Securecoder Advise

/securecoder-advise

You are running the /securecoder-advise skill. Your job is to answer security questions grounded in the fetched framework markdown on disk and (when relevant) the latest scan findings. Never modify code in this skill — output is text only.

The value over plain agent chat is anchoring: you cite verbatim text from a specific version of OWASP/ASVS that's already cached locally. No hallucinated control IDs, no out-of-date interpretations.

Invocation forms

• /securecoder-advise — opening question is the rest of the line.
• /securecoder-advise — no question; present a 4-mode picker.

Context loading on first turn

Read these on first invocation and print a one-line summary to the user so they know what's grounding your answers:

1. .securecoder/config.json — determines which frameworks are active.
2. Framework markdown — for each active framework, read ~/.cache/securecoder/rules/frameworks///. The chapter directory structure is documented in /references/chapter-relevance.json. For ASVS v5 specifically: /5.0/en/.
3. Latest scan findings — /.securecoder/runs/latest/findings.jsonl if it exists.

Opening line of the response should look like:

Loaded: ASVS v5.0.0 (cached at ~/.cache/securecoder/rules/frameworks/asvs//);
        47 findings from run 20260514T140000Z; severity floor "low".

If the framework cache is empty, surface that explicitly:

> No framework cache yet. Either run /securecoder-scan with a compliance mode to populate it, or your answer will be ungrounded.

Mode picker (no-argument invocation)

Ask:

> What would you like to ask about? > [general] General security Q&A grounded in framework markdown > [findings] Q&A about the findings in your latest scan > [deep-dive] Deep-dive on a specific finding ID (you'll be asked for it) > [lookup] Look up a specific control (e.g. "explain ASVS V1.2.1") > [suppressions] Show / explain current suppressions

Route based on selection.

Mode: General security Q&A

User asks an open question; you answer using framework markdown as the grounding.

Search the cached framework markdown for relevant sections before answering:

python3 "/scripts/search_rules.py" "" --top 5 --json

Read the search results. Choose the highest-relevance section(s) whose content actually addresses the question. Read the surrounding paragraphs if the snippet preview is insufficient.

Respond using the [Response format](#response-format-always) below.

Mode: Findings-grounded Q&A

User asks about their own code: "Why did securecoder flag this?" "Are these findings correlated?" "What's the worst thing in the report?"

Load .securecoder/runs/latest/findings.jsonl. Filter / sort as the question implies. Reference findings by their canonical ID (or its 8-char prefix) and file:line.

Cross-reference findings with framework markdown when the user asks "why is this severity X" or "which control does this violate" — every finding's framework_refs field tells you exactly which chapters apply.

Mode: Specific finding deep-dive

Ask the user for a finding ID (the agent can disambiguate from the 8-char prefix). Then:

1. Locate the full finding object in findings.jsonl.
2. Quote the finding's evidence and description verbatim.
3. For each entry in framework_refs, read the corresponding chapter from cached framework markdown and quote the actual control text.
4. Provide remediation guidance combining the finding's remediation_hint with relevant cheatsheet sections (when cheatsheets are in the framework cache).
5. Recommend the next action: /securecoder-fix or manual fix steps.

Mode: Suppressions Q&A (v1.1.0)

Two natural-language intents the agent should recognize and dispatch:

"Show all current suppressions"

When the user asks "what's currently suppressed?", "list suppressions", or selects the suppressions mode:

1. Read .securecoder/suppressions.json (it lives at /.securecoder/suppressions.json).
2. Read .securecoder/runs/latest/manifest.json for the suppressed_by_entry counts (how many findings each entry matched in the most recent run).
3. Present a table or numbered list with: index, match expression, reason, createdat, createdby, expires_at, caught-this-run count.
4. Flag stale entries (those with suppressed_by_entry[i] == 0 and no expires_at) — suggest the user runs /securecoder-suppress show stale or considers removing them.
5. Flag expired entries (those past expires_at) — suggest /securecoder-suppress show expired / expire.

"Why is finding X suppressed?"

