model.context_length global config override is not re-read after /model switch

[bonaluo]

Bug Description

model.context_length in config.yaml is read once at agent initialization in agent/agent_init.py (line 1314) and stored as agent._config_context_length. When the user switches models via /model, switch_model() in agent/agent_runtime_helpers.py (line 1426) sets agent._config_context_length = None to clear the old value, but never re-reads the config file to restore the global model.context_length override.

Resolution chain after switch:

get_model_context_length(
    config_context_length=None,       # step 0: global override LOST
    custom_providers=custom_providers_list,  # step 0b: per-model override works
)

• Step 0 (config_context_length): Always None after switch, so the global model.context_length is lost.
• Step 0b (custom_providers per-model): Works correctly — the config file is re-read and custom_providers[i].models.<model>.context_length is honored.
• If the user only set the global model.context_length (not per-model in custom_providers) and the custom endpoint can't auto-detect via /models -> falls through to 256K default.

This creates an inconsistent experience: the global model.context_length works at startup but silently disappears after any /model command, even when switching back to the same model.

Steps to Reproduce

1. Configure in ~/.hermes/config.yaml:

   model:
     provider: custom:my-api
     model: deepseek-v4-pro
     context_length: 128000

2. Start a session — context length correctly shows 128K.
3. Run /model mini-max-m2.7 then /model deepseek-v4-pro (back to original).
4. Now context length shows auto-detected value (e.g. 256K), not 128K.

Expected Behavior

The global model.context_length should be re-read from config.yaml after every /model switch, so it stays effective for any model that doesn't have a per-model override.

Alternatively, if the per-model custom_providers override is the intended approach, the docs should clarify that the global model.context_length is a one-time init-only setting and per-model overrides in custom_providers.models.<name>.context_length are required for persistence across model switches.

Additional Context

• The per-model custom_providers override (step 0b) already works in both init and switch_model paths (fix Bug: /model switch to named custom provider ignores custom_providers model context_length #15779).
• get_custom_provider_context_length() in hermes_cli/config.py:3896 is the shared helper used by both paths.
• The gap is specifically that step 0 (global config override) is not re-applied.

[waefrebeorn]

Reproduced this on current main and confirmed the root cause is exactly as described: switch_model() in agent/agent_runtime_helpers.py sets agent._config_context_length = None unconditionally and never re-reads config, while the per-model custom_providers override IS re-read live (#15779) — so the documented global pin silently dies on the first /model and stays dead even after switching back to the configured default.

Hit this in production via the gateway: model.context_length: 262144 capping poolside/laguna-s-2.1 (1M catalog window, degrades past ~256K) — one /model round-trip and the compressor was sizing against 1M again.

Opened PR #70210 with a fix that re-resolves the pin from live config at switch time, scoped with the same normalize_model_for_provider + _context_route_mismatch route-pin semantics agent_init.py already uses — restored only when the switched-to model+route equals the configured default, so it can't leak onto other models (#62152) and the stale-inheritance protection the clearing originally provided is preserved. 5 new behavior-contract tests + all 78 existing switch_model tests green.

[GottZ]

This was generated by AI during triage.

Summary

Twenty-five PRs address or reference this issue complex. The core #40979 diffs either restore the explicit model.context_length after /model, preserve it only for the configured model and route, or change adjacent persistence, Desktop, fallback, insights, Bedrock, and Codex-runtime behavior without fixing that exact post-switch resolution path.

