CodeSeeq v0.4.0

v0.3.9 - 2026-08-17

Fixed

• Parallel codeseeq now binds ports reliably. The bridge no longer relies
  on a connect-probe (TOCTOU race) to pick a port. When CODESEEQ_BRIDGE_PORT
  is unset it performs a real bind() starting at CODESEEQ_OPENRESPONSES_PORT
  and increments up to CODESEEQ_OPENRESPONSES_PORT_SCAN_LIMIT, writing the
  chosen port to CODESEEQ_BRIDGE_PORT_FILE. Multiple simultaneous invocations
  now obtain distinct ports without crashing on "address already in use".
• Decoupled container bridge ports. The standalone bridge container now
  binds a fixed internal port (CODESEEQ_BRIDGE_CONTAINER_PORT, default 8080)
  and is mapped to the auto-selected host port, eliminating the
  host-port == container-port collision.
• CODESEEQ_STREAM_IDLE_TIMEOUT_MS is now actually enforced. The bridge
  applies it as an idle (between-chunk) read timeout on streaming responses, so
  a stalled upstream stream or client can no longer hang a uvicorn task
  indefinitely (a timeout now yields a response.failed event).
• Per-model env overrides now work for every model. The bridge previously
  generated per-model env keys from the full provider@model slug
  (e.g. CODESEEQ_QWIBUS_QWIBUS_QWIKK_BASE_URL instead of the documented
  CODESEEQ_QWIBUS_QWIKK_BASE_URL), so qwibus (and other) per-model overrides
  were silently ignored. The provider prefix is now stripped, and explicit
  env/config values correctly take precedence over the catalog.
• Container runtime no longer drops qwibus/provider config. The env
  collection now forwards generic base-URL overrides and all per-model
  CODESEEQ_<MODEL>_* knobs into the Codex container, so qwibus endpoints can
  be configured in container mode.
• Removed dead session-tracking code (_active_sessions,
  _session_is_known, _prune_expired_sessions, CODESEEQ_SESSION_TTL_SECONDS)
  which had no callers.

Added

• JSON configuration alternative. Every environment variable can now be
  supplied through a JSON config file (keys are the literal env-var names).
  Precedence is environment variable > JSON config > built-in default. Config
  path is CODESEEQ_CONFIG_JSON (default ~/.config/codeseeq/config.json,
  or /home/codeseeq/.config/codeseeq/config.json in the container).
• Catalog clarification. config/model-catalog.json (bridge) and
  config/codex-model-catalog.json (Codex TUI) are documented as two distinct,
  both-used catalogs; neither is redundant.

Changed

• Health endpoint reports version 0.3.9.
• Version bump to v0.3.9. All docs, configs, and internal version strings
  updated.

---

v0.3.8 - 2026-08-17

Added

• Two new qwibus models. qwibus-qwikk (quick, no reasoning) and
  qwibus-qmplx (heavy reasoning/thinking) run against a local
  OpenAI-compatible gateway (default http://127.0.0.1:1337) and require no
  API key.
• Per-model endpoint + sampling configuration. The bridge now supports
  per-model base_url, chat_url, temperature, top_p, top_k,
  context_window, max_output_tokens, timeout_seconds, enable_thinking,
  and system_prompt. Defaults are sourced from config/model-catalog.json
  (single source of truth), with per-model CODESEEQ_<KEY>_* env overrides and
  generic OPENAI_BASE_URL / DEEPSEEK_BASE_URL / CODESEEQ_BASE_URL
  fallbacks.
• DeepSeek per-model defaults. deepseek-v4-flash timeout 60s,
  deepseek-v4-pro 120s, deepseek-v4-flash-thinking 600s, and
  deepseek-v4-pro-thinking 1200s, all with 1M context and 384K max output.
  enable_thinking is true on the two thinking variants.

Changed

• Health endpoint reports version 0.3.8.
• Version bump to v0.3.8. All docs, configs, and internal version strings
  updated.

