Measured with the unmodified Codex binary on 10 Oct 2026:
- real context windows per model (NanoGPT list, local table, combo = its
smallest member, capped at the summarizer's window) so Codex compacts by
itself when switching to a smaller model; an overflow guard trims
oversized old parts and reports what still does not fit
- classic tools instead of code mode for external models: GLM 0/3 -> 3/3,
Qwen 0/3 -> 2/3 on simple tool tasks, Claude/cx/Gemini unchanged
- phase (commentary / final_answer) on external assistant messages
- an answer cut by the client is carried to the next turn once
- images stay for vision-capable targets
- summary brake stops tight loops, not frequent compaction
- permanent gateway errors and tool-less targets fail at once with a
readable message instead of five reconnects
- a lone "." (used to apply a model change in voice) is answered locally
- the bridge listens within seconds; the IDE waits 45 s for it
Release: AITURK IDE 1.0.0-beta.26.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
User decision (10 Oct 2026): combos keep their own name; single models
read as short provider/model with the version kept (cc/opus-5-5,
agy/gemini-3.8-flash-high, bekci/kurnaz). No-think copies show their
real provider with a "(no-think)" suffix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The size-proportional minimum (input/250, cap 4000) rejected a correct summary:
a 646,955 byte thread (225k tokens, five labelled facts in a long log) was
summarized by the subscription model in 1501-1701 characters, the bridge demanded
2587, and Codex could not continue the thread on the selected model. The floor now
only catches broken output (input/2000, cap 300). Memory loss is prevented by who
writes the summary, not by its length.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Model switch lost the conversation: on 10 Oct a 1 MB thread was compacted by the
newly selected small model into 2201 characters. The compaction summary for a bridge
model is now written by a subscription model (ozet_kopru.ozet_modeli, default
gpt-6.1-sol); if that request is refused before it starts, the target model is used
as before. A summary far shorter than the conversation warrants is not accepted:
Codex gets response.failed and keeps the history.
The line under each added model no longer repeats its name. Teams show their
members, fallback chains their order, single models their provider and the traits
readable from the name; the text fits one phone row (30 characters).
Also normalizes tool call ids in the streamed response so Codex persists ids the
subscription backend accepts (fc_ -> ctc_), and bumps the bridge protocol to 6.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
sirala_codex re-sorted the whole list and rewrote every priority, which pushed the
native subscription models to the bottom and put same-named AITURK copies
(GPT-6-Astra) above them. Native models now keep their own order, priority and
name; bridge models are appended after them and a copy that shares a native name
is shown with an (AITURK) suffix.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex only received combos and the models named in the IDE config, so none of
the models behind the providers added to 9Router showed up. Those single
models are now always added; 9Router's combos stay hidden. The copy of the
same model that comes through OmniRoute is still added only when the IDE
config names it, so each model is listed once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRDGW1LiKcDUZYUdKBoWYW
Combos that come from 9Router, directly or through the gateway's 9router
provider, are no longer added to Codex's model list; the gateway's own combos
keep their plain names. The IDE no longer leaves a config.toml backup behind:
turning the setting off already restores the file byte for byte.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRDGW1LiKcDUZYUdKBoWYW
Codex compresses request bodies with zstd. The agent venv is Python 3.11, which
has no zstd in its standard library, so every Codex request failed on a fresh
install. The matching zstandard build now ships beside the bridge; the status
endpoint reports ok=false when bodies cannot be read, and the IDE then leaves
Codex's config alone. A running bridge from an older build is replaced, and
"Sync now" makes the bridge re-read the IDE session at once.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRDGW1LiKcDUZYUdKBoWYW
Settings → Codex Bağlama runs the bundled bridge for an OpenAI Codex install on
the same computer: the IDE's models and combos appear in Codex and switching
models keeps the conversation. The IDE points Codex's openai_base_url at the
bridge only after the bridge answers, keeps it running across logins, and
removes every trace when the setting is turned off. A Codex that already uses
another base URL is left untouched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRDGW1LiKcDUZYUdKBoWYW
Adds the IDE's models and combos to Codex's model list and keeps conversation
summaries readable across providers, so switching models in Codex no longer
drops the conversation. Standard-library Python, run on the agent venv.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HRDGW1LiKcDUZYUdKBoWYW
A one-shot run (hermes chat --query-file ... -Q, stdin closed, no window)
still installs the prompt_toolkit approval callback but never starts the
prompt UI. The protected instruction-file gate (AGENTS.md / CLAUDE.md /
SOUL.md ...) took the registered callback for a reachable human and waited
out approvals.timeout on a panel painted nowhere. With the profile value of
300000 s the stage job hung for good.
- Gate: in a one-shot run, or when only an automatic approver
(subagent_auto_approve) is attached, return BLOCKED / APPROVAL REQUIRED
immediately instead of prompting. Human consent is still required every
time; yolo and approvals.single_query_mode still do not open this gate.
