fix(file_tools): strip doubled base prefix for cwd-shaped relative paths (#67185) by tusharui · Pull Request #67318 · NousResearch/hermes-agent

tusharui · 2026-07-19T04:47:11Z

Problem

When a model emits a relative path that textually mirrors the working directory (e.g. home/user/dev/notes/file.md — an absolute path missing its leading /), _resolve_path_for_task silently creates a doubled path like /home/user/dev/home/user/dev/notes/file.md instead of failing or writing to the intended location.

Fix

Add _strip_doubled_base_prefix() to detect and strip the duplicated prefix using two strategies (common prefix + suffix-matching) that work on both POSIX and Windows. Integrated into _resolve_path_for_task for both path branches. Also adds a heuristic warning in _path_resolution_warning for bare-absolute paths whose first segment matches known root directories (home, Users, tmp, var, etc.).

Files changed

tools/file_tools.py — new helper + integration into resolve + heuristic warning (~60 LOC)
tests/tools/test_file_tools_doubled_path.py — new test file (17 tests)

Tests

_strip_doubled_base_prefix: full strip, partial strip, no-strip for normal/absolute/single-segment paths, POSIX path
_resolve_path_for_task: cwd-shaped path, deep nesting, normal relative, absolute, tilde
_path_resolution_warning: bare-absolute home/... and tmp/... warn, normal relative/absolute don't

Fixes #67185

…ths (NousResearch#67185) When a model emits a relative path that textually mirrors the working directory (e.g. home/user/dev/notes/file.md — an absolute path missing its leading /), _resolve_path_for_task silently creates a doubled path like /home/user/dev/home/user/dev/notes/file.md instead of failing. Add _strip_doubled_base_prefix() to detect and strip the duplicated prefix using two strategies (common prefix + suffix-matching) that work on both POSIX and Windows. Integrate into _resolve_path_for_task for both path branches. Also add heuristic warning in _path_resolution_warning for bare-absolute paths whose first segment matches known root directories (home, Users, tmp, var, etc.). Adds 17 tests covering full/partial strip, no-strip for normal paths, integration with _resolve_path_for_task, and bare-absolute warnings.

alt-glitch · 2026-07-19T04:54:41Z

This was generated by AI during triage.

Related to #67185 and the open #67216/#67220/#67232 cluster. This branch auto-corrects paths, whereas the others use warning-only or different rewrite contracts; a maintainer choice is needed.

teknium1

Thanks for the focused reproduction and regression coverage. The underlying bug is present on current main: tools/file_tools.py:397-398 joins a non-absolute path to the task base, while :428-429 suppresses the existing warning for an in-workspace doubled target.

Problems

tools/file_tools.py:419-431 treats a two-segment suffix match as a duplicated base. With base /home/user/dev, legitimate relative user/dev/file.md is silently redirected to /home/user/dev/file.md; the PR's tests/tools/test_file_tools_doubled_path.py:43-47 codifies analogous partial stripping.
The container branch returns at tools/file_tools.py:456-457 before the new host-branch calls at :471 and :479, so it does not fix the same input for container backends.
tools/file_tools.py:512-520 reports that correction occurred for any home/tmp-style first segment even when _strip_doubled_base_prefix() returned None.

Suggested changes

Prefer a warning-only, full-root-replay detector rather than rewriting ambiguous relative paths; preserve valid workspace subtrees.
Cover container and explicit Windows resolution flows, plus suffix-only and root-named relative false positives.

Automated hermes-sweeper review.

teknium1 · 2026-07-19T04:58:01Z

+            ):
+                overlap = max(overlap, suffix_len)
+                break
+        if overlap < 2:


A two-segment overlap is not sufficient evidence of a duplicated cwd. With base /home/user/dev, legitimate relative user/dev/file.md matches the base suffix and is rewritten to /home/user/dev/file.md, silently dropping user/dev. Require a full normalized replay or retain the path and emit a warning.

