20 — Comparison Matrix and Tier List

SpecStory is the strongest reusable baseline, but its two-of-six roster overlap and lack of a host-side purge-surviving archive leave the core Déjà Vu gap open.

Summary

Ranks session-capture options by agent coverage, capture robustness, content fidelity, search, sharing, health, and license. Verification cutoff: 2026-07-03. Tools come from 01/02 (SpecStory), 11 (OSS), and 12 (platforms). Tier value means usefulness to a jackin❯ session-capture capability: roster coverage × mechanism robustness × output/search/share quality × maintenance health × license reusability. It does not rank personal developer products.

Question and scope

Which candidates best fit jackin❯ by roster coverage, extraction mechanism, output, health, and license?

Method

The matrix scores candidates against evidence in the agent-store, exporter, platform, and SpecStory chapters using the stated capability and mechanism criteria.

Findings

Capability matrix (the contenders)

Capture = how conversation content is acquired. Live = captures during the session vs post-hoc. Content = full conversation (incl. tool calls) vs metrics only.

Tool	Capture	Live	Content	Output	Search	Share	Turn↔code link	License	Health (2026-07-03)
SpecStory CLI v2	native stores, fsnotify	✅ (watch) + post-hoc	full incl. thinking, tools, usage	markdown in-repo	FTS5, cross-project	cloud links (opt-in)	prototype (provenance engine, hidden flag)	Apache-2.0	active, v2.0.0 4 days old
SpecStory IDE ext	Cursor `state.vscdb` / Copilot `chatSessions` polling	✅	full (known gaps: Cmd-K, images)	markdown	via cloud	anon + cloud links	❌	closed	active; breakage history
Mantra	native stores	✅	full + per-turn git snapshots	desktop timeline replay	✅ FTS	web publish	✅ shipped (diff/restore)	closed freeware	active (rel. 2026-07-01)
claude-replay	native stores (5 agents)	post-hoc	full	self-contained HTML replay	❌	file = artifact	❌	MIT	active
coding_agent_session_search	native stores (11+ providers)	post-hoc	full (indexed)	TUI/CLI results	✅ best-in-class breadth	❌	❌	nonstandard	very active
claude-code-log	Claude JSONL	post-hoc	full	HTML/MD	project browse	❌	❌	OSS	very active
claude-code-transcripts	Claude JSONL	post-hoc	full	multi-page HTML	❌	gist / SSO sites	❌	OSS	slowing
aichat search (claude-code-tools)	native stores (Claude, Codex…)	post-hoc	full (Tantivy index)	CLI/TUI	✅	❌	❌	MIT	active
claude-trace	fetch-patch wire tap	✅	full + system prompt + raw API	JSONL + HTML	❌	file	❌	OSS	low activity
Langfuse (hook route)	Claude hooks → traces	✅	full conversations (this route only)	web platform	✅	team platform	❌	OSS core + SaaS	active
Happy	wraps CLI, E2E relay	✅	live session state	mobile/web viewer	❌	device sync, not links	❌	MIT	very active
vibe-kanban	owns launched sessions	✅	per-attempt logs	web kanban	❌	❌	per-task diffs	Apache-2.0	active
Amp threads (native)	server-side by design	✅	full	web threads	✅	workspace/unlisted URLs	❌	closed SaaS	active
opencode `/share` (native)	server publish	✅	full	web page	❌	public links	❌	OSS agent, closed share svc	active
ccusage (for contrast)	Claude JSONL	post-hoc	metrics only	CLI reports	❌	leaderboards	❌	MIT	16.8k★, very active

Coverage vs jackin❯'s roster

jackin❯ runs Claude Code, Codex, Amp, Kimi, OpenCode, Grok (30). Coverage of that roster by the cross-agent tools:

