opencode-llmstack 0.7.1__tar.gz → 0.7.3__tar.gz

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opencode-llmstack
-Version: 0.7.1
+Version: 0.7.3
 Summary: Multi-tier local LLM stack: llama-swap + FastAPI auto-router + opencode wiring.
 Author: llmstack
 License: MIT
@@ -313,10 +313,15 @@ py -3 -m venv .venv
 #    `start` falls back to spawning a PowerShell subshell.
 .venv\Scripts\llmstack start
 
-# 4. Auto-activate per project from any new PowerShell window:
-Invoke-Expression (& llmstack activate powershell | Out-String)
-# or persist (writes ~/.powershell_llmstack_hook + sources it on every shell):
-"Invoke-Expression (& llmstack activate powershell | Out-String)" | Add-Content $PROFILE
+# 4. Auto-activate per project from any new PowerShell window. The hook
+#    file is a .ps1 (PowerShell won't dot-source it without that
+#    extension) and dot-sourcing it requires script execution to be
+#    allowed -- if you see "running scripts is disabled on this
+#    system", run once:
+#       Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
+llmstack activate powershell | Out-String | Invoke-Expression
+# or persist (writes ~/.powershell_llmstack_hook.ps1 + sources it on every shell):
+"llmstack activate powershell | Out-String | Invoke-Expression" | Add-Content $PROFILE
 ```
 
 Notes:

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/README.md

@@ -278,10 +278,15 @@ py -3 -m venv .venv
 #    `start` falls back to spawning a PowerShell subshell.
 .venv\Scripts\llmstack start
 
-# 4. Auto-activate per project from any new PowerShell window:
-Invoke-Expression (& llmstack activate powershell | Out-String)
-# or persist (writes ~/.powershell_llmstack_hook + sources it on every shell):
-"Invoke-Expression (& llmstack activate powershell | Out-String)" | Add-Content $PROFILE
+# 4. Auto-activate per project from any new PowerShell window. The hook
+#    file is a .ps1 (PowerShell won't dot-source it without that
+#    extension) and dot-sourcing it requires script execution to be
+#    allowed -- if you see "running scripts is disabled on this
+#    system", run once:
+#       Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
+llmstack activate powershell | Out-String | Invoke-Expression
+# or persist (writes ~/.powershell_llmstack_hook.ps1 + sources it on every shell):
+"llmstack activate powershell | Out-String | Invoke-Expression" | Add-Content $PROFILE
 ```
 
 Notes:

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/llmstack/commands/activate.py

@@ -28,20 +28,42 @@ def _print_help() -> None:
     print("usage: llmstack activate <zsh|bash|powershell>", file=sys.stderr)
 
 
+def _is_powershell(shell: str) -> bool:
+    return shell in ("powershell", "pwsh")
+
+
 def _hook_path(shell: str) -> Path:
     """``~/.<shell>_llmstack_hook`` -- ``pwsh`` is normalised to ``powershell``
-    so the user doesn't end up with two redundant files."""
-    name = "powershell" if shell in ("powershell", "pwsh") else shell
-    return Path.home() / f".{name}_llmstack_hook"
+    so the user doesn't end up with two redundant files.
+
+    PowerShell additionally needs a ``.ps1`` suffix or the host won't
+    dot-source it -- without the extension Windows hands the file to
+    the OS shell file-association (Notepad, etc.) instead of running
+    it as a script.
+    """
+    if _is_powershell(shell):
+        return Path.home() / ".powershell_llmstack_hook.ps1"
+    return Path.home() / f".{shell}_llmstack_hook"
 
 
 def _source_line(shell: str, path: Path) -> str:
     """Shell-specific incantation to load the hook file."""
-    if shell in ("powershell", "pwsh"):
+    if _is_powershell(shell):
         return f". '{path}'"
     return f'source "{path}"'
 
 
+def eval_line(shell: str) -> str:
+    """The one-shot the user pastes / adds to their rc to install the hook.
+
+    POSIX shells use ``eval "$(...)"``; PowerShell has no ``eval`` and
+    needs ``Invoke-Expression`` over the captured stdout.
+    """
+    if _is_powershell(shell):
+        return f"llmstack activate {shell} | Out-String | Invoke-Expression"
+    return f'eval "$(llmstack activate {shell})"'
+
+
 def write_hook(shell: str) -> tuple[Path, str]:
     """Render the hook for ``shell``, write it to disk, return ``(path, source_line)``.
 