Acknowledged - 2-segment suffix match is too ambiguous. I'm replacing this with a warning-only full-root-replay detector that only flags when the entire normalized base path is repeated in the resolved result, not partial suffix matches

teknium1 · 2026-07-19T04:58:01Z

-    return resolved.resolve()
+    base = _resolve_base_dir(task_id, container_paths=False)
+    resolved = (base / p).resolve()
+    stripped = _strip_doubled_base_prefix(base.resolve(), resolved)


This integration only covers the POSIX host branch. The container branch above returns at line 457 before reaching this call, so the same cwd-shaped relative input remains doubled for Docker/container backends.

the fix only covers the POSIX host branch. I'll restructure so the detection runs after all three resolution paths (container, Windows, POSIX) instead of only the last one.

teknium1 · 2026-07-19T04:58:01Z

            return None
+        # Heuristic: detect bare-absolute paths (missing leading /).
+        first_seg = Path(filepath).parts[0] if Path(filepath).parts else ""
+        if first_seg in _BARE_ABSOLUTE_ROOT_SEGS:


This warning claims the doubled prefix was auto-stripped without checking whether stripping happened. A normal relative tmp/cache.txt under a base with no matching tmp prefix remains unchanged but receives the correction warning.

the warning fires without checking if stripping actually happened. I'm removing this first-segment heuristic. The new approach only warns when the full-root replay detector matches, so no false positives for legitimate relative paths like tmp/cache.txt.

GottZ · 2026-08-03T16:25:22Z

This was generated by AI during triage.

Summary

Eight PRs address #67185: #67216 and #67232 preserve ambiguous relative-path semantics and return a warning for an exact workspace-prefix replay, while the other six use path-rewrite variants with narrower coverage, broader matching, or warning-propagation gaps. The visible diffs show that #67216 and #67232 share the same production mechanism, with #67216 adding a warning-precedence regression test.

Related pull requests

fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216 best fix — (+345/-7) — n/a: Adds a warning-only, normalized full-base-replay detector and returns the diagnosis through write and patch results without rewriting the supplied path; tests cover POSIX, containers, symlink roots, normalization, Windows UNC, and warning precedence. This supports its high-salvageability keep_open verdict and the recorded best-fix selection.
fix(file_tools): strip duplicated prefix for cwd-shaped relative paths (#67185) #67218 [closed] partial — (+0/-0) — n/a: This closed PR remains relevant as the earlier POSIX prefix-stripping approach, but the documented implementation silently rewrote ambiguous paths, omitted native Windows, and bundled unrelated QQBot work. Its COMMENTED keep_open review required splitting that work and replacing the rewrite with a cross-platform surfaced warning, which fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216 provides.
fix(write_file): detect and correct cwd-shaped relative paths (#67185) #67220 partial — (+58/-1) — n/a: The exact POSIX echo branch addresses the reported cause, but the additional root-directory heuristic can redirect legitimate relative paths such as lib/utils.py or tmp/foo to filesystem-root locations, while native Windows and regression tests are absent. Despite its COMMENTED keep_open review, those diff-visible blockers require author rework rather than retention as the preferred fix.
fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67232 fixes — (+326/-7) — n/a: Implements the same warning-only production change as fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216 with POSIX, container, symlink, normalization, write/patch, and Windows UNC coverage. Despite its high-salvageability keep_open verdict, the production hunk is reported byte-identical to fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216 and its tests omit fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216's additional warning-precedence case, making it the duplicate branch.
fix(file_tools): strip doubled base prefix for cwd-shaped relative paths (#67185) #67318 partial — (+271/-5) — n/a: Auto-strips full and partial base overlaps, so its two-segment suffix rule can silently retarget a legitimate user/dev/file.md; the container branch bypasses the correction and the broad warning can claim a correction that did not occur. Despite its COMMENTED keep_open review, these visible correctness and coverage issues require a warning-only rework or a split.
fix(file): detect absolute-shaped relative paths missing leading / #67421 partial — (+26/-0) — n/a: Rewrites paths based on generic root-like first segments without proving that they replay the active base, unconditionally changing container and Windows relative-path semantics, and adds no tests. Despite its COMMENTED keep_open review, the diff needs replacement with an exact base-prefix detector and targeted regression coverage.
fix(write_file): detect cwd-shaped relative paths missing leading slash #67426 related — (+143/-3) — n/a: The current visible diff moves detection into _resolve_path_for_task, but accepts base-suffix matches down to one segment, silently coerces matches to filesystem-root paths, leaves container and native-Windows branches without equivalent handling, and labels a resolver-only check as write-file integration. Despite its COMMENTED keep_open review and later non-contributor update claims, the visible diff still requires a full-root, warning-only rework.
fix(file_tools): strip duplicated prefix for cwd-shaped relative paths (review fixes) #67572 partial — (+128/-6) — n/a: Detects and rewrites doubled POSIX-style prefixes across resolver branches, but base.relative_to("/") and slash parsing do not implement drive-aware Windows behavior, and warnings.warn() is not returned in the tool JSON seen by the model. Despite its COMMENTED keep_open review, those concrete gaps require author action before this branch can serve as the cross-platform fix.

