CodeSeeq v0.3.4

v0.3.4 - 2026-07-07

Changed

• Version bump to v0.3.4.

v0.3.3 - 2026-07-07

Added

• CODESEEQ_REASONING_EFFORT environment variable. Controls the
  reasoning effort level forwarded to DeepSeek for thinking-enabled
  models. Accepted values: minimal, low, medium, high, xhigh,
  max. The bridge maps low/medium → high and xhigh → max
  to match DeepSeek's supported levels.
• Reasoning summary wiring in the launcher. When thinking is enabled,
  the codeseeq launcher now exports MODEL_SUPPORTS_REASONING_SUMMARIES=true
  and MODEL_REASONING_SUMMARY=auto, corresponding to the updated model
  catalog entries. When thinking is disabled, both are set to false/none.

Fixed

• Bridge reasoning-effort resilience. The bridge now handles missing or
  non-dict reasoning fields in thinking-enabled requests without crashing.
  When reasoning is absent, falls back to the CODESEEQ_REASONING_EFFORT
  environment variable, defaulting to an empty string (conservative). Also
  adds a truthiness guard so effort="" does not trigger the mapping logic.
• Thinking-model catalog entries updated. deepseek-v4-flash-thinking
  and deepseek-v4-pro-thinking now correctly declare
  supports_reasoning_summaries: true with reasoning_summary: "auto" and
  supported_reasoning_levels: [{level: "high"}, {level: "max"}]. This
  fixes Codex treating these models as non-thinking-capable.

Changed

• Model catalog formatting. config/codex-model-catalog.json entries
  reformatted for consistency (expanded inline-objects to multi-line).

---

v0.3.2 - 2026-06-29

Fixed

• tty: true forced on all shell/exec tool calls. The bridge now
  unconditionally sets tty: true on every shell, exec_command, bash,
  and related tool call before forwarding to Codex. This keeps stdin open
  for subsequent write_stdin calls, preventing "stdin is closed for this
  session" errors when the model pipes input to a running command. Non-
  interactive commands work fine with tty: true — they just complete
  normally.
• update_plan explanation stripped before Codex forwarding. Codex's
  Rust tool router rejects explanation at the top level of update_plan
  arguments as an unknown field (its struct doesn't include it despite the
  prompt mentioning it). The bridge now strips explanation from
  normalized arguments before forwarding, so Codex can parse the call
  without "unknown field explanation, expected step or status" errors.
• Tool steering instruction updated. The injected steering message now
  documents that exec_command always requires tty: true.

---

v0.3.1 - 2026-06-27

Added

• Uncensored mode (-U / --uncensored-mode). Loads the Dolfino (Dirty D)
  uncensored system prompt from config/uncensored.md for unrestricted interactions
  with no content filtering or moralizing guardrails. Also settable via
  CODESEEQ_UNCENSORED_MODE=true environment variable. The flag is safe to use
  alongside any other CodeSeeq flags and subcommands.
• Privacy hardening (default-on). Every generated Codex config now includes:
  • web_search = "live" — live web search enabled
  • [analytics] enabled = false — analytics disabled
  • [feedback] enabled = false — feedback disabled
  • [otel] exporter/metrics_exporter/trace_exporter = "none" — all OpenTelemetry
    pipelines disabled
  • [otel] log_user_prompt = false — raw prompt content not logged
  • [history] persistence = "none" — history persistence disabled
• Upstream Codex command blocking. Commands that contact OpenAI/ChatGPT services
  (login, logout, cloud, app, app-server, plugin, update, features,
  remote-control) are blocked by default with a clear error message. Override via
  CODESEEQ_ALLOW_UPSTREAM_CODEX_SERVICES=true.
• OPENAI_API_KEY no longer auto-populated from DEEPSEEK_API_KEY. Direct
  DeepSeek-only auth; no implicit key aliasing for OpenAI-shaped tooling.
• Codex version pinned to 0.130.0. Both Dockerfile (ARG CODEX_NPM_VERSION)
  and Makefile (CODEX_NPM_VERSION) default to a pinned release instead of latest.
  Installer no longer auto-fetches @openai/codex@latest without
  CODESEEQ_ALLOW_LATEST_RELEASE=true.

