00 — Executive Summary

Research conducted: 2026-07-03. Mission per the brief: how SpecStory works, what else exists, what it means for jackin❯. This dossier is the evidence layer under the existing Déjà Vu design proposal.

The landscape in six sentences

Session capture — persisting every AI coding conversation into durable, searchable artifacts — has one production-grade cross-agent implementation: SpecStory, whose open-source (Apache-2.0) Go CLI reads six agents' native on-disk stores via fsnotify, normalizes them into one schema, and renders markdown into .specstory/history/ (01, 02). The entire market — SpecStory, the closed competitor Mantra, and ~40 OSS exporters/viewers — has converged on the same mechanism: read the agent's own session files; nobody PTY-scrapes, and hooks/OTel/wire-taps appear only as side channels (11, 12, 20). Every agent jackin❯ runs persists a watchable local transcript except Amp, which requires live --stream-json capture — and the two agents nobody covers, Kimi and Grok, turn out to have good native surfaces (versioned wire.jsonl; documented session dirs + ACP streams) (10). Capture itself is commoditized to $0 by its own vendor; value has moved up-stack to search, cross-agent resume, skill-mining (Lore), and turn↔code attribution (Mantra's git time-travel, SpecStory's provenance engine). Four days before this research, SpecStory v2.0.0 shipped cross-project cross-agent resume — reconstructing a session into a different agent's native format — proving demand for exactly the cross-agent layer Déjà Vu targets. As of 2026-07-03 no product ships a cross-agent, host-side, purge-surviving archive: Déjà Vu's slot in the market is still empty.

Headline findings

1. Déjà Vu's core bets are confirmed by the whole market. Native-file reading first, vendor streams as hedge, screen-scrape used by literally nobody (20 §mechanism scorecard).
2. SpecStory's implementation is a reusable blueprint, not just inspiration: 58k LOC Apache-2.0 Go — provider SPI (11 methods), unified SessionData schema, disposable FTS5 index fingerprinted by size+mtime+version, full-rewrite idempotent markdown rendering, cross-agent reconstruction with MigrationNote (02).
3. Best-case off-the-shelf coverage of jackin❯'s six agents is 2/6 (SpecStory) to 3/6 (claude-replay); zero tools cover Kimi or Grok — first-mover ground for jackin❯ (20 §coverage).
4. OTel cannot cover the fleet: only Claude/Codex/Gemini export it at all; for Amp, OpenCode, Kimi, Grok there is no telemetry channel — files and vendor APIs are the only fleet-wide capture path (10).
5. The failure modes are documented, not hypothetical: editor-update breakage, CPU-burning polls, silent incompleteness, un-gitignored transcripts leaking secrets — each with SpecStory issue numbers to design against (01 §complaints).
6. jackin❯'s substrate advantage is real and already shipped: agent homes are restored host-side under ~/.jackin/data/<container>/, so capture for five of six agents is a read of paths jackin❯ already owns — while today eject/prune deletes those very histories (30).

What changes for Déjà Vu

Confirmed: vector ranking, provider architecture, storage shape, market-gap claim. New inputs: the roster is six agents now (Grok added; several docs still say five); SpecStory shipped cross-agent resume ahead of plan — mechanism worth studying; attribution (turn↔git-state) has moved from "distant non-goal" to "two competitors are on it"; Kimi/Grok store layouts are now mapped externally but need in-container verification. Full delta in 30.