dredge

name	dredge
description	Explicit-request macro. Activate only when the user directly requests this macro; never infer activation from task characteristics. Skill — Search prior coding-agent chat transcripts (Claude Code, Codex, ...) for context. Use when the user explicitly activates this skill or asks to recall or dig up something from past chats.

Honor every skill explicitly activated in the user's request exactly once. If another activated skill is not yet loaded and the host provides a skill-loading mechanism, load it through that mechanism. Do not reload an active skill.

Do NOT re-invoke this skill recursively. Execute the workflow below once, then stop.

Task

Find and surface context from prior coding-agent chat transcripts relevant to the user's query.

Where transcripts live

Each coding agent keeps its own transcript store.

• Claude Code — ~/.claude/projects/<project-slug>/*.jsonl, where <project-slug> is the absolute project path with / replaced by - (e.g. /Users/x/repo → -Users-x-repo). The current project's slug can be derived from pwd.
• Codex and other agents — each has its own layout, indexed by AgentsView. The grep fallback only knows Claude Code's layout.
• Each .jsonl file is one session; each line is a JSON message event with fields like role, content, timestamp.
• On the grep backend, use file mtime for time-window filtering. (The AgentsView backend filters on per-message timestamps instead — see "Search backend".)

Argument handling

• With a query — use the text supplied with macros:dredge as the search focus.
• No query — infer the topic from the current conversation.

Scope

Default to the current project's transcripts (slug derived from pwd). Widen based on natural-language hints in the query:

• "across all projects" / "anywhere" / "in any chat" → search every project under ~/.claude/projects/.
• Names a different repo ("in the craken project") → resolve to that project's slug and search there. If the slug is ambiguous, list candidates under ~/.claude/projects/ and pick by name match.
• Time hints ("yesterday", "last week", "earlier today", "recent") → filter by file mtime.

Search backend

dredge has two backends. Resolve config by merging these files per key (first existing wins; precedence: project override > user-local > user):

1. .agents/skill-configs/dredge/config.yaml — per-repo override (optional)
2. ~/.agents/skill-configs/dredge/config.local.yaml — user-local
3. ~/.agents/skill-configs/dredge/config.yaml — user (global)

Legacy fallback (older installs): the same files under ~/.claude/skill-configs/dredge/ and .claude/skill-configs/dredge/. If config is found only at a legacy path, use it and offer to move it to the new location.

With no config, behave as backend: auto. See config.example.yaml for all keys.

• AgentsView (when backend is auto and agentsview is on PATH, or backend is agentsview):
  1. If sync_before_search (default true), run agentsview sync first (incremental; ~sub-second when warm). This is required — search hits a SQLite index that is stale until synced. The sync prints a noisy progress bar; ignore it (only the final Sync complete line matters).
  2. agentsview session search "<query>" --format json --in <search_in> --limit 30, adding the scope/time/agent flags below.
  3. For each promising match, fetch context with agentsview session messages <session_id> --from <ordinal> --limit <n> --format json (do NOT read the whole .jsonl). The ordinal from session search indexes directly into session messages --from. Use agentsview session export <session_id> only when the full raw transcript is needed.
• grep (when backend is grep, or auto/agentsview and the binary is missing): the file-scan approach in "Approach" below.

Never hard-fail: if AgentsView is unavailable or errors, warn briefly and fall back to grep.

AgentsView flag mapping

On the AgentsView backend, translate the natural-language scope/time hints (see "Scope") to flags:

• Current project → resolve basename(pwd), confirm it appears exactly once in agentsview projects --json, and pass --project <name> only if it resolves unambiguously. If two projects share that basename (collision), omit --project and fall back to slug-grep for that scope rather than returning wrong-project hits.
• "across all projects" / "anywhere" → omit --project.
• Named repo → match its basename against projects --json (same collision rule).
• Time hints ("yesterday", "last week", "earlier today") → --date-from / --date-to / --date <YYYY-MM-DD> / --active-since <RFC3339> (per-message timestamps, not file mtime).
• Filter to specific agents with --agent <name> per the agents config (e.g. --agent claude --agent codex); omit --agent entirely when agents: all (the default) to search every indexed coding agent. --include-children only if include_children is true.
• Concept recall → --fts (fast tokenized search over messages); exact string / error message → --regex or the default substring match.

Approach

The grep fallback (also the AgentsView path's last resort):

1. Shortlist candidate .jsonl files with rg / grep using keywords from the query. Combine with find -newer / mtime for time windows.
2. Read the top candidates to pull the actual context (not just the matching lines).

Output

• Answer the query directly.
• Cite each session you drew from by agent + <session-id>, with that agent's resume command when you know it (e.g. claude --resume <session-id> for Claude Code, codex resume <session-id> for Codex). The AgentsView session_id is the agent's native session id; the grep path cites <project-slug>/<session-id>.jsonl.
• If you don't know an agent's resume command, cite the agent name + session id alone.
• If nothing relevant turns up, say so plainly — don't fabricate.