[diegosouzapw]

What

Compression backlog F2.1 + Item 2, delivered together. The pure RTK miners (discoverRepeatedNoise / suggestFilter) already shipped; what was missing — and what got Item 2 deferred as "no usable sample source" — was the sample source. This wires it end to end.

Meta-finding

Both the miners AND a capture store already existed: maybePersistRtkRawOutput persists redacted, opt-in command output to DATA_DIR/rtk/raw-output/*.log. The only real gap was an adapter + the routes + UI. So F2.1 was not a from-scratch sub-project.

Backend (commit 1)

• Capture enhancement (backward-compatible): maybePersistRtkRawOutput now writes a <base>.meta.json sidecar carrying the full command — the .log filename only had a lossy slug. The .log stays pure output (the raw-output recovery route is unchanged); the sidecar write is best-effort.
• Adapter listRtkCommandSamples({limit}) reads the store into CommandSample[] — exact command from the sidecar, falling back to the filename slug for legacy captures; newest-first, empty/corrupt entries skipped, capped to bound memory.
• Routes (management-auth, read-only, suggestions-only):
  • GET /api/context/rtk/discover → ranked noise candidates over the samples.
  • GET /api/context/rtk/learn?command=X → a suggested filter draft for X, learned only from that command's samples (commandToId match); 400 without a command.
• Barrel re-exports the miners + adapter + commandToId.

UI (commit 2)

A Learn & Discover card on Context & Cache → RTK: scan for repeated noise; type a command → preview a suggested filter draft. Suggestions only — the operator reviews and saves via the existing filter trust path. Fail-soft.

Tests (TDD)

• 8 sample-source (sidecar write, exact/slug command recovery, limit, empty/skip).
• 5 route (discover happy/empty, learn happy/400/command-filtering).
• 4 vitest card (controls, discover fetch+render, learn fetch+render, fail-soft).
• 91 rtk regression tests still green; typecheck:core clean; lint / file-size / cycles / test-discovery / any-budget clean.

i18n note

14 EN keys added under contextRtk. Locale translation is deferred to the release i18n:run — the repo's established flow (locales already sit at ~79% pre-existing coverage debt, and i18n-ui-coverage is a PR→main gate, not a release fast-gate). i18n:sync-ui was deliberately not run here because it would dump the entire pre-existing 42k-key __MISSING__ backlog into this PR.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[gemini-code-assist]

Code Review

This pull request implements the 'Learn & Discover' feature for RTK filters, enabling operators to mine captured command outputs for recurring noise and filter suggestions. It introduces sidecar metadata files to preserve exact commands, adds API routes for discovery and learning, and integrates a new UI card component. The review feedback correctly identifies a correctness and performance issue in listRtkCommandSamples where applying the limit before filtering by command can cause older matching samples to be missed, and reading .log files for non-matching commands is inefficient. The reviewer suggests introducing a commandFilter option to filter commands before reading log files and applying the limit.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Applying the limit before filtering by command can lead to a correctness bug where matching samples are missed if they are older than the most recent limit samples. Additionally, reading the .log file content for every sample before checking if it matches the target command is a performance bottleneck.

By introducing a commandFilter option, we can check the command (retrieved from the sidecar or slug) first, and skip reading the .log file entirely for non-matching commands. This resolves the correctness issue and significantly optimizes disk I/O.

export function listRtkCommandSamples(opts: { limit?: number; commandFilter?: (command: string) => boolean } = {}): CommandSample[] {
  const dir = path.join(dataDir(), "rtk", "raw-output");
  if (!fs.existsSync(dir)) return [];
  const limit = Math.max(1, Math.floor(opts.limit ?? 500));

  let logs: string[];
  try {
    logs = fs.readdirSync(dir).filter((f) => f.endsWith(".log"));
  } catch {
    return [];
  }
  // Newest first: the filename is timestamp-prefixed, so a reverse lexical sort works.
  logs.sort((a, b) => (a < b ? 1 : a > b ? -1 : 0));

  const samples: CommandSample[] = [];
  for (const fileName of logs) {
    if (samples.length >= limit) break;
    const fullPath = path.join(dir, fileName);
    if (!fullPath.startsWith(dir)) continue;

    let command = "";
    try {
      const metaRaw = fs.readFileSync(fullPath.replace(/\.log$/, ".meta.json"), "utf8");
      const meta = JSON.parse(metaRaw) as { command?: unknown };
      if (typeof meta.command === "string" && meta.command.trim()) command = meta.command.trim();
    } catch {
      // No/!invalid sidecar → fall back to the filename slug below.
    }
    if (!command) command = commandFromSlug(fileName) || "tool-output";

    if (opts.commandFilter && !opts.commandFilter(command)) {
      continue;
    }

    let output: string;
    try {
      output = fs.readFileSync(fullPath, "utf8");
    } catch {
      continue;
    }
    if (output.trim().length === 0) continue;

    samples.push({ command, output });
  }
  return samples;
}

[gemini-code-assist]

Filter the command samples during the retrieval phase using the new commandFilter option. This avoids the correctness bug where matching samples are missed if they are older than the most recent limit samples, and significantly improves performance by skipping unnecessary disk reads.

Suggested change

	const targetId = commandToId(command);
	const matching = listRtkCommandSamples({ limit }).filter(
	(sample) => commandToId(sample.command) === targetId
	);
	const targetId = commandToId(command);
	const matching = listRtkCommandSamples({
	limit,
	commandFilter: (cmd) => commandToId(cmd) === targetId,
	});