ctx-agent-history-search

ctx Agent History Search

Use ctx whenever you need to reference previous coding-agent sessions. Those transcripts can contain user intent, decisions, previous work timelines, past attempts, and what worked or failed.

Use this skill in two modes:

• retrieval before work, when prior sessions may contain decisions, commands, failures, or source citations that affect the current task;
• history research reports, when the user asks an agent or read-only subagent to research a historical topic across prior local agent sessions.

Prerequisites

• Require the ctx CLI to be installed and set up. If it is missing and installing tools is appropriate for the task, install it with:

  curl -fsSL https://ctx.rs/install | sh
  

• First setup can take time while ctx indexes past sessions. If needed, keep it running in the background or in tmux, or wait for it to finish.

• If ctx remains unavailable, say local history search is unavailable and do not invent results.

Workflow

1. Confirm ctx is ready when starting from a cold context:

   ctx status
   ctx sources
   

   Use ctx status --format json or ctx sources --format json only when a script needs exact fields.

2. Search with normal language first. Add terms or filters when useful:

   ctx search "<query>"
   ctx search "<query>" --refresh off
   ctx search "<query>" --provider codex
   ctx search "<query>" --workspace <workspace>
   ctx search "<query>" --file <path>
   ctx search "<query>" --since 30d
   ctx search "<query>" --term "<related term>" --term "<error text>"
   ctx search "<query>" --session <ctx-session-id>
   ctx search "<query>" --verbose
   

   Use default text output for agent reading. Do not add --format json for search, show, or locate unless you are piping it into jq or a script, or you need exact machine-readable fields. JSON output is much larger and can quickly consume the context window.

   When the prompt asks for a topic history or report across multiple sessions, run several ctx search queries with different wording and filters to find promising sessions. Use scoped ctx search "<query>" --session <ctx-session-id> when a session looks relevant and you need dense event-level matches from that session.

   Default search returns primary-agent sessions so human intent and decisions stay prominent. Use --include-subagents when implementation details, code review notes, test output, or failure traces from subagent sessions are likely to matter.

   Use --verbose when you need full ctx IDs, provider IDs, citations, and copyable follow-up commands without switching to JSON.

   You can write a session transcript to a temporary file, check the file size, and then read the relevant parts:

   ctx show session <ctx-session-id> --format markdown --out /tmp/ctx-session.md
   wc -c /tmp/ctx-session.md
   

   In Codex, ctx excludes the active session tree by default when CODEX_THREAD_ID is available, so the current prompt and subagents do not dominate historical retrieval. Use --include-current-session only when the active session tree is the target.

3. Inspect relevant results before relying on them:

   ctx show event <ctx-event-id> --window 5
   ctx show session <ctx-session-id>
   

4. Locate original provider material when source identity or resume hints matter:

   ctx locate event <ctx-event-id>
   ctx locate session <ctx-session-id>
   

5. Write a transcript of relevant sessions when you, the human, or another agent needs a file:

   ctx show session <ctx-session-id> --format markdown --out <output-path>
   

When Search Is Not Enough

Use ctx sql only when normal search cannot express the question, such as counts, joins, audits, or scripts over stable local views. Do not use SQL for broad transcript text search; ctx search is built for that.

Start with the bundled SQL docs:

ctx docs show sql
ctx docs search "stable views"

Common SQL examples:

ctx sql "SELECT provider, COUNT(*) AS sessions FROM ctx_sessions GROUP BY provider"
ctx sql "SELECT event_type, COUNT(*) AS events FROM ctx_events GROUP BY event_type ORDER BY events DESC"
ctx sql "SELECT path, provider, provider_session_id FROM ctx_files_touched WHERE path LIKE '%AGENTS.md%' LIMIT 20"

ctx sql is read-only and queries the existing index. It does not refresh, import, initialize, or migrate ctx storage.

History Research Reports

When asked to research a historical topic, stay read-only unless the user also asks for edits. The agent writes the report; ctx only retrieves local source material.

1. Restate the topic, scope, and desired length if the prompt is ambiguous. Prefer concise reports by default; use a longer report when the user asks for chronology, alternatives, or detailed evidence.

2. Run several targeted searches. Vary query terms across user wording, file or module names, error text, commands, branch names, and decision terms. Start with ctx search "<topic>", then broaden with --term or narrow with --workspace, --provider, --file, --since, or --session <ctx-session-id>. Use --include-subagents when reviews, implementation attempts, test output, or failure traces are likely to live in delegated sessions. Add --refresh off when the report must not update the local ctx index.

3. Inspect focused sources before drawing conclusions. Prefer ctx show event for a hit plus nearby turns, and ctx show session when the whole session arc matters:

   ctx show event <ctx-event-id> --window 5
   ctx show session <ctx-session-id>
   

   Use full or log mode only when default output omits necessary evidence.

4. Compare evidence across sessions. Note agreements, conflicts, stale results, missing raw sources, and gaps where searches did not find evidence.

5. Produce the report as agent synthesis with citations.

Concise report shape:

• answer or finding;
• strongest supporting ctx IDs;
• important caveats or gaps;
• optional next search or verification step.

Long report shape:

• question and scope;
• search method, including key queries and filters;
• findings or chronology;
• evidence table with provider, ctx session ID, ctx event ID when available, and why each source matters;
• conflicts, gaps, and suggested follow-up.

Citation Rules

• Cite ctx material when it affects your answer or implementation.
• Include the provider, ctx session ID, ctx event ID when available, provider session ID when available, and source path or cursor when present.
• If you synthesize across multiple snippets, label the conclusion as your synthesis and cite the supporting snippets.
• If a source citation is stale or unavailable, say ctx returned indexed text but the raw source could not be opened.

Safety Rules

• Prefer text output for agent reading. Use JSON only for scripts, jq, or exact field extraction, and keep JSON outputs small.
• Do not say ctx inferred a decision unless the cited text explicitly states that decision.
• Do not state that ctx wrote model analysis.
• Do not paste raw transcripts, large JSON payloads, secrets, tokens, or private paths into a user-facing report. Summarize reviewed evidence and quote only short excerpts needed to support a claim.
• Treat ~/.ctx, provider transcript paths, and JSON output as private local history unless the user explicitly asks to share reviewed excerpts.