Review before you close. Bundles the aspiration-self-monitoring skill (the 'what': after the last change, review the result in the medium it is used in - read the document, view the image, run the code, play the audio - and compare it, point by point, with a source the agent did not write - the original it reproduces, the specification it follows, the real input, the owner's words; meets only when that comparison shows it, a difference still there is a defect, a source out of reach is unverified) with per-host hooks (the 'when') that record which edited files nobody reviewed since their change and scan the final message for a completion claim and the [ASPIRATION CHECK] block. Non-blocking by default; ASPMON_STRICT=1 blocks once. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopself-monitoringmetacognitionParts-ledger discipline for multi-part or hard tasks. Bundles the coverage-self-monitoring skill (the 'what': write the ledger first - the parts, which is hardest and why, hardest first - and close each part as done, blocked with an observed reason, or returned to the owner) with per-host hooks (the 'when') that count the stub / placeholder / TODO markers written per turn, ask for a ledger when the prompt enumerates three or more items, and scan the final message for deferred work with no [COVERAGE CHECK]. Never blocks. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopcoverageself-monitoringObservation-vs-conjecture discipline for coding agents. Bundles the epistemic-self-monitoring skill (the 'what': claims carry their evidence and a named falsifier; only verified claims become facts or closures) with per-host hooks (the 'when'). Non-blocking by default; an opt-in strict mode gates closures on a complete [EPISTEMIC CLOSE] block. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
verificationagent-loopself-monitoringStructured-handoff discipline for the final message of a turn. Bundles the handoff-self-monitoring skill (the 'what': a close modelled on the SBAR / I-PASS handoff protocols - Status first, Situation in the reader's terms, the fork as Options with a default, one Next action - so a reader who has the message and not the trace can act on it) with per-host hooks (the 'when') that inject the format on the first closing-shaped tool call of a turn and scan the final message for a decision named but not handed off - an offer, a fork, a closing question, a returned coverage part - and for the [HANDOFF] block. Non-blocking by default, except a blocked close in a turn whose search found nothing, which blocks once; HANDMON_STRICT=1 blocks once on any finding. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
handoffagent-loopself-monitoringKeep a change to what the request names. Bundles the hygiene-self-monitoring skill (the 'what': list the public behaviour a change reaches - an exported function, a helper other callers use, a sibling fixed for consistency, a documented default - mark what the request names, and put back or offer to the owner what it does not; size is not the measure) with per-host hooks (the 'when') that record which public functions the session's edits changed, file and line, and read the [HYGIENE CHECK] at the close against that record. Non-blocking by default, except a block that contradicts itself (Outside kept, or put back while the code still has it), which blocks once; HYGMON_STRICT=1 blocks once on any finding. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopself-monitoringpublic-apiReal-result discipline for work that cannot be done as asked. When a service cannot be reached, a key is missing or two tests want different answers, the easy route is a result that only looks done: a fallback that returns made-up values, a table standing in for the service, code that answers its caller differently, a machine changed until the run goes green. Bundles the integrity-self-monitoring skill (the 'what': the result must be real and the route to it legitimate; "cannot be done as asked, because X" is a complete answer; close with an [INTEGRITY CHECK] - Result, Route, Outside the task, Told the user) with per-host hooks (the 'when') that read each edit to product code for those shapes - only in a file that calls a service - and each shell command for a change to the machine or a server started, and say what it means for the user once per finding. No message on the prompt. Never blocks. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopself-monitoringshortcutPersist-or-quit self-check for coding agents. Bundles the persistence-self-monitoring skill (the 'what': name the hypothesis held, count the variants, name a rival, decide) with per-host hooks (the 'when') that count what an agent cannot feel - the same file edited N times, the same command or error failing N times, tool calls piling up since the user's last message - and speak only when a threshold is crossed. Never blocks. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopself-monitoringmetacognitionCross-session ledger discipline for work that outlives the session. Bundles the progress-self-monitoring skill (the 'what': one file in the project, .agent/progress.md, that holds what the next session cannot recover from the code - the parts still blocked or returned with their observed reason, and the next action - written by the agent and re-opened before substantive work) with per-host hooks (the 'when') that announce the ledger's open-item count and age when a session opens or continues after a compaction, and notice a turn that edited files and left a ledger with open items untouched. The hooks read that one file's metadata and count, never its text, and never write it. Never blocks. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
resumeagent-loopledgerCheckable-reason discipline for stopping, deferring, narrowing or softening. Bundles the termination-self-monitoring skill (the 'what': a state-shaped reason - fatigue, clock, context budget, confidence as a feeling, difficulty as a mood, a run of apologies - is replaced by a checkable one: gate-not-run, owner-choice, budget-spent, limit-observed, or none and continue) with per-host hooks (the 'when') that scan the final message for those persona artifacts and for the [TERMINATION CHECK] block. Non-blocking by default; TERMMON_STRICT=1 blocks once. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).
agent-loopself-monitoringmetacognition
