tkersey/learnings

Learnings

Overview

Capture durable lessons as soon as evidence appears, not only at explicit retrospective requests. Also use the same CLI for interactive browsing and filtered search so prior learnings can be loaded deliberately instead of defaulting every read to recall. Write each validated learning as JSONL so future agents can reuse successful patterns and avoid footguns quickly.

Zig CLI Iteration Repos

When iterating on the Zig-backed learnings/append_learning helper CLI path, use these two repos:

• skills-zig ($HOME/workspace/tk/skills-zig): source for the learnings and append_learning Zig binaries, build/test wiring, and release tags.
• homebrew-tap ($HOME/workspace/tk/homebrew-tap): Homebrew formula updates/checksum bumps for released learnings binaries.

Trigger cues

• $learnings
• "browse learnings" / "show recent learnings" / "search learnings"
• "what do we already know about X" / "summarize learnings about X"
• "lessons learned" / "takeaways" / "wrap up" / "handoff"
• Before commit/PR/ship; after a proof signal changes state (fail->pass)
• "strategy pivot" / "footgun" / "gotcha" / "retry loop"

Load Modes

• recall: implementation preflight for a concrete task. Use the active request text as the query and treat hits as constraints/invariants.
• recent: interactive browsing. Use when the user wants to look around, skim the latest learnings, or orient before deciding what matters.
• query: filtered or aggregated browsing. Use when the user wants search, ranking, grouping, or summaries by status/tag/path/time.
• Routing rule: if the user asks to browse, search, summarize, or rank learnings, do not default to recall; start with recent or query, then switch to recall only when transitioning into implementation.
• Context-gathering rule: after the first user request is available and before substantial edits, run recall during context gathering rather than on empty SessionStart hooks; if early exploration materially narrows the slice, rerun recall with the refined query.

Operating Contract (Auto-Capture)

• Allowed to run automatically at the end of an implementation turn (success or paused) and at delivery boundaries (commit/PR/handoff).
• Best-effort and non-blocking: never require extra context; if evidence is thin, either write nothing or persist review_later with placeholders.
• Noise control: if no Capture Checkpoint occurred or nothing decision-shaping emerged, append 0 records; otherwise prefer 1 record (max 3).
• Completion proof: after an implementation turn with a checkpoint or an intentional no-write, produce one exact proof line: appended: id=..., duplicate-skip: <reason>, or 0 records appended: <reason>.
• Stop-hook compatibility: when append dedupes or you intentionally skip capture, keep the proof line on one line with the exact duplicate-skip: or 0 records appended: prefix so hooks can distinguish a valid no-write from a missed capture.
• Helper-first write path: use learnings append/run_learnings_tool append for normal writes; keep append_learning available as a compatibility binary, and treat manual .learnings.jsonl edits as emergency-only with explicit rationale.
• Repo-root fail-closed rule: only append from a verified git repo root. If git rev-parse --show-toplevel fails or the active task lives in a non-repo workspace such as ~/.codex/memories, skip capture and emit 0 records appended: non-repo cwd.
• Path-drift rule: append from the verified repo root with --path .learnings.jsonl; if the CLI reports a different target path, treat that append as failed, repair the stray write if needed, and reappend correctly before closing.
• Commit coupling: before any Codex-made commit after learning capture, check .learnings.jsonl even if the append happened in an earlier turn. If it has pending changes and git check-ignore -v --no-index .learnings.jsonl does not report .git/info/exclude, include those learnings in the next commit by default. If the file is local-only by .git/info/exclude, or the user explicitly asks for a commit that excludes learnings, leave it unstaged and state that boundary in the proof. If the shared file also holds unrelated fresh rows, stage only the session-owned rows with an index patch rather than leaving all learnings for a follow-up commit.

Quick Start

CODEX_SKILLS_HOME="${CODEX_HOME:-$HOME/.codex}"
CLAUDE_SKILLS_HOME="${CLAUDE_HOME:-$HOME/.claude}"
LEARNINGS_SPECS_DIR="$CODEX_SKILLS_HOME/skills/learnings/specs"
[ -d "$LEARNINGS_SPECS_DIR" ] || LEARNINGS_SPECS_DIR="$CLAUDE_SKILLS_HOME/skills/learnings/specs"

run_learnings_tool() {
  local subcommand="${1:-}"
  if [ -z "$subcommand" ]; then
    echo "usage: run_learnings_tool <datasets|query|recent|recall|codify-candidates|quality-audit|value-report|memory-digest|append> [args...]" >&2
    return 2
  fi
  shift || true