@@ -62,10 +84,24 @@ def run(args: list[str]) -> int:
 
     path, src = write_hook(shell)
 
-    eval_line = f'eval "$(llmstack activate {shell})"'
+    line = eval_line(shell)
     print(f"[OK] hook written: {path}", file=sys.stderr)
     print( "     activate in this shell now (and for every new shell:", file=sys.stderr)
-    print(f"     paste into your rc):  {eval_line}", file=sys.stderr)
+    print(f"     paste into your rc):  {line}", file=sys.stderr)
+    if _is_powershell(shell):
+        # PowerShell's default `Restricted` policy on Windows blocks
+        # dot-sourcing any .ps1; surface the one-time fix so the
+        # `Invoke-Expression` line above doesn't fail with "running
+        # scripts is disabled on this system".
+        print(
+            "     PowerShell execution policy must allow local scripts; "
+            "if dot-sourcing fails, run once:",
+            file=sys.stderr,
+        )
+        print(
+            "         Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned",
+            file=sys.stderr,
+        )
 
     print(src)
     return 0

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/llmstack/commands/setup.py

@@ -18,7 +18,7 @@ import shutil
 import subprocess
 
 from llmstack._platform import IS_WINDOWS, shell_family
-from llmstack.commands.activate import write_hook
+from llmstack.commands.activate import eval_line, write_hook
 from llmstack.download.binary import install_llama_swap
 from llmstack.download.ggufs import download_all, wait_for_downloads
 from llmstack.paths import is_remote, remote_url
@@ -86,16 +86,25 @@ def run(args: list[str]) -> int:
         rc_hint = "your shell rc file"
         hook_arg = None
 
-    eval_line: str | None = None
+    activate_line: str | None = None
     if hook_arg is not None:
         path, _src = write_hook(hook_arg)
-        eval_line = f'eval "$(llmstack activate {hook_arg})"'
+        activate_line = eval_line(hook_arg)
         print(f"[OK] hook installed: {path}")
         print()
         print("To turn it on in this shell now (and persist across new shells, paste")
         print(f"the same line into {rc_hint}):")
         print()
-        print(f"    {eval_line}")
+        print(f"    {activate_line}")
+        if family == "powershell":
+            # PowerShell needs script execution allowed before
+            # dot-sourcing the .ps1; flag the one-time fix here so the
+            # Invoke-Expression line above doesn't silently fail.
+            print()
+            print("    PowerShell execution policy must allow local scripts;")
+            print("    if dot-sourcing fails with \"running scripts is disabled\", run once:")
+            print()
+            print("        Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned")
 
     print()
     print("[5/5] checking opencode...")
@@ -134,8 +143,8 @@ def run(args: list[str]) -> int:
     print("[OK] setup complete.")
     print()
     print("Next steps:")
-    if eval_line is not None:
-        print(f"  1. Run (and paste into {rc_hint} for persistence): {eval_line}")
+    if activate_line is not None:
+        print(f"  1. Run (and paste into {rc_hint} for persistence): {activate_line}")
     else:
         print("  1. Source the generated hook in your shell rc (see above)")
     print("  2. llmstack install     # generate .llmstack/ configs for this project")

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/llmstack/models.ini

@@ -51,7 +51,6 @@
 [DEFAULT]
 host         = 127.0.0.1
 router_port  = 10101        ; FastAPI auto-router (what opencode hits)
-swap_port    = 10102        ; llama-swap manager UI + raw model endpoints
 n_gpu_layers = 999          ; offload everything to Metal on Apple Silicon
 flash_attn   = on
 jinja        = true
@@ -72,11 +71,26 @@ ctx_size     = 131072       ; native 32k extended via YaRN (factor 4)
 rope_scaling = yarn (scale=4, orig_ctx=32768)
 size_gb      = 2.5
 quant        = Q5_K_M
-status       = downloading      ; queued by `llmstack.sh download`
-opencode_use = small_model + auto-fast tier
 sampler      = temp=0.2, top_p=0.95, top_k=40, min_p=0.05      ; deterministic
 description  = Qwen2.5-Coder 3B - autocomplete / FIM / quick Q&A
 
+; Bedrock alternative for code-fast -- comment out the [code-fast] block above
+; and uncomment the block below to swap to a hosted fast tier (Claude Haiku
+; 4.5: cheapest + fastest Anthropic model with tool calling, sub-second TTFT).
+; See "BEDROCK NOTES" at the bottom of this file for profile / sampler /
+; access-form details.
+;
+; [code-fast]
+; tier         = code
+; role         = fast
+; backend      = bedrock
+; aws_model_id = eu.anthropic.claude-haiku-4-5-20251001-v1:0
+; aws_region   = eu-central-1
+; aws_profile  = bedrock-prod
+; ctx_size     = 200000
+; sampler      = temp=0.2     ; deterministic; Haiku 4.5 accepts ONE of temp / top_p
+; description  = Claude Haiku 4.5 on Bedrock - hosted fast tier for autocomplete / FIM / quick Q&A
+
 [code-smart]
 tier         = code
 role         = agent
@@ -88,11 +102,25 @@ size_gb      = 45
 size_gb_next = 50
 quant        = Q4_K_M
 quant_next   = UD-Q4_K_XL
-status       = ready (Q4_K_M); UD-Q4_K_XL queued
-opencode_use = agent.build + auto-agent tier
 sampler      = temp=0.5, top_p=0.85, top_k=20, min_p=0.05, rep_pen=1.05  ; balanced agent
 description  = Qwen3-Coder-Next 80B-A3B MoE - heavy coder for agent loops
 
