fix(agent): initialize Bedrock clients on model switch

[shaoohh]

Summary

Fixes #41296 for the remaining initialization gap when switching to AWS Bedrock during a live session.

• route Bedrock Claude switches to AnthropicBedrock with the region parsed from the Bedrock runtime URL
• route Bedrock Converse switches to the SDK-only path and reload configured guardrails
• restore Bedrock region and guardrail state if the client rebuild fails

Validation

• git diff --check
• parsed the changed Python modules with ast.parse
• verified the Bedrock runtime URL regex against an ap-northeast-1 endpoint
• attempted the repository-prescribed scripts/run_tests.sh target; this Windows environment has no installed WSL distribution, and its Python lacks the project test dependencies (pytest, PyYAML), so the targeted pytest suite could not run locally.

[alt-glitch]

This was generated by AI during triage.

Related to #41296 and the closed-unmerged predecessors #65433 and #67948. This is the active refile; the live patch restores regional Bedrock initialization and rollback state rather than duplicating an open implementation.

[teknium1]

Thanks for preserving the Bedrock initialization and rollback state. The underlying runtime gap is still present on current main, but a current pipeline change prevents the Claude half of this patch from being used by normal /model switches.

Problems

• hermes_cli/runtime_provider.py:2120-2133 resolves Claude-on-Bedrock as anthropic_messages, but hermes_cli/model_switch.py:1475-1477 then unconditionally applies the Bedrock host mode from hermes_cli/providers.py:601-602, changing it to bedrock_converse. Consequently, the new AnthropicBedrock branch is not reached for a standard Claude /model switch.
• The added tests pass api_mode directly to agent.switch_model(), so they do not exercise that resolution-and-override path.

Suggested changes

• Preserve the resolved Claude Bedrock anthropic_messages mode through the model-switch pipeline, while retaining Converse for non-Claude targets.
• Add an end-to-end regression covering both Claude and non-Claude Bedrock switches through hermes_cli.model_switch.switch_model() and the live-agent runtime switch.

Automated hermes-sweeper review.

[teknium1]

This branch is correct only if api_mode reaches it as anthropic_messages. On current main, the normal /model pipeline resolves Claude Bedrock as Messages in hermes_cli/runtime_provider.py:2120-2133 but then overwrites it to bedrock_converse in hermes_cli/model_switch.py:1475-1477. Please preserve the resolved Claude mode and add a pipeline-level regression; the direct helper test alone cannot catch this.

[GottZ]

This was generated by AI during triage.

Summary

Twenty-five PRs address or reference this issue complex across context-length ownership and persistence, Desktop assignment, /insights attribution, Bedrock switching, and Codex replay state. The diffs range from clearing or restoring model.context_length during runtime changes to fixing separate model-switch surfaces and state, so consolidation should preserve the issue-specific canonical branches rather than combine them indiscriminately.