  local mode=""
  local bin=""
  local marker=""
  case "$subcommand" in
    append|append-learning|append_learning)
      mode="append"
      bin="learnings"
      marker="append_learning.zig"
      ;;
    datasets|query|recent|recall|codify-candidates|quality-audit|value-report|memory-digest)
      mode="learnings"
      bin="learnings"
      marker="learnings.zig"
      ;;
    *)
      echo "unknown learnings subcommand: $subcommand" >&2
      return 2
      ;;
  esac

  learnings_repo_root() {
    git rev-parse --show-toplevel 2>/dev/null
  }

  run_learnings_append() {
    local repo_root=""
    local expected_path=""
    local output=""
    local status=0
    local actual_path=""

    repo_root="$(learnings_repo_root)" || {
      echo "0 records appended: non-repo cwd" >&2
      return 3
    }
    expected_path="$repo_root/.learnings.jsonl"
    output="$(
      cd "$repo_root" &&
      "$bin" append --path .learnings.jsonl "$@"
    )"
    status=$?
    printf '%s\n' "$output"
    [ $status -ne 0 ] && return $status

    actual_path="$(printf '%s\n' "$output" | sed -n 's/.* path=//p' | tail -n 1)"
    if [ -n "$actual_path" ] && [ "$actual_path" != "$expected_path" ]; then
      echo "append-path-mismatch: expected $expected_path got $actual_path" >&2
      return 4
    fi
  }

  install_learnings_direct() {
    local repo="${SKILLS_ZIG_REPO:-$HOME/workspace/tk/skills-zig}"
    if ! command -v zig >/dev/null 2>&1; then
      echo "zig not found. Install Zig from https://ziglang.org/download/ and retry." >&2
      return 1
    fi
    if [ ! -d "$repo" ]; then
      echo "skills-zig repo not found at $repo." >&2
      echo "clone it with: git clone https://github.com/tkersey/skills-zig \"$repo\"" >&2
      return 1
    fi
    if ! (cd "$repo" && zig build -Doptimize=ReleaseSafe); then
      echo "direct Zig build failed in $repo." >&2
      return 1
    fi
    mkdir -p "$HOME/.local/bin"
    for direct_bin in learnings append_learning; do
      if [ ! -x "$repo/zig-out/bin/$direct_bin" ]; then
        echo "direct Zig build did not produce $repo/zig-out/bin/$direct_bin." >&2
        return 1
      fi
      install -m 0755 "$repo/zig-out/bin/$direct_bin" "$HOME/.local/bin/$direct_bin"
    done
  }

  local os="$(uname -s)"
  if command -v "$bin" >/dev/null 2>&1 && (
    [ "$mode" != "append" ] && "$bin" --help 2>&1 | grep -q "$marker" ||
    [ "$mode" = "append" ] && "$bin" append --help 2>&1 | grep -q "$marker"
  ); then
    if [ "$mode" = "append" ]; then
      run_learnings_append "$@"
    else
      "$bin" "$subcommand" "$@"
    fi
    return
  fi

  if [ "$os" = "Darwin" ]; then
    if ! command -v brew >/dev/null 2>&1; then
      echo "homebrew is required on macOS: https://brew.sh/" >&2
      return 1
    fi
    if ! brew install tkersey/tap/learnings; then
      echo "brew install tkersey/tap/learnings failed." >&2
      return 1
    fi
  elif ! (
    command -v "$bin" >/dev/null 2>&1 && (
      [ "$mode" != "append" ] && "$bin" --help 2>&1 | grep -q "$marker" ||
      [ "$mode" = "append" ] && "$bin" append --help 2>&1 | grep -q "$marker"
    )
  ); then
    if ! install_learnings_direct; then
      return 1
    fi
  fi

  if command -v "$bin" >/dev/null 2>&1 && (
    [ "$mode" != "append" ] && "$bin" --help 2>&1 | grep -q "$marker" ||
    [ "$mode" = "append" ] && "$bin" append --help 2>&1 | grep -q "$marker"
  ); then
    if [ "$mode" = "append" ]; then
      run_learnings_append "$@"
    else
      "$bin" "$subcommand" "$@"
    fi
    return
  fi
  echo "learnings binary missing or incompatible after install attempt: $bin" >&2
  if [ "$os" = "Darwin" ]; then
    echo "expected install path: brew install tkersey/tap/learnings" >&2
  else
    echo "expected direct path: SKILLS_ZIG_REPO=<skills-zig-path> zig build -Doptimize=ReleaseSafe" >&2
  fi
  return 1
}

Capture Checkpoints

Capture at least once per coding turn when any of these checkpoints occurs:

1. Validation transition: a signal changes state (fail->pass, pass->fail, timeout->stable).
2. Strategy pivot: an approach is abandoned, replaced, or simplified.
3. Footgun discovery: hidden risk, brittle assumption, or recurring trap is observed.
4. Momentum discovery: a pattern repeatedly accelerates implementation or debugging.
5. Delivery boundary: immediately before commit/PR/final handoff.