Fixed

• Bash 3.2 compatibility for empty array expansion. The _CODESEEQ_KEPT_ARGS
  array expansion at end of flag pre-parse loop now uses the ${array[@]+"${array[@]}"}
  pattern to avoid "unbound variable" errors on macOS (bash 3.2) and other older
  shells. Fixes ./codeseeq --uncensored-mode and other invocations that consume
  CodeSeeq-specific flags without passing any remaining arguments to Codex.
• Uncensored mode recursion guard. The uncensored mode code now sets
  CODESEEQ_UNCENSORED_DONE=true and unsets CODESEEQ_UNCENSORED_MODE in the
  child process to prevent infinite recursion when "$SELF_PATH" system add -f
  re-enters the launcher with the same env var set.

Changed

• Installer no longer auto-installs @openai/codex. The install subcommand
  no longer runs npm install -g @openai/codex. Users install Codex manually with
  npm install -g @openai/codex@0.130.0 (pinned version shown in the error message).
• Host runtime Codex version check prints pinned version.
  run_host_codex() now tells users to install @openai/codex@0.130.0 instead of
  the unversioned package.
• check.sh extended. New assertions validate privacy hardening config content,
  blocked upstream commands, pinned CODEX_NPM_VERSION in both Dockerfile and
  Makefile, and that OPENAI_API_KEY is not auto-exported from DEEPSEEK_API_KEY.

---

v0.2.9 - 2026-05-12

Fixed

• Flags after run now parsed before runtime dispatch.
  ./codeseeq run --runtime-mode host --bridge-mode process "hello" correctly routes to host runtime.
• Host diagnostics no longer require container.
  models, doctor, and config work without Docker/Podman in host runtime mode.
• Process bridge cleanup verified. Owned bridges stopped on all exit paths (EXIT/INT/TERM/HUP).
• Prompt temp files no longer leak. TMP_FILES properly tracked after command substitution.
• External mode supports /v1/ trailing slash.
• 0.0.0.0 bridge host now writes 127.0.0.1 in client config (warning printed).
• make bridge-process-smoke no longer leaks processes. POSIX-safe script with cleanup trap.
• Container bridge smoke updated with correct model IDs and bridge host binding.
• Doctor output cleaned up — "Bridge URL" instead of "OpenResponses", runtime/bridge mode fields.
• Package hygiene strengthened — make clean-artifacts, make package-check, docs warn against manual zips.
• No Codex source modified. No upstream open-responses runtime dependency.

Added

• Host-native bridge process mode. CODESEEQ_BRIDGE_MODE=process starts
  bin/codeseeq-bridge.py as a direct child process on the host with no
  Docker or Podman required. All four bridge modes are supported: process,
  container, external, and auto (default, prefers process when Python
  dependencies are available, falling back to container).
• Host runtime mode. CODESEEQ_RUNTIME_MODE=host launches Codex directly
  on the host alongside a process bridge, without going through the container
  runtime at all. container and auto (default) remain available.
• External bridge mode. CODESEEQ_BRIDGE_MODE=external with
  CODESEEQ_BRIDGE_BASE_URL lets CodeSeeq talk to a pre-existing bridge
  without starting anything locally.
• Bridge reuse. CODESEEQ_BRIDGE_REUSE=1 causes CodeSeeq to check for a
  healthy bridge at the configured port and reuse it instead of starting a
  new one.
• CLI flags for bridge configuration. --bridge-mode, --bridge-url,
  and --bridge-port are now accepted on the codeseeq launcher and the
  run subcommand so bridge mode can be set per-invocation without
  environment variables.
• Unified CI release pipeline. Release step merged into ci.yml — runs
  only on tag pushes (v*) after all checks pass.

