AgenC 0.19.9

AgenC 0.19.9

AgenC 0.19.9 lets a command that the sandbox cuts off from the network run
after one approval, and fixes two small problems found while testing 0.19.8.

Network commands in the sandbox

• In the default permission mode, when the sandbox blocked a command's network
  access (for example npm install, git clone or pip install), AgenC told
  the model that approval was unavailable. The model then gave up, or asked for
  narrower permissions that still failed. Now it retries once with your
  approval and the command runs.
• curl -s and wget -q print nothing when the network is blocked, and npm
  on macOS reports ENOTFOUND. AgenC now recognizes these as well.
• With --bypass-approvals, AgenC tells the model that the retry runs without
  asking, because you turned approvals off.

Fixes

• The first Browser call opens its page even when the model names tab 1
  before any tab is open. Before, that first call failed and the model had to
  try again.
• On a machine without git, /status says "git not installed". Before, it
  showed the raw error "spawn git ENOENT" and flagged it for attention.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.9. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.11

AgenC 0.19.11

AgenC 0.19.11 shows what your model calls really cost. Before, a call to a
model AgenC had no price for was shown at the highest rate AgenC knows, so
an OpenRouter call that cost a third of a cent could show 88 cents.

Real costs

• OpenRouter: AgenC now shows the charge OpenRouter reports with every
  response. With your own provider key on OpenRouter, the provider's charge
  is added to OpenRouter's fee.
• Every model in the model picker now has a price. New prices: all Amazon
  Bedrock models (US East on-demand rates), Ollama Cloud, more Qwen and
  GitHub Copilot models, and GLM-5 Turbo on Z.AI.
• Calls that cost nothing per token now show $0: NVIDIA NIM's free API, the
  Qwen Token Plan, the GLM Coding Plan, and Grok Composer through Grok Build.
• Budget limits still reserve conservatively before a call and settle at the
  real price after it.

Model list

• Hidden because the provider retired them or no longer sells them: six Qwen
  models Alibaba retired on 2026-10-10 (qwen-coder-plus, qwen-turbo,
  qwen3-8b, qwen3-14b, qwen3-30b-a3b, qwen3-235b-a22b), three Copilot models
  GitHub retired on 2026-10-02 (Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7
  Code), and three Ollama Cloud models missing from Ollama's price list.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.11. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.10

AgenC 0.19.10

AgenC 0.19.10 names DeepSeek's new Flash model, adds it on OpenRouter, and
fixes reopening a session while its reply is still streaming.

DeepSeek V4.1 Flash-1017

• DeepSeek now serves DeepSeek-V4.1-Flash-1017 under deepseek-flash, which
  is AgenC's default DeepSeek model, so DeepSeek sessions already use it with
  nothing to change. The model picker now shows its name. Context, output
  limit, effort levels and prices are the same as before.
• On OpenRouter it has its own id, deepseek/deepseek-v4.1-flash-1017, now in
  AgenC's OpenRouter catalog.
• On a small "make the failing tests pass" task (median of 4 runs), AgenC took
  17.4 s and $0.0058 with 1017. The day before, with the previous V4.1 Flash,
  it took 25.2 s and $0.0068.

Fixes

• Closing the terminal while a reply streams and then running
  agenc --continue could fail with "canonical resume source ... changed
  during authorization", because the daemon was still writing the session.
  AgenC now accepts what the daemon appended; a replaced or rewritten session
  file is still refused.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.10. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.9 candidate run 39016266427

Permanent candidate escrow for agenc-core run 39016266427.

AgenC 0.19.11 candidate run 39037459994

Permanent candidate escrow for agenc-core run 39037459994.

AgenC 0.19.10 candidate run 39023835805

Permanent candidate escrow for agenc-core run 39023835805.

AgenC 0.19.8

AgenC 0.19.8

AgenC 0.19.8 starts a TUI session on Windows in the default mode, and fixes
two small problems found while testing 0.19.7.

Windows

• AgenC has no sandbox on Windows yet, so the default mode refused to start a
  session there (sandbox_required_unavailable). Now the TUI asks once,
  before the first session, whether to continue without a sandbox.
• Continue saves sandbox_mode = "danger-full-access" in your config.toml.
  Approvals still apply, so AgenC asks before each command and file edit.
  Quit saves nothing, and the question comes back on the next start. To
  answer again, run agenc config unset sandbox_mode.
• The question never replaces a sandbox mode that you, a profile or a
  repository chose.
• A run without a terminal (agenc -p) never asks. Until you have answered,
  it refuses as before. It cannot answer approvals either, so pass
  --bypass-approvals for unattended runs.
• The trust card now says when the sandbox is off for the run. On a host
  without a sandbox, --bypass-approvals also says so in the TUI footer.

Fixes

• The bundled simplify skill now says it is only for when you ask to review
  or clean up code. In one 0.19.7 test run the model started it on its own
  after its fix already passed, and the run took twice as long. /simplify
  and review requests work as before.
• An error that stops a session from starting no longer ends with two
  periods.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.8. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.7

AgenC 0.19.7

AgenC 0.19.7 makes every request in a standard session smaller again, and
retries a provider outage that answers with an HTML error page.