Auto-trigger rule: if none of these checkpoints occurred and nothing decision-shaping emerged, do not append anything.

Essentialness Test

Before writing a learning, require all three:

1. Decision delta: would this change what you do next time?
2. Transferability: does it generalize beyond this one exact file/run path?
3. Counterfactual: if ignored, is there a predictable failure, cost, or missed leverage?

If any answer is no, skip capture (or use review_later only when uncertainty itself is decision-shaping).

Workflow

1. Gather evidence.
   • Inspect changed surface (git status -sb, git diff --stat, targeted git diff).
   • Collect executed validation signals and outcomes.
   • Collect failed attempts only when evidenced by commands, logs, diffs, or test output.
2. Distill session essence first.
   • Compress the turn to four lines: objective, inflection, proof, transferable rule.
   • Keep only lessons that alter future decisions.
3. Distill candidate learnings.
   • Convert narrative into rule form (When X, prefer Y because Z).
   • Prefer one essential learning over several procedural notes.
   • Avoid path-bound/session-bound wording unless the path itself is the lesson.
4. Assign semantic status.
   • Use a concise action status in snake_case.
   • Prefer do_more or do_less when they fit.
   • Choose a more precise status when needed (for example investigate_more, codify_now, avoid_for_now).
5. Run pre-append quality gate.
   • learning contains explicit condition + action (When/If ... prefer/do ...).
   • evidence includes at least one concrete anchor (command outcome, commit SHA, run ID, file path, or exact error string).
   • application states what to do on the next similar task.
   • status reflects intent (avoid defaulting to do_more when codify_now/investigate_more/avoid_for_now is more accurate).
   • If working under temp paths (/tmp, /var/folders/...), either skip or mirror only high-signal lessons into a durable repo .learnings.jsonl.
6. Persist immediately.
   • Verify the target repo root first with git rev-parse --show-toplevel; if that fails, skip capture with 0 records appended: non-repo cwd unless the user explicitly names another durable target.
   • Append each accepted learning to .learnings.jsonl in that repo root (0 records is valid when nothing qualifies).
   • Use learnings append via run_learnings_tool append, which changes into the verified repo root and passes --path .learnings.jsonl.
   • If the append output reports path=..., require an exact match with $repo_root/.learnings.jsonl; treat any mismatch as a failed append and repair/reappend before finishing.
   • Before any Codex-made commit after learning capture, check .learnings.jsonl explicitly even if the append happened in an earlier turn.
   • If .learnings.jsonl has pending changes, stage the publishable learning rows into that next commit by default.
   • If git check-ignore -v --no-index .learnings.jsonl reports .git/info/exclude, treat the file as local-only, leave it unstaged, and never use git add -f to satisfy commit coupling, even if .learnings.jsonl is already tracked.
   • If the user explicitly asks for a commit that excludes learnings, leave .learnings.jsonl unstaged and state that boundary in the proof.
   • If unrelated rows are present in a shared .learnings.jsonl, use a targeted index patch instead of postponing the learnings write.
7. Report concise highlights in chat.
   • Summarize the 1-3 highest leverage learnings (or say none).
   • Reference the write result with one exact proof line (appended: id=..., duplicate-skip: <reason>, or 0 records appended: <reason>).

JSONL Contract

Write one JSON object per line:

{
  "id": "lrn-20260207T173422Z-a91f4e2c",
  "captured_at": "2026-02-07T17:34:22Z",
  "status": "do_more",
  "learning": "Boundary parsing eliminated downstream guard duplication.",
  "evidence": [
    "uv run pytest tests/parser_test.py::test_rejects_invalid passed after boundary parse refactor"
  ],
  "application": "Parse and refine request payloads once at API boundaries.",
  "tags": [
    "api",
    "testing"
  ],
  "context": {
    "repo": "owner/repo",
    "branch": "main",
    "paths": [
      "src/parser.py",
      "tests/parser_test.py"
    ]
  },
  "related_ids": [
    "lrn-20260130T120000Z-deadbeef"
  ],
  "supersedes_id": "lrn-20260101T090000Z-cafebabe",
  "source": "skill:learnings",
  "fingerprint": "a91f4e2c6b5d3f10"
}

Required keys:

• id
• captured_at
• status
• learning
• evidence
• application
• source
• fingerprint

Optional keys:

• context
• tags
• related_ids
• supersedes_id

Write Procedure

Use one append call per learning:

run_learnings_tool append \
  --status do_more \
  --learning "Boundary parsing eliminated downstream guard duplication." \
  --evidence "uv run pytest tests/parser_test.py::test_rejects_invalid passed after boundary parse refactor" \
  --application "Parse and refine request payloads once at API boundaries." \
  --tag api \
  --tag testing

Call the wrapper from a real repo checkout only; for non-repo tasks such as memory-folder maintenance, default to 0 records appended: non-repo cwd unless the user explicitly names a target repo/file.