---

v0.3.7 - 2026-07-14

Changed

• Venice image backend auto-detection. When VENICE_API_KEY is set,
  CodeSeeq automatically enables the Venice backend without requiring
  CODESEEQ_IMAGE_BACKEND=venice. The launcher, entrypoint, and bridge
  all auto-detect the key and activate Venice. Explicit CODESEEQ_IMAGE_BACKEND
  still works as an override.
• Default image output to current directory. bin/codeseeq-venice-image.py
  now saves generated images to the current working directory instead of a
  codeseeq-images/ subdirectory.
• ping-image no longer requires explicit CODESEEQ_IMAGE_BACKEND=venice.
  Setting VENICE_API_KEY is sufficient; auto-detection handles the rest.

Removed

• OpenAI image_gen skill deleted. The .codeseeq/skills/.system/imagegen/
  skill (OpenAI image_gen tool, GPT image models, OPENAI_API_KEY dependency)
  has been completely removed. CodeSeeq uses Venice.ai exclusively for image
  generation. No OpenAI API keys or endpoints are ever contacted for images.

Changed

• Health endpoint now reports version 0.3.7, shows effective image backend
  (including auto-detected venice), and Venice API key state.
• Version bump to v0.3.7. All docs, configs, and internal version strings
  updated.

---

v0.3.6 - 2026-07-14

Added

• Venice.ai image generation backend. New CODESEEQ_IMAGE_BACKEND configuration
  option (default none). Set to venice to enable image generation via the
  Venice.ai API using VENICE_API_KEY.
  • Supports all Venice /image/generate parameters: model selection (auto or
    specific models like z-image-turbo, gpt-image-2, nano-banana-pro),
    aspect ratio, resolution (1K/2K/4K), format (jpeg/png/webp), variants (1–4),
    safe mode, watermark control, negative prompts, seed, and CFG scale.
  • New standalone script: bin/codeseeq-venice-image.py for direct CLI usage
    without Codex.
  • New bridge endpoint: POST /v1/images/generations for OpenAI-compatible
    image generation proxied through Venice.
  • New ping-image diagnostic command.
  • New --image-backend and --venice-image-model CLI flags.
  • Doctor output now includes image backend status and Venice API key state.
  • Health endpoint now reports image_backend status.
• Comprehensive environment variable configuration for all Venice image
  parameters: CODESEEQ_VENICE_IMAGE_MODEL, CODESEEQ_VENICE_IMAGE_ASPECT_RATIO,
  CODESEEQ_VENICE_IMAGE_RESOLUTION, CODESEEQ_VENICE_IMAGE_FORMAT,
  CODESEEQ_VENICE_IMAGE_VARIANTS, CODESEEQ_VENICE_IMAGE_SAFE_MODE,
  CODESEEQ_VENICE_IMAGE_HIDE_WATERMARK, CODESEEQ_VENICE_IMAGE_SEED,
  CODESEEQ_VENICE_IMAGE_RETURN_BINARY, CODESEEQ_VENICE_IMAGE_NEGATIVE_PROMPT,
  CODESEEQ_VENICE_IMAGE_CFG_SCALE, CODESEEQ_VENICE_IMAGE_STEPS.

Changed

• Version bump to v0.3.6.
• Health endpoint now reports image_backend status and Venice configuration.
• Dockerfile now includes bin/codeseeq-venice-image.py.
• All documentation updated with image backend section and configuration reference.

---

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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 us...

CodeSeeq v0.3.9

v0.3.9 - 2026-08-17

Fixed

• Parallel codeseeq now binds ports reliably. The bridge no longer relies
  on a connect-probe (TOCTOU race) to pick a port. When CODESEEQ_BRIDGE_PORT
  is unset it performs a real bind() starting at CODESEEQ_OPENRESPONSES_PORT
  and increments up to CODESEEQ_OPENRESPONSES_PORT_SCAN_LIMIT, writing the
  chosen port to CODESEEQ_BRIDGE_PORT_FILE. Multiple simultaneous invocations
  now obtain distinct ports without crashing on "address already in use".
• Decoupled container bridge ports. The standalone bridge container now
  binds a fixed internal port (CODESEEQ_BRIDGE_CONTAINER_PORT, default 8080)
  and is mapped to the auto-selected host port, eliminating the
  host-port == container-port collision.
• CODESEEQ_STREAM_IDLE_TIMEOUT_MS is now actually enforced. The bridge
  applies it as an idle (between-chunk) read timeout on streaming responses, so
  a stalled upstream stream or client can no longer hang a uvicorn task
  indefinitely (a timeout now yields a response.failed event).
• Per-model env overrides now work for every model. The bridge previously
  generated per-model env keys from the full provider@model slug
  (e.g. CODESEEQ_QWIBUS_QWIBUS_QWIKK_BASE_URL instead of the documented
  CODESEEQ_QWIBUS_QWIKK_BASE_URL), so qwibus (and other) per-model overrides
  were silently ignored. The provider prefix is now stripped, and explicit
  env/config values correctly take precedence over the catalog.
• Container runtime no longer drops qwibus/provider config. The env
  collection now forwards generic base-URL overrides and all per-model
  CODESEEQ_<MODEL>_* knobs into the Codex container, so qwibus endpoints can
  be configured in container mode.
• Removed dead session-tracking code (_active_sessions,
  _session_is_known, _prune_expired_sessions, CODESEEQ_SESSION_TTL_SECONDS)
  which had no callers.