Lighter requests

• The part of each request that AgenC resends on every model call is about
  40 percent smaller than in 0.19.6. On DeepSeek the first request of a task
  went from about 8,950 tokens to about 5,260.
• The system prompt states the same rules in fewer words. That includes the
  memory instructions, the permission and sandbox rules, and the section on
  helper agents.
• Tool descriptions are shorter. The shell tool and Grep are sent with the
  fields models use most. Every other field still works, and the full
  description loads when the model asks for it.
• AskUserQuestion and the plan mode tools load on demand. In plan mode the
  model can still ask you questions and present its plan for approval.
• The skill list shows at most 120 characters of each skill's description.
  The skill itself keeps its full text.
• On a small "make the failing tests pass" task with DeepSeek V4 Flash
  (median of 4 runs each), 0.19.6 took 25.9 s and 94,000 tokens. 0.19.7
  takes 25.2 s and 74,000 tokens.
• On a set of 10 everyday coding tasks, tokens fell about 37 percent, wall
  time about 12 percent and cost about 13 percent compared with 0.19.6. On a
  separate set of 12 held-out tasks, tokens fell about 42 percent, wall time
  about 11 percent and cost about 17 percent. All runs were solved on both
  sets, as with 0.19.6.
• Light sessions, and sessions with defer_rare_tools = false, are
  unchanged.

Fixes

• During an outage, the gateway in front of an OpenAI-compatible provider can
  answer 502, 503 or 504 with an HTML error page. AgenC now retries that like
  any other server error. Before, the turn stopped and showed the page's HTML
  as the error.
• The system prompt no longer says that a turn continues until a token target
  such as "+500k" is reached. No part of AgenC did that.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.7. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.6

AgenC 0.19.6

AgenC 0.19.6 makes the default TUI faster and much lighter on tokens, and
fixes problems in the Venice and Ledger menus and in long sessions.

Faster and lighter

• Each request is about half the size. Standard mode now sends 12 tools on
  every call instead of 33. Helper agents, background process tools, web
  tools and rarely used tools load on demand when a task needs them. On
  DeepSeek the first request went from about 17,500 tokens to about 8,600.
• The model is asked to batch independent reads and commands, and to keep its
  final report short. On small tasks it no longer starts a separate
  verification helper.
• On a small "make the failing tests pass" task with DeepSeek V4 Flash,
  0.19.5 took 30.6 s and 203,000 tokens. 0.19.6 takes 24.1 s and 103,000
  tokens. With GPT-6 Luna the same task went from 241.7 s and 755,000 tokens
  to about 80 s and 130,000 tokens on runs without provider stalls.
• On a set of 10 everyday coding tasks, wall time fell 8 percent, cost 10
  percent and tokens 42 percent. On a separate held-out set of 12 tasks, wall
  time fell 17 percent, cost 13 percent and tokens 46 percent. The same tasks
  were solved: 127 of 128 runs against 128 of 128 before.
• When you type a prompt, the first request goes out about 0.1 s after you
  press Enter. The TUI now prepares its session while you type, and AgenC no
  longer waits for a provider's model list before the first request.
• On macOS each shell command starts about 35 ms sooner, because the sandbox
  helpers for the next command are prepared while the current one runs.
• On Linux, AgenC forces its session database to disk about half as often
  during a turn, and file reads wait less behind those disk flushes.
  Everything is still flushed at the end of each turn. After a power loss or
  system crash, resuming continues from the last saved step, and a command
  that was running is marked for review instead of being run again.
• In large repositories, checking whether the shell already showed a file now
  happens only when an edit needs it, not after every command.

Fixes

• The Venice provider menu lists every Venice model that can call tools,
  about 130, instead of 3. It refreshes after you save, replace or remove a
  key, and any listed model can be selected. Venice models use Venice's
  published prices for the cost estimate and the budget check. A model with
  no published price shows "price unknown".
• Coding prompts that mention a ledger, such as a file named ledger.txt or
  accounting code, no longer open the Ledger hardware-wallet screen or check
  for a device. The screen opens for the Ledger commands and for questions
  about your Ledger device. While it is open, Esc, /exit and Ctrl+C work.
• A prompt too large for the model's context is refused before it is sent or
  saved, with its size and the limit.
• When a model's stream is cut off before any tool ran, AgenC retries it
  instead of failing the turn.
• A partly streamed answer stays visible when you detach from a session and
  reopen it.
• A model stream that turns into repeated nonsense is stopped early and
  retried once, instead of running and billing for minutes. Set
  stream_degenerate_guard = false to turn this off.
• Terminal color codes are removed from the command output the model sees.
• A mistyped model name on a provider with a fixed model list is refused with
  a suggestion for the closest name.
• Quitting the TUI before sending a prompt no longer leaves an empty session
  behind, and agenc --continue never resumes an empty session.
• With a custom endpoint, a session switches to the endpoint's real context
  and output limits at its next turn. Until then it uses the built-in limits
  for known models. Limits you set in the config always apply.

Upgrade notes

Standalone installations update with agenc update; npm installations with
npm install -g @tetsuo-ai/agenc@0.19.6. Then run agenc daemon restart.
On a Mac, if a dialog asks whether AgenC's node may access your Documents,
Desktop or Downloads folder, allow it so AgenC can reopen your sessions there.

AgenC 0.19.8 candidate run 38972049808

Permanent candidate escrow for agenc-core run 38972049808.