+; Bedrock alternative for code-smart -- comment out the [code-smart] block
+; above and uncomment the block below to swap to a hosted heavy coder
+; (Claude Sonnet 4.6: agent-loop workhorse, heavy tool calling, multi-file
+; edits). See "BEDROCK NOTES" at the bottom of this file.
+;
+; [code-smart]
+; tier         = code
+; role         = agent
+; backend      = bedrock
+; aws_model_id = eu.anthropic.claude-sonnet-4-6
+; aws_region   = eu-central-1
+; aws_profile  = bedrock-prod
+; ctx_size     = 200000
+; sampler      = temp=0.5     ; Sonnet 4.6 accepts ONE of temp / top_p; pick `temp` for agent work
+; description  = Claude Sonnet 4.6 on Bedrock - heavy coder for agent loops
+
 ; Top-tier hosted coder. Shipped disabled because it requires boto3 +
 ; AWS Bedrock access. `llmstack install` auto-uncomments the block
 ; below (by stripping the leading "; " from each line and dropping
@@ -107,10 +135,9 @@ description  = Qwen3-Coder-Next 80B-A3B MoE - heavy coder for agent loops
 ; role         = ultra
 ; backend      = bedrock
 ; aws_model_id = global.anthropic.claude-opus-4-7   ; global.* cross-region inference profile
-; aws_region   = us-east-1                          ; API anchor region; global.* auto-routes inference cross-region
-; aws_profile = bedrock-prod                        ; uncomment + set your own profile name; falls back to default cred chain otherwise
+; aws_region   = eu-central-1                       ; API anchor region; global.* auto-routes inference cross-region (set EU as the anchor for residency)
+; aws_profile  = bedrock-prod                       ; conventional profile name; configure once with `aws configure --profile bedrock-prod` (or change to your own and run `llmstack install`)
 ; ctx_size     = 200000
-; opencode_use = on-demand top-tier coder for hard agent tasks
 ; ; NB: no `sampler =` line. Claude Opus 4.7 explicitly rejects all
 ; ;     sampler params (temperature, top_p, top_k) -- per the Bedrock
 ; ;     model card, "the recommended migration path is to omit these
@@ -134,11 +161,26 @@ size_gb      = 9.2
 size_gb_next = 12.1
 quant        = Q4_K_M
 quant_next   = Q6_K
-status       = ready (Q4_K_M); Q6_K queued
-opencode_use = agent.plan + auto-plan tier
 sampler      = temp=0.7, top_p=0.9, top_k=40, min_p=0.05       ; creative thinking
 description  = Qwopus GLM 18B - planning, design discussions, architecture
 
+; Bedrock alternative for plan -- comment out the [plan] block above and
+; uncomment the block below to swap to a hosted planner (Claude Opus 4.6:
+; deep reasoning for design discussions and architecture). Opus 4.6 still
+; accepts both temperature and top_p (unlike 4.7), so the local sampler
+; maps over cleanly. See "BEDROCK NOTES" at the bottom of this file.
+;
+; [plan]
+; tier         = chat
+; role         = plan
+; backend      = bedrock
+; aws_model_id = eu.anthropic.claude-opus-4-6-v1
+; aws_region   = eu-central-1
+; aws_profile  = bedrock-prod
+; ctx_size     = 200000
+; sampler      = temp=0.7, top_p=0.9     ; creative; Opus 4.6 accepts both
+; description  = Claude Opus 4.6 on Bedrock - planning, design discussions, architecture
+
 [plan-uncensored]
 tier         = chat
 role         = plan-uncensored
@@ -150,11 +192,51 @@ size_gb      = 13
 size_gb_next = 20
 quant        = i1-Q4_K_M
 quant_next   = i1-Q6_K
-status       = ready (i1-Q4_K_M); i1-Q6_K queued
-opencode_use = agent.plan-nofilter + auto via [nofilter] trigger
 sampler      = temp=0.85, top_p=0.95, top_k=50, min_p=0.05     ; max exploration
 description  = Mistral-Small 3.2 24B Heretic - no-filter planning
 
