fix(agent): stop doubling the profile path in the system-prompt hint (#72894)

[jeff-mettel]

What & why

Fixes #72894.

The named-profile branch of the "Active Hermes profile" hint appended /profiles/{active_profile} to get_hermes_home(). But _resolve_active_profile_name() (agent/file_safety.py) returns a non-default name only when get_hermes_home() has already resolved under <root>/profiles/ — that's precisely how it derives the name. Both scoping mechanisms satisfy that:

• HERMES_HOME=<root>/profiles/<name> in the environment
• the multiplexer's contextvar override (gateway/run.py _profile_runtime_scope → set_hermes_home_override)

So in that branch the home is always already profile-scoped and the appended suffix always doubles. The same branch then used get_hermes_home() for the default profile's data pointers, where the root was intended — placing them inside the active profile.

Verified live

On a real 4-profile install (upstream/main @ 8defb9f):

active profile      : via
get_hermes_home()   : /Users/…/.hermes/profiles/via
get_default_hermes_root() : /Users/…/.hermes

hint says session writes    : /Users/…/.hermes/profiles/via/profiles/via/   ← doubled
hint says default data lives: /Users/…/.hermes/profiles/via/skills/          ← inside the active profile

actual session home         : /Users/…/.hermes/profiles/via/
actual default data         : /Users/…/.hermes/skills/

Every named-profile session ships a prompt naming directories that don't exist, and telling the model this profile's own skills/plugins/cron/memories belong to the default profile — the exact cross-profile confusion the hint and classify_cross_profile_target exist to prevent.

The change

Use get_hermes_home() directly as the profile home (it already is <root>/profiles/<name> whenever this branch runs), and get_default_hermes_root() — already import-safe, and documented to return <root> in profile mode — for the default-profile pointers.

The default-profile branch is untouched.

Tests

TestNamedProfileHint in tests/agent/test_system_prompt.py:

• test_session_home_is_not_doubled — asserts …/profiles/mac/ and explicitly that …/profiles/mac/profiles/mac never appears
• test_default_profile_data_points_at_the_root — root pointers, and that none land inside the active profile
• test_default_profile_hint_is_unchanged — the default branch still renders as before

13 passed  tests/agent/test_system_prompt.py

Note on the full-directory run

These three tests pass standalone and for the whole file. In a full pytest tests/agent/ run they fail — but so do two pre-existing tests in this same file on unmodified main (test_build_system_prompt_records_stable_prefix, test_coding_prompt_preserves_legacy_workspace_order), with the same signature: module-level monkeypatch.setattr on agent.system_prompt stops taking effect once other files in the directory have run, so the real identity/paths leak through.

Measured on this machine, pytest tests/agent/:

	failures
unmodified main	87
this branch	90 (the 3 above)

I could not isolate the polluting module within a reasonable timebox, and did not want to reshape unrelated tests to chase it. Flagging it rather than papering over it — if CI is green on that directory, the pollution is local to my environment and these will pass there too; if it isn't, the pre-existing pair is the better place to fix the root cause, and I'm happy to follow up separately.

Platforms

macOS 15 (Darwin 25.5.0), Python 3.11. Path handling goes through the existing hermes_constants helpers, which already handle native Windows roots.

Duplicate check

Related history, none of it landed:

• fix(agent): correct named-profile paths in system prompt hint #68117 — the closest prior attempt; closed unmerged after triage first called it a duplicate of fix(agent): use Hermes root in profile prompt #52676 and then retracted that twice ("Current main still has the named-profile nesting defect").
• fix(agent): use Hermes root in profile prompt #52676 — also closed unmerged; corrected configured-root display, not the profile-home interpretation.
• fix(system-prompt): use platform-native path for profile references (#66196) #66725 / the [Bug]: System prompt hardcodes ~/.hermes, misdirecting the agent under a non-default HERMES_HOME #66450 fix (2bae4df) — only swapped a hardcoded ~/.hermes for get_hermes_home(), which is what introduced the doubling for named profiles.

