---

name: memory-safety-and-progress description: "Preserve truthful experiment context with bounded memory, append-only history, safety signals, gates, and read-only progress inspection." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

Memory, safety, and progress

Use this sub-skill when an experiment loop needs to remember what happened, avoid repeating dead ends, detect unsafe/stale progress, or report status. It is an operating guide for the project state contract; it is not the loop orchestration guide and it does not launch experiments, inspect GPUs, install skills, or mutate an Obsidian vault.

Operating rules

1. Treat PROJECT_BRIEF.md as frozen Tier 1 context. Never rewrite it as a side effect of logging or compaction.
2. Use workspace/MEMORY_LOG.md for bounded milestones and recent decisions. Use the ledger and journals for durable, append-only detail; do not pretend that a compacted memory log is the complete history.
3. Record every cycle outcome in workspace/experiments.jsonl when the ledger is enabled. One JSON object per line is the recovery boundary.
4. Keep DEAD_ENDS.md and INSIGHTS.md append-only. Rotation moves the full old file to a uniquely named .bak; never delete an archive to make a rotation appear successful.
5. Interpret safety, stagnation, and phase-gate results as signals. They are advisory: stop, back off, change direction, or ask for a human decision; do not fabricate a result from a signal.
6. A known failed terminal outcome stays failed everywhere. In particular, Slurm TIMEOUT, CANCELLED, OUT_OF_MEMORY, NODE_FAIL, and similar terminal states must remain failed in state, ledger, and dashboards.
7. For inspection, run the bundled read-only script. It reads a supplied project and emits checks; it never writes state, journals, exports, or archives.

First response to a progress request

• Read the brief, memory log, counter, state, ledger tail, journal tails, and cycle-time file if present.
• Report missing, malformed, stale, or unreadable files explicitly.
• Separate observed values (ledger/state) from derived values (best metric, stagnation, gate, rate-limit wait).
• If state.status == "failed", show the terminal state and do not downgrade it to completed, idle, or “probably okay”.
• If history is malformed, preserve the original bytes and use the valid records only; recovery is append-first, not rewrite-first.

Quick inspection

From this skill directory, use:

python scripts/inspect_progress_state.py --project <project>

Useful bounded variants are --format json, --ledger-tail 10, --journal-tail 1200, --metric-key acc, --metric-direction higher_better, --gate-threshold 0.8, --max-cycles-per-hour 6, and --now <unix-seconds>. Add --workspace <relative-or-absolute-workspace> when the project does not use the default workspace directory. The script only reads; it is safe for triage and synthetic fixtures.

Choose the right state mechanism

Need	Mechanism	Truth boundary
Frozen goal and constraints	PROJECT_BRIEF.md	Human/project input
Small context for the next think step	MEMORY_LOG.md	Bounded, lossy view
Every cycle and outcome	experiments.jsonl	Append-only event history
“Do not retry this”	DEAD_ENDS.md	Append-only, rotated
Durable observation	INSIGHTS.md	Append-only, rotated
Current process/result snapshot	state.json	Mutable current state
Restart-safe cycle number	.cycle_counter	Plain integer
Anti-burn starts	.cycle_times	JSON list of timestamps

Read the linked references for exact constructors, fields, formulas, output formats, recovery actions, and failure handling:

• memory-ledger-journal.md
• safety-and-gates.md
• progress-export.md
• troubleshooting.md

Handoff boundaries

Route core THINK/EXECUTE/REFLECT orchestration to autonomous-experiments. Route GPU discovery, utilization, and resource reservation to gpu-and-resource-operations. Route source skill installation or importing to skills-and-installation. This sub-skill may interpret their reported state, but must not silently take over those responsibilities.

When a report is requested, include: project goal; cycle count; current state; latest valid ledger records; metric direction and best value; stagnation and phase-gate verdicts with their parameters; journal tails and archive presence; active violations and rate-limit wait; export target/status; and unresolved file or backend errors. A missing metric is “no metric yet”, not zero. An indeterminate backend outcome is “unknown”, not success.