Directive-Braindump Triage — execute only the residual
Given an operator directive-braindump (a *.braindump.md / prompt-aux-N-style scratch of
mixed directives), produce/update ONE provenance ledger recording WHICH corpus artifact
already fulfills each directive — so a future amnesic agent runs only the verified residual.
Canonical output (single file): a themed provenance ledger — default in the host's
planning/scratch dir as <theme>-provenance-ledger.md; repo-scoped if the braindump is repo-bound.
Why: a braindump without a provenance ledger gets re-learned. A directive without
application is a dead file. This is the cure for "we keep re-learning what we already knew."
Input: path to a directive-braindump file. Optional arg = an existing ledger path to update.
Distinct-from-siblings (DRY — composes, never duplicates)
agentic-tool-intake(skill) decides whether to ADOPT an external tool → this classifies operator DIRECTIVES already in hand. Different input, different verdict space.protocols/agentic-tool-lifecycle.md(protocol) governs distill→promote of tools → this triages intentions, then feeds the lifecycle when a residual warrants a new tool.corpus-firing-audit(sibling skill): audits the standing corpus for firing-vs-theater (artifact → application). This triages a braindump file (directive → artifact). Opposite traversal; pair the two skills, never merge them.- anti-theater grounding (host rule, if present) → used in P6 (no hallucinated provenance).
Idempotency contract (non-negotiable)
- Probe before acting — read the braindump's
PROCESSEDbanner (if any) + any existing ledger FIRST; never re-run a triage that already converged. - Update in place — append/refresh the SAME ledger; never spawn a parallel doc (DRY).
- Convergence — no corpus change ⇒ semantically identical ledger (only generated-at differs).
- Read-only corpus — the ONLY writes are the ledger + a
PROCESSEDbanner on the source braindump. Never re-execute already-DONE directives.
Steps
P0 — Recon (read-only, observe-before-assume). OBSERVE before assuming: read the braindump; probe the auto-loaded corpus (the host's rules dir, memory index, tickets via the project tracker / MCP) AND the environment relevant to the directives. Read any existing ledger. Treat felt-certainty over an unprobed corpus as a smell (a Dunning-Kruger guard), not a signal.
P1 — Decompose + harmonize. Restate the braindump as N atomic directives — edited, corrected, expanded — grouped by ontological family (e.g. Posture · Validation · Provenance · Governance · Meta-learning · Strategy). This is the polished canonical form; preserve the raw scratch as source.
P2 — Classify × provenance. For each directive emit: status ∈ {DONE · OPEN · DROP-EXPLICIT · COVERED} + the artifact that fulfills it (rule / memory / ticket / PR / commit / skill).
Anti-re-learning rule: a directive already satisfied by the corpus is DONE, not pending.
Session meta-directives (slash-commands, --auto-merge, motivational asides) are EXCLUDED —
they are execution authorizations, not reusable content.
P3 — Inter-dependency graph (mermaid DAG): recursive · sequential · inter-dependent · independent · blocked. Mark the critical residual chain + transversal/independent items.
P4 — Residual roadmap (Eisenhower 2×2). ONLY the verified-residual gets a disposition: Q1 ticket+execute · Q2 ticket+schedule · Q3 ticket+delegate · Q4 note-or-drop-explicit. Drop decisions are logged with rationale (drop-explicit > unpaid memory-debt).
P5 — Anti-re-learning note + persist. Record the meta-lesson (provenance + sharpen-an-existing-
fire-point > add a passive rule). Stamp the source braindump with a PROCESSED <YYYY-MM-DD>
(ISO-8601) banner pointing at the ledger. The pair is atomic (rubric C1, pilot #354):
banner-without-ledger-row OR ledger-row-without-banner = the triage FAILED — an amnesic
future session will either re-learn the dump (no banner) or re-hunt the dump (no row).
If either write fails, mark the run incomplete and fix the missing half before closing.
P6 — Verify + brief. Confirm every "DONE" cites a real, locatable artifact (no hallucinated provenance — every claim points at something a reader can open); emit a short briefing (N DONE / OPEN / DROP + the residual next-action).
Example (input → output)
Input: directive-braindump-triage docs/scratch/posture.braindump.md
Output — a provenance ledger (excerpt):
# posture — provenance ledger (generated 2026-06-22)
| # | directive (atomic) | status | fulfilling artifact |
|---|--------------------------------|---------------|--------------------------------|
| 1 | recon the env before assuming | DONE | rule: <env-recon rule> |
| 2 | stamp agent-authored tickets | DONE | rule: <ticket-provenance rule> |
| 3 | add a posture self-test | OPEN | — (residual → Q2 ticket+sched) |
| 4 | rename every internal variable | DROP-EXPLICIT | rationale: out-of-scope, low value |
Residual: only #3. DAG: #3 independent. Eisenhower: Q2.
Re-running with no corpus change ⇒ the same ledger (only the generated date differs) — the
idempotency contract. Directives #1/#2 are never re-executed (already DONE); #4 is logged-dropped,
not silently abandoned.
Guard-rails
Recon-before-assume · Dunning-Kruger guard · never re-execute already-DONE directives · escalate
to a human only post-recon when genuinely beyond reach · stamp any created ticket as agent-authored ·
auto-merge only within the host's standing-authorization guardrails · transcribe the delegation
heritage (see protocols/agent-delegation.md) on any delegation · keep outputs secret-clean.
Provenance
Genesis: distilled as the form-correspondent of an operator agentic-posture braindump in a
downstream project's prompt-library, then elevated to a user-scope skill. Promoted here via the
protocols/agentic-tool-lifecycle.md distill→promote path after clearing the ADR-005
≥2-dogfood-cycle gate — two ratified cycles (the triage analyses of an agentic-posture braindump
and a repo-inventory braindump). Cross-link slug: [[directive-braindump-triage]].