The defect is still present on 8defb9fd6, as the live output above shows.

---

Authored by an AI agent (Claude Opus 5) operating autonomously on @jeff-mettel's behalf: the defect was traced, the patch written, and the tests run and verified end-to-end before submission.

[teknium1]

Thanks for tracing this to the profile-home/root distinction. The current-main premise is confirmed: agent/system_prompt.py:423-425 appends /profiles/{active_profile} to get_hermes_home() and uses that same value for default-profile paths, while agent/file_safety.py:402-411 only returns a named profile when the active home is already under <root>/profiles.

Problems

• The new regression tests mock all three parts of the resolution chain (get_hermes_home, get_default_hermes_root, and _resolve_active_profile_name). They validate template rendering but do not execute the real profile-resolution relationship that causes the defect.

Suggested changes

• Add one temporary-HERMES_HOME integration-style prompt test for <root>/profiles/<name> without mocking those resolvers. hermes_constants.py:153-190 already defines the required root behavior, and tests/test_hermes_constants.py:47-55 covers its custom-root form.

The production hunk is narrow and its current-main context remains present; this is an automated hermes-sweeper review.

[teknium1]

Please add a companion test using a real temporary HERMES_HOME=<root>/profiles/<name> instead of mocking all three resolvers. This regression depends on their composition, not only on the prompt template.

[jeff-mettel]

Accepted and fixed in 8235d7c8b.

The mocked tests are replaced with ones that set a real HERMES_HOME under a tmp root and mock no resolver. They assert the chain before the prompt:

assert _resolve_active_profile_name() == "coder"
assert get_hermes_home() == profile_home          # <root>/profiles/coder
assert get_default_hermes_root() == root          # <root>

then that the rendered hint uses the profile home once, never <profile_home>/profiles/coder, and points default-profile data at the root. The default-profile branch is covered the same way.

This also resolves the order-dependent failures flagged in the PR description, and the cause turns out to be the same thing the review identified. Module-attribute monkeypatching stops taking effect once other files in tests/agent/ have run — which is why the mocked tests failed in a full-directory run while passing standalone, and why two pre-existing tests in this file fail the same way on unmodified main. Tests that mock nothing are immune.

Measured on tests/agent/:

	failures
unmodified main	87
this branch, previous mocked tests	90
this branch now	87, none in this file

Only the integration test fails without the production hunk:

FAILED TestNamedProfileHintIntegration::test_real_hermes_home_under_profiles_renders_correct_paths

The production change itself is unchanged from the original review.

---

Filed by an AI agent (Claude Opus 5) operating autonomously on @jeff-mettel's behalf. Failure counts were measured with and without the patch against c9de69c6d before posting.

[Maxson-dev]

Thanks — I narrowed #76155 to the non-overlapping cron side only. It now leaves the system-prompt fix and resolver-chain tests to this PR, while adding active-profile cron script guidance and create/update preflight for missing files. No overlapping files remain between the PRs.

[jeff-mettel]

Thanks @Maxson-dev — confirmed on my side: #76155 now touches only the cron subcommand, cron tools, their tests, and docs, while this PR remains agent/system_prompt.py + tests/agent/test_system_prompt.py. No overlapping files between the two, so they can land independently. The split makes sense — preflighting missing scripts at create/update time is a different defect than the doubled profile path in the prompt hint.

---

Filed by an AI agent (Claude Fable 5) operating autonomously on @jeff-mettel's behalf. File lists were compared via the GitHub API before posting.

[GottZ]

This was generated by AI during triage.

Summary

Twenty-three PRs address or reference this issue complex. The core profile-prompt lineage is #52670/#52676/#57223/#66219/#66453/#66725/#67192/#68117/#74979: #67192 removed the literal ~/.hermes but used the profile-scoped home as a root, while #74979 directly corrects the resulting doubled active-profile path and misplaced default-profile pointers; the remaining PRs cover adjacent path surfaces, symlink detection, or unrelated fixes bundled into #67192.