+; Bedrock alternative for plan-uncensored -- comment out the [plan-uncensored]
+; block above and uncomment ONE of the blocks below. Anthropic models on
+; Bedrock are filtered, so for the uncensored slot we pick the largest
+; open-weights model on Bedrock: Llama 3.1 405B has minimal safety post-
+; training and matches the spirit of the local Heretic tier. NOTE: Meta
+; models do NOT require the AWS use-case form, so this swap unblocks
+; plan-uncensored on a fresh AWS account.
+;
+; REGION CAVEAT: unlike the other tiers above, Llama 3.1 405B has NO
+; cross-region inference profile (no eu.* / global.*) and is only
+; deployed in US regions. Pin to us-west-2 even when the rest of the
+; stack is anchored in eu-central-1. If EU residency is mandatory for
+; this tier, switch to one of the eu.anthropic.* IDs at the cost of
+; losing the "uncensored" property. See "BEDROCK NOTES" at the bottom
+; of this file.
+;
+; [plan-uncensored]
+; tier         = chat
+; role         = plan-uncensored
+; backend      = bedrock
+; aws_model_id = meta.llama3-1-405b-instruct-v1:0
+; aws_region   = us-west-2     ; Llama 405B has no EU deployment; keep on US
+; aws_profile  = bedrock-prod
+; ctx_size     = 128000
+; sampler      = temp=0.85, top_p=0.95   ; max exploration
+; description  = Llama 3.1 405B on Bedrock - no-filter planning
+;
+; ...or, if your org locks Bedrock access to a VPC endpoint, use this
+; variant instead (same model + sampler, with aws_endpoint_url set):
+;
+; [plan-uncensored]
+; tier             = chat
+; role             = plan-uncensored
+; backend          = bedrock
+; aws_model_id     = meta.llama3-1-405b-instruct-v1:0
+; aws_region       = us-west-2     ; Llama 405B has no EU deployment
+; aws_profile      = bedrock-prod
+; aws_endpoint_url = https://bedrock-runtime.us-west-2.vpce.amazonaws.com
+; ctx_size         = 128000
+; sampler          = temp=0.85, top_p=0.95
+; description      = Llama 3.1 405B on Bedrock (VPC) - no-filter planning
+
 ;------------------------------------------------------------------------------
 [ROUTING]
 ; STEP-DOWN ladder: start at the top of the fidelity ladder for short
@@ -203,30 +285,46 @@ uncensored_triggers   = [nofilter], [uncensored], [heretic], "uncensored:", "nof
 ultra_triggers        = [ultra], [opus], "ultra:", "opus:" (line start)
 