Added

• JSON configuration alternative. Every environment variable can now be
  supplied through a JSON config file (keys are the literal env-var names).
  Precedence is environment variable > JSON config > built-in default. Config
  path is CODESEEQ_CONFIG_JSON (default ~/.config/codeseeq/config.json,
  or /home/codeseeq/.config/codeseeq/config.json in the container).
• Catalog clarification. config/model-catalog.json (bridge) and
  config/codex-model-catalog.json (Codex TUI) are documented as two distinct,
  both-used catalogs; neither is redundant.

Changed

• Health endpoint reports version 0.3.9.
• Version bump to v0.3.9. All docs, configs, and internal version strings
  updated.

---

v0.3.8 - 2026-08-17

Added

• Two new qwibus models. qwibus-qwikk (quick, no reasoning) and
  qwibus-qmplx (heavy reasoning/thinking) run against a local
  OpenAI-compatible gateway (default http://127.0.0.1:1337) and require no
  API key.
• Per-model endpoint + sampling configuration. The bridge now supports
  per-model base_url, chat_url, temperature, top_p, top_k,
  context_window, max_output_tokens, timeout_seconds, enable_thinking,
  and system_prompt. Defaults are sourced from config/model-catalog.json
  (single source of truth), with per-model CODESEEQ_<KEY>_* env overrides and
  generic OPENAI_BASE_URL / DEEPSEEK_BASE_URL / CODESEEQ_BASE_URL
  fallbacks.
• DeepSeek per-model defaults. deepseek-v4-flash timeout 60s,
  deepseek-v4-pro 120s, deepseek-v4-flash-thinking 600s, and
  deepseek-v4-pro-thinking 1200s, all with 1M context and 384K max output.
  enable_thinking is true on the two thinking variants.

Changed

• Health endpoint reports version 0.3.8.
• Version bump to v0.3.8. All docs, configs, and internal version strings
  updated.

---

v0.3.7 - 2026-07-14

Changed

• Venice image backend auto-detection. When VENICE_API_KEY is set,
  CodeSeeq automatically enables the Venice backend without requiring
  CODESEEQ_IMAGE_BACKEND=venice. The launcher, entrypoint, and bridge
  all auto-detect the key and activate Venice. Explicit CODESEEQ_IMAGE_BACKEND
  still works as an override.
• Default image output to current directory. bin/codeseeq-venice-image.py
  now saves generated images to the current working directory instead of a
  codeseeq-images/ subdirectory.
• ping-image no longer requires explicit CODESEEQ_IMAGE_BACKEND=venice.
  Setting VENICE_API_KEY is sufficient; auto-detection handles the rest.

Removed

• OpenAI image_gen skill deleted. The .codeseeq/skills/.system/imagegen/
  skill (OpenAI image_gen tool, GPT image models, OPENAI_API_KEY dependency)
  has been completely removed. CodeSeeq uses Venice.ai exclusively for image
  generation. No OpenAI API keys or endpoints are ever contacted for images.

Changed

• Health endpoint now reports version 0.3.7, shows effective image backend
  (including auto-detected venice), and Venice API key state.
• Version bump to v0.3.7. All docs, configs, and internal version strings
  updated.