Related pull requests

• fix: replace hardcoded ~/.hermes with display_hermes_home() in agent-facing text #10285 [merged] related — (+14/-8) — adjacent merged reference: replaces hardcoded ~/.hermes in tool schemas and agent-facing skill/cron/TTS text via display_hermes_home(), but does not change the active-profile hint where the doubled-path defect occurs.
• fix(aux): resolve key_env: prefix in auxiliary task config #24088 [closed] fixes — (+199/-0) — superseded auxiliary-credentials fix: resolves task-level key_env/api_key_env with tests, but is unrelated to profile-path construction; it was superseded by fix(aux): resolve key_env/api_key_env in auxiliary task config #36080 and ultimately implemented through merged fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(aux): resolve key_env/api_key_env in auxiliary task config #28766 [closed] fixes — (+199/-0) — duplicate of fix(aux): resolve key_env: prefix in auxiliary task config #24088: contains the same auxiliary key_env resolution and test suite, with no system-prompt path change; it remains relevant only as duplicate history for the implementation later carried by fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(aux): resolve key_env/api_key_env in auxiliary task config #36080 [closed] best fix — (+91/-0) — implemented auxiliary-credentials reference: adds tested key_env/api_key_env fallback in _resolve_task_provider_model(), unrelated to the reported profile-path cause. Despite its keep_open review, contributor closure is supported by the diff-equivalent implementation merged through fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(system_prompt): use dynamic HERMES_HOME paths instead of hardcoded ~/.hermes #52670 [closed] fixes — (+22/-13) — superseded root-based prompt fix: replaces the literal prompt paths with get_default_hermes_root(), directly addressing custom roots without treating a profile-scoped home as the root; it was closed in favor of the more thoroughly tested fix(agent): use Hermes root in profile prompt #52676.
• fix(agent): use Hermes root in profile prompt #52676 [closed] best fix — (+174/-20) — closed reference implementation: introduces display_hermes_root() backed by get_default_hermes_root() and behavior tests for default, named, custom, and home-relative layouts. Despite its keep_open review, it was closed after 2bae4df8bb29243fe6b627cb3942aa272edcffc2; that replacement used get_hermes_home() and left the named-profile defect now targeted by fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979.
• fix(windows): show native profile path in system prompt #57223 fixes — (+254/-9) — partial Windows-specific alternative: adds a separate platform-display module and Windows prompt tests, but hardcodes %LOCALAPPDATA%\hermes and therefore still misreports an explicit Windows HERMES_HOME. Its keep_open review supports salvaging native-Windows coverage, not retaining the duplicate resolver implementation unchanged.
• fix(profiles): resolve symlinked profiles dir in active-profile detection #63982 related — (+85/-4) — related but distinct profile-safety fix: resolves a symlinked profiles/ directory consistently so active-profile detection and cross-profile classification remain correct. The keep_open review identified only Windows test portability, which the author addressed in cfd6be6fd; this does not duplicate fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979's prompt-home/root correction.
• fix: use OS-aware path for profile folder in system prompt (#66196) #66219 [closed] best fix — (+22/-10) — closed platform-root variant: derives prompt paths from get_default_hermes_root(), directly avoiding the profile-home-as-root error, but retains a prohibited source-reading test instead of behavioral Windows coverage. Its keep_open review requested that test replacement; contributor closure cited unresolved merge conflicts rather than invalidating the production approach.
• fix: replace hardcoded ~/.hermes with get_hermes_home() in system prompt (#66450) #66453 [closed] fixes — (+5/-4) — superseded source of the regression: replaces literals with get_hermes_home() but then appends /profiles/{active_profile} and locates default data beneath the same profile-scoped home. Despite its keep_open review, the review itself documents this defect; the change was salvaged into fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192 and is precisely what fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979 now corrects.
• fix(cli): sanitize UTF-16 .env without corrupting the first key #66475 [closed] best fix — (+339/-5) — merged-reference UTF sanitizer fix: safely distinguishes UTF-32 before decoding UTF-16 and adds extensive rewrite/non-mutation coverage, but is unrelated to profile paths. Despite its keep_open review, it was closed after its implementation and authorship were absorbed into merged fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(cli): quote .env values with internal whitespace in save_env_value #66483 [closed] best fix — (+210/-1) — merged-reference .env serializer fix: quotes internal whitespace and tests dotenv and shell round trips, unrelated to the prompt defect. Despite its keep_open review's Windows test-portability concern, the production fix and selected coverage were absorbed into merged fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(save_env_value): quote values with internal spaces to prevent shell-sourcing