The helper script:

• Normalizes status to snake_case.
• Defaults status to review_later when omitted.
• Enforces strict quality by default (condition+action learning statement, anchored evidence, concrete application, and non-temporary repo path unless explicitly allowed).
• Refuses append in non-repo cwd instead of guessing a target .learnings.jsonl.
• Runs append from the verified repo root and passes --path .learnings.jsonl.
• Treats reported target-path drift as failure instead of silently accepting the wrong file.
• Supports explicit escape hatches:
  • --quality-mode best_effort for intentional thin capture.
  • --allow-temp-path when temporary-repo capture is unavoidable.
• In best_effort mode only, backfills missing evidence/application with placeholders.
• Captures a best-effort repo slug from remote.origin.url (or falls back to repo dir name).
• Captures branch and changed paths from git when available.
• Computes a fingerprint for duplicate detection.
• Appends to .learnings.jsonl in repo root by default.

Browsing, Recall, and Codify Loop

Use the miner script both for interactive browsing and for task-start loading once the user prompt is available.

Routing:

• Browse latest learnings: run_learnings_tool recent --limit 10
• Search or rank learnings: run_learnings_tool query --spec "@$LEARNINGS_SPECS_DIR/top-tags.json"
• Search by path concentration: run_learnings_tool query --spec "@$LEARNINGS_SPECS_DIR/top-paths.json"
• Refresh disposable memory consolidation digest: run_learnings_tool memory-digest (defaults to $CODEX_HOME/memories/extensions/learnings/resources/latest_learnings_digest.md, falling back to $HOME/.codex/... when CODEX_HOME is unset)
• Implementation preflight during context gathering: distill the request to a compact topical query first (roughly 4-8 task-defining terms; skip boilerplate, pasted skill blocks, and AGENTS text), then run run_learnings_tool recall --query "<focused task terms>" --limit 5 --drop-superseded
• Refined-scope preflight: rerun run_learnings_tool recall once after parse, seq, or early file reads only if those steps materially narrow the implementation slice; tighten the query rather than replaying the full prompt.
• If browse intent is ambiguous, start with recent before escalating to query or recall.

CLI:

run_learnings_tool datasets
run_learnings_tool query --spec "@$LEARNINGS_SPECS_DIR/status-rank.json"
run_learnings_tool query --spec "@$LEARNINGS_SPECS_DIR/top-tags.json"
run_learnings_tool query --spec "@$LEARNINGS_SPECS_DIR/top-paths.json"
run_learnings_tool recent --limit 15
run_learnings_tool recall --query "fix flaky pre-commit hook" --limit 5
run_learnings_tool codify-candidates --min-count 3 --limit 20
run_learnings_tool quality-audit --since 2026-02-01 --until 2026-03-05 --format json
run_learnings_tool value-report --sessions-root "$HOME/.codex/sessions" --since 2026-03-01 --comparator impl_nonrecall --format json
run_learnings_tool memory-digest --limit-candidates 12

Promotion rule of thumb:

• If a learning is repeated (theme appears >= 3 times) or status is codify_now, promote it into durable docs (for example codex/AGENTS.md or a relevant skill doc).
• After codifying, append a follow-up learning referencing the durable anchor (use --related-id/--supersedes-id and a codified tag).

Runtime bootstrap policy for learnings mirrors cas: use Zig binaries only (learnings with append plus append_learning compatibility), default to Homebrew install on macOS, and fallback to direct Zig install from skills-zig on non-macOS.

Guardrails

• Ground every learning in observed evidence; do not infer hidden causes as facts.
• Do not write pure changelog bullets; write decision-shaping rules.
• Keep status action-oriented and semantically meaningful for the situation.
• Capture essence, not chronology: avoid storing every step if one inflection explains the outcome.
• Do not let append target resolution depend on ambient non-repo cwd; verify repo root first or skip capture.
• If a JSON learning row lands in source/docs or any file other than the intended .learnings.jsonl, revert it immediately and reappend through the verified repo-root CLI path.
• If .learnings.jsonl matches .git/info/exclude, treat it as local-only state even if the file is already tracked: keep capture local, and do not force-add, stage, or commit it unless the user explicitly requests publication.
• Avoid duplicate entries; the helper script skips exact duplicates by fingerprint.
  • If you have materially new evidence for an existing learning, append a follow-up record (adjust status and/or make the learning statement more specific) or re-run with --allow-duplicate intentionally.
  • Do not hand-edit existing JSONL lines in place.
• If evidence is weak, persist with review_later and placeholders, then enrich later.