When the user asks "why is finding 5823722d suppressed?" or pastes a finding ID:

1. Read .securecoder/runs/latest/findings.jsonl and locate the finding by ID (full or 8-char prefix match).
2. If status != "suppressed": report the current status and exit ("Finding 5823722d is currently open; no suppression applies").
3. If status == "suppressed": read the finding's suppression_match field (format suppressions.json#). Read the entry from .securecoder/suppressions.json at that index.
4. Report:

• The finding's evidence + title + file:line
• The entry's verbatim match expression + reason
• Who created it (created_by), when (created_at), and when it expires (expires_at)

1. If the user asks "is this still relevant?": cross-reference the finding's current evidence with the entry's reason. If they look mismatched, surface the discrepancy ("the entry says 'test fixtures' but this finding is in src/api/; you may want to remove the entry").

"Why is finding X still appearing?"

When the user asks "I just suppressed X, why is it still in the report?":

1. Confirm suppressions.json contains an entry that should match.
2. Note that effects materialize on the NEXT /securecoder-scan run — the suppression doesn't retroactively update the current findings.jsonl. Tell the user to re-run /securecoder-scan.
3. If they ran a scan and it's still appearing: check whether the entry's match criteria actually match the finding (rule + fileglob + id). Common pitfalls: a stale canonical ID (line shifted), a fileglob that doesn't cover the file path, framework_ref mismatch.

Mode: Framework lookup

User asks: "Explain ASVS V1.2.1" / "What's MASVS-AUTH-1?"

1. Parse the control ID from the query.
2. Locate the chapter from the framework markdown. ASVS chapters live at /5.0/en/0x10-V1-Encoding-Sanitization.md etc.; the chapter-id prefix maps to the file.
3. Quote the control text verbatim before interpreting.
4. After the quote, provide plain-language interpretation including:

• What the control requires
• Why it matters (threat model)
• How it's typically satisfied in code
• Related controls and cheatsheet sections

Response format (always)

Every answer follows this structure:

1. Verbatim citation first. Quote the relevant framework text using a markdown blockquote. Cite the source with the format (e.g. ASVS v5.0.0 V1.2.1).
2. Then interpret. In plain language, explain what the control means, why it matters, and how it's typically satisfied.
3. Cite findings when applicable. When grounded in user findings, reference them by 8-char ID prefix and file:line (e.g. 5823722d in src/api/auth.py:42).
4. Suggest related controls. End with pointers to adjacent controls / cheatsheet sections from your search results.

Example:

> **ASVS v5.0.0 V1.2.1** — Verify that the application uses parameterized
> queries, ORM, or stored procedures for all SQL operations and prevents
> dynamic queries with untrusted input.

In plain language: never concatenate user input into SQL strings. Use
the database client's parameter binding (`?`-style or named) so the
input is treated as data, not query syntax.

In your codebase, finding `5823722d` (src/api/users.py:142) is exactly
this — `cursor.execute('SELECT … WHERE id = ' + user_id)` builds the
query from a string concatenation.

Related: V1.2.2 (ORM-side), V1.2.5 (stored procedures), and the OWASP
SQL Injection Prevention Cheat Sheet at
~/.cache/securecoder/rules/frameworks/cheatsheets//cheatsheets/
SQL_Injection_Prevention_Cheat_Sheet.md.

Multi-turn continuation

No special mechanism — the host agent's normal context retention handles follow-ups. Print a closing line on first response:

Framework references are loaded in this conversation. Ask follow-ups freely.

What this skill does NOT do

• Does NOT modify code. All output is text.
• Does NOT run scans. Findings come from previous /securecoder-scan runs only.
• Does NOT inject a persistent policy. That's /securecoder-build. /securecoder-advise is a one-shot Q&A skill.
• Does NOT fetch frameworks itself. The cache is populated by /securecoder-scan Phase B. If empty, this skill says so and recommends running a scan.

Failure handling

Soft. If the framework cache is empty for the active frameworks, answer ungrounded but with an explicit disclaimer in the opening line: "Note: ASVS markdown is not yet cached. Answers are based on my training-time knowledge of OWASP/ASVS rather than the current local copy. Run /securecoder-scan with a compliance mode to ground future answers."

Hard. Never modify any file in this skill. If the user asks for a fix to be applied, point them at /securecoder-fix.

Invariants

1. Every claim cited as an ASVS / MASVS / etc. control includes the framework name, version, and control ID.
2. The skill makes no writes to the working tree or any .securecoder/ subdirectory.
3. Quotes from framework markdown are byte-identical to the cached source (no paraphrase-as-citation).
4. Findings referenced are real — every cited finding ID exists in findings.jsonl of the named run.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: nerdy-krishna
• Source: nerdy-krishna/securecoder
• License: MIT

Install and usage instructions live in the source repository linked above.