RLM profiles, history and budgets¶

The native Python RLM remains the default. It keeps repository context as data, uses a persistent Python namespace and supports semantic queries and recursive coding children. Version 2.6 adds experimental profiles without replacing that workflow. No model quality or cost improvement has been established yet.

Choose a profile¶

In the TUI, open :connect, choose Connect a harness with your model, then RLM. Existing host, Docker, Monty and optional A2A choices remain available. New choices include Python+Bash, Python+Bash in Docker, selective observations and RLM settings. You can also open :rlm settings directly.

Settings select the tool surface, execution boundary, observation strategy and inference limits. Save profile writes a project HarnessSpec. Start new session branches into a new session with that spec and continues model connection. Changing these settings requires a new session; an active worker retains its original policy. Existing A2A settings are preserved when saving a profile.

Direct connection shortcuts:

:connect rlm-host
:connect rlm-docker
:connect rlm-monty
:connect rlm-hybrid
:connect rlm-hybrid-docker
:connect rlm-selective
:connect rlm-settings
:connect rlm-a2a

Built-in presets:

Python+Bash needs POSIX process groups on the host. On Windows, select Docker. Bash must be installed in the selected environment. Monty remains a read-only research profile with no shell, writes or coding children. It supports history reads, context exploration and semantic queries. A combined Monty and Docker execution profile is future work.

Harness configuration¶

These fields belong under a native RLM HarnessSpec's runtime.config:

runtime:
  backend: rlm
  config:
    sandbox: docker
    allow_network: false
    tool_surface: python-bash
    observations: selective
    recent_messages: 12
    observation_chars: 2000
    budget:
      max_calls: 100
      max_tokens: 200000
      max_cost_usd: 5.0

Limits of zero mean unlimited. These limits apply to the complete session family: root inference, recursive coding children and semantic queries. Existing semantic-subcall quotas remain additional limits. The shared allowance persists across worker restarts and explicit child continuations.

The call allowance is checked atomically before inference. Token and USD limits are thresholds over provider-reported usage, not absolute billing caps: the current request can cross a threshold. With either threshold enabled, concurrent inference waits for a live request to settle before admission. An interrupted request requires explicit usage reconciliation. Missing token or price reports block further calls under the corresponding threshold. Known cost and unknown reports are shown separately, including in run statistics.

A2A stays off by default. It is selected separately through :rlm a2a, with explicit peers and key environment references. Hosted credits have their own admission policy and are not converted into model USD. Remote token and USD usage remains unknown. This release requires no Cloud Run changes for native profiles; hosted paid execution still depends on the service's existing setup.

History and selective observations¶

The persistent Python namespace offers branch-scoped history:

matches = history.search('failing assertion', limit=10)
page = history.read(matches[0]['id'], start=0, size=4000)
history.stats()

History includes original tool output and pre-compaction messages from the current branch. Handles from another branch cannot be read. Selective mode projects a provider-facing copy: large tool results become compact receipts and bounded previews, and older assistant responses become history references. It preserves user instructions, roles and tool-call/result identities. Full session records remain intact and readable through history handles. This is an experimental retrieval strategy; reducing visible output can also hurt quality.

Durable Bash jobs¶

Bash and Python share the same workspace, job ledger and sandbox. Python can manage a job without placing all its output into the conversation:

job = commands.run('python3 -m pytest -q', request_id='check-auth', timeout=120)
job.status()
job.wait(timeout=5)
job.read(stream='stdout', start=0, size=4000)

The Bash tool exposes run/start, status, read, wait, cancel and list. A command returns a compact receipt; read retrieves a bounded page. Captured output has a per-stream bound and reports omitted characters. Reusing a request ID with the same inputs returns the existing job; changed inputs are refused. Timeouts and cancellation stop the process group, including descendants.

Mutating commands serialize across the family. Explicit read_only jobs can run together; this is a caller declaration, not filesystem isolation. Convenience workspace writes also wait for a clear command lease. Raw Python filesystem or subprocess operations and the legacy shell.run bypass this coordination. Host Python is unrestricted. Use Docker for a boundary; these APIs are guardrails.

The broker filters credential-named environment variables. Host Python can still access the SuperQode process environment directly. Docker keeps provider credentials and inference on the host, and Monty exposes neither the filesystem nor subprocesses. An interrupted job retains an unknown outcome and a workspace lease; SuperQode never automatically retries its side effects.

Inspect and recover from the TUI¶

:rlm profile
:rlm budget
:rlm budget unresolved [offset]
:rlm usage
:rlm history failing assertion
:rlm history read <id> [start] [size]
:rlm jobs
:rlm job status <id>
:rlm job read <id> stdout [start] [size]
:rlm job cancel <id>

Operational inspection remains available during a model turn. Outstanding or unknown jobs prevent acceptance until they are settled. Inspect partial writes before explicitly recording a verified interrupted outcome:

:rlm job reconcile <id> <returncode> <verification reason>
:rlm budget reconcile <call-id> <tokens> <USD> <verification reason>

A surviving command must be cancelled first. Budget reconciliation refuses a live inference request; stop it and verify its provider usage before recording values. Reconciliation is an audit record, not a retry or a billing refund.

Compare profiles before claiming savings¶

The coding pilot example compares profiles using fresh workspaces, fixed model selection, alternating profile order, independent graders inside the selected execution boundary and per-attempt family usage. Offline mode uses scripted edits and only checks execution plumbing. The bundled tasks are small bug fixes; broader repository tasks and repeated live measurements are needed before claiming a quality or cost advantage.

Continual harness refinement, speculative execution and persistent agency are outside this release. The research notes record the design space that informed these changes.