Changed

• Dockerfile no longer pulls open-responses npm dependency. The
  upstream open-responses package was removed from the container image
  build. The actual runtime bridge is entirely bin/codeseeq-bridge.py.
• Bridge mode architecture rewrite. The launcher now has a unified
  bridge_start() abstraction that selects between process, container,
  external, and auto modes with consistent health-check and cleanup
  behavior. Process-mode owned bridges are stopped on launcher exit.

Notes

• No Codex source was modified. The bridge remains a drop-in
  Responses-compatible API that Codex talks to exactly as it would talk to
  any OpenAI-compatible provider.
• wire_api = "responses" in the generated Codex config stays because Codex
  expects that value; it does not mean the upstream open-responses package
  is used.

---

0.2.8 - 2026-05-12

Fixed

• Version bump and release process improvements.

---

0.2.7 - 2026-05-08

Added

• codeseeq nuke subcommand. Uninstalls all CodeSeeq user artifacts — the
  installed snapshot at ~/.config/codeseeq, the launcher at ~/bin/codeseeq,
  and any leftover ~/.codeseeq state — with a confirmation prompt before
  removal. The local repo checkout and container images are left untouched.

Changed

• README.md description rewritten. The first paragraph now describes
  CodeSeeq as a "drop-in launcher and CLI wrapper" rather than a "command
  switch", better reflecting its role as a full launcher/substitute that
  manages container runtime, bridge lifecycle, configuration, and DeepSeek
  model wiring.
• CodeSeeq splash image. The codeseeq.jpg image was added to the README
  between the introductory paragraph and the version/release-notes section.
• Version bumped to 0.2.7. Updated VERSION, README.md, and all
  doc-version references in docs/ARCHITECTURE.md, docs/SECURITY.md, and
  docs/TROUBLESHOOTING.md.

Fixed

• Typo --sanbox in README.md and doc examples. The documented alias
  examples in README.md, docs/TROUBLESHOOTING.md, and docs/SECURITY.md
  were using --sanbox instead of the correct --sandbox. Fixed to show the
  proper flag spelling while keeping --sanbox as an accepted internal alias.

---

0.2.6 - 2026-05-08

Added

• Container-launcher rewrite. The ./codeseeq launcher was substantially
  rewritten with robust configuration defaults, helper functions (warn,
  bool_normalize), and expanded environment-variable plumbing for
  CODESEEQ_MODEL, CODESEEQ_THINKING, CODESEEQ_APPROVAL_POLICY,
  CODESEEQ_SANDBOX_MODE, CODESEEQ_OPENRESPONSES_PORT,
  CODESEEQ_OPENRESPONSES_URL, CODESEEQ_CONTEXT_WINDOW,
  CODESEEQ_HOST_CODEX_HOME, and CODESEEQ_SYSTEM_PROMPT_FILE.
• System prompt injection. New CODESEEQ_SYSTEM_PROMPT_FILE (default
  ${WORKDIR}/.codeseeq/system-prompt.md) is read and injected into Codex's
  TOML config as a quoted string. Helper functions (system_prompt_present,
  system_prompt_state, system_prompt_bytes, system_prompt_lines) report
  prompt state at startup.
• Workspace banner. Entrypoint prints a summary banner showing the
  workspace path, version, model, approval policy, sandbox mode, key
  configuration hash, and system-prompt state on each launch.
• .env.example template. Documented all supported environment variables
  with their defaults, so users can copy .env.example to .env and customize.
• Expanded smoke-test suite. scripts/smoke-all.sh now runs container
  smoke tests and host-cli smoke tests. New scripts/runtime.sh checks for
  container and GPU host capabilities.
• Bridge binary on codeseeq-bridge.py. The bridge now lives at its own
  path (bin/codeseeq-bridge.py) in the container, launched side-by-side with
  Codex rather than being embedded.

