File dump

File Dump

Task

Write the requested content — whatever the user asked to dump: an analysis, review, report, comparison, plan, or output produced in this session — to exactly one file whose name and location follow the standard dump convention: {description}-{model}-{date}-{session} plus an extension, saved under {save_dir} (default .ai-coding-artifacts/dumps/ in the workspace root). This skill owns the naming convention; trajectory-snapshot follows the same convention for session-trajectory dumps.

Deferral: when the thing being dumped is the session trajectory (the ordered timeline of what happened), use the trajectory-snapshot skill instead; when it is resumable session state for continuing work in another chat or tool, use the save-session-state command.

Parameters

• {description} — kebab-case slug naming the content; optional, default: derived from the user's request and the content itself (1–5 lowercase words joined by hyphens, e.g. api-error-analysis). MUST match [a-z0-9]+(-[a-z0-9]+)*.
• {save_dir} — directory the dump is written into; optional, default: .ai-coding-artifacts/dumps/, resolved relative to the current workspace root (MAY be an absolute path). An explicit directory or full path in the user's request overrides the default.
• {filename_format} — filename template; optional, default: {description}-{model}-{date}-{session}.md. Field order in the default is exactly description, model, date, session. The default renders like api-error-analysis-claude-4.5-opus-2026-07-12T21-04-30Z-a1b2c3d4.md. When the content is natively a non-markdown format (JSON, YAML, CSV, JSONL, patch, plain log text), the .md extension is replaced by the content's native extension. When the user supplies an explicit filename, it is used as given and no tokens are imposed on it.
• {model} — slug of the model powering this agent session (e.g. claude-4.5-opus, gpt-5.2); optional, default: auto-detected from the session's own metadata or system context, normalized to lowercase with spaces and slashes replaced by hyphens (dots are kept); unknown-model when undeterminable.
• {date} — filesystem-safe UTC timestamp from date -u +%Y-%m-%dT%H-%M-%SZ (e.g. 2026-07-12T21-04-30Z); not user-set.
• {session} — first 8 characters of the current session id; optional, default: auto-detected from the tool's transcript location — Cursor: ~/.cursor/projects/<workspace-slug>/agent-transcripts/<session-id>/; Claude Code: ~/.claude/projects/<munged-cwd>/<session-id>.jsonl; when the id is unknown, the most recently modified transcript in that directory is the current session (it is being appended live); unknown when undeterminable.

Success Criteria

• Parameters resolve to explicit or default values before any write; a resolved {description} that does not match [a-z0-9]+(-[a-z0-9]+)* aborts with an explanatory error and no filesystem side effects.
• The dump exists at the absolute path resolved from {save_dir} plus the rendered {filename_format} against the workspace root, parent directories created as needed, and is non-empty.
• The filename is rendered with all tokens substituted; no unresolved {…} tokens remain; undeterminable {model} / {session} render as unknown-model / unknown rather than being omitted.
• No pre-existing file is overwritten: on collision, a -2 (then -3, …) suffix is appended before the extension.
• An explicit path or filename in the user's request wins over the convention: it is used as given (resolved against the workspace root when relative), and only the collision rule still applies.
• Markdown dumps open with the provenance header from Output Format; non-markdown dumps contain the content raw with no header or wrapper.
• The dumped content is exactly what was asked for — nothing fabricated, nothing silently truncated.
• Secrets that appeared in the session (tokens, private keys, .env contents, password= / Authorization:-style literals) appear in the dump only as <redacted>.
• The chat reply is a one-line confirmation with the absolute dump path; the file body is not duplicated into chat.

Guardrails

• MUST NOT overwrite any existing file; resolve collisions with numeric suffixes.
• MUST create missing parent directories for the resolved output path.
• MUST honor an explicit user-supplied path or filename over every convention default.
• MUST NOT fabricate content: the dump contains what the session actually produced; unknown facts are recorded as unknown, not invented.
• MUST NOT inline secrets; redact with <redacted>.
• MUST NOT commit, push, or modify any file other than the dump file itself (including .gitignore — leave untracked dumps untracked).
• Scope: write one requested dump file. Out of scope: session trajectories (use trajectory-snapshot), resumable session state (use save-session-state), uploading artifacts anywhere.

Workflow

1. Classify the request: session trajectory → trajectory-snapshot; resumable session state → save-session-state; anything else continues here.
2. Resolve {description}: take the user's slug when given; otherwise derive a 1–5-word kebab-case slug from the request and content; abort if the resolved value fails the slug pattern.
3. Detect {model} from the session's own metadata or system context and normalize it; detect {session} from the transcript location; take {date} from date -u +%Y-%m-%dT%H-%M-%SZ; fall back to unknown-model / unknown rather than guessing.
4. Pick the extension: .md for markdown or prose content; the content's native extension otherwise.
5. Render the filename from {filename_format}; resolve it against {save_dir} (or the user's explicit path) into an absolute path; create missing parent directories; on collision append -2, -3, … before the extension until the name is free.
6. Write the file: provenance header plus content for markdown dumps, raw content otherwise; redact secrets.
7. Reply with the one-line confirmation.

Output Format

Chat reply — one line: Dump written: <absolute path> (<bytes> bytes).

File when the dump is markdown — provenance header, then the content:

# {Human-readable title of the dump}

- **Description:** {description}
- **Model:** {model}
- **Captured at (UTC):** {iso_8601_utc}
- **Session:** {session_id_or_unknown}

---

{content}

File when the dump is not markdown — the content exactly as produced, in its native format; no header, wrapper, or commentary.