 ;------------------------------------------------------------------------------
-; BEDROCK EXAMPLES (commented out -- copy / uncomment to adopt)
+; BEDROCK NOTES  (referenced by the commented-out alternatives above)
 ;------------------------------------------------------------------------------
-; To swap one of the local GGUF tiers above for an AWS Bedrock model, COMMENT
-; OUT the existing tier of the same name and uncomment one of these. The router
-; auto-detects backend=bedrock from the presence of `aws_model_id` -- no other
-; flag needed. llama-swap won't load it; the router calls Bedrock directly via
-; boto3 (`pip install 'llmstack[bedrock]'`).
+; Each tier section above carries a "Bedrock alternative for <tier>" block
+; directly underneath it (commented out by default). To swap a tier:
+;
+;   1. comment out the active local section (GGUF by default);
+;   2. uncomment the Bedrock-alternative block beneath it;
+;   3. run `llmstack install` (and `llmstack restart` if the tier was
+;      already loaded -- bedrock creds aren't picked up live).
 ;
-; Credentials: this file ONLY names a profile. The actual keys / SSO /
-; role chaining live in the standard AWS config files. One-time setup:
+; The router auto-detects backend=bedrock from `aws_model_id`, but every
+; alternative block also sets `backend = bedrock` explicitly so the intent
+; is obvious. llama-swap won't load bedrock tiers; the router calls
+; Bedrock directly via boto3 (`pip install 'llmstack[bedrock]'`).
+;
+; PROFILE: every alternative uses `aws_profile = bedrock-prod`, the
+; conventional profile name for this stack. The actual keys / SSO /
+; role chaining live in the standard AWS config files (this file ONLY
+; names a profile -- never put credentials here). One-time setup:
 ;
 ;     aws configure --profile bedrock-prod
-;     # for SSO:    aws configure sso --profile bedrock-prod
-;     # for role chaining, edit ~/.aws/config and add a profile with:
+;     # SSO:           aws configure sso --profile bedrock-prod
+;     # role chaining: edit ~/.aws/config and add:
+;     #   [profile bedrock-prod]
 ;     #   role_arn       = arn:aws:iam::123456789012:role/llmstack-bedrock
-;     #   source_profile = bedrock-prod
+;     #   source_profile = bedrock-prod-base
+;
+; To use a different profile name, edit the `aws_profile` line. To fall
+; back on boto3's default chain (env vars, default profile, instance
+; role), remove the line entirely.
 ;
-; Then reference the profile name from your tier with `aws_profile = ...`.
-; If you omit `aws_profile`, boto3's default chain applies (env vars,
-; default profile, instance role -- whatever boto3 normally finds).
+; UPGRADE PRE-STAGING: optional `aws_model_id_next` (+ `aws_region_next`)
+; is the queued upgrade target -- mirrors gguf `hf_file_next`. The router
+; uses it only when `llmstack start --next` is in effect; permanent
+; promotion is the same as gguf: edit `aws_model_id` and re-run
+; `llmstack install`.
 ;
-; SAMPLER NOTE: the `sampler = temp=..., top_p=..., top_k=..., ...`
-; line on each tier is the SINGLE SOURCE OF TRUTH for sampling, but how
-; it gets applied depends on the backend:
+; SAMPLER: the `sampler = temp=..., top_p=..., top_k=..., ...` line on
+; each tier is the SINGLE SOURCE OF TRUTH for sampling, but how it gets
+; applied depends on the backend:
 ;
 ;   * gguf tiers -- the llama-swap generator bakes the sampler keys
 ;     into the llama-server startup command line as `--temp`,
@@ -248,57 +346,21 @@ ultra_triggers        = [ultra], [opus], "ultra:", "opus:" (line start)
 ; opencode.json is sampler-free in both cases by design (the
 ; opencode.json generator never emits sampler params on agents).
 ;
-; Per-Bedrock-family rules (as of 2026):
+; Per-Bedrock-family sampler rules (as of 2026):
 ;
 ;   * Claude Opus 4.7+ -- rejects all sampler params; OMIT `sampler =`
 ;     entirely (the router will then pass requests through untouched).
-;   * Claude Sonnet 4.5 / Haiku 4.5 -- accept `temp` OR `top_p`, never
-;     both; pick one.
+;   * Claude Sonnet 4.5 / 4.6 / Haiku 4.5 -- accept `temp` OR `top_p`,
+;     never both; pick one.
 ;   * Claude Opus 4.x (4.1, 4.5, 4.6) -- accept `temp` and `top_p`.
-;   * Llama / Titan / Cohere / etc. -- accept `temp` + `top_p`; check
-;     the model card if in doubt.
+;   * Llama / Titan / Mistral / Cohere / Nova / etc. -- accept `temp`
+;     + `top_p`; check the model card if in doubt.
 ;
-; Example A: top-tier coder on Bedrock (us-west-2), default cred chain.
-; Optional `aws_model_id_next` (and optional `aws_region_next`) is the
-; queued upgrade target -- mirrors gguf `hf_file_next`. The router uses
-; it only when `--next` is in effect; permanent promotion is the same
-; as gguf: edit `aws_model_id` and re-run `llmstack install`.
-;
-; [code-smart]
-; tier              = code
-; role              = agent
-; backend           = bedrock
-; aws_model_id      = anthropic.claude-sonnet-4-5-20250929-v1:0
-; aws_region        = us-west-2
-; aws_model_id_next = anthropic.claude-sonnet-5-20260201-v1:0   ; queued
-; aws_region_next   = us-east-1                                  ; (optional) different region for the new model
-; ctx_size          = 200000
-; sampler           = temp=0.5      ; Sonnet 4.5 accepts ONE of temp / top_p; pick `temp` for agent work
-; description       = Claude Sonnet 4.5 on Bedrock - heavy coder for agent loops
-;
-; Example B: planner in a different AWS account, accessed via a named
-; profile that itself uses role-chaining + SSO under ~/.aws/config.
-; (Different tier => different profile name; different account/region.)
-;
-; [plan]
-; tier         = chat
-; role         = plan
-; aws_model_id = us.anthropic.claude-opus-4-1-20250805-v1:0
-; aws_region   = us-east-1
-; aws_profile  = bedrock-planning
-; ctx_size     = 200000
-; sampler      = temp=0.7, top_p=0.9
-; description  = Claude Opus 4.1 on Bedrock - planning, design discussions
-;
-; Example C: large model behind a VPC endpoint.
-;
-; [plan-uncensored]
-; tier             = chat
-; role             = plan-uncensored
-; aws_model_id     = meta.llama3-1-405b-instruct-v1:0
-; aws_region       = us-west-2
-; aws_profile      = bedrock-prod
-; aws_endpoint_url = https://bedrock-runtime.us-west-2.vpce.amazonaws.com
-; ctx_size         = 128000
-; sampler          = temp=0.85, top_p=0.95
-; description      = Llama 3.1 405B on Bedrock - max-exploration planning
+; ACCESS: Anthropic Claude on Bedrock requires a one-time use-case-form
+; approval per AWS account (Bedrock console -> Model catalog -> pick the
+; model -> fill the form). Approval is account-level and persists; once
+; granted, every Claude variant works (bare ID, us./eu./global. cross-
+; region profile, application inference profile ARN). To skip the form
+; entirely, use the Llama 3.1 405B variant under [plan-uncensored] (Meta
+; models don't require the form) or pick another non-Anthropic family
+; (Amazon Nova, Mistral, Cohere, Titan).