Changed

• Launcher becomes dual-purpose. ./codeseeq now detects the install
  subcommand automatically and delegates to scripts/install-local.sh. Running
  without subcommand starts the container with all configuration variables
  forwarded.
• CODESEEQ_WORKDIR_HOST now resolves symlinks. Uses pwd -P instead of
  plain $PWD so bind-mount paths are canonical.
• Documentation overhaul. README.md, docs/ARCHITECTURE.md,
  docs/SECURITY.md, and docs/TROUBLESHOOTING.md were rewritten with
  up-to-date configuration references, container-runtime instructions, and
  security/architecture guidance.
• Scripts polished. scripts/check.sh extended with bridge-extraction
  regression tests; scripts/package.sh streamlined; scripts/install-local.sh
  updated for the new launcher layout.

Fixed

• codeseeq binary made executable in-repo. The root codeseeq file now
  has the executable bit set so it runs directly without bash codeseeq.

---

0.2.5 - 2026-05-07

Fixed

• Fixed split display-mangled DSML such as <____DSML____tool_calls>...
  leaking into the Codex UI after a successful tool call. The streaming buffer
  now normalizes obfuscated DSML after chunk reassembly, so the block is either
  converted into a tool call or suppressed instead of being shown as assistant
  text.
• Added missing output_index metadata to streamed message/tool lifecycle
  events so current Codex builds keep output deltas attached to their active
  items instead of logging orphaned OutputTextDelta diagnostics.
• Fixed Responses top-level function tools being collected for steering but not
  forwarded to DeepSeek's nested Chat Completions tools shape. This keeps
  DeepSeek able to emit actual structured tool calls instead of plain bash
  snippets.
• Updated README, quickstart, state docs, bridge docs, and CI build metadata to
  reflect the current single-container local-bridge runtime.
• Added workspace/ to .gitignore so the local repro clone does not break
  git add ..

---

0.2.4 - 2026-05-07

Fixed

• Fixed regular danger-full-access launches emitting both --ask-for-approval ... and --dangerously-bypass-approvals-and-sandbox. Codex rejects that
  combination. The launcher now omits --ask-for-approval whenever it emits the
  bypass flag.

---

0.2.3 - 2026-05-07

Changed

• codeseeq --yolo and codeseeq -y now only add Codex launch switches
  --dangerously-bypass-approvals-and-sandbox and, for codex exec paths,
  --skip-git-repo-check.
• Yolo mode no longer injects --ask-for-approval never, no longer injects
  --sandbox danger-full-access, and no longer rewrites
  CODESEEQ_APPROVAL_POLICY / CODESEEQ_SANDBOX_MODE config values.

---

0.2.2 - 2026-05-07

Added

• codeseeq --yolo and codeseeq -y wrapper flags. They force
  CODESEEQ_APPROVAL_POLICY=never and CODESEEQ_SANDBOX_MODE=danger-full-access,
  and launch Codex with --ask-for-approval never plus --sandbox danger-full-access
  and --dangerously-bypass-approvals-and-sandbox.
• Direct run/prompt shortcuts keep using codex exec --skip-git-repo-check.
  codeseeq --yolo codex exec ... also injects --skip-git-repo-check when it is
  not already present.

---

0.2.1 - 2026-05-07

Malformed XML compatibility patch for DeepSeek tool-use output.

Fixed

• Recognizes model-invented outer tool tags such as
  <exec_command><command>...</command></exec_command>, <bash>...</bash>,
  and <tool_call name="...">...</tool_call> as real Codex function calls
  instead of streaming them as assistant text.
• Extends streaming buffering to hold those malformed tags until the closing
  tag arrives, preventing visible XML leakage in the Codex UI.
• Normalizes common XML argument aliases against the registered Codex tool
  schema, including command -> cmd for exec_command/unified shell tools.
• Adds focused bridge extraction regression coverage and wires it into
  scripts/check.sh.

