agent-loop plugins

by David Giordano

Parts-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).

by David Giordano

Parts-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Observation-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).

by David Giordano

Observation-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Plan-anchored drift self-check for coding agents. Bundles the executive-self-monitoring skill (the 'what': re-ground in the active plan/gate before substantive work) with a per-host cadence hook (the 'when'). Not a blocker - a periodic nudge to avoid investing effort outside the plan. Portable core: the skill only; cadence hooks are host-specific (see hooks/ for Claude Code and Codex, cursor/ for Cursor).

by David Giordano

Plan-anchored drift self-check for coding agents. Bundles the executive-self-monitoring skill (the 'what': re-ground in the active plan/gate before substantive work) with a per-host cadence hook (the 'when'). Not a blocker - a periodic nudge to avoid investing effort outside the plan. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Structured-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; HANDMON_STRICT=1 blocks once. Portable core: the skill only; hooks are host-specific (hooks/ for Claude Code and Codex, cursor/ for Cursor).

by David Giordano

Structured-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; HANDMON_STRICT=1 blocks once. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Real-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).

by David Giordano

Real-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Persist-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).

by David Giordano

Persist-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Cross-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).

by David Giordano

Cross-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.

by David Giordano

Checkable-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).

by David Giordano

Checkable-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. Codex: the same hooks/ adapter as Claude Code - same events, same payload, same output envelope; the hooks run once you trust them with /hooks.