Related pull requests

• fix: preserve config_context_length when switching models #6844 [closed] partial — (+118/-1) — partial fix, closed but relevant as an earlier reference merged through fix(model): salvage model/provider resolution PRs #6860 #6844 #6863 #6943 #7187: it retained the initialization-time _config_context_length when rebuilding the compressor, but did not re-read live config or prevent the pin from leaking to a different destination model.
• fix(model): warn on stale context_length after /model --global #10366 unrelated — (+319/-1) — unrelated to model.context_length global config override is not re-read after /model switch #40979: it warns about a stale override after --global rather than restoring a lost override at runtime. Keep open only with the salvage path from its keep_open review: use provider-aware context resolution, remove the nonexistent /context advice, and cover the current CLI, gateway, and TUI paths.
• fix: sync model.context_length on --global model switch #17245 [closed] unrelated — (+3/-0) — unrelated, closed and superseded by fix: resolve 7 identified issues [automated] #17246: it writes an auto-resolved window into the highest-priority explicit config field, addressing persistence after --global rather than re-reading the user's existing pin.
• fix: sync context_length on model switch and show updated limit #17269 unrelated — (+24/-0) — unrelated to the root cause: it persists detected context metadata and displays the updated compressor limit. Its keep_open review must be resolved by separating explicit caps from detected windows and retargeting the moved gateway/TUI paths before this work is salvageable.
• fix: clear stale config context_length override on model switch #21509 [closed] unrelated — (+6/-1) — unrelated, closed predecessor: it intentionally clears the override so a new model can auto-resolve, which is the behavior that exposes model.context_length global config override is not re-read after /model switch #40979 when returning to the configured default.
• fix: keep insights model metadata in sync after /model switch #22282 [closed] unrelated — (+111/-6) — unrelated, closed because /insights attribution is already implemented on main by cb7f6bbb2e5696bae2c595ca8a3a52cefb6a4a7e, as documented by the maintainer-bot from hermes_state.py:2593-2660, agent/insights.py:524-534, and tests/agent/test_insights.py:323-350; its one-row overwrite would also lose mixed-model attribution.
• fix(agent): clear stale config context_length on model switch #22387 [closed] unrelated — (+10/-0) — unrelated, closed but relevant as the source salvaged into merged fix(agent): clear stale config context_length on model switch #24724: it established the stale-inheritance clearing behavior for model switches and fallbacks that a model.context_length global config override is not re-read after /model switch #40979 fix must preserve for non-owning models.
• fix: clear stale _config_context_length when switching providers #24079 [closed] partial — (+114/-0) — partial fix, closed as already implemented on main through fix(agent): clear stale config context_length on model switch #24724: the maintainer-bot cites commit b6d48ebfde63f325143bc601cbc622d89d428f1a and agent/agent_runtime_helpers.py:1426, while this diff's provider-only comparison is weaker than model-and-route ownership.
• fix(agent): clear stale config context_length on model switch #24724 [merged] unrelated — (+10/-0) — unrelated but relevant merged reference implementation: it clears stale _config_context_length during explicit switches and fallback activation, establishing the non-leakage invariant that the model.context_length global config override is not re-read after /model switch #40979 restoration must retain.
• fix(agent): preserve global model.context_length through switch_model and fallback #32413 fixes — (+28/-14) — fixes the loss broadly by re-reading the global override for every switch and fallback, but its keep_open verdict identifies conflicts and missing switch/fallback regression tests. Salvage only the live-config resolution while defining whether the pin is truly global or owned by the configured model and route.
• fix(agent): preserve config context_length when fallback/switch uses same model #32452 [closed] unrelated — (+18/-2) — unrelated, closed narrow workaround for Context window changes to 256K after interrupted compaction and resume #32423: it restores the saved override only when the model name is unchanged, so it misses a switch away and back and does not validate provider or endpoint identity.
• fix: persist resolved context_length after /model switch #36199 [closed] partial — (+7/-0) — partial fix, closed and explicitly superseded by fix(config): resolve provider-level context_length on /model switch #37712 despite an earlier non-contributor approval: it writes a detected window back into _config_context_length, conflating resolved metadata with the user's explicit pin and potentially persisting the wrong value.
• fix(agent): restore global model.context_length override on /model switch #41010 [closed] best fix — (+12/-0) — fixes model.context_length global config override is not re-read after /model switch #40979 and remains the best existing fix by the Verify verdict: it minimally re-reads model.context_length in the existing live-config block. It is closed and remains only a useful code reference, not a verified reference implementation, because the supplied diff contains no regression test or recorded execution result and applies the pin without model-and-route ownership checks.