---

0.2.0 - 2026-05-07

DSML/tool-calling correctness pass for bin/codeseeq-bridge.py. The bridge now
properly streams tool calls to Codex CLI, normalizes display-mangled DSML in
history, and remaps emitted tool names onto whatever the client actually
registered. Drop-in replacement; no changes required to Codex, Dockerfile,
entrypoint, or scripts.

Fixed

• DSML leakage during streaming. Raw <function_calls>...</function_calls>
  XML was being streamed verbatim to the Codex TUI before the post-stream
  extractor ran. Replaced with StreamingDsmlBuffer that detects DSML inline,
  emits only safe text deltas, and surfaces tool-call blocks as soon as their
  closing tag is seen. Buffer uses depth tracking so a nested </invoke> inside
  an outer <function_calls> wrapper does not terminate prematurely.
• call_id: None on response.output_item.added. The added event was fired
  before the tool name and call id were known, then back-filled. Now deferred
  until the call has a real name and id, so Codex never sees a partial item.
• Wrong delta event for function tools. Used
  response.custom_tool_call_input.delta for function-typed tools; modern Codex
  listens on response.function_call_arguments.delta. Both are now emitted
  (modern + legacy) so older Codex builds keep working.
• Missing response.function_call_arguments.done. Now emitted, in the proper
  place in the lifecycle.
• Broken DSML extraction lifecycle. Post-stream DSML extraction emitted only
  output_item.done. Now emits the full sequence: output_item.added ->
  function_call_arguments.delta -> function_call_arguments.done ->
  output_item.done (plus legacy custom_tool_call_input.delta).
• Duplicate response.completed. Deduplicated to a single emission.
• Display-mangled DSML in history. Codex's TUI obfuscates < to
  <____DSML____ for safe display. When that text fed back as history, DeepSeek
  imitated the malformed format. Added normalize_dsml_display() applied to ALL
  inbound message content so the model only ever sees clean XML or, ideally,
  structured tool_calls.

Added

• Tool-name aliasing. Flat TOOL_NAME_ALIASES map — emitted name -> ordered
  tuple of preferred replacements. resolve_tool_name() does exact ->
  case-insensitive -> alias-prefs (only those actually registered) -> fuzzy
  (difflib, cutoff 0.7) -> first preference fallback. Common variants covered:
  bash/sh/execute_command/exec_command/run_command -> shell;
  write/write_file/create_file -> apply_patch/write_file;
  edit/patch/str_replace_editor -> str_replace/apply_patch;
  read_file/view_file/cat -> view; etc. Toggle with env
  CODESEEQ_BRIDGE_TOOL_ALIAS_FUZZY (default on).
• Tool-use steering system message. When tools are present in the request, a
  small system message is injected telling the model to emit structured
  tool_calls rather than XML. Toggle via env CODESEEQ_BRIDGE_TOOL_STEERING
  (default on).
• Stricter error handling for upstream stream. httpx.RemoteProtocolError,
  httpx.ReadError, and asyncio.CancelledError are caught separately so the
  bridge logs and surfaces the right SSE error type rather than 500ing.

Notes

• No schema changes to /v1/responses, /v1/models, or /health.
• No changes to Codex configuration, Dockerfile, container entrypoint, or smoke
  scripts; behavior is fully on the bridge side.
• CODESEEQ_BRIDGE_DEBUG_LOG=1 continues to dump full request/response payloads
  to /tmp/codeseeq-bridge.log for diagnostics.

---

0.1.0 - 2026-05-07

• Initial public version of CodeSeeq single-container CLI workflow.
• Added root ./codeseeq launcher and container entrypoint path.
• Added DeepSeek/OpenResponses bridge runtime wiring and smoke scripts.
• Added VERSION file with starting semantic version.
• Switched license from AGPL-3.0 to Apache 2.0 (updated LICENSE, COPYRIGHT,
  README.md).