- One-shot exit prints an "approval_required: {json}" line on stderr
(session_id, tool_call_id, request_id, targets, reason, applied=false)
so the launcher can tell "blocked, needs a human" from "still working".
- execute_code: an approval that arrives after the calling script stopped
waiting (300 s reply timeout, script run timeout, retired kernel cell)
is not applied.
- Same invisible wait removed from the computer_use and memory-write
prompts in a one-shot run. The one-shot marker is no longer inherited by
child processes.
- BLOCKED text no longer breaks on a target path containing braces.
Long-running tools are untouched: no general tool timeout was shortened.
Kart: turkservis-20261007-1421-2
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01EwSpve9HuXMy7dHf4RDMzX
Uzun sure acik kalan IDE'de "Telefon ekle" ilk istegi olu bir keep-alive
baglantisinda kaybedebiliyordu; bulusma kaydi yayinlanamayinca kod yolu bes
dakika boyunca "Kodla baglanma su an kullanilamiyor" kaliyordu (QR calisiyordu).
- publishPairOffer: en cok 4 deneme (0,5 / 1,5 / 3 sn); ag hatasi, 5xx ve 429
yeniden denenir, 422 kalicidir; yaniti kaybolan denemeden sonraki 409 kaydin
yerinde oldugu anlamina gelir (iptal jetonu yoktur, kayit kendiliginden biter).
- Pencerede "Yeniden dene": ayni kod ve QR ile kaydi yeniden yayinlar.
- Role oturumu kaydi ve "Giris yap" baslangici icin tek ek deneme (zaman asimi
yeniden denenmez).
- Gunluge yalniz "pair offer published / failed after N tries" duser; kod,
lookup ve jeton yazilmaz.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B8ppceaR5s6cvMnm7uGRnf
The phone link no longer depends on the local network. The IDE dials out to
turkservis.online and bridges tunnel streams to its own TLS proxy on
loopback; the phone runs TLS to the pinned IDE certificate inside the
tunnel, so the relay only carries ciphertext. The proxy binds to 127.0.0.1
by default and enabling the feature never asks for a firewall rule; direct
same-network access is an explicit option. The window shows the real relay,
listener and firewall state, and pairing no longer needs local addresses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B8ppceaR5s6cvMnm7uGRnf
Settings gets a Phone link window: when enabled, the IDE main process runs a
small TLS proxy on the local network (self-signed certificate, pinned by its
fingerprint) in front of the loopback backend and pairs a phone by QR code
or an 8-character code. Each phone gets its own revocable token; the backend
session token never leaves the computer. The root page, terminal sockets and
secret or settings-changing endpoints are not reachable from a phone. The
old WebRTC pairing window is replaced.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B8ppceaR5s6cvMnm7uGRnf
The onboarding and provider settings get a Sign in button: the IDE shows a
short code, opens the turkservis.online approval page, and after approval
saves the TurkServis provider itself and hands a separate member key to the
AITURK CLI over stdin (aiturk login --key-stdin --keep-other-account).
Keys never reach the renderer, logs, process arguments or URLs. Pasting an
API key remains as the fallback.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01B8ppceaR5s6cvMnm7uGRnf
Owner decision (3 Oct 2026): where the AITURK CLI is available, TurkServis
models are reached through it, not MCP. The backend no longer writes a
`turkservis-media` MCP server into config.yaml on startup or when a member
endpoint is saved. Existing entries, including disabled ones, are left exactly
as they are. A bundled `aiturk-media` skill teaches the agent to list and run
media with `aiturk media`.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_012wwqJMAuhMBMWvsvNLFGb9
A clean npm ci hoists qrcode-generator and jsqr to the workspace root, so the bundler now follows Node resolution from apps/mobile instead of assuming apps/mobile/node_modules.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
onnxruntime (wake word) and ctranslate2 (local speech-to-text) link against
msvcp140.dll and msvcp140_1.dll, which the managed CPython does not ship.
On Windows without the system VC++ Redistributable both imports failed with
"DLL load failed: The specified module could not be found", so "Hey Hermes"
and voice chat transcription never worked.
- Bundle the Microsoft-signed VC143 CRT set (14.44.35211, matching the
managed CPython's vcruntime140) as a Windows extra resource with a SHA-256
manifest.
- Before every local backend start, copy missing or mismatching CRT files
beside the managed python.exe (the process image directory, searched first),
verifying digests; never overwrite the interpreter's own vcruntime and never
block boot. Existing installs heal on their next launch.
- Report an unloadable native voice wheel as a VC++ runtime problem in the
wake-word requirements, local transcription and `hermes doctor`, instead of
treating installed package metadata as proof the wheel loads.
- Bump AITURK IDE to 1.0.0-beta.12.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
