New-agent support roadmap: OpenClaw, Pi, goose, Crush, Cline, OpenHands, Hermes, Copilot, Gemini, Kilo, and qualification backlog

[ImL1s]

Summary

This is the execution umbrella for expanding Portable Resume beyond the original eight sources/destinations.

It links the scalable capability architecture, first-wave implementation issues, phased destination-first profiles, compatibility/migration support, and the second-wave research backlog. It does not replace the source-correctness umbrella #18 or destination/installer-hardening umbrella #30; implementation PRs must satisfy those existing safety and transaction contracts.

Research baseline: current upstream projects/docs reviewed on 2026-07-26.

Product boundary

Every new integration remains:

• local/offline for source reads
• immutable toward source stores
• inert/untrusted at handoff
• stdlib-only in the bundled Portable Resume runtime
• no source CLI/server/gateway/cloud attachment
• no execution of recovered commands/tool calls
• no auth/settings/credential migration
• no native live-process continuation claim

Source adapter, destination Skill, native package, activation and marketplace support are independent axes.

Foundation

• Next-wave support architecture: decouple source adapters, destination hosts, and native package surfaces #36 — decouple source adapters, destination hosts and native package surfaces; dynamic rectangular matrices and evidence