pymaxim 1.3.3

pymaxim 1.3.3 — "Release pipeline"

Released 2026-10-24 (UTC). pip install --upgrade pymaxim

1.3.3 ships the fixes found in the Console by maxim-pulse (the separate project that builds the Console), the Console's live stream from a clean install, and the
first pymaxim release CI publishes from a signed tag. It makes no new behavioural claim. Some changes break on
purpose; each is marked below with what to do.

Security: upgrade every maxim on a host together

• Every llama-cpp-server maxim spawns now requires a key, solo included
  (#1314). Before, only a leader's server was keyed, so any local
  process or user on a shared host could use a solo's inference. A non-leader's server now uses a per-host key,
  ~/.config/maxim/llm_server_key (0600; %APPDATA%\maxim\llm_server_key on Windows), created on first start.
  • Upgrade every maxim on a host together. An older maxim reuses the keyed server and gets 401s on every call.
  • Stop any llama-cpp-server an older maxim started. 1.3.3 refuses a server that answers without a key; the
    next start spawns a keyed one. Nothing is killed automatically.
  • A lane you pointed by hand at a server on this host needs MAXIM_LANE_<TIER>_REMOTE_API_KEY. maxim doctor
    prints the command.
  • A container with a read-only home needs a writable ~/.config/maxim for the key file.
• The agent can no longer change the operator's internet switch
  (#1345). internet_access_toggle can only lower internet for
  its own session; only the operator turns it on.
• The file tools refuse to write git configuration (#957):
  anything under .git, a git dir by identity, and the user- and system-level git config. This narrows a
  code-execution path through git filters; it does not close it, because bash can still write there.
• Sim reports no longer record an endpoint's credentials, path or query
  (#997). report.json's language_endpoint and aut_endpoint
  keep scheme://host[:port] only; a script that read the path from those fields must stop.
• An unusable leader key is a typed refusal on every Python, never a traceback or a silent "no key"
  (#1311, #1312).

The Console

• pymaxim[console] serves the live stream from a clean install. The extra now installs websockets, so /ws
  upgrades instead of returning 404 (#1347).
• Piper speaks again (#1332). In 1.3.2 every synthesis failed
  on every installable piper version, so spoken-code pairing played nothing. Correction to the 1.3.2 notes:
  "verified Piper voice provisioning" overclaimed. 1.3.2 verified the voice's files and that it loaded, not that it
  could speak.
• Pairing arms only on a voice that speaks (#1334). A robot
  whose voice loads but cannot speak stays on the paste sign-in.
• Talk is offered only the tools it may use, and a refused tool call goes back to the model instead of idling to
  the timeout (#1412,
  #1413). A single message in Talk no longer creates a new memory
  every half second (#1414), so memory counts from Console sessions
  are not comparable to 1.3.2. A setup change made through the Console reaches the next talk without a restart
  (#1411); a maxim config set from another process still needs one.
• maxim serve no longer hangs up to 15 s on shutdown after a Console tab closes.
• The wheel bundles the maxim-pulse Console v0.4.2 (commit bcd0d1e; 1.3.2 bundled v0.4.0). The Console's API
  is unchanged (contract 0.6.0); typed payloads come in 1.3.4.

Release integrity

• 1.3.3 is the first release CI publishes. A signed tag starts the release workflow. It requires a green nightly
  on that exact commit, smoke-tests the exact wheel in a clean environment, waits for the owner's approval, and
  publishes through PyPI trusted publishing, with no stored token.
• The vendored Console must be the pinned, newest pulse release for its contract; the build refuses any other
  bundle.
• Two optional extras, tts and console, run against their real libraries. Each is installed alone, at the
  lowest versions the extra declares and at the latest, in every nightly, and the release waits for them.

Behaviour changes and what to do

• NAc no longer forgets while switched off. A saved agent resumes with its learning intact, however long it was
  off. A learned fear, including one imported in a bundle, therefore never fades. To restore the old offline decay: NACConfig(offline_wall_decay=True). NAc.load(apply_decay=...) is
  deprecated and is removed in 1.5.0; --resume-sim still passes it internally, so ignore that warning.
• A missing local model downloads only with consent, on every path (#1343). A first maxim.run() in a notebook
  no longer downloads a model silently: maxim config set llm.auto_download true, or call
  maxim.download_model(name).
• maxim --clear-memory clears the whole data home and asks first (#1337). Off a terminal it refuses with exit
  2 unless you pass --yes; declining now exits 1 (it exited 0).
• A benchmark passes only strictly (#1310): every expectation met and at least one threshold applied, and
  maxim --sim benchmark exits 1 otherwise. The shipped cognitive_suite cannot pass yet, for any model, and now
  says so.
• A refused tool call is fed back to the model (#1413). An LLM-driven sim can make up to two extra model calls
  each time the model picks a tool it isn't allowed, so action counts, cost and blocked counts are not comparable to
  1.3.2.
• An agent id holding : is refused when the agent is named (InvalidAgentIdError, from maxim.create.agent,
  maxim.load.agent and the agent builders). On 1.3.2 such an agent's learning was filed under the wrong agent when
  it reloaded; 1.3.3's NAc refuses to record it. To keep an agent saved under such a name, rename its directory under
  ~/.maxim/agents/ to a name without : and load it by the new name.
• The model is no longer offered execute_file, which was always refused. Account for it when comparing
  prompts across versions.
• maxim oasis serve --insecure with an unusable (not missing) key refuses with exit 2; it used to serve
  without auth (#1311).
• Python API: InternetAccessTool no longer takes set_internet_access_fn=, and
  maxim.utils.set_internet_access defaults to source="operator" (#1345). NAc's credit methods raise ValueError
  on a NaN or infinite input, and a loaded nac.json has its values clamped to range, with a warning (#1102).

Evidence status

No experiment result changed status. The one result whose evidence is out of date, Exp 37 (ledger row T1-2), stays
marked STALE, under a recorded owner exception (release-1.3.3-T1-2): its cross-model "Goldilocks" map may only be
cited with its caveat. Its re-run is a new dose-response experiment planned before 1.4.0. The
ledger is
authoritative.

Not claimed

No new behavioural claim. The memory, fear and drive mechanisms are bio-inspired models, not neuroscience
simulations.

The full list, with each fix's guard, is in the
CHANGELOG. Each file's sha256 is shown on the
GitHub Release and
PyPI.

Author: Denny Schaedig

---

vendored console: pulse v0.4.2 (bcd0d1e, contract 0.6.0)