---

v0.3.6 - 2026-07-14

Added

• Venice.ai image generation backend. New CODESEEQ_IMAGE_BACKEND configuration
  option (default none). Set to venice to enable image generation via the
  Venice.ai API using VENICE_API_KEY.
  • Supports all Venice /image/generate parameters: model selection (auto or
    specific models like z-image-turbo, gpt-image-2, nano-banana-pro),
    aspect ratio, resolution (1K/2K/4K), format (jpeg/png/webp), variants (1–4),
    safe mode, watermark control, negative prompts, seed, and CFG scale.
  • New standalone script: bin/codeseeq-venice-image.py for direct CLI usage
    without Codex.
  • New bridge endpoint: POST /v1/images/generations for OpenAI-compatible
    image generation proxied through Venice.
  • New ping-image diagnostic command.
  • New --image-backend and --venice-image-model CLI flags.
  • Doctor output now includes image backend status and Venice API key state.
  • Health endpoint now reports image_backend status.
• Comprehensive environment variable configuration for all Venice image
  parameters: CODESEEQ_VENICE_IMAGE_MODEL, CODESEEQ_VENICE_IMAGE_ASPECT_RATIO,
  CODESEEQ_VENICE_IMAGE_RESOLUTION, CODESEEQ_VENICE_IMAGE_FORMAT,
  CODESEEQ_VENICE_IMAGE_VARIANTS, CODESEEQ_VENICE_IMAGE_SAFE_MODE,
  CODESEEQ_VENICE_IMAGE_HIDE_WATERMARK, CODESEEQ_VENICE_IMAGE_SEED,
  CODESEEQ_VENICE_IMAGE_RETURN_BINARY, CODESEEQ_VENICE_IMAGE_NEGATIVE_PROMPT,
  CODESEEQ_VENICE_IMAGE_CFG_SCALE, CODESEEQ_VENICE_IMAGE_STEPS.

Changed

• Version bump to v0.3.6.
• Health endpoint now reports image_backend status and Venice configuration.
• Dockerfile now includes bin/codeseeq-venice-image.py.
• All documentation updated with image backend section and configuration reference.

---

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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 us...

CodeSeeq v0.3.8

v0.3.8 - 2026-08-17

Added

• Two new qwibus models. qwibus-qwikk (quick, no reasoning) and
  qwibus-qmplx (heavy reasoning/thinking) run against a local
  OpenAI-compatible gateway (default http://127.0.0.1:1337) and require no
  API key.
• Per-model endpoint + sampling configuration. The bridge now supports
  per-model base_url, chat_url, temperature, top_p, top_k,
  context_window, max_output_tokens, timeout_seconds, enable_thinking,
  and system_prompt. Defaults are sourced from config/model-catalog.json
  (single source of truth), with per-model CODESEEQ_<KEY>_* env overrides and
  generic OPENAI_BASE_URL / DEEPSEEK_BASE_URL / CODESEEQ_BASE_URL
  fallbacks.
• DeepSeek per-model defaults. deepseek-v4-flash timeout 60s,
  deepseek-v4-pro 120s, deepseek-v4-flash-thinking 600s, and
  deepseek-v4-pro-thinking 1200s, all with 1M context and 384K max output.
  enable_thinking is true on the two thinking variants.

Changed

• Health endpoint reports version 0.3.8.
• Version bump to v0.3.8. All docs, configs, and internal version strings
  updated.

---

v0.3.7 - 2026-07-14

Changed

• Venice image backend auto-detection. When VENICE_API_KEY is set,
  CodeSeeq automatically enables the Venice backend without requiring
  CODESEEQ_IMAGE_BACKEND=venice. The launcher, entrypoint, and bridge
  all auto-detect the key and activate Venice. Explicit CODESEEQ_IMAGE_BACKEND
  still works as an override.
• Default image output to current directory. bin/codeseeq-venice-image.py
  now saves generated images to the current working directory instead of a
  codeseeq-images/ subdirectory.
• ping-image no longer requires explicit CODESEEQ_IMAGE_BACKEND=venice.
  Setting VENICE_API_KEY is sufficient; auto-detection handles the rest.

