sast-prioritise
Skill25 starsChanged 4 months agoApache-2.0View on GitHub
The stage between SAST enumeration and the finder. Takes the stateful cell QUEUE (algorithmically pre-scored by measured priors) and spends a small agent budget deciding what is WORTH READING and what question to ask — never whether a bug exists, which is the finder's job. Writes its judgement back into the queue so the corpus is worked through incrementally: these cells have been looked at, these have not. Use when asked to "prioritise the cells", "triage-lite", "what should the finder read first", or before spending a finder pass on a large corpus.
In this repository21 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- customizeSkill
- patchSkill
- quickstartSkill
- sast-drivenSkill
- sast-prioritiseSkill
- threat-modelSkill
- triageSkill
- variant-scanSkill
- verifySkill
- vuln-scanSkill
- customizeSkill
- patchSkill
- quickstartSkill
- sast-drivenSkill
- threat-modelSkill
- triageSkill
- variant-scanSkill
- verifySkill
- vuln-scanSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 49
If the queue does not exist, build it first — never prioritise from a raw cells.jsonl, because the result would have nowhere to live.
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
/sast-prioritise
The SAST layer enumerates; the finder reads; this stage decides reading order. It exists because the two numbers on either side of it do not meet: a corpus pass produces thousands of cells, a finder pass costs an agent call each, and without a stage in between the only lever anyone reaches for is budget — which conflates what is worth reading with what we can afford.
This stage never decides whether a defect exists. It decides whether a cell deserves a finder's attention and, more usefully, what question the finder should arrive with. A cell it ranks low stays in the queue; see the discipline at the end.
Step 0 — Read the queue, do not rebuild it
The queue is state, not a run artifact. scripts/build_sast_queue.py (re)builds it from
cells.jsonl and preserves every human/agent decision already recorded. Load it and report:
cellstotal,by_status,by_class, and how many carry a decision already.- The prior range, and the fact that priors are priors — measured directions, not mined rules.
- What is NOT yet minable: until cells carry outcomes, no predictor can be fit. That is the point of writing decisions back.
If the queue does not exist, build it first — never prioritise from a raw cells.jsonl, because the
result would have nowhere to live.
Step 1 — Understand the three cell families before judging any of them
They are not the same kind of object and must not be judged by one rubric. This is a correction from
measurement, not taste: the first prior blended them and put limit-pair cells with a SINGLE
declaration at the top of the queue, which is exactly backwards.
| family | what a cell means | what makes it worth reading |
rule cell (resource, index-surface, unsafe, …) | N rule hits of one class in one module | RARITY. E7c: 1.9 hits/crate → 11.1 % discrimination vs 32.5 → 0.5 %. In the h2 pass the two SMALLEST cells produced both real leads; the five largest (527 hits) produced none |
limit-pair | cap DECLARATIONS vs ENFORCEMENT sites in one module | the ASYMMETRY. object: 287 declared / 13 enforced overall, concentrated (src/macho.rs 51/0). A 1-declaration cell is the least interesting, not the most |
coverage-gap | a shipped file no rule pointed at at all | density of LOGIC-shaped sinks (alloc / unsafe / unwrap). Not index/cast/arith_sub: lopdf's glyphnames.rs scored 3,281 of those and is a pure data table |
Step 2 — Select the slice, and say what you are NOT looking at
Take the top --top N by prior among status: unread, honouring --crate / --class filters.
Then state, in the same breath, how many cells were left unread and at what prior — a slice that does
not name its own tail reads as a complete pass.
Step 3 — One agent per cell
Spawn subagent_type: "general-purpose", all in one message. The brief below is derived from
/vuln-scan's review brief and /triage's ranking prompt, cut down to the one judgement this stage
makes.
You are deciding whether a static-analysis CELL is worth a security reviewer's time, and what
question that reviewer should arrive with. You are NOT deciding whether a bug exists — do not
attempt a verdict, and do not report a finding.
You may Read/Grep the repository at {REPO_ROOT}. Do NOT execute anything.
CELL
id: {cell_id}
crate: {crate}
class: {class} module: {module} data path: {data_path or "unclassified"}
evidence: {hits} hits from {engines}
rules (rarest first): {rules}
sites: {sites} ← every site, not a sample
prior: {prior} from {prior_terms} ← measured directions, NOT a mined rule
FAMILY RUBRIC — use the one that matches `class`:
· rule cell → is this class of operation, in THIS module, on a path that consumes untrusted
input? A rare rule firing in a parser is worth more than a common rule anywhere.
· limit-pair → a cap is declared {declared}× and enforced {enforced}× here. Is there a path
that reaches the guarded operation WITHOUT consulting the cap? That gap is the
whole question; a module that declares one cap and enforces none may simply have
nothing to guard.
· coverage-gap → NO rule fired anywhere in this file. Does it contain reachable logic (parsing,
arithmetic on parsed values, indexing, unsafe), or is it data/boilerplate?
────────────────────────────────────────────────────────────────────
STEP 1. Read enough of the module to answer: what does this code DO, and is it on a path that
sees attacker-controlled bytes? Name the entry point you traced to, or say you could not find one.
STEP 2. Assign PRIORITY, and be willing to say `low`:
high — reachable from untrusted input AND the cell's shape could plausibly break there
medium — reachable, but the shape looks guarded or the class is weak here
low — not attacker-reachable, or the sites are structurally uninteresting (data tables,
generated code, pure boilerplate)
unknown — you could not establish reachability from what you read; say what is missing
STEP 3. Write the QUESTION a finder should arrive with — one sentence, specific to this module.
"Check for bugs" is a non-answer; "does `parse_header` bound `len` before the `vec![0; len]` at
:214" is an answer.
STEP 4. State what you did NOT read, and what would change your priority.
Return JSON only:
{"cell_id": "...", "priority": "high|medium|low|unknown",
"reachable_from": "entry point or null", "question": "...",
"reason": "2-3 sentences", "unread": "what you did not look at"}Assume nothing from the prior. The prior chose the reading order; it is not evidence about the code. An agent that reproduces the prior instead of reading has done nothing.
Step 4 — Write decisions back into the queue
For each judged cell set agent_priority, agent_reason, question, status: "in-review",
updated, by. Never drop a row. Then report:
- The priority histogram, and specifically how often the agent DISAGREED with the prior — that is the first evidence about whether the priors are any good, and the seed of the future mined predictor.
- Cells the agent could not establish reachability for (
unknown): a queue that hides those is hiding its own blind spot. - What remains
unread, with the prior at the cut.
Constraints
- Priority is an ORDER, never a CUT. A
lowcell stays in the queue with its reason recorded.deprioritisedis not terminal and nothing is deleted — otherwise this stage re-creates the silent zero ("we never looked" reading as "there was nothing") one layer higher. - No verdicts. This stage may not mark a cell
candidateorrefuted; only a finder pass with an independent PoC may. - Priors are priors. They are measured directions (E7c rarity, the class × engine grid, the limit-pair asymmetry), not mined rules. Do not present them as learned.
- Every judgement is written back. An agent call whose output is not persisted converts budget into nothing and leaves the corpus exactly as unlabelled as before.
- Say what was not read. At both levels: the agent names what it skipped inside a cell, and the stage names the tail of the queue it never reached.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