breakage (#66482) #66583 [closed] fixes — (+1/-0) — duplicate serializer fix: adds the same any(c.isspace()) predicate as fix(cli): quote .env values with internal whitespace in save_env_value #66483 but no regression test. Despite its keep_open review, the diff has no unique salvage over test-backed fix(cli): quote .env values with internal whitespace in save_env_value #66483, which was carried by fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(env): detect UTF-16 BOM before sanitizing .env to prevent silent key corruption (#66474) #66585 [closed] fixes — (+120/-2) — superseded mixed-scope fix: combines whitespace quoting with a two-byte UTF-16 detector, but misclassifies UTF-32 LE and bypasses the established atomic sanitizer path. Its keep_open review records those defects; fix(cli): sanitize UTF-16 .env without corrupting the first key #66475 supplied the safer implementation later merged through fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(config): quote .env values containing internal whitespace (#66482) #66625 [closed] fixes — (+2/-0) — duplicate serializer fix: detects spaces and tabs but adds no regression coverage. Despite its keep_open review, that review gives salvageability none and explicitly points to test-backed fix(cli): quote .env values with internal whitespace in save_env_value #66483, whose implementation landed through fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(env): detect and convert UTF-16 .env files to UTF-8 (#66474) #66627 [closed] partial — (+26/-0) — duplicate UTF-16 fix: atomically transcodes BOM-marked input but its two-byte check also matches UTF-32 LE and it adds no tests. Despite its keep_open review, the review explicitly favors fix(cli): sanitize UTF-16 .env without corrupting the first key #66475's ordered BOM handling and coverage, subsequently merged via fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(auxiliary): resolve key_env in _resolve_task_provider_model (#66641) #66655 [closed] fixes — (+7/-0) — superseded auxiliary-credentials fix: adds the missing task-level key_env/api_key_env fallback without tests, unrelated to profile paths. Despite its keep_open review, it was closed after the exact production change was salvaged with authorship into merged fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192; fix(aux): resolve key_env/api_key_env in auxiliary task config #36080 supplied the requested test cases.
• fix(config): quote .env values with internal whitespace (#66482) #66711 [closed] fixes — (+71/-0) — duplicate serializer fix: applies the same internal-whitespace predicate as fix(cli): quote .env values with internal whitespace in save_env_value #66483 with focused unit coverage, but does not affect the system prompt. It was closed as a duplicate, and the canonical serializer repair was already absorbed into fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(env_loader): transcode UTF-16 BOM .env instead of corrupting first key (#66474) #66716 [closed] partial — (+111/-1) — duplicate UTF-16 sanitizer fix: transcodes LE/BE BOM files and tests key preservation, but lacks fix(cli): sanitize UTF-16 .env without corrupting the first key #66475's necessary UTF-32-before-UTF-16 guard. It was closed as duplicate history after the broader implementation was selected for fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192.
• fix(system-prompt): use platform-native path for profile references (#66196) #66725 [closed] fixes — (+19/-9) — closed duplicate path fix: replaces prompt literals with get_hermes_home(), which is correct only when that value is a root; in a named profile it recreates the same doubled-path and misplaced-default-data defect. It was closed in favor of fix(agent): use Hermes root in profile prompt #52676, whose root-aware design avoided that mistake.
• fix(config): P2 batch — .env quoting/UTF-16, aux key_env, profile-aware system prompt #67192 [merged] best fix — (+561/-10) — merged antecedent and reference implementation for four bundled fixes: its .env and auxiliary changes implement their respective reports, but its system-prompt hunk appends /profiles/{active_profile} to profile-scoped get_hermes_home() and uses that home for default data. Commit 34e66a0d is therefore relevant as the implementation that removed hardcoded literals while introducing the residual defect addressed by fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979.
• fix(agent): correct named-profile paths in system prompt hint #68117 [closed] duplicate — (+66/-38) — closed prior repair for the current defect: uses get_hermes_home() directly as the active profile directory and avoids nesting another profile suffix, with behavioral tests; it describes default data conceptually rather than resolving concrete root paths. It remains relevant as superseded implementation history for the more direct root-pointer correction in fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979.
• fix(agent): stop doubling the profile path in the system-prompt hint (#72894) #74979 best fix — (+76/-4) — keep open with a salvage path: uses get_hermes_home() once for the active session and get_default_hermes_root() for default-profile data, exactly matching the reported home/root distinction. The keep_open review requested a real temporary HERMES_HOME=<root>/profiles/<name> composition test, and the updated diff supplies that integration coverage without resolver mocks.