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/llmstack/shell_env.py

@@ -261,14 +261,15 @@ _ZSH_HOOK = r"""# --- llmstack auto-activation hook (zsh) ----------------------
 #
 # Tool-availability gate: before activating, we verify the tools needed
 # for this channel are present:
-#   - `llmstack` (always required)
-#   - `llama-swap`     (only for local channels: current / next)
-#   - `llama-server` or `llama-cli` (likewise local-only)
-# external-mode projects skip the local-tool checks because llama-swap
-# and llama-server live on the remote. If any required tool is missing
-# we print "folder detected but tool not available" + install hints and
-# DON'T activate -- the env stays clean so opencode keeps using the
-# user's global config until they install the missing piece.
+#   - `llmstack`   (always required -- blocker)
+#   - `llama-swap` (only for local channels: current / next -- blocker)
+#   - `llama-server` / `llama-cli` (local-only, *warning* not blocker --
+#     a Bedrock-only models.ini activates fine without llama-server;
+#     local GGUF rows would fail to start, hence the heads-up)
+# external-mode projects skip all local-tool checks because llama-swap
+# and llama-server live on the remote. Blockers print "folder detected
+# but tool not available" + install hints and skip activation; warnings
+# print a one-shot hint and activate anyway.
 #
 # Marker file format (one line):
 #   <channel>[ <url>]
@@ -322,13 +323,18 @@ _llmstack_find_swap() {
 }
 
 _llmstack_check_tools() {
-    # Populates _llmstack_missing array. Returns 0 iff nothing is missing.
+    # Populates _llmstack_missing (blockers) and _llmstack_warnings
+    # (non-blockers). Returns 0 iff there are no blockers; warnings
+    # never block activation.
     _llmstack_missing=()
+    _llmstack_warnings=()
     command -v llmstack >/dev/null 2>&1 || _llmstack_missing+=("llmstack")
     if [[ "${1:-current}" != "external" ]]; then
         _llmstack_find_swap || _llmstack_missing+=("llama-swap")
+        # llama-server is a soft requirement: bedrock-only models.ini
+        # files don't need it, so missing == warn-and-continue.
         if ! command -v llama-server >/dev/null 2>&1 && ! command -v llama-cli >/dev/null 2>&1; then
-            _llmstack_missing+=("llama-server")
+            _llmstack_warnings+=("llama-server")
         fi
     fi
     (( ${#_llmstack_missing[@]} == 0 ))
@@ -343,7 +349,7 @@ _llmstack_install_hint() {
 }
 
 _llmstack_warn_missing() {
-    # $1 = project root; uses _llmstack_missing.
+    # $1 = project root; uses _llmstack_missing (blockers only).
     print -r -- ""
     print -P -- "%F{220}[llmstack]%f detected $1/.llmstack but missing local tool(s):"
     local t
@@ -354,6 +360,19 @@ _llmstack_warn_missing() {
     print -r -- ""
 }
 
+_llmstack_warn_optional() {
+    # $1 = project root; uses _llmstack_warnings. One-shot per activation
+    # (the LLMSTACK_WORK_DIR idempotency guard suppresses repeats).
+    print -r -- ""
+    print -P -- "%F{220}[llmstack]%f $1: activating without optional local tool(s):"
+    local t
+    for t in "${_llmstack_warnings[@]}"; do
+        _llmstack_install_hint "$t"
+    done
+    print -r -- "    bedrock-only models.ini works fine; local GGUF rows will fail to start."
+    print -r -- ""
+}
+
 _llmstack_deactivate() {
     if [[ -n "${LLMSTACK_WORK_DIR:-}" ]]; then
         unset OPENCODE_CONFIG LLMSTACK_WORK_DIR LLMSTACK_ACTIVE LLMSTACK_CHANNEL LLMSTACK_REMOTE_URL
@@ -393,12 +412,17 @@ _llmstack_activate() {
     fi
     : "${_ch:=current}"
 
-    # Tool gate -- bail before exporting anything if requirements aren't met.
+    # Tool gate -- bail before exporting anything if a *blocker* is
+    # missing. Non-blocking warnings (e.g. llama-server for a
+    # bedrock-only setup) print a hint but proceed.
     if ! _llmstack_check_tools "$_ch"; then
         _llmstack_warn_missing "$found"
         export _LLMSTACK_WARNED_FOR="$found"
         return 0
     fi
+    if (( ${#_llmstack_warnings[@]} > 0 )); then
+        _llmstack_warn_optional "$found"
+    fi
 
     export OPENCODE_CONFIG="$found/.llmstack/opencode.json"
     export LLMSTACK_WORK_DIR="$found"
@@ -441,14 +465,15 @@ _BASH_HOOK = r"""# --- llmstack auto-activation hook (bash) --------------------
 #
 # Tool-availability gate: before activating we verify the tools needed
 # for this channel are present:
-#   - `llmstack` (always required)
-#   - `llama-swap`     (only for local channels: current / next)
-#   - `llama-server` or `llama-cli` (likewise local-only)
-# If any required tool is missing we print a one-shot "folder detected
-# but tool not available" warning + install hints and DON'T activate
-# (env stays clean). The warning is suppressed on subsequent prompts in
-# the same project via the _LLMSTACK_WARNED_FOR guard so we don't spam
-# every PROMPT_COMMAND tick.
+#   - `llmstack`   (always required -- blocker)
+#   - `llama-swap` (only for local channels: current / next -- blocker)
+#   - `llama-server` / `llama-cli` (local-only, *warning* not blocker --
+#     a Bedrock-only models.ini activates fine without llama-server;
+#     local GGUF rows would fail to start, hence the heads-up)
+# Blockers print a one-shot "folder detected but tool not available"
+# warning + install hints and DON'T activate (env stays clean). Warnings
+# print a hint and activate anyway. The _LLMSTACK_WARNED_FOR guard
+# suppresses repeat warnings on subsequent PROMPT_COMMAND ticks.
 #
 # Marker file format (one line):
 #   <channel>[ <url>]
@@ -501,12 +526,17 @@ _llmstack_find_swap() {
 }
 