• fix(config): re-read model.context_length from config after /model switch (#40979) #41270 best fix — (+138/-3) — fixes model.context_length global config override is not re-read after /model switch #40979 and extends restoration through primary-runtime snapshot/fallback state, with switch and restore tests. Keep open with a salvage path consistent with its keep_open review: retain the initialization snapshots, fallback preservation, and temporary-HERMES_HOME regression now visible in the diff, while narrowing the context pin to the final destination-resolution semantics and dropping unrelated release attribution.
• fix(bedrock): preserve regional runtime on model switch #41322 [closed] unverified — (+131/-2) — unverified and closed as stale, but relevant as the original Bedrock regional-runtime implementation: it fixes target-model routing and _bedrock_region, not the global context-pin loss, and was later salvaged by fix(bedrock): preserve regional runtime on model switch #65433/fix(agent): initialize Bedrock clients on model switch #67948/fix(agent): initialize Bedrock clients on model switch #67970.
• Preserve context_length when switching desktop models #41965 unverified — (+32/-3) — unverified Desktop-side companion: it stops the model assignment helper from deleting an explicit override, but does not repair the in-memory /model resolution in model.context_length global config override is not re-read after /model switch #40979. Follow its keep_open verdict by rebasing and replacing the contradictory direct-helper test before salvage.
• fix: preserve configured context length on model changes #44947 fixes — (+97/-16) — fixes the runtime loss and a partial-dashboard-save path, but its keep_open review records salvageability none until maintainers decide global-cap versus per-model semantics; the diff also misses the provider-inference assignment path. Keep open only for that semantics decision rather than using it as the consolidation target.
• [codex] preserve context length on model re-pick #59125 unverified — (+38/-5) — unverified but high-salvage adjacent fix for [Bug]: POST /api/model/set always drops model.context_length, even on a same-model/same-provider re-pick #59050: it preserves context_length only on a same-model/same-provider Dashboard re-pick and clears it on a real switch. Its maintainer-bot keep_open verdict is compatible with retaining it as the canonical Desktop helper fix, separate from model.context_length global config override is not re-read after /model switch #40979.
• fix(config): preserve model.context_length on same-model re-pick #59191 [closed] unverified — (+7/-1) — unverified, closed as duplicate of [codex] preserve context length on model re-pick #59125: it makes the same same-model/same-provider comparison but lacks [codex] preserve context length on model re-pick #59125's regression tests, normalized incoming model handling, and updated contract documentation.
• fix(agent): switch_model resets codex reasoning replay flag for xai-oauth/codex #61552 unverified — (+121/-0) — unrelated to context length: it snapshots and restores the Codex reasoning-replay flag across model switches and fallback restoration. Its keep_open review requested a fallback-restore regression, and the supplied diff now contains that test, so it should remain an independent salvage candidate.
• fix: preserve model.context_length on model switch #62552 fixes — (+97/-24) — fixes the agent-side loss by re-reading and coercing model.context_length, including quoted numeric values, while preserving the no-override path. Keep open with the salvage path from its review, but remove the incorrect closure of Desktop issue [Desktop] Model picker clears model.context_length when switching models in a fresh session #41944 and align ownership semantics with the configured model and route.
• fix(bedrock): preserve regional runtime on model switch #65433 [closed] unverified — (+105/-0) — unverified, closed Bedrock salvage of fix(bedrock): preserve regional runtime on model switch #41322: it restores regional SDK state and rollback coverage but does not address model.context_length global config override is not re-read after /model switch #40979's context pin; it remains relevant as a predecessor to the active Bedrock work.
• fix(agent): initialize Bedrock clients on model switch #67948 [closed] unrelated — (+160/-0) — unrelated, closed predecessor to fix(agent): initialize Bedrock clients on model switch #67970: it initializes Bedrock clients, region, guardrails, and rollback state, but its targeted tests were not executed in the reported environment and it does not restore model.context_length.
• fix(agent): initialize Bedrock clients on model switch #67970 unrelated — (+160/-0) — unrelated active Bedrock refile: it repairs regional client initialization and rollback, not model.context_length global config override is not re-read after /model switch #40979. Keep open with the contributor review's required salvage path: preserve Claude's resolved anthropic_messages mode through hermes_cli.model_switch and add a pipeline-level regression instead of relying on direct helper tests.
• fix(agent): restore model.context_length pin when /model returns to the configured default #70210 best fix — (+237/-3) — fixes model.context_length global config override is not re-read after /model switch #40979 with the strongest explicit model-and-route pin semantics and focused behavior-contract tests. Follow its keep_open review and contributor feedback: author action is required to rebase onto current main and move the ownership decision into the post-swap destination-context block, because the current early assignment is overwritten before reaching the compressor.

