You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Add Claude Code, OpenCode, and Grok observation #2
The same CLI and event contract observe all four requested harnesses on this Mac.
Blockers
Task 1's reviewed contract and exact candidate.
Acceptance criteria
Three native adapters retain usage, tools, skill/read evidence, compaction, and lifecycle data at each source's actual granularity.
Preserve model/effort evidence and source-specific reasoning/cache semantics. Do not invent per-response counters from aggregate-only sources.
Publish a capability matrix by installed harness version, separating observed, unreported, and unsupported event families.
Fill necessary recording gaps with bounded passive hooks where supported. Verify their actual delivery and expose capture failures.
Claude shared-session ancestry, OpenCode parent-message links, and Grok merged-turn/synthetic-message cases behave according to the parent spec.
Non-goals
New routing choices, credential changes, broad parser forks, global hook installation, and Rocky.
Required proof
Sanitized native fixtures for each harness, replay/idempotency checks, separate ccusage token reconciliation, and real verification of each hook used. Existing four-agent evidence supplies fixtures; do not rerun those agents just to manufacture usage.
Objective contribution
Make ordinary work comparable across the user's actual harnesses.
Shipped in v0.2.0 through #20 (component PRs #10 to #19, each independently reviewed to approval), promoted by Toolybara in toolboxmd/marketplace#72 and installed on this Mac from the marketplace (Codex and Grok Build). Evidence: Claude Code (#10), OpenCode (#13) and Grok Build (#14, #17) adapters; on this Mac 82 Claude transcripts, 278 OpenCode sessions and 1,129 Grok sessions imported with each source's own counter semantics; agent-observer trace --capabilities publishes the capability matrix. Not in 0.2.0: passive recording hooks, split to #21.
Activity
wip: Claude Code adapter with human-input kinds, skill loads and inst…
wip: OpenCode adapter privacy, pending tools, mutable rows and schema…
wip: Grok adapter privacy, prompt-id pairing, late usage, method vali…
wip: OpenCode excerpts exclude child sessions, direction blocks remov…
wip: Grok privacy fails closed, fixed error categories, whitelisted e…
wip: OpenCode adapter conforms to the ledger privacy spec (#2)
wip: Grok event details type-checked, replay errors deduplicated (#2)
wip: OpenCode adapter reading the OpenCode session database (#2)
wip: OpenCode adapter privacy, pending tools, mutable rows and schema…
wip: OpenCode excerpts exclude child sessions, direction blocks remov…
wip: OpenCode adapter conforms to the ledger privacy spec (#2)
wip: Grok adapter reading Grok session directories (#2)
wip: Grok adapter privacy, prompt-id pairing, late usage, method vali…
34 remaining items
wip: Grok event details type-checked, replay errors deduplicated (#2)
wip: Grok adapter uses the shared privacy module (#2)
wip: Grok child detection before replay, late model reconciliation, m…
wip: Grok model reconciliation keeps the stronger evidence; tokens re…
wip: Grok model provenance, proven authorship only, safe session ids (#2
wip: Grok keeps models without provenance; only native text content p…
Grok Build adapter (#14)
wip: Grok parent index per operation including source-scoped syncs; u…
wip: Grok events-only changes still reconcile response models (#2)
wip: Grok sync is linear: one parse per session, one parent index per…
wip: Grok parent index per operation including source-scoped syncs; u…
wip: Grok events-only changes still reconcile response models (#2)
Grok sync in linear time (#17)
lukemaj commentedon Sep 24, 2026
Shipped in v0.2.0 through #20 (component PRs #10 to #19, each independently reviewed to approval), promoted by Toolybara in toolboxmd/marketplace#72 and installed on this Mac from the marketplace (Codex and Grok Build). Evidence: Claude Code (#10), OpenCode (#13) and Grok Build (#14, #17) adapters; on this Mac 82 Claude transcripts, 278 OpenCode sessions and 1,129 Grok sessions imported with each source's own counter semantics;
agent-observer trace --capabilitiespublishes the capability matrix. Not in 0.2.0: passive recording hooks, split to #21.🤖 Generated with Claude Code