Tool	Claude	Codex	Amp	Kimi	OpenCode	Grok	Roster score
SpecStory CLI	✅	✅	❌	❌	❌ (open request #146)	❌	2/6
Mantra	✅	❌	❌	❌	❌	❌	1/6
claude-replay	✅	✅	❌	❌	✅	❌	3/6
coding_agent_session_search	✅	✅	❌ `[UNVERIFIED — provider list not fully published]`	❌	✅ likely	❌	~3/6
vibe-kanban (captive)	✅	✅	✅	❌	✅	❌	4/6, not exportable
Nobody	—	—	—	—	—	—	0 tools cover Kimi or Grok

Two consequences: (a) no off-the-shelf tool solves jackin❯'s roster — best case 3/6; (b) Kimi and Grok providers would be first-in-market, and both turn out to have good native surfaces (Kimi versioned wire.jsonl + ACP; Grok documented session dirs + ACP updates.jsonl — 10).

Mechanism scorecard (matches Déjà Vu's vector taxonomy)

Mechanism	Used by	Verdict from field evidence
A. Native-store read	SpecStory CLI, Mantra, all OSS exporters	Wins everywhere it exists: lossless, passive, post-hoc capable. Risk = format churn; mitigated by fingerprinted disposable indexes and vendor-blessed hedges
B. Hooks	Langfuse route, disler's observability, claude_telemetry	Good liveness signal; only Claude hands back `transcript_path`; nobody uses hooks as sole capture
C. OTel	Grafana official, claude-code-otel, Logfire	Metrics/cost only; content redacted by design; never a transcript source
D. Wire tap	claude-trace, LeanMCP proxy	Only source of system prompts + guaranteed thinking; operationally invasive; niche
E. Vendor stream/API	Amp `--stream-json`, OpenCode `serve`, Codex `--json`	The correct answer where no file exists (Amp) and the hedge where formats are declared unstable (Codex)
F. Screen scrape	nobody in the entire market	Confirms Déjà Vu's last-resort ranking — no shipping product relies on it

Tier list

Tier	Tool	Rationale (one line each)
S	SpecStory CLI	Only production-grade cross-agent capture pipeline with reusable Apache-2.0 source; 2/6 roster coverage but the mechanism blueprint (SPI, schema, index, reconstruct) transfers wholesale
A	Mantra	Proves the turn↔git-state layer users want next; closed source caps it at "study the UX, not the code"
A	claude-replay	Best output artifact in the market (portable HTML replay), 3/6 roster, MIT
A	coding_agent_session_search	Best retrieval breadth (11+ providers), same-day maintenance; pairs with any capture layer
A	claude-code-log	The Claude Code archival workhorse; parse rules worth mirroring
B	claude-code-transcripts	Best publishing/sharing workflow (gist, SSO sites); single-agent, slowing
B	aichat search / claude-history / claude-code-history-viewer	Healthy single-purpose retrieval/browsing tools; validate index-over-store design
B	Amp threads · opencode `/share`	Native benchmarks any cross-agent surface must match (and interop with, since jackin❯ runs both agents)
B	Langfuse hook route	Only observability platform that lands full coding-agent conversations; team features for free; Claude-only ingestion today
C	claude-trace · LeanMCP	Wire-depth when system prompts/thinking must be exact; too invasive as default
C	ccusage / usage monitors / viberank / vibe-log	Different job (metrics); proves multi-product value of native stores
C	Happy · vibe-kanban · Crystal · Conductor	Orchestrator-captive history; category evidence, not reusable components
F	cursor-chat-export, cclogviewer, sniffly, claude-code-otel, opcode, omnara, cui, claude-code-exporter	Stale, archived, or upstream-deleted — the graveyard that shows maintenance is the hard part

What the matrix says in one paragraph

The market has one production cross-agent capture pipeline (SpecStory, open, 2/6 roster fit), one closed product pointing at the next layer (Mantra's git time-travel), a healthy single-agent Claude ecosystem, thin-to-nonexistent everything else, and zero coverage of Kimi and Grok. Capture mechanism is a solved question (native stores + vendor-stream hedges); differentiation now lives in normalization quality, retrieval, sharing surfaces, and turn↔code attribution. For jackin❯ that reads: nothing to buy, one blueprint to port, two first-mover providers to write, and a defensible position from container ownership that no market player shares (30).