Removed

• OpenAI image_gen skill deleted. The .codeseeq/skills/.system/imagegen/
  skill (OpenAI image_gen tool, GPT image models, OPENAI_API_KEY dependency)
  has been completely removed. CodeSeeq uses Venice.ai exclusively for image
  generation. No OpenAI API keys or endpoints are ever contacted for images.

Changed

• Health endpoint now reports version 0.3.7, shows effective image backend
  (including auto-detected venice), and Venice API key state.
• Version bump to v0.3.7. All docs, configs, and internal version strings
  updated.

---

v0.3.6 - 2026-07-14

Added

• Venice.ai image generation backend. New CODESEEQ_IMAGE_BACKEND configuration
  option (default none). Set to venice to enable image generation via the
  Venice.ai API using VENICE_API_KEY.
  • Supports all Venice /image/generate parameters: model selection (auto or
    specific models like z-image-turbo, gpt-image-2, nano-banana-pro),
    aspect ratio, resolution (1K/2K/4K), format (jpeg/png/webp), variants (1–4),
    safe mode, watermark control, negative prompts, seed, and CFG scale.
  • New standalone script: bin/codeseeq-venice-image.py for direct CLI usage
    without Codex.
  • New bridge endpoint: POST /v1/images/generations for OpenAI-compatible
    image generation proxied through Venice.
  • New ping-image diagnostic command.
  • New --image-backend and --venice-image-model CLI flags.
  • Doctor output now includes image backend status and Venice API key state.
  • Health endpoint now reports image_backend status.
• Comprehensive environment variable configuration for all Venice image
  parameters: CODESEEQ_VENICE_IMAGE_MODEL, CODESEEQ_VENICE_IMAGE_ASPECT_RATIO,
  CODESEEQ_VENICE_IMAGE_RESOLUTION, CODESEEQ_VENICE_IMAGE_FORMAT,
  CODESEEQ_VENICE_IMAGE_VARIANTS, CODESEEQ_VENICE_IMAGE_SAFE_MODE,
  CODESEEQ_VENICE_IMAGE_HIDE_WATERMARK, CODESEEQ_VENICE_IMAGE_SEED,
  CODESEEQ_VENICE_IMAGE_RETURN_BINARY, CODESEEQ_VENICE_IMAGE_NEGATIVE_PROMPT,
  CODESEEQ_VENICE_IMAGE_CFG_SCALE, CODESEEQ_VENICE_IMAGE_STEPS.

Changed

• Version bump to v0.3.6.
• Health endpoint now reports image_backend status and Venice configuration.
• Dockerfile now includes bin/codeseeq-venice-image.py.
• All documentation updated with image backend section and configuration reference.

---

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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....

CodeSeeq v0.3.7

v0.3.7 - 2026-07-14

Changed

• Venice image backend auto-detection. When VENICE_API_KEY is set,
  CodeSeeq automatically enables the Venice backend without requiring
  CODESEEQ_IMAGE_BACKEND=venice. The launcher, entrypoint, and bridge
  all auto-detect the key and activate Venice. Explicit CODESEEQ_IMAGE_BACKEND
  still works as an override.
• Default image output to current directory. bin/codeseeq-venice-image.py
  now saves generated images to the current working directory instead of a
  codeseeq-images/ subdirectory.
• ping-image no longer requires explicit CODESEEQ_IMAGE_BACKEND=venice.
  Setting VENICE_API_KEY is sufficient; auto-detection handles the rest.

Removed

• OpenAI image_gen skill deleted. The .codeseeq/skills/.system/imagegen/
  skill (OpenAI image_gen tool, GPT image models, OPENAI_API_KEY dependency)
  has been completely removed. CodeSeeq uses Venice.ai exclusively for image
  generation. No OpenAI API keys or endpoints are ever contacted for images.

Changed

• Health endpoint now reports version 0.3.7, shows effective image backend
  (including auto-detected venice), and Venice API key state.
• Version bump to v0.3.7. All docs, configs, and internal version strings
  updated.