Duplicates

The direct #40979 restoration family is #41010, #41270, #62552, and #70210, with #32413 and #44947 as broader semantic variants; #41010 is the minimal closed reference, while #70210 contributes the strongest route-scoped tests and #41270 contributes primary-runtime/fallback state coverage. #21509, #22387, #24079, and merged #24724 form the opposite stale-inheritance-clearing lineage that the final fix must preserve for non-owning destinations. #59191 is a closed duplicate of canonical #59125. #41322, #65433, #67948, and #67970 are the separate Bedrock chain; #17245/#17269, #41965, and #36199 are adjacent persistence/display variants rather than exact duplicates.

Suggested consolidation

#41010 remains the Verify-selected best existing fix, but it is closed and its test-free 12-line diff is not sufficient as the actionable consolidation target. Author action: rebase #70210 onto current main and move its model-and-route-scoped global-pin decision into the final post-swap destination-context resolution so the value actually reaches the compressor; this follows, rather than overrides, the keep_open review on #70210. Keep #41270 open with a salvage path as a complementary verify-confirmed fix: preserve its initialization/fallback snapshots and real-config regression, then coordinate those lifecycle changes with #70210's ownership semantics instead of closing it as excess scope. Likewise, do not close #32413, #44947, or #62552 over their visible keep_open reviews until the global-versus-route-owned semantics are resolved; once one corrected implementation carries both destination-resolution and lifecycle tests, the remaining open runtime variants can be reassessed as explicit duplicates. The already closed duplicate/predecessor chains #21509/#22387/#24079 and #59191 need no reopening.

Complex graph

flowchart TD
    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
    I40979(["issue #40979 (open)"])
    I41296(["issue #41296 (open)"])
    I41944(["issue #41944 (open)"])
    I59050(["issue #59050 (open)"])
    P6844["PR #6844 (closed)"]
    P24079["PR #24079 (closed)"]
    P36199["PR #36199 (closed)"]
    P41322["PR #41322 (closed)"]
    P41965["PR #41965 (open)"]
    P44947["PR #44947 (open)"]
    P59125["PR #59125 (open)"]
    P70210["PR #70210 (open)"]
    subgraph Dup32413 ["PRs duplicating each other"]
        P32413["PR #32413 (open)"]
        P41010["PR #41010 (closed)"]
        P41270["PR #41270 (open)"]
        P62552["PR #62552 (open)"]
    end
    More["+13 more PRs not shown"]
    MoreI(["+3 more issues not shown"])
    P6844 -.->|partial| I40979
    P24079 -.->|partial| I40979
    P32413 -->|fixes| I40979
    P36199 -.->|partial| I40979
    P41010 -->|best fix| I40979
    P41270 -->|best fix| I40979
    P41322 -.->|unverified| I40979
    P41965 -.->|unverified| I40979
    P44947 -->|fixes| I40979
    P59125 -.->|unverified| I40979
    P62552 -->|fixes| I40979
    P70210 -->|best fix| I40979
    P41322 -->|best fix| I41296
    P41965 -->|best fix| I41944
    P44947 -.->|partial| I41944
    P62552 -.->|partial| I41944
    P41965 -.->|partial| I59050
    P44947 -.->|partial| I59050
    P59125 -->|best fix| I59050
    class I40979 open
    class I41296 open
    class I41944 open
    class I59050 open
    class I40979 target
    class P6844 closed
    class P24079 closed
    class P32413 open
    class P36199 closed
    class P41010 closed
    class P41270 open
    class P41322 closed
    class P41965 open
    class P44947 open
    class P59125 open
    class P62552 open
    class P70210 open
    class P41010 best
    class P41270 best
    class P41322 best
    class P41965 best
    class P59125 best
    class P70210 best
    class More closed
    class MoreI closed
    click I40979 "https://github.com/NousResearch/hermes-agent/issues/40979"
    click I41296 "https://github.com/NousResearch/hermes-agent/issues/41296"
    click I41944 "https://github.com/NousResearch/hermes-agent/issues/41944"
    click I59050 "https://github.com/NousResearch/hermes-agent/issues/59050"
    click P6844 "https://github.com/NousResearch/hermes-agent/pull/6844"
    click P24079 "https://github.com/NousResearch/hermes-agent/pull/24079"
    click P32413 "https://github.com/NousResearch/hermes-agent/pull/32413"
    click P36199 "https://github.com/NousResearch/hermes-agent/pull/36199"
    click P41010 "https://github.com/NousResearch/hermes-agent/pull/41010"
    click P41270 "https://github.com/NousResearch/hermes-agent/pull/41270"
    click P41322 "https://github.com/NousResearch/hermes-agent/pull/41322"
    click P41965 "https://github.com/NousResearch/hermes-agent/pull/41965"
    click P44947 "https://github.com/NousResearch/hermes-agent/pull/44947"
    click P59125 "https://github.com/NousResearch/hermes-agent/pull/59125"
    click P62552 "https://github.com/NousResearch/hermes-agent/pull/62552"
    click P70210 "https://github.com/NousResearch/hermes-agent/pull/70210"

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 25 pull requests and 7 issues in this complex. Each diff was read against this issue; Assessment working set: 137 kB of PR diffs, 70 kB of issue/PR text, 32 kB of discussion (47 comments), 54 verify verdicts. verdicts reflect diff content, not PR titles. Part of an automated triage batch.