Related pull requests

• fix: preserve config_context_length when switching models #6844 [closed] partial — (+118/-1) — n/a: Preserves the initialization-time context override across switches without destination-model or route scoping; closed but relevant because its commits were cherry-picked through fix(model): salvage model/provider resolution PRs #6860 #6844 #6863 #6943 #7187.
• fix(model): warn on stale context_length after /model --global #10366 best fix — (+319/-1) — n/a: Recorded best existing fix for Feature: Warn on context_length mismatch after /model --global switch #10157 because it warns instead of overwriting the override; keep its reviewed salvage path to use provider-aware resolution, remove the nonexistent /context advice, and cover current CLI, gateway, and TUI paths.
• fix: sync model.context_length on --global model switch #17245 [closed] partial — (+3/-0) — n/a: Auto-persists the resolved destination window as the explicit override rather than warning; closed but relevant as the narrow precursor superseded by fix: resolve 7 identified issues [automated] #17246.
• fix: sync context_length on model switch and show updated limit #17269 [closed] partial — (+24/-0) — n/a: Extends auto-persistence with gateway and compressor feedback but can pin inferred metadata as a user override; closed but still relevant because its keep_open review identifies the required semantic split and moved gateway/TUI paths.
• fix: clear stale config context_length override on model switch #21509 [closed] related — (+6/-1) — n/a: Clears stale context state only in switch_model(), preventing inheritance but losing a valid configured pin; closed and subsumed by the broader fix(agent): clear stale config context_length on model switch #22387/fix(agent): clear stale config context_length on model switch #24724 clearing implementation.
• fix: keep insights model metadata in sync after /model switch #22282 [closed] best fix — (+111/-6) — n/a: Recorded best historical fix for Bug: /insights shows wrong model name — displays Gemini instead of actual active model (GLM-5.1 via zai provider) #3988, but its single-row rewrite would collapse mixed-model usage; closed because main has the stronger per-call implementation in cb7f6bbb2e5696bae2c595ca8a3a52cefb6a4a7e.
• fix(agent): clear stale config context_length on model switch #22387 [closed] related — (+10/-0) — n/a: Clears stale context state in both switch and fallback paths; closed but relevant because its implementation was salvaged and merged as fix(agent): clear stale config context_length on model switch #24724.
• fix: clear stale _config_context_length when switching providers #24079 [closed] partial — (+114/-0) — n/a: Selectively re-resolves context by provider but does not establish model-and-route ownership; closed as already implemented for stale-shadowing purposes through fix(agent): clear stale config context_length on model switch #24724 at b6d48ebfde63f325143bc601cbc622d89d428f1a.
• fix(agent): clear stale config context_length on model switch #24724 [merged] related — (+10/-0) — n/a: Merged reference implementation for clearing stale overrides during switches and fallbacks, establishing the non-leakage behavior that newer restoration fixes must retain.
• fix(agent): preserve global model.context_length through switch_model and fallback #32413 partial — (+28/-14) — n/a: Broadly restores the global override in switch and fallback paths but lacks regression tests and destination scoping; keep open with the reviewed salvage path of rebasing the current helpers and testing global preservation versus stale per-model clearing.
• fix(agent): preserve config context_length when fallback/switch uses same model #32452 [closed] best fix — (+18/-2) — n/a: Recorded best existing fix for Context window changes to 256K after interrupted compaction and resume #32423 because it restores the saved override for same-model switch/fallback cases; closed but relevant as a narrow treatment that does not cover the full compaction/resume failure complex.
• fix: persist resolved context_length after /model switch #36199 [closed] partial — (+7/-0) — n/a: Writes a resolved runtime window back into _config_context_length, avoiding some 256K displays while conflating metadata with config intent; despite its approval, it is closed and superseded by fix(config): resolve provider-level context_length on /model switch #37712.
• fix(agent): restore global model.context_length override on /model switch #41010 [closed] best fix — (+12/-0) — n/a: Recorded best existing fix for model.context_length global config override is not re-read after /model switch #40979 and the smallest direct live-config re-read; closed but relevant as an implementation source without regression coverage or model-and-route scoping.
• fix(config): re-read model.context_length from config after /model switch (#40979) #41270 best fix — (+138/-3) — n/a: Recorded best existing fix for model.context_length global config override is not re-read after /model switch #40979, connecting live config re-read with initialization snapshots, custom-provider state, fallback restoration, and tests; keep open for rebase and current-head reconciliation rather than discarding that lifecycle coverage.
• fix(bedrock): preserve regional runtime on model switch #41322 [closed] best fix — (+131/-2) — n/a: Recorded best existing fix for [Bug]: Mid-session /model switch to Bedrock never sets _bedrock_region — regional inference profiles (jp.*/eu.*/apac.*) fail with ValidationException "The provided model identifier is invalid" #41296 because it combines target-model Bedrock routing with regional runtime initialization; author-closed as stale but still relevant as the source later salvaged into fix(bedrock): preserve regional runtime on model switch #65433 and the active refile fix(agent): initialize Bedrock clients on model switch #67970.
• Preserve context_length when switching desktop models #41965 best fix — (+32/-3) — n/a: Recorded best existing fix for [Desktop] Model picker clears model.context_length when switching models in a fresh session #41944 because it changes the Desktop assignment helper that deletes the override; keep open with the reviewed salvage path of rebasing and replacing the contradictory helper test.
• fix: preserve configured context length on model changes #44947 fixes — (+97/-16) — n/a: Mixes broad runtime-pin preservation with partial Dashboard-save preservation, but misses the provider-inference deletion path and conflicts with current non-leakage semantics; keep open only pending the contributor-requested ownership decision.
• [codex] preserve context length on model re-pick #59125 best fix — (+38/-3) — n/a: Recorded best existing fix for [Bug]: POST /api/model/set always drops model.context_length, even on a same-model/same-provider re-pick #59050, preserving the override only for the same normalized provider/model assignment and testing both preservation and real-switch clearing; keep open as the canonical high-salvageability branch.
• fix(config): preserve model.context_length on same-model re-pick #59191 [closed] fixes — (+7/-1) — n/a: Implements the same [Bug]: POST /api/model/set always drops model.context_length, even on a same-model/same-provider re-pick #59050 condition without [codex] preserve context length on model re-pick #59125's tests and docstring update; despite its automated keep_open verdict, contributor review and repeated closure establish it as a closed duplicate of [codex] preserve context length on model re-pick #59125.
• fix(agent): switch_model resets codex reasoning replay flag for xai-oauth/codex #61552 related — (+121/-0) — n/a: Separately resets, snapshots, and restores Codex reasoning-replay state across switches and fallback restoration, with the requested restore-path test now present; keep open on its independent salvage path.
• fix: preserve model.context_length on model switch #62552 partial — (+97/-24) — n/a: Re-reads integer and quoted-numeric context overrides with invalid/absent-value tests but applies the pin to every destination; keep open for ownership scoping and do not use it as a substitute for Desktop fix Preserve context_length when switching desktop models #41965.
• fix(bedrock): preserve regional runtime on model switch #65433 [closed] best fix — (+105/-0) — n/a: Recorded best existing fix for [Bug]: Mid-session /model switch to Bedrock never sets _bedrock_region — regional inference profiles (jp.*/eu.*/apac.*) fail with ValidationException "The provided model identifier is invalid" #41296 and a current-main salvage of fix(bedrock): preserve regional runtime on model switch #41322's regional Bedrock runtime plus rollback tests; closed but relevant as the direct predecessor to fix(agent): initialize Bedrock clients on model switch #67970.
• fix(agent): initialize Bedrock clients on model switch #67948 [closed] duplicate — (+160/-0) — n/a: Adds Bedrock SDK initialization, guardrail reload, and rollback state but only helper-level validation; closed and superseded by the active refile fix(agent): initialize Bedrock clients on model switch #67970.
• fix(agent): initialize Bedrock clients on model switch #67970 best fix — (+160/-0) — n/a: Recorded best existing fix for [Bug]: Mid-session /model switch to Bedrock never sets _bedrock_region — regional inference profiles (jp.*/eu.*/apac.*) fail with ValidationException "The provided model identifier is invalid" #41296 and the active Bedrock runtime refile, but the blocking contributor review shows normal /model still overwrites Claude's resolved anthropic_messages mode and the direct-helper tests cannot detect that pipeline failure.
• fix(agent): restore model.context_length pin when /model returns to the configured default #70210 best fix — (+237/-3) — n/a: Recorded best existing fix for model.context_length global config override is not re-read after /model switch #40979's model-and-route ownership semantics, but its early pin restoration is overwritten by the later destination-context block; keep open with the reviewed salvage path of moving the decision into final resolution and testing the live config path.