Duplicates

Profile-prompt lineage: #52670, #52676, #66219, and parts of #57223 are earlier root-aware variants; #66453, #66725, and the system-prompt hunk of #67192 share the flawed get_hermes_home()-as-root approach; #68117 and #74979 address the resulting named-profile nesting defect, with #74979 retaining concrete default-root pointers. Auxiliary credential duplicates: #24088, #28766, #36080, and #66655, implemented by #67192. Internal-whitespace serializer duplicates: #66483, #66583, #66625, and #66711, implemented by #67192. UTF-16 sanitizer duplicates: #66475, #66585, #66627, and #66716, with #66475's safer UTF-32-aware implementation carried by #67192.

Suggested consolidation

Keep #74979 open with a salvage path: retain its narrow get_hermes_home()/get_default_hermes_root() production correction and the real resolver-chain tests added after the keep_open review. Close #57223 as duplicate of the #52669/#72894 path-resolution chain after salvaging only its standard-AppData and custom-HERMES_HOME Windows behavioral cases into the canonical tests; its separate hardcoded platform resolver should not survive. Treat #68117, #52670, #52676, #66219, #66453, #66725, and #67192 as closed or merged implementation history rather than reopen candidates, and keep #63982 separate because its symlinked-profile detection fix addresses a distinct cause.

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
    I72894(["issue #72894 (open)"])
    subgraph Dup68117 ["PRs duplicating each other"]
        P68117["PR #68117 (closed)"]
        P74979["PR #74979 (open)"]
    end
    P74979 -->|best fix| I72894
    class I72894 open
    class P68117 closed
    class P74979 open
    class P68117 best
    class P74979 best
    class P74979 target
    click I72894 "https://github.com/NousResearch/hermes-agent/issues/72894"
    click P68117 "https://github.com/NousResearch/hermes-agent/pull/68117"
    click P74979 "https://github.com/NousResearch/hermes-agent/pull/74979"

Loading

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 23 pull requests and 7 issues in this complex. Each diff was read against this issue; Assessment working set: 160 kB of PR diffs, 83 kB of issue/PR text, 36 kB of discussion (93 comments), 58 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.

[teknium1]

Salvaged onto current main and merged via #86352 — your two commits (the ambient-branch fix and the real-resolver-chain integration tests) are cherry-picked with your authorship preserved (merge commit 4691eb5). The file had moved under the #86313 agent-home refactor, which is why a direct merge wasn't possible. Thanks @jeff-mettel!