---

v0.3.6 - 2026-07-14

Added

• Venice.ai image generation backend. New CODESEEQ_IMAGE_BACKEND configuration
  option (default none). Set to venice to enable image generation via the
  Venice.ai API using VENICE_API_KEY.
  • Supports all Venice /image/generate parameters: model selection (auto or
    specific models like z-image-turbo, gpt-image-2, nano-banana-pro),
    aspect ratio, resolution (1K/2K/4K), format (jpeg/png/webp), variants (1–4),
    safe mode, watermark control, negative prompts, seed, and CFG scale.
  • New standalone script: bin/codeseeq-venice-image.py for direct CLI usage
    without Codex.
  • New bridge endpoint: POST /v1/images/generations for OpenAI-compatible
    image generation proxied through Venice.
  • New ping-image diagnostic command.
  • New --image-backend and --venice-image-model CLI flags.
  • Doctor output now includes image backend status and Venice API key state.
  • Health endpoint now reports image_backend status.
• Comprehensive environment variable configuration for all Venice image
  parameters: CODESEEQ_VENICE_IMAGE_MODEL, CODESEEQ_VENICE_IMAGE_ASPECT_RATIO,
  CODESEEQ_VENICE_IMAGE_RESOLUTION, CODESEEQ_VENICE_IMAGE_FORMAT,
  CODESEEQ_VENICE_IMAGE_VARIANTS, CODESEEQ_VENICE_IMAGE_SAFE_MODE,
  CODESEEQ_VENICE_IMAGE_HIDE_WATERMARK, CODESEEQ_VENICE_IMAGE_SEED,
  CODESEEQ_VENICE_IMAGE_RETURN_BINARY, CODESEEQ_VENICE_IMAGE_NEGATIVE_PROMPT,
  CODESEEQ_VENICE_IMAGE_CFG_SCALE, CODESEEQ_VENICE_IMAGE_STEPS.

Changed

• Version bump to v0.3.6.
• Health endpoint now reports image_backend status and Venice configuration.
• Dockerfile now includes bin/codeseeq-venice-image.py.
• All documentation updated with image backend section and configuration reference.

---

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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 ...

CodeSeeq v0.3.6

v0.3.6 - 2026-07-14

Added

• Venice.ai image generation backend. New CODESEEQ_IMAGE_BACKEND configuration
  option (default none). Set to venice to enable image generation via the
  Venice.ai API using VENICE_API_KEY.
  • Supports all Venice /image/generate parameters: model selection (auto or
    specific models like z-image-turbo, gpt-image-2, nano-banana-pro),
    aspect ratio, resolution (1K/2K/4K), format (jpeg/png/webp), variants (1–4),
    safe mode, watermark control, negative prompts, seed, and CFG scale.
  • New standalone script: bin/codeseeq-venice-image.py for direct CLI usage
    without Codex.
  • New bridge endpoint: POST /v1/images/generations for OpenAI-compatible
    image generation proxied through Venice.
  • New ping-image diagnostic command.
  • New --image-backend and --venice-image-model CLI flags.
  • Doctor output now includes image backend status and Venice API key state.
  • Health endpoint now reports image_backend status.
• Comprehensive environment variable configuration for all Venice image
  parameters: CODESEEQ_VENICE_IMAGE_MODEL, CODESEEQ_VENICE_IMAGE_ASPECT_RATIO,
  CODESEEQ_VENICE_IMAGE_RESOLUTION, CODESEEQ_VENICE_IMAGE_FORMAT,
  CODESEEQ_VENICE_IMAGE_VARIANTS, CODESEEQ_VENICE_IMAGE_SAFE_MODE,
  CODESEEQ_VENICE_IMAGE_HIDE_WATERMARK, CODESEEQ_VENICE_IMAGE_SEED,
  CODESEEQ_VENICE_IMAGE_RETURN_BINARY, CODESEEQ_VENICE_IMAGE_NEGATIVE_PROMPT,
  CODESEEQ_VENICE_IMAGE_CFG_SCALE, CODESEEQ_VENICE_IMAGE_STEPS.

Changed

• Version bump to v0.3.6.
• Health endpoint now reports image_backend status and Venice configuration.
• Dockerfile now includes bin/codeseeq-venice-image.py.
• All documentation updated with image backend section and configuration reference.

---

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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...

CodeSeeq v0.3.5

v0.3.5 - 2026-07-08

Changed

• Version bump to v0.3.5.

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...

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-...

CodeSeeq v0.3.3

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, an...

CodeSeeq v0.3.2

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...

CodeSeeq v0.3.1

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...