Duplicates

#67216 and #67232 are substantive duplicates: their tools/file_tools.py production hunks are byte-identical according to the current-head comparison, while #67216 carries the additional test_doubled_warning_precedes_out_of_workspace_check regression. The other PRs overlap on #67185 but implement materially different rewrite contracts.

Suggested consolidation

Keep #67216 open with a salvage path: retain its focused warning-only production change and broad regression coverage, consistent with the recorded best-fix selection and the maintainer-bot keep_open verdict. Close #67232 as duplicate of #67216 despite its keep_open verdict because its production hunk is byte-identical and #67216 has the extra precedence test; keep #67218 closed as superseded, and request author action on #67220, #67318, #67421, #67426, and #67572 to rebase onto the warning-only contract or split out any independently salvageable tests, explicitly addressing their respective false-positive, backend-coverage, and warning-propagation blockers.

Complex graph

flowchart LR
    classDef open fill:#dbeafe,stroke:#1d4ed8,color:#1e3a8a
    classDef merged fill:#dcfce7,stroke:#15803d,color:#14532d
    classDef closed fill:#e5e7eb,stroke:#6b7280,color:#1f2937
    classDef unverified fill:#f3f4f6,stroke:#9ca3af,color:#374151
    classDef best stroke-width:3px,stroke:#b45309
    classDef target stroke-width:3px,stroke:#4338ca
    I67185(["issue #67185 (open)"])
    P67318["PR #67318 (open)"]
    P67318 -.->|partial| I67185
    class I67185 open
    class P67318 open
    class P67318 target
    click I67185 "https://github.com/NousResearch/hermes-agent/issues/67185"
    click P67318 "https://github.com/NousResearch/hermes-agent/pull/67318"

Graph: solid arrow = fixes / best fix, dashed arrow = partial or unverified (see edge label); boxed group = PRs duplicating each other; amber border = best fix; indigo border = target; gray node = closed (state tag in the node label).

Cross-PR triage: Reviewed 8 pull requests and 1 issue in this complex. Each diff was read against this issue; Assessment working set: 74 kB of PR diffs, 20 kB of issue/PR text, 25 kB of discussion (36 comments), 9 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

teknium1 reviewed Jul 19, 2026

View reviewed changes

teknium1 added the sweeper:blast-moderate Sweeper blast radius: moderate — a subsystem or single platform label Jul 19, 2026

This was referenced Jul 19, 2026

fix(file): detect absolute-shaped relative paths missing leading / #67421

Open

fix(write_file): detect cwd-shaped relative paths missing leading slash #67426

Open

fix(file_tools): strip duplicated prefix for cwd-shaped relative paths (review fixes) #67572

Open

Sora-bluesky mentioned this pull request Jul 28, 2026

fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67216

Open

13 tasks

GottZ mentioned this pull request Jul 30, 2026

write_file: cwd-shaped relative path (absolute path missing leading /) silently resolves to doubled path #67185

Open

This was referenced Aug 3, 2026

fix(write_file): detect and correct cwd-shaped relative paths (#67185) #67220

Open

fix(tools): warn when a cwd-shaped relative path resolves to a doubled tree #67232

Closed

Navigation Menu

fix(file_tools): strip doubled base prefix for cwd-shaped relative paths (#67185) - #67318

fix(file_tools): strip doubled base prefix for cwd-shaped relative paths (#67185)#67318
tusharui wants to merge 1 commit into
NousResearch:mainfrom
tusharui:fix/write-file-doubled-path

tusharui commented Jul 19, 2026

Uh oh!

alt-glitch commented Jul 19, 2026

Uh oh!

teknium1 left a comment

Uh oh!

teknium1 Jul 19, 2026

Uh oh!

tusharui Jul 19, 2026

Uh oh!

teknium1 Jul 19, 2026

Uh oh!

tusharui Jul 19, 2026

Uh oh!

teknium1 Jul 19, 2026

Uh oh!

tusharui Jul 19, 2026

Uh oh!

GottZ commented Aug 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Conversation

tusharui commented Jul 19, 2026

Problem

Fix

Files changed

Tests

Uh oh!

alt-glitch commented Jul 19, 2026

Uh oh!

teknium1 left a comment

Choose a reason for hiding this comment

Problems

Suggested changes

Uh oh!

teknium1 Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

tusharui Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

tusharui Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

teknium1 Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

tusharui Jul 19, 2026

Choose a reason for hiding this comment

Uh oh!

GottZ commented Aug 3, 2026

Summary

Related pull requests

Duplicates

Suggested consolidation

Complex graph

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants