Seed Code CLI v9.3.2

Seed Code CLI v9.3.2 — Release Notes

Version: 9.3.2 · Tag: v9.3.2 · Release date: 2026-11-03
Release page: https://github.com/Alshahriar-07/seedcode-cli/releases/tag/v9.3.2

v9.3.2 is a reliability and interface-polish release on top of the v9.3.x agent
architecture: real workspace execution is now enforced by evidence, the
terminal interface is quieter, and a timestamp that appeared beside AI
responses is gone.

What's new in v9.3.2

• Minimal thinking display. The bottom strip is now one small dim
  Thinking... label followed by the reasoning summaries the connected model
  or provider actually exposes. The former full-width "LIVE AI BRIEFING"
  dashboard — heading, model/provider identity, status fields, LIVE badges,
  metadata — was removed from the terminal. The strip:
  • hides itself entirely when there is no reasoning to show;
  • keeps a compact bounded height and never squeezes a short terminal;
  • follows the newest update automatically, while a manual scroll-up keeps
    the reader's position;
  • preserves completed reasoning after the turn (Ctrl+B toggles it);
  • records the honest line Reasoning stream unavailable for this model.
    when the provider exposes nothing — reasoning is never fabricated.
    The execution-status object still exists as internal backend state (phase,
    task, activity, retry, error) for logs and debugging; it is simply no longer
    rendered into the conversation.
• No timestamp beside AI responses. The activity log no longer prefixes
  each line with a bracketed [HH:MM:SS] wall-clock stamp. That prefix,
  printed between streamed assistant output during a turn, is what appeared as
  a timestamp next to AI responses. Timestamps remain in the backend
  ActivityEvent record for sorting, logs and session management; activity
  lines keep their real state markers and messages.
• Guaranteed real workspace execution. A task that must modify files can
  no longer complete on a printed claim. Completion is gated on observed
  evidence — the target file really created/changed, or the command really
  run — and the session re-instructs the agent with the unmet criteria when
  only text was produced. Read-only tasks (research, explanation, review)
  verify from real succeeded-inspection evidence instead, so the gate
  understands intent rather than forcing writes.
• Explicit execution rules in the system prompts. Code belongs in files; a
  printed tool call is not an executed tool call; chat output is not file
  output; verify before reporting; never claim success that did not happen.
  These rules are in both the native JSON-tool and text-protocol preambles.
• File tools hardened.
  • read_file refuses probable binaries (NUL probe) with actionable
    guidance instead of feeding mojibake into the model; line ranges remain
    available for large files.
  • All writes (write_file, edit_file, insert_in_file, append_file)
    are atomic — temp file → flush → fsync → replace — so an interruption
    cannot leave a truncated file, and a probable binary target is refused
    rather than overwritten with text.
  • edit_file gained an explicit replace_all for several independent
    occurrences; the default stays exactly-one with ambiguous targets
    rejected, and unrelated content is preserved.
  • search_text gained ignore_case and exclude; path globs are
    path-aware (src/*.py is one level, src/**/*.py recurses — fnmatch
    no longer lets * cross separators); the workspace walker honors
    .gitignore.
• TODO / task orchestration improvements carried in this release line.
  Hierarchical plans keep their nesting, dependencies schedule
  dependency-first, repeated failures cannot be marked complete, and the
  right-side panel always reflects real backend task state.

Carried over from v9.3.1 / v9.2.x

• One unified agent with a deterministic intent classifier: every request
  is one agent turn (no Chat/Agent/Assist/Code modes).
• Clean activity log with real markers (⚙ started, ✓ succeeded,
  ✗ failed, ℹ informational) and a spinner that shows the current
  operation with a live elapsed timer.
• Real terminal and file reporting: exit codes, command output, file
  diffs, one operation per line.
• Right-side task plan panel driven only by orchestrator task events.
• Full tool surface including create_file, copy_file, run_tests,
  get_command_output, check_exit_code, inspect_process, and the
  contract aliases list_directory / search_files / search_content.
• Ollama, OpenRouter, FreeModel, AeroLink, default and custom
  OpenAI-compatible providers, each with its own key and model.

