boshu2__agentops/MEMORY.md

Project Memory

Architecture

• Harness Reuse: Long-running RPI improvements should extend existing phase/result/handoff artifacts instead of creating a second run ledger (source: .agents/learnings/2026-03-28-last-week-commits.md)
• Trust-Gated Context: Startup quality improves when canonical findings and planning rules outrank raw notes and packet families stay experimental until health is proven (source: .agents/learnings/2026-04-01-session-intelligence-trust-gates.md)
• Sidecar Single Writer: Parallel read lanes should emit sidecars while one reducer owns durable queue writes (source: .agents/learnings/2026-04-26-dream-generator-sidecar-single-writer.md)

Process

• Proof Before Close: Child bead closeout should record scoped files or proof artifacts so closure-integrity stays replayable (source: .agents/learnings/2026-03-28-last-week-commits.md)
• Remote CI Counts: Release completion requires the first green remote validate run, not only the local release gate (source: .agents/learnings/2026-03-22-v2.29.0-full-release.md)

Debugging

• Tracker Skew First: When tracker probes fail with schema errors, check the tracker binary version and cache state before blaming repo code. For br, the SQLite cache can go stale under concurrent ledger writes; the JSONL is the source of truth (originally learned on the retired bd; source: .agents/learnings/2026-03-24-codex-hookless-followup.md)

Patterns

• CLI-Owned Lifecycle (superseded): skills should delegate lifecycle state to the CLI instead of parsing state files themselves, but the specific commands this taught (ao codex ensure-start / ensure-stop) are now deprecated legacy shims; see AGENTS-RUNTIME.md (source: .agents/learnings/2026-03-24-codex-hookless-followup.md)

Key Lessons

• Validation Before Autonomy: Prove code and agent output before expanding orchestration or autonomy. (source: .agents/learnings/2026-06-16-validation-before-autonomy.md)
• RPI Run Context Needs Run-ID Scope: Phase handoff readers should filter by current run ID before injecting summaries, or unrelated prior RPI runs can contaminate implementation and validation prompts (source: .agents/learnings/2026-05-01-rpi-run-context-needs-run-id-scope.md)
• CI Scanner Syntax Parity: Contract validators that mirror each other need paired fixtures for every accepted syntax form, not only comments saying the regexes agree (source: .agents/learnings/2026-04-27-ci-contract-scanners-need-syntax-parity.md)
• Audit Parser Reality: Mechanical auditors must parse File: prose, anchors, and examples correctly or they create noisy false failures (source: .agents/learnings/2026-03-28-last-week-commits.md)
• Output Mode Orthogonality: --json must preserve normal command side effects and change only serialization (source: .agents/learnings/2026-04-10-output-modes-must-not-change-command-side-effects.md)
• Pair Command Refactors With Tests: Production command refactors under cli/cmd/ao/ should ship with direct test diffs so the command/test-pairing gate is designed for, not rediscovered at push time (source: .agents/learnings/2026-04-14-command-refactors-need-paired-tests.md)
• Scrub RPI Runtime Env in Raw Go Checks: Raw validation of internal/rpi is not trustworthy on this machine unless AGENTOPS_RPI_RUNTIME is explicitly scrubbed first (source: .agents/learnings/2026-04-14-scrub-rpi-runtime-from-raw-validation.md)
• Corroborate Before Council: Dream long-haul should spend budget on cheap packet corroboration before paying for council; strong corroboration can make the expensive lane unnecessary (source: .agents/learnings/2026-04-14-dream-longhaul-postmortem.md)
• Durable Closure Evidence: Closed beads should cite committed discovery artifacts or explicit proof packets, not ephemeral seed paths, so closure audits can replay mechanically (source: .agents/learnings/2026-04-14-dream-longhaul-postmortem.md)
• Timeout-Aware Generators: Pluggable generators need context-aware IO and stall tests before joining Dream fanout (source: .agents/learnings/2026-04-26-generator-timeout-contract.md)
• Stack PR Queues Base-First: Drain stacked PR queues by merging foundations first, then rebasing remaining branches onto the updated target with explicit force-with-lease proof (source: .agents/learnings/2026-04-30-pr-queue-stack-drain.md)