 _llmstack_check_tools() {
+    # Populates _llmstack_missing (blockers) and _llmstack_warnings
+    # (non-blockers). Returns 0 iff there are no blockers.
     _llmstack_missing=()
+    _llmstack_warnings=()
     command -v llmstack >/dev/null 2>&1 || _llmstack_missing+=("llmstack")
     if [[ "${1:-current}" != "external" ]]; then
         _llmstack_find_swap || _llmstack_missing+=("llama-swap")
+        # llama-server is a soft requirement: bedrock-only models.ini
+        # files don't need it, so missing == warn-and-continue.
         if ! command -v llama-server >/dev/null 2>&1 && ! command -v llama-cli >/dev/null 2>&1; then
-            _llmstack_missing+=("llama-server")
+            _llmstack_warnings+=("llama-server")
         fi
     fi
     [[ ${#_llmstack_missing[@]} -eq 0 ]]
@@ -530,6 +560,18 @@ _llmstack_warn_missing() {
     printf '    not activating. install the missing tool(s) and `cd` back in to retry.\n\n'
 }
 
+_llmstack_warn_optional() {
+    # $1 = project root; uses _llmstack_warnings. One-shot per
+    # activation thanks to LLMSTACK_WORK_DIR idempotency.
+    printf '\n'
+    printf '\033[38;5;220m[llmstack]\033[0m %s: activating without optional local tool(s):\n' "$1"
+    local t
+    for t in "${_llmstack_warnings[@]}"; do
+        _llmstack_install_hint "$t"
+    done
+    printf '    bedrock-only models.ini works fine; local GGUF rows will fail to start.\n\n'
+}
+
 _llmstack_deactivate() {
     if [[ -n "${LLMSTACK_WORK_DIR:-}" ]]; then
         unset OPENCODE_CONFIG LLMSTACK_WORK_DIR LLMSTACK_ACTIVE LLMSTACK_CHANNEL LLMSTACK_REMOTE_URL
@@ -565,11 +607,16 @@ _llmstack_activate() {
     fi
     : "${_ch:=current}"
 
+    # Blockers only -- non-blocking warnings (e.g. llama-server for a
+    # bedrock-only setup) print a hint but don't skip activation.
     if ! _llmstack_check_tools "$_ch"; then
         _llmstack_warn_missing "$found"
         export _LLMSTACK_WARNED_FOR="$found"
         return 0
     fi
+    if [[ ${#_llmstack_warnings[@]} -gt 0 ]]; then
+        _llmstack_warn_optional "$found"
+    fi
 
     export OPENCODE_CONFIG="$found/.llmstack/opencode.json"
     export LLMSTACK_WORK_DIR="$found"
@@ -618,12 +665,25 @@ _POWERSHELL_HOOK = r"""# --- llmstack auto-activation hook (PowerShell) --------
 #
 # Tool-availability gate: before activating we verify the tools needed
 # for this channel are present:
-#   - llmstack (always required)
-#   - llama-swap (only for local channels: current / next)
-#   - llama-server or llama-cli (likewise local-only)
-# If any required tool is missing we print a one-shot warning + install
-# hints and DON'T activate. The _LLMSTACK_WARNED_FOR guard suppresses
-# the warning on subsequent prompts in the same project.
+#   - llmstack   (always required -- blocker)
+#   - llama-swap (only for local channels: current / next -- blocker)
+#   - llama-server / llama-cli (local-only, *warning* not blocker --
+#     a Bedrock-only models.ini activates fine without llama-server;
+#     local GGUF rows would fail to start, hence the heads-up)
+# Blockers print a one-shot warning + install hints and DON'T activate.
+# Warnings print a hint and activate anyway. The _LLMSTACK_WARNED_FOR
+# guard suppresses repeats on subsequent prompts in the same project.
+#
+# Note: this hook file MUST be saved with a `.ps1` extension or
+# PowerShell won't dot-source it (it'll try to open the file via the
+# Windows shell file-association instead). `llmstack activate
+# powershell` writes ~/.powershell_llmstack_hook.ps1 for that reason.
+#
+# Note: PowerShell's default execution policy on Windows (`Restricted`)
+# blocks loading any .ps1 from disk. If you see "running scripts is
+# disabled on this system", allow signed local scripts once with:
+#     Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
+# (or `-ExecutionPolicy Bypass` for an even looser policy).
 #
 # Add to your $PROFILE (one time):
 #     llmstack activate powershell | Out-String | Invoke-Expression
@@ -691,16 +751,20 @@ function global:_LlmstackFindSwap {
 