Backend changes

• verify_task takes a read-only intent switch; the session classifies a task
  as read-only only when it has no modify verbs and at least one research
  verb. Mutating intent is the default — enforcement stays strict for
  anything ambiguous.
• A file-modification task whose evidence contains only reads is rejected
  with an explicit unmet criterion; the claim alone can never complete it.
• File edits and writes are atomic and fsynced; a locked target fails with an
  actionable error and leaves no partial file or temp residue behind.
• The search walker applies a conservative .gitignore subset (names, paths,
  directory rules; negations skipped) to reduce noise without hiding source.
• Version 9.3.2 across package metadata, npm launcher, IRM installers, Setup
  script and docs.

UI changes

• Bottom strip: Thinking... + dim reasoning, directly above the composer.
  No heading, no model/provider metadata, no LIVE badge, no status fields.
• Activity lines no longer carry a bracketed wall-clock prefix.
• The right-side TODO panel, the left chat/activity pane, the composer and the
  hint row are unchanged; resizing still re-fits every region without
  clipping.

Provider changes

• Thinking streams remain provider-specific: Ollama's thinking capability
  (probed via GET /api/show) and OpenAI-compatible reasoning /
  reasoning_content / reasoning_details deltas surface as
  ThinkingDelta events. A failed request is reported as a failure;
  automatic failover stays opt-in (auto_failover, off by default).

Bug fixes

• Bracketed [HH:MM:SS] timestamp appearing next to AI responses (activity
  log prefix removed; the response renderer itself never added one).
• Large "LIVE AI BRIEFING" dashboard with model/provider/status metadata
  replaced by the minimal thinking strip.
• Binary files decoded into model context and binary overwrites corrupting
  files (both refused).
• Non-atomic writes that could leave a truncated file when interrupted.
• search_text globs matching across path separators (src/*.py matching
  src/gen/x.py).
• File-modification tasks completing on a printed claim with no file
  operation.

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

Windows users can also download SeedCode-CLI-Setup-9.3.2.exe (Setup wizard) or
SeedCode-CLI-9.3.2-windows-x64.exe (portable) directly from the release page.

Verify afterwards:

seedcode --version    # -> Seed Code CLI 9.3.2

Release artifacts

File	Size (bytes)	Description
seedcode_cli-9.3.2-py3-none-any.whl	see SHA256SUMS.txt	Python wheel (any OS)
seedcode_cli-9.3.2.tar.gz	see SHA256SUMS.txt	Python source distribution
SeedCode-CLI-9.3.2-windows-x64.exe	see SHA256SUMS.txt	Standalone Windows x64 executable
seedcode-cli-9.3.2.tgz	see SHA256SUMS.txt	npm package tarball (launcher)
SeedCode-CLI-Setup-9.3.2.exe	see SHA256SUMS.txt	Windows Setup installer (Inno Setup)
SHA256SUMS.txt	—	SHA-256 of all five artifacts above

SHA-256 verification

cd dist/release/v9.3.2
sha256sum -c SHA256SUMS.txt

# Windows PowerShell
Get-FileHash .\SeedCode-CLI-9.3.2-windows-x64.exe -Algorithm SHA256

Both remote installers download SHA256SUMS.txt and abort on a missing entry,
an unreachable checksum file, or a mismatch. Verification is never bypassed.

Known limitations

• Cloud providers were not live-tested in this environment. No valid cloud
  keys were available; their request/response paths are covered by the mocked
  provider suite.
• No prebuilt Linux/macOS binary is published in this release; the shell
  installer uses the verified wheel.
• The npm tarball is a launcher, not a self-contained runtime: it downloads
  the official binary from the release on first run.
• The IRM installers were not executed against a published GitHub release
  because publishing is not authorized; their logic is covered by the
  installer test suites.
• The GitHub release has not been published by this build. All artifacts
  are prepared and verified locally in dist/release/v9.3.2/; publishing is a
  separate, explicit step.