell layer can't disagree about which file is edited.
- Responses now report the absolute resolved_path and files_modified, so a
wrong-cwd mismatch is visible on the first call.
- _path_resolution_warning(): emits a _warning when a relative path resolves
OUTSIDE the live terminal cwd (e.g. a worktree session writing into main).
Validation: 11 new unit tests + 43 live E2E assertions (worktree routing,
mid-session cd, V4A patches, divergence warning, absolute paths, consecutive
patches); 466 existing file/path/terminal tests green.
* remove Vercel AI Gateway provider and Vercel Sandbox terminal backend
Both Vercel-hosted integrations are removed end-to-end. Users on the AI
Gateway should switch to OpenRouter or one of the other aggregators
(Nous Portal, Kilo Code). Users on the Vercel Sandbox backend should
switch to Docker, Modal, Daytona, or SSH.
What's removed:
- `plugins/model-providers/ai-gateway/` provider plugin
- `hermes_cli/vercel_auth.py` Vercel-Sandbox auth helper
- `tools/environments/vercel_sandbox.py` terminal backend
- `ai-gateway` provider wiring across auth, doctor, setup, models,
config, status, providers, main, web_server, model_normalize, dump
- `vercel_sandbox` backend wiring across terminal_tool, file_tools,
code_execution_tool, file_operations, approval, skills_tool,
environments/local, credential_files, lazy_deps, prompt_builder,
cli, gateway/run
- `AI_GATEWAY_BASE_URL` constant, `_AI_GATEWAY_HEADERS` auxiliary-client
header set, run_agent base-URL header/reasoning special-cases
- `[vercel]` pyproject extra and `vercel`/`vercel-workers` from uv.lock
- env vars: `AI_GATEWAY_API_KEY`, `AI_GATEWAY_BASE_URL`, `VERCEL_TOKEN`,
`VERCEL_PROJECT_ID`, `VERCEL_TEAM_ID`, `VERCEL_OIDC_TOKEN`,
`TERMINAL_VERCEL_RUNTIME`
- Tests: deletes test_ai_gateway_models.py and
test_vercel_sandbox_environment.py; scrubs references across 23
surviving test files (no entire tests deleted unless they were
dedicated to AI Gateway / Sandbox)
- Docs: provider tables, env-var reference, setup guides, security
notes, tool config, terminal-backend tables — English plus zh-Hans
i18n parity
- `hermes-agent` skill: provider table entry and remote-backend list
What stays (intentional):
- `popular-web-designs/templates/vercel.md` — CSS design reference,
unrelated to Vercel-the-AI-product
- `x-vercel-id` in `stream_diag.py` headers — generic Vercel CDN
response header, useful diag signal on any Vercel-hosted endpoint
- `vercel-labs/agent-browser` URL in browser config — lightpanda
browser project, different OSS effort
- `userStories.json` historical contributor entry mentioning Vercel
Sandbox — archive, not active docs
Validation:
- 1153 tests in the 22 targeted files pass (`scripts/run_tests.sh`)
- Full repo `py_compile` clean
- Live import of every touched module + invariant check (no
`ai-gateway` in `PROVIDER_REGISTRY`, no `_AI_GATEWAY_HEADERS`, no
`vercel_sandbox` in `_REMOTE_TERMINAL_BACKENDS`)
* test: convert profile-count check from change-detector to invariant
The hardcoded "== 34" assertion broke when ai-gateway was removed.
Per AGENTS.md change-detector-test guidance, assert the relationship
(registry count >= number of plugin dirs) instead of a literal count.
Counts shift when providers are added/removed; that's expected.
Adolanium
Teknium
kshitij
Kent
dsad
ethernet
Brooklyn Nicholson
Gille
AhmetArif0
zapabob
0xbyt4
Ben Barclay
zccyman
Tranquil-Flow
Eugeniusz Gilewski
Brandon Zarnitz
Sworntech-dev
Andres Sommerhoff
whyhkzk
Brian D. Evans
sbw2025