Duplicates

#21509 is the narrower duplicate of #22387, whose implementation was merged through #24724; #17245 is the narrow precursor to the broader #17269 approach; #59191 is a closed duplicate of canonical #59125; and #41322 → #65433 → #67948 → #67970 is the Bedrock salvage/refile chain, with #67948 directly superseded by #67970.

Suggested consolidation

Author action: rebase #67970 onto main, preserve Claude Bedrock's resolved anthropic_messages mode through hermes_cli.model_switch, and add pipeline-level Claude and non-Claude regressions, or split that pipeline correction from the already useful regional initialization and rollback changes. Keep the issue-specific canonical branches #10366, #41270/#70210, #41965, and #59125 on their documented salvage paths; the duplicate chains are already closed except for active canonical endpoints, so no additional duplicate closure is supported by the supplied states.

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
    I41296(["issue #41296 (open)"])
    subgraph Dup41322 ["PRs duplicating each other"]
        P41322["PR #41322 (closed)"]
        P65433["PR #65433 (closed)"]
        P67948["PR #67948 (closed)"]
        P67970["PR #67970 (open)"]
    end
    P67970 -->|best fix| I41296
    class I41296 open
    class P41322 closed
    class P65433 closed
    class P67948 closed
    class P67970 open
    class P41322 best
    class P65433 best
    class P67970 best
    class P67970 target
    click I41296 "https://github.com/NousResearch/hermes-agent/issues/41296"
    click P41322 "https://github.com/NousResearch/hermes-agent/pull/41322"
    click P65433 "https://github.com/NousResearch/hermes-agent/pull/65433"
    click P67948 "https://github.com/NousResearch/hermes-agent/pull/67948"
    click P67970 "https://github.com/NousResearch/hermes-agent/pull/67970"

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: 136 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.