Skip to content

pymaxim 1.3.1

Choose a tag to compare

@dennys246 dennys246 released this 27 Sep 22:09
· 2035 commits to main since this release
v1.3.1
7e695a5

pymaxim 1.3.1 — "Hardening"

Released 2026-09-27 (UTC — PyPI upload_time). pip install --upgrade pymaxim

1.3.1 makes no new behavioural claim. It fixes what the v1.3.0 re-score found, and each fix ships
with the guard that holds it (a test, a lint, a CI lane or a required check), because the score card
credits only what is enforced. If you share substrate, run sandboxed tools, or call the Python API,
read Upgrading below: several fixes change behaviour on purpose.

Security

  • The sandbox runs what was approved, and only inside the sandbox (#800–#802).
    • Python scripts ran at all for the first time: the restricted wrapper used to block its own imports.
    • Containment compares resolved paths, not a string prefix, and no longer follows a symlink out.
    • What runs is the content that was approved, not whatever is at the path later.
    • SUPERVISED approval fails closed: with no approver attached, a script that needs approval
      does not run.
  • A mode's limits are enforced when a tool runs (#826). They used to be applied only to the
    prompt: a tool the mode excludes still ran when the model named it. The executor now checks the
    live mode at every dispatch. Passive mode, which observes and proposes, refuses tools that act on
    the host.
  • The agent can't switch itself into singularity (#821, #828). The mode tool and the CLI refuse
    a self-requested switch into any mode flagged to execute code (today, singularity). So does a spoken
    phrase, since any audio in the room could say it. Passive → active is unchanged: active runs shell
    and sandbox tools under approval, which non-interactive runs auto-approve, so a passive agent can
    still leave passive on its own. That gap is tracked with the approval surface (#924, #922).
  • Autonomy requests no longer pretend (#827). With no human approver attached, a request fails
    with that reason instead of waiting forever. The approval surface itself is #922.
  • A model-chosen web fetch connects only to the public address it checked (#824), which closes
    DNS rebinding. The fetch byte cap now bounds the download itself (#825).
  • Tool output reaches the model fenced as untrusted data (#823). A fetched page's own
    instructions block no longer looks like the prompt's. Fencing does not stop a model from choosing to
    follow injected text.

Sharing substrate

  • Public bundle format 1 is frozen. A downloaded release is a compatibility promise: every 1.x
    reads both bundle shapes as published, and changing the format needs a recorded decision.
    docs/plans/public_format_freeze.md says what is promised and what is not.
  • Oasis release format v2 (bundle schema 3, signed releases only).
    • A signed release carries a detached signature over every member, a signed entry index, its
      signer, a per-key release sequence and its license.
    • A receiver journals what it admitted. It refuses a second payload under the same key and sequence
      (equivocation), and a legacy v1 bundle from a key whose v2 release it already accepted (downgrade).
    • Unsigned bundles stay schema 2, readable by 1.3.0.
  • Merging no longer destroys your own causal links (#913). Ingest paired links by outcome alone,
    so a receiver's links that differed only in context overwrote each other: one real state went from
    607 to 443 links with an empty donor. Donor situations that align with yours now fold together
    instead of the last one winning (#914).
  • Export privacy. The export scrub is an allowlist at every level, and a signed release carries no
    local agent id.

The Python API does what its docs say

  • AgentInstance.export_memories() counts what the store holds. It reported 0 for every agent.
  • The maxim.create.agent example runs as written, and capture() names the fix for a wrong
    argument type.
  • maxim.diagnose() runs the same checks as maxim doctor.
  • maxim.campaign(prompt_handler=...) works.
  • A pip install gets the foundational preamble. It is read from the Constitution, which now ships in
    the wheel, and it carries every §1 hard constraint, including the actuator-speed one the old
    preamble had omitted.
  • --session <id> on maxim substrate, maxim hive pull and maxim roy diff finds a simulation's
    session, and for ingest / hive pull a maxim.create.agent() name. Before, a bare ID was looked
    up in a directory nothing writes to.

Release and research integrity

  • A release waits for green nightlies. The release build refuses unless a nightly passed on the
    exact commit being released. The model-cache nightly, red for 16 nights, is green again.
  • The test process cannot reach the network. A measured run made 52 real outbound attempts; a
    conftest guard now makes them fail (a subprocess a test spawns is not covered).
  • The prereg lint governs 1.3's own experiments. Exp 58–62, R2 and R3 were outside it; it now
    checks 52 records, all passing.
  • The ledger's 1.3.1 trigger walk (window v1.3.0..042b7d90; the one later merge, #933, fires no
    trigger). Every graduation-ledger row whose re-run trigger fired carries a dated note:
    • Exp 10 (cross-session memory) was re-run on the rig: MAINTAINED, narrow. Both resumes reloaded
      the saved store exactly (100 memories). The fields the persistence change added came back unchanged
      on all 100, and 3 memories surfaced on every resume turn observed. But every run stopped early on
      a known planning defect (D13, #935): one turn per resumed phase (1–3 across all five sessions),
      against eight in August. The record
      says so, and discloses the attempts that were not used.
    • Exp 42, 45, 53b, 56, 60, 61 and 62, the pain cascade and the reflexes were discharged with
      offline evidence. That includes checking that the new mode gate cannot refuse anything in the
      survival experiments, which run in active mode.
    • Exp 37's prompt trigger fired, since the preamble text changed. It was not re-fired; the row
      carries no claim.

