Collect Transcripts
Ingest raw session transcripts from supported coding-agent harnesses via vendor-specific adapters, normalize them to a common event schema, triage for struggle signals, and run deep analysis on flagged sessions. Findings are written to episodic memory using the D4 tag schema with source:transcript-mined.
Arguments
$ARGUMENTS - Optional flags:
--adapter <name> (claude_code_cli, claude_code_web, codex_cli, codex_web, antigravity_cli, grok_cli, pi_cli; default: all available)--threshold <float> (composite score threshold for deep analysis; default: 5.0)--dry-run (print planned operations without API calls; default in CI)--enable (opt-in to actually run LLM analysis; required to make API calls)
Adapters
| Adapter |
Source |
Schema Version |
claude_code_cli |
~/.claude/projects/<encoded-cwd>/<session-id>.jsonl |
1.0 |
claude_code_web |
CLI bridge via claude --teleport <session-id> |
1.0 |
codex_cli |
$CODEX_HOME/sessions/YYYY/MM/DD/rollout-*.jsonl |
rollout-v1 |
codex_web |
CLI bridge via codex cloud |
rollout-v1 |
antigravity_cli |
~/.antigravity/projects/<encoded-cwd>/<session-id>.jsonl |
1.0 |
grok_cli |
~/.grok/sessions/session-*.jsonl |
grok-session-v1 |
pi_cli |
~/.pi/sessions/session-*.ndjson |
pi-ndjson-v1 |
All adapters fail soft (log warning, skip) when their source is unavailable.
Pipeline
1. Discover sessions (per adapter)
2. Normalize to NormalizedEvent schema
3. Sanitize (secrets, entropy, paths) — BEFORE any LLM sees content
4. Triage (score for struggle signals: retries, errors, scope violations, corrections)
5. Deep analyze (flagged sessions only — heuristic or LLM)
6. Write findings to episodic memory (D4 tag schema, source:transcript-mined)
How It Works
- Discovery: Each adapter enumerates available sessions from its source
- Normalization: Raw vendor events -> NormalizedEvent (common schema in
references/event-schema.md)
- Sanitization: Reuses
session-log sanitizer extended for tool-call arguments and tool-result outputs
- Triage: Scores each session on retry_count, tool_error_count, scope_violation_count, user_correction_count
- Deep Analysis: Runs on flagged sessions (composite score >= threshold), extracts structured findings
- Memory: Findings written with D4 tags (
failure_type:*, capability_gap:*, etc.) + source:transcript-mined
Model Resolution
- Triage: archetype
analyst (standard tier), configurable via config.yaml: triage.archetype
- Deep analysis: archetype
reviewer (premium tier), configurable via config.yaml: deep_analysis.archetype
- Both resolve via
agents_config.resolve_model() from the coordinator
Prerequisites
- Python 3.11+
- At least one harness's sessions directory accessible on disk
- For web adapters: vendor CLI installed and authenticated
Steps
1. Discover and Normalize
python3 <agent-skills-dir>/collect-transcripts/scripts/adapters/claude_code_cli.py
2. Triage
python3 <agent-skills-dir>/collect-transcripts/scripts/triage.py \
--events-dir docs/transcripts/$(date +%Y-%m-%d)/ \
--threshold 5.0 \
--dry-run
3. Deep Analysis (flagged sessions only)
python3 <agent-skills-dir>/collect-transcripts/scripts/deep_analyze.py \
--events-file docs/transcripts/2026-06-01/session-abc.jsonl \
--dry-run
1---2name: collect-transcripts3description: Ingest raw session transcripts from coding-agent harnesses via vendor-specific adapters, normalize to a common event schema, triage for struggle signals, and write structured findings to episodic memory4category: Git Workflow5tags: [transcripts, session-mining, capability-gaps, harness, multi-vendor]6triggers:7 - "collect transcripts"8 - "mine transcripts"9 - "transcript mining"10 - "session transcripts"11---1213# Collect Transcripts1415Ingest raw session transcripts from supported coding-agent harnesses via vendor-specific adapters, normalize them to a common event schema, triage for struggle signals, and run deep analysis on flagged sessions. Findings are written to episodic memory using the D4 tag schema with `source:transcript-mined`.1617## Arguments1819`$ARGUMENTS` - Optional flags:20- `--adapter <name>` (claude_code_cli, claude_code_web, codex_cli, codex_web, antigravity_cli, grok_cli, pi_cli; default: all available)21- `--threshold <float>` (composite score threshold for deep analysis; default: 5.0)22- `--dry-run` (print planned operations without API calls; default in CI)23- `--enable` (opt-in to actually run LLM analysis; required to make API calls)2425## Adapters2627| Adapter | Source | Schema Version |28|---------|--------|---------------|29| `claude_code_cli` | `~/.claude/projects/<encoded-cwd>/<session-id>.jsonl` | 1.0 |30| `claude_code_web` | CLI bridge via `claude --teleport <session-id>` | 1.0 |31| `codex_cli` | `$CODEX_HOME/sessions/YYYY/MM/DD/rollout-*.jsonl` | rollout-v1 |32| `codex_web` | CLI bridge via `codex cloud` | rollout-v1 |33| `antigravity_cli` | `~/.antigravity/projects/<encoded-cwd>/<session-id>.jsonl` | 1.0 |34| `grok_cli` | `~/.grok/sessions/session-*.jsonl` | grok-session-v1 |35| `pi_cli` | `~/.pi/sessions/session-*.ndjson` | pi-ndjson-v1 |3637All adapters fail soft (log warning, skip) when their source is unavailable.3839## Pipeline4041```421. Discover sessions (per adapter)432. Normalize to NormalizedEvent schema443. Sanitize (secrets, entropy, paths) — BEFORE any LLM sees content454. Triage (score for struggle signals: retries, errors, scope violations, corrections)465. Deep analyze (flagged sessions only — heuristic or LLM)476. Write findings to episodic memory (D4 tag schema, source:transcript-mined)48```4950## How It Works51521. **Discovery**: Each adapter enumerates available sessions from its source532. **Normalization**: Raw vendor events -> NormalizedEvent (common schema in `references/event-schema.md`)543. **Sanitization**: Reuses `session-log` sanitizer extended for tool-call arguments and tool-result outputs554. **Triage**: Scores each session on retry_count, tool_error_count, scope_violation_count, user_correction_count565. **Deep Analysis**: Runs on flagged sessions (composite score >= threshold), extracts structured findings576. **Memory**: Findings written with D4 tags (`failure_type:*`, `capability_gap:*`, etc.) + `source:transcript-mined`5859## Model Resolution6061- **Triage**: archetype `analyst` (standard tier), configurable via `config.yaml: triage.archetype`62- **Deep analysis**: archetype `reviewer` (premium tier), configurable via `config.yaml: deep_analysis.archetype`63- Both resolve via `agents_config.resolve_model()` from the coordinator6465## Prerequisites6667- Python 3.11+68- At least one harness's sessions directory accessible on disk69- For web adapters: vendor CLI installed and authenticated7071## Steps7273### 1. Discover and Normalize7475```bash76python3 <agent-skills-dir>/collect-transcripts/scripts/adapters/claude_code_cli.py77```7879### 2. Triage8081```bash82python3 <agent-skills-dir>/collect-transcripts/scripts/triage.py \83 --events-dir docs/transcripts/$(date +%Y-%m-%d)/ \84 --threshold 5.0 \85 --dry-run86```8788### 3. Deep Analysis (flagged sessions only)8990```bash91python3 <agent-skills-dir>/collect-transcripts/scripts/deep_analyze.py \92 --events-file docs/transcripts/2026-06-01/session-abc.jsonl \93 --dry-run94```
jankneumann