Seed Code CLI v10.2.1

Seed Code CLI v10.2.1 — Release Notes

Version: 10.2.1 · Tag: v10.2.1 · Release date: 2026-11-04
Release page: https://github.com/Alshahriar-07/seedcode-cli/releases/tag/v10.2.1

v10.2.1 is the execution-reliability release. The model still makes the
decisions; SeedCode now owns the execution rules that keep a turn honest:
model-aware prompt/step budgets, a strict tool-call protocol with validation,
automatic continuation of truncated responses, a completion gate backed by
real evidence, and privacy-safe failover. The interface was also rebuilt
around a pure-black / Seed-Green identity with a full-width conversation and
an inline task plan.

What's new in v10.2.1

Agent execution reliability

• Model-aware execution profiles (seedcode/core/model_profiles.py):
  small, standard and advanced profiles control prompt style, step and
  tool-call budgets, output caps, history compaction and the preferred tool
  protocol. Resolution uses an explicit prompt_profile, then a user
  model_profiles map (exact names or globs), then local-provider parameter
  metadata; it never relies on a hardcoded model-name list. Local models with
  a declared small parameter size get the compact profile; everything else
  keeps the standard behaviour.

• Compact small-model context. A compact profile receives short,
  sequential instructions with explicit completion rules instead of the full
  coding-agent preamble. Standard and advanced models keep the existing,
  detailed context.

• Text tool protocol with TOOL_CALL: markers. Local models that cannot
  use native function calling emit:

  TOOL_CALL: {"name": "list_dir", "arguments": {"path": "src/"}}
  

  The parser also accepts the legacy fenced ```tool block and
  OpenAI-style name/arguments JSON. A marker inside a documentation/code
  fence is treated as an example, never executed.

• Strict pre-execution validation. Tool name, required arguments and
  integer argument types are checked before dedupe, permission checks or
  execution. Malformed JSON is reported to the model and never executed.

• No raw tool JSON in the transcript. Recognized native calls and text
  markers are removed from the assistant text shown to the user.

• Automatic response continuation. Providers now report their terminal
  finish reason (FinishEvent). When generation stops at the output-token
  limit, SeedCode preserves the generated prefix, asks the model to continue
  exactly where it stopped, and returns the concatenated answer. Continuation
  is bounded by the active profile. Text-protocol replies are also inspected
  for an unterminated fence or a call marker with no complete payload.

• Completion requires evidence. A reply that claims completion after a
  failed tool call is sent back for one verification step (bounded once per
  turn) instead of being accepted. The persistent session's existing
  evidence-based task verification is unchanged.

• Duplicate-call protection (carried from v10.1.0) still replays a
  previously answered provider call id rather than executing it twice.

Provider and model recovery

• Bounded exponential backoff with jitter for transient failures:
  approximately 1s, 3s, then 7s, capped at 30s. A provider's Retry-After
  hint is still honoured within that cap, and retries remain bounded.
• Model failover through the new fallback_models configuration: when the
  active model fails transiently, the next configured model on the same
  provider is tried before any provider switch, with the task, history and
  context preserved.
• Model management screen (/models). Switch from the live catalogue, use
  a custom model id when a gateway exposes no catalogue, maintain the ordered
  fallback list, clear the active model, and choose or pin an execution
  profile (globally or per model). Every action persists to config.json.
• Provider management screen (/provider manage). Select a provider, run
  a real connection test, inspect every visible provider's status/model/key
  state, toggle automatic failover, manage custom providers, and manage the
  active API key. /provider test and /provider failover on|off are direct
  shortcuts; the failover toggle updates the live engine immediately.
• Local-to-cloud privacy gate. A selected local provider (Ollama) is never
  silently moved to a cloud provider. Crossing that boundary requires the
  explicit allow_local_to_cloud opt-in (default false); the switch is
  announced when it happens.
• Provider failover remains opt-in through auto_failover.

Terminal interface

• Pure black background and Seed Green identity. The Seed palette now uses
  #22C55E (signature), #16A34A (secondary), #4ADE80 (bright),
  #FAFAFA (primary text), #A1A1AA / #71717A (secondary/muted),
  #27272A (border), #EF4444 (error), #F59E0B (warning) and #3B82F6
  (information). The prompt_toolkit base style sets a black background.
• Full-width conversation. The permanent right-side task panel and its
  reserved column are gone; the conversation uses the complete terminal width.
• Inline task progress. The live plan renders in the conversation timeline
  as Task Progress, with the existing real-state glyphs and verified-work
  progress. It is driven by the task graph, not by UI animation.
• The composer, streaming, permission panel, thinking strip, scrolling and
  resize handling are unchanged.

Configuration

New config.json fields (all optional):

Field	Default	Purpose
prompt_profile	"auto"	auto, small, standard or advanced.
model_profiles	{}	Per-model profile overrides ({"qwen3:4b": "small"}).
max_agent_steps	0	Explicit per-turn step ceiling (0 = profile).
max_tool_calls_per_step	0	Explicit calls per step (0 = profile).
continue_truncated	true	Continue output-limit responses.
fallback_models	[]	Model failover candidates for the active provider.
allow_local_to_cloud	false	Permit local→cloud failover explicitly.

Backend changes

• seedcode/core/model_profiles.py: profile definitions and resolution.
• seedcode/core/agent.py: compact prompt, marker parser, strict validation,
  visible-text filtering, continuation, completion nudge, profile budgets.
• seedcode/core/providers/base.py: FinishEvent stream event.
• seedcode/core/providers/{openrouter,ollama,freemodel,aerolink}.py: real
  finish reasons for output-limit continuation.
• seedcode/core/models.py: new configuration fields and the small-model
  output cap in effective_max_tokens.
• seedcode/core/chat.py: jittered bounded backoff, model failover and the
  local-to-cloud privacy gate.
• seedcode/ui/theme.py, seedcode/ui/tui.py: Seed palette, black base
  style, full-width layout and inline task progress.
• Version 10.2.1 across package metadata, npm launcher, IRM installers, Setup
  script and docs.

Installation

# Windows (recommended, per-user, no admin, verifies SHA-256)
irm https://seedcode-cli.vercel.app/install.ps1 | iex

# Linux / macOS
curl -fsSL https://seedcode-cli.vercel.app/install.sh | bash

# Any OS with Python 3.10+
pip install seedcode-cli

Verify afterwards:

seedcode --version    # -> Seed Code CLI 10.2.1

Verification

The full test suite is run with python -m pytest -q. New coverage in
tests/test_v102_agent_reliability.py locks in profile resolution, the
TOOL_CALL: protocol (including fenced examples being ignored), strict
argument validation, raw-JSON filtering, truncated-response continuation,
the completion gate, model/provider failover privacy, the palette and the
inline plan. tests/test_v102_management.py covers the /models command and
its fallback/profile helpers, the provider failover toggle, and finish-reason
parity for FreeModel Claude/Codex and AeroLink.

Known limitations

• No release artifacts were built or published as part of this change.
  Publishing is a separate, explicit step.
• The TypeScript terminal UI option was evaluated and not adopted in this
  release. The Python prompt_toolkit application was retained and extended;
  no second service or IPC layer was introduced.
• Agent, session, tool, theme, settings, configuration and help screens are
  unchanged; provider and model management now have dedicated real actions
  (/provider manage, /models).
• Cloud providers were not live-tested in this environment; their request
  paths remain covered by mocked provider tests.
• The inline plan is a live view of the current plan; it is not a separately
  scrollable transcript region.