 function global:_LlmstackCheckTools {
     param([string]$Channel)
-    $missing = @()
+    # Returns @{ missing = <blockers>; warnings = <non-blockers> }.
+    $missing  = @()
+    $warnings = @()
     if (-not (Get-Command llmstack -ErrorAction SilentlyContinue)) { $missing += "llmstack" }
     if ($Channel -ne "external") {
         if (-not (_LlmstackFindSwap)) { $missing += "llama-swap" }
+        # llama-server is a soft requirement: bedrock-only models.ini
+        # files don't need it, so missing == warn-and-continue.
         if (-not (Get-Command llama-server -ErrorAction SilentlyContinue) -and `
             -not (Get-Command llama-cli    -ErrorAction SilentlyContinue)) {
-            $missing += "llama-server"
+            $warnings += "llama-server"
         }
     }
-    return ,$missing
+    return @{ missing = ,$missing; warnings = ,$warnings }
 }
 
 function global:_LlmstackInstallHint {
@@ -723,6 +787,20 @@ function global:_LlmstackWarnMissing {
     Write-Host ""
 }
 
+function global:_LlmstackWarnOptional {
+    # Non-blocking: tool isn't on PATH but a bedrock-only models.ini
+    # would still work. One-shot per activation thanks to the
+    # LLMSTACK_WORK_DIR idempotency guard.
+    param([string]$Found, [string[]]$Warnings)
+    Write-Host ""
+    $esc = [char]27
+    Write-Host -NoNewline "${esc}[38;5;220m[llmstack]${esc}[0m "
+    Write-Host "${Found}: activating without optional local tool(s):"
+    foreach ($t in $Warnings) { Write-Host (_LlmstackInstallHint $t) }
+    Write-Host "    bedrock-only models.ini works fine; local GGUF rows will fail to start."
+    Write-Host ""
+}
+
 function global:_LlmstackDeactivate {
     if ($env:LLMSTACK_WORK_DIR) {
         Remove-Item Env:OPENCODE_CONFIG     -ErrorAction SilentlyContinue
@@ -751,12 +829,15 @@ function global:_LlmstackActivate {
     $marker  = if (Test-Path -LiteralPath $live) { _LlmstackReadMarker $live } else { _LlmstackReadMarker $intent }
     $channel = if ($marker.channel) { $marker.channel } else { "current" }
 
-    $missing = _LlmstackCheckTools $channel
-    if ($missing.Count -gt 0) {
-        _LlmstackWarnMissing $found $missing
+    $tools = _LlmstackCheckTools $channel
+    if ($tools.missing.Count -gt 0) {
+        _LlmstackWarnMissing $found $tools.missing
         $env:_LLMSTACK_WARNED_FOR = $found
         return
     }
+    if ($tools.warnings.Count -gt 0) {
+        _LlmstackWarnOptional $found $tools.warnings
+    }
 
     $env:OPENCODE_CONFIG   = Join-Path $found ".llmstack/opencode.json"
     $env:LLMSTACK_WORK_DIR = $found

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/opencode_llmstack.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: opencode-llmstack
-Version: 0.7.1
+Version: 0.7.3
 Summary: Multi-tier local LLM stack: llama-swap + FastAPI auto-router + opencode wiring.
 Author: llmstack
 License: MIT
@@ -313,10 +313,15 @@ py -3 -m venv .venv
 #    `start` falls back to spawning a PowerShell subshell.
 .venv\Scripts\llmstack start
 
-# 4. Auto-activate per project from any new PowerShell window:
-Invoke-Expression (& llmstack activate powershell | Out-String)
-# or persist (writes ~/.powershell_llmstack_hook + sources it on every shell):
-"Invoke-Expression (& llmstack activate powershell | Out-String)" | Add-Content $PROFILE
+# 4. Auto-activate per project from any new PowerShell window. The hook
+#    file is a .ps1 (PowerShell won't dot-source it without that
+#    extension) and dot-sourcing it requires script execution to be
+#    allowed -- if you see "running scripts is disabled on this
+#    system", run once:
+#       Set-ExecutionPolicy -Scope CurrentUser -ExecutionPolicy RemoteSigned
+llmstack activate powershell | Out-String | Invoke-Expression
+# or persist (writes ~/.powershell_llmstack_hook.ps1 + sources it on every shell):
+"llmstack activate powershell | Out-String | Invoke-Expression" | Add-Content $PROFILE
 ```
 
 Notes:

{opencode_llmstack-0.7.1 → opencode_llmstack-0.7.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "opencode-llmstack"
-version = "0.7.1"
+version = "0.7.3"
 description = "Multi-tier local LLM stack: llama-swap + FastAPI auto-router + opencode wiring."
 readme = "README.md"
 requires-python = ">=3.11"