fix: copy conversation_history to prevent caller mutation - #235
Closed
atian8179 wants to merge 1 commit into
Closed
Conversation
Replace `messages = conversation_history or []` with `messages = list(conversation_history) if conversation_history else []` to avoid mutating the caller's list. The previous code created an alias to the caller's list, so every append during the conversation also modified the original list unexpectedly. Closes NousResearch#228
Contributor
|
Closing in favor of PR #229 which has the same fix plus tests. Thanks for the contribution! |
rhCat
added a commit
to rhCat/Maria
that referenced
this pull request
Jul 26, 2026
…th not rediscovering A resume-here note spanning both repos (Maria + cyberware), written because most of what was learned this session is not recoverable from the diffs: it is measurements against the live fleet and a list of things that silently do nothing. Records what shipped (cyberware NousResearch#232-NousResearch#235 merged, 5 commits unreleased since v1.8.0; four commits on feat/ed25519-assertions here), what is deployed and proven (Maria authenticating with a fresh Ed25519 assertion per claim, verified offline, a 5-turn LLM round passing 4/4), and exactly where it stopped — v1.9.0 untagged and this branch without a PR. Blockers that need a decision or a machine unreachable from here: the alembic worker re-serving cached evidence under the REQUESTED commit hash; cibatio rendering no judgment because INTEL_* never reaches the engine; /run-conservation-stack exiting 4; whether selfmod should be approvable at all; and the fact that a caged agent has no reachable approval channel — 45 push_backs, 9 approvals, all synthetic. Also the four bugs found and fixed (the 403/409 transport collapse that had made Maria read-only; a config key that registered nothing; .strip() truncating ~5% of binary keys; a None mirror_dir crash), and the traps worth not paying for twice — GOVD_AUTH_VERIFIER must be env because the entrypoint regenerates govd.json; docker-compose.maria.yml is UNTRACKED; the documented launch line has two wrong values; the worker is on 11080 and needs a commit SHA not a branch; commits need --no-gpg-sign; and `docker exec … python3 - "$ARG" <<PY` runs nothing and exits 0. Ends with the pattern that generated most of the work: seven times something reported success while achieving nothing. Exit codes were honest about crashes and silent about doing nothing, so every green result here was earned by checking content — and guard tests are mutation-verified, because a guard test never seen failing is not yet evidence. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
rhCat
added a commit
to rhCat/Maria
that referenced
this pull request
Aug 1, 2026
…roved nothing BRANCHING.md covers how to sync. This records what happened on the first one: upstream refactored the block-decision logic out of both executors into a single _run_agent_tool_execution_middleware, so our two gate hooks collapsed into one call site — a tighter waist (11 dispatch paths funnel through it) that also runs on final_args, after Relay rewrites, so ARGS_DIGEST finally covers what actually executes. The part worth keeping: tests/test_govern_gate.py passed 38/38 before any conflict was resolved, and would pass with the gate deleted — it never references tool_executor. That is cyberware NousResearch#235's shape exactly. Records the direct wiring proof and its mutation check, and makes 'the gate is still reached' step 3 of the next sync rather than an assumption. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
rhCat
added a commit
to rhCat/Maria
that referenced
this pull request
Aug 1, 2026
One govd on the VPS, many agents. Each agent authenticates as its own named principal with its own key and capability scope; all write to one shared ledger. The board shows one node — agents are the `principal` column on run rows, not board entries. This is the shape governance-pm drives at (M3-T03-retire-peragent); the per-agent govd is what that milestone retires. Covers the govd node, turning on the ed25519 identity scheme, an identity per agent, the container (Hermes' own provider config rather than a proxy, TUI/CLI as the surface), and verification. The section worth reading first is the failure mode: FIVE distinct faults all present as "agent healthy, every tool call denied", and they are indistinguishable from inside the container and from the agent's logs. The node's ledger is the only discriminator — rows-all-rejected is an AUTHORISATION problem (subject mismatch, narrow ACL, expiry); no-rows-at-all is a WIRING problem (stale chip, auth_verifier unset, image pre-NousResearch#235, or govd unreachable). Every one of these was hit at least once during this deployment. Records the derivation that makes the identity step portable: the subject is "ed25519:" + sha256(raw_pub).hexdigest()[:16], verified byte-identical against infra.cwp.sign.keyid — so a node whose cyberware clone is older than the image can still register a principal. And the traps, each of which cost real time once: services bind the tailnet IP rather than loopback (three separate services, three separate false "it's down"); a failed credential lookup makes compose no-op while looking like a successful deploy; deploy-node.sh overwrites run/ from etc/ and rewrites govd.json, silently reverting auth_verifier; no agent key is ever created for you; the agent-principal and monitor tokens are distinct and easily swapped. The govd half has been run against a live node. The container launch and first-claim verification have not been completed on a VPS yet, and the doc says so. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Problem
run_conversation()inrun_agent.pycreates an alias instead of a copy of theconversation_historyparameter:Since
messagesis the same object as the caller's list, everymessages.append()during the conversation also modifies the caller's original list. This is a classic Python mutable argument pitfall.Reported in #228.
Solution
This creates a shallow copy when a non-empty list is passed, preserving the caller's original list. When
conversation_historyisNoneor empty, a new list is created as before.Changes
run_agent.pyline 2572: 1 line changedTesting
Noneand[]inputs still produce an empty listCloses #228