Memory strength: recorded, not yet used

The memory-strength phases land as recording. Under the default retention strategy they change
nothing:

  • every memory records when it happened, how strongly it encoded, the body's drive pressure and
    relief at the time, and the situation it happened in;
  • a strong moment tags the memories just before it in the same situation;
  • situation recall is wired but nothing consumes it yet.

maxim config set memory.strategy strength opts into the storage-strength model. It is an
uncalibrated placeholder, not campaign-ready.

Corrections to 1.3.0

  • Exp 61's bundles were unsigned. The 1.3.0 notes said the fear transferred "through the shipped
    signed-bundle path". The harness exported without --sign. The claim rests on the shipped export and
    ingest path, not on signing; signing is covered by its own tests.
  • The maxim substrate invalidate upgrade step was incomplete. Corrected on the v1.3.0 release and
    in its notes: maxim substrate invalidate --session <ID> prints the census, then
    maxim substrate invalidate --session <ID> --modality world --drop-geometry <stale-tag> --apply.
  • Exp 37/38's NAc-bias-off arm is void (#889). It left the reward bias on. No EARNED claim rested
    on it.

Not claimed

Exp 62, the fear carried by the same body into a different pool, is EARNED on the ledger
(2026-09-20). It is not a 1.3.1 claim until its different-reader pass is recorded.

Upgrading

pip install --upgrade pymaxim. Behaviour that changes on purpose, most likely first:

  • Python memory API: Hippocampus.capture(), capture_from_loop() (and its async form) and
    store() now require encoding= (keyword-only); capture_from_loop also requires
    situation=. A 1.3.0 call without it raises TypeError. Pass what you measured, or
    encoding=EncodingSignals.unmeasured("api") (from maxim.memory.encoding import EncodingSignals).
    capture() also raises TypeError for a wrong argument type.
  • The plain CLI agent and maxim.run() run in passive mode by default (unchanged), and passive is
    now enforced: bash, edit_file, git_commit, run_tests, execute_file and the other
    host-acting tools are refused there. Say or type "maxim active" to switch.
  • Sandbox: a script that needs approval and has no approver attached is refused
    (APPROVAL_UNAVAILABLE), under SUPERVISED and whenever no autonomy controller is present. A
    working_dir outside the sandbox, or a script over 120 KiB, is refused.
  • Modes and autonomy: the agent cannot switch itself, or be switched by a phrase, into
    singularity; autonomy requests with no approver fail.
  • Internet policy now takes effect (#822): a previously-off internet toggle or a hand-written
    util/internet_policy.json now applies, and an unreadable or ill-typed policy fails closed.
    Model-chosen fetches no longer use HTTP(S)_PROXY (#824), so they fail behind a mandatory proxy.
  • Signing releases: maxim substrate export --sign needs --license SPDX-ID, and takes its
    sequence from the signing key's per-host counter. A key you signed with before 1.3.1 needs
    --release-sequence N once
    , since the counter has never seen it.
  • Receiving and contributing:
    • a v2 release needs maxim hive pull --receiver-agent-id;
    • hive pull / hive contribute to a remote or LAN Oasis need --api-key (the local leader key is
      used for a loopback Oasis only);
    • a newly added Oasis refuses legacy v1 signatures by default (maxim hive trust --accept-v1 to
      allow), and hive add refuses a non-canonical or duplicate key;
    • maxim oasis publish needs --queen-key, and release ids are now signed-payload digests (re-pin
      any --release <old id>).
  • Exporting: an unsigned substrate export ships only your own learning, not material you
    ingested; --no-identity-filter no longer keeps free-text signatures; contributor_id /
    signer_identity must match [A-Za-z0-9_.@:-]{1,128}.
  • --session: a bare ID is looked up in ~/.maxim/sim_reports/ (and, for ingest / hive pull,
    ~/.maxim/agents/); an ID in more than one place, or an empty argument, is refused.
  • Validation now refuses bad inputs: a pain or reaction intensity outside [0, 1] (or non-finite),
    a sensor reflex declaring an absolute value: (custom reflex YAMLs use delta:), a negative
    damage_component amount, and an unknown memory.strategy.
  • Doctor and diagnose: on a machine configured as a peer, diagnose() now runs the peer probes,
    which make real network calls. maxim doctor --as peer <url> and diagnose(peer=url) send your
    configured key only to its own leader URL, so another URL now fails the auth check.
  • Python API: export_memories() raises when the store can't be read;
    campaign(interactive=True, prompt_handler=...) raises ValueError.
  • Removed or now-required internals: Executor.get_last_rpe and
    SupervisionPolicy.allowed_mode_transitions are removed; propose_via_substrate requires
    situation_cue=.
  • Config downgrade: a config.json written by 1.3.1 carries a memory section that older builds
    refuse.
  • Every agent's prompt carries the Constitution's four hard constraints word for word, and the
    preamble header lost its "(from AGENTS.md)" suffix.

The full list, with each fix's guard, is in the
CHANGELOG.

Score card

A blind re-score is taken at the v1.3.1 tag, as for 1.3.0. Its evidence agents work from a
firewalled copy with the earlier cards removed. The card lands in docs/limits/score_cards/ beside
the 1.3.0 cards.