oh-my-pi

Type: types/note.md

Evidence basis: source code and shipped doctrine at commit be6cb8217cd4c1dafcc86793ae5d809ea4d7396a (2026-09-05; package version 18.1.10), analysed 2026-09-26. No runtime or causal experiment was executed. This is a whole-system runtime analysis with explicit limits, not current-tip coverage. The exact analysis retains the evidence, canonical records and bounded memory profile.

Runtime and control

oh-my-pi encloses the coding loop: CLI/SDK entry, session and context assembly, model dispatch, tool execution, steering, task workers and retained state. The model proposes content and tool calls; symbolic code governs scheduling and continuation. Interactive, print and client-protocol modes share substantial session machinery. Tools return observations and effects into later turns; successful execution alone does not certify the resulting work. Exact records: RTE-1, RTE-3 and RTE-5.

The tool wrapper checks policy again after an extension revises arguments. Required approval without an interactive UI fails. Task workers, however, switch to unattended yolo mode while retaining explicit tool policies. Extensions also expose direct command execution. These are distinct control boundaries: a registered-tool approval check does not establish global process isolation. Optional worktree apply-back checks run status and Git applicability, with recovery artifacts on failure; neither check proves task correctness. Exact records: RTE-2, RTE-3 and RTE-4; wrapper source.

Retained context and memory

Session continuity reconstructs a selected branch around checkpoints. Maintenance alternatives include prose summaries, handoff text, rasterized conversation archives, mechanical shrinking with recovery references, and opaque provider-native replacement history. A readable display summary therefore does not identify everything the model consumes. Exact records: OBJ-12, OBJ-13, RTE-15 and RTE-16.

Cross-session backends differ in authority and selection. Local rollout processing derives project summaries and lessons; detailed files can be requested. Mnemopi retains scoped facts/episodes with lexical/vector access and ranking metadata. Hindsight exposes remote recall/reflect and curated bank material whose internal algorithms remain opaque. Sharpshooter extracts friction-selected project decisions and supplies them as instructions. Local memory explicitly asks for verification against current repository evidence; decision-memory wording carries stronger instructional force. Exact records: RTE-17, RTE-18, RTE-19 and RTE-20.

Durable trace-derived material has wired later consumers, satisfying the analysis profile's trace_learning write/read-back criterion. That does not establish improved capacity. Representation, selection and transformation aggregates remain uncertain where included payloads or services are opaque. The profile explicitly excludes auxiliary saved-rule, managed-skill and experiment routes; the whole-system account covers them separately.

Revision and warrant

The /omfg path turns a complaint into a stream rule, checks its matcher against assistant history, and feeds a failed candidate plus stated failure into another proposal. This wires all four theory-builder conditions for matcher-applicability repair. Human adoption remains part of the broader save route. A match does not prove that the rule detected the complained-of error, that its corrective instructions work, or that future work improves. Exact records: RTE-8 and RTE-9; controller source.

Managed skills provide another instruction-revision path, with structural/path checks and authored-skill precedence. Autoresearch instead changes an external project: it runs a benchmark, records results, accepts an agent-supplied keep/discard decision, commits or reverts, and supplies notes/results to later rounds. A supplied metric differing from the parsed value produces a warning with both values retained. The machinery enforces repository consequences; it does not independently enforce improvement or correctness. Exact records: RTE-10, RTE-11, RTE-12, RTE-13 and RTE-14.

The source supports bounded reflection through prior assistant behavior, saved rules and later control, and a standing complaint-responsive self-change pathway. It supplies no observed learning gain or occurrent self-improvement. Benchmarking external product changes does not itself improve the harness.

Scope

Provider weights, external service internals, actual deployment grants and candidate-linked runs were not inspected. Review/advisor judgments, checker results, retained summaries and successful application have different warrant. Evidence that would change the assessment includes a retained rule-repair/adoption/use trace, explicit experiment criticism shaping subsequent hypotheses, or controlled dependence on recalled material.