feat(parser): add gptme agent parser

[TimeToBuildBob]

Summary

Adds gptme as a supported agent in AgentsView.

gptme is an open-source terminal AI assistant (Spring 2023, one of the earliest agent CLIs) supporting Anthropic, OpenAI, Google, DeepSeek, and local models via OpenRouter. Sessions are stored as JSONL files with one JSON object per message.

Session format

{"role": "system", "content": "...", "timestamp": "2026-06-13T10:00:00.000000", "pinned": true}
{"role": "user", "content": "Write a hello world program in Python.", "timestamp": "..."}
{"role": "assistant", "content": "I'll write...", "timestamp": "...", "metadata": {"model": "openrouter/anthropic/claude-sonnet-4-6", "usage": {"input_tokens": 120, "output_tokens": 42, "cache_read_tokens": 80, "cache_creation_tokens": 0}}}
{"role": "tool", "content": "Saved file: hello.py", "timestamp": "..."}

Session discovery path: ~/.local/share/gptme/logs/*/conversation.jsonl

Changes

• internal/parser/gptme.go — DiscoverGptmeSessions, FindGptmeSourceFile, ParseGptmeSession with full token accounting (input/output/cache read/cache write tokens)
• internal/parser/types.go — register AgentGptme in the Registry with GPTME_DIR env var and gptme_dirs config key
• internal/parser/gptme_test.go — tests for parse, discover, find, and session-name-to-project extraction
• internal/parser/testdata/gptme/ — fixture JSONL with 7 messages covering all role types and token_usage fields

Implementation notes

• Tool messages ("role": "tool") are mapped to RoleUser with IsSystem: true, following the same pattern used for other agents' tool output
• System messages are skipped (they're context injections, not conversation turns)
• Session name format: YYYY-MM-DD-topic or YYYY-MM-DD-HHMMSS-topic — the parser extracts the topic part as the project name
• Token field names: gptme uses input_tokens/output_tokens/cache_read_tokens/cache_creation_tokens (vs. Anthropic API's cache_read_input_tokens)

Testing

go test ./internal/parser/... -run "TestParseGptme|TestDiscoverGptme|TestFindGptme|TestGptme" -v

All 4 tests pass.

[roborev-ci]

roborev: Combined Review (0a02eca)

High-level verdict: changes are not ready; gptme is registered but not wired into sync, and its fixture timestamp format is not parsed.

High

• Location: internal/parser/types.go:456
• Problem: The new agent is registered for discovery, but the sync engine has no AgentGptme case or processGptme path. Discovered gptme files will fall into the engine's unknown-agent branch instead of being parsed and indexed.
• Fix: Add sync engine support analogous to the other file-backed JSONL agents and cover it with a sync integration test.

Medium

• Location: internal/parser/gptme.go:97
• Problem: The parser uses parseTimestamp, but the added fixture uses timestamps like 2026-06-13T10:00:00.000000, which that helper does not accept. Those gptme records will persist with nil session/message timestamps, breaking chronological ordering and date-based filtering for this format.
• Fix: Add a gptme-specific timestamp parser or extend accepted layouts, then assert parsed session/message timestamps in the test.

---

Panel: ci_default_security | Synthesis: codex, 8s | Members: codex_default (codex/default, done, 2m51s), codex_security (codex/security, done, 1m35s) | Total: 4m34s

[TimeToBuildBob]

Addressed the two roborev findings in 832fa1c:

High — sync engine missing AgentGptme dispatch
Added a classifyOnePath case for AgentGptme (iterates e.agentDirs[parser.AgentGptme], matches <logsDir>/<session-dir>/conversation.jsonl) and wired processGptme into the processFile switch, following the same pattern as other file-backed JSONL agents.

Medium — microsecond timestamp format not parsed
Extended timestampLayouts with "2006-01-02T15:04:05.000000" to handle gptme's timezone-less microsecond timestamps. Added assertions in TestParseGptmeSession for sess.StartedAt, sess.EndedAt, and msgs[0].Timestamp to confirm end-to-end parsing. All 4 parser tests pass.

[roborev-ci]

roborev: Combined Review (832fa1c)

No issues found.

---

Panel: ci_default_security | Synthesis: codex | Members: codex_default (codex/default, done, 8m59s), codex_security (codex/security, done, 3m14s) | Total: 12m13s

[roborev-ci]

roborev: Combined Review (746bb5e)

No Medium, High, or Critical findings were reported.

The only findings were Low severity and are omitted per instructions.

---

Panel: ci_default_security | Synthesis: codex, 5s | Members: codex_default (codex/default, done, 5m54s), codex_security (codex/security, done, 2m17s) | Total: 8m16s

[roborev-ci]

roborev: Combined Review (d76fab2)

Medium: gptme tool output is being hidden from visible transcripts.

• Medium: internal/parser/gptme.go:164 stores gptme tool messages with IsSystem: true, causing UI/export/search paths to treat visible tool output, such as command results, as hidden system messages. Store gptme tool output as visible transcript content, or attach it as tool-result data if gptme provides a stable tool call ID.

---

Panel: ci_default_security | Synthesis: codex, 8s | Members: codex_default (codex/default, done, 5m27s), codex_security (codex/security, done, 2m21s) | Total: 7m56s

[TimeToBuildBob]

Addressed the latest roborev Medium finding in c91507a5.

gptme does not expose a stable tool-call ID in these JSONL lines, so I did not synthesize ToolResults. Instead, I changed standalone role:"tool" entries to stay visible as assistant transcript content rather than hidden IsSystem messages. That keeps command output like Saved file: hello.py and Hello, World! visible in UI/export/search without inflating UserMessageCount.

Verified with:

• go test ./internal/parser -run "TestParseGptmeSession|TestDiscoverGptmeSessions|TestFindGptmeSourceFile|TestGptmeProjectFromSessionName"

[roborev-ci]

roborev: Combined Review (c91507a)

Verdict: No Medium, High, or Critical findings to report.

The only reported issue was Low severity and has been omitted per instructions.

---

Panel: ci_default_security | Synthesis: codex, 7s | Members: codex_default (codex/default, done, 6m30s), codex_security (codex/security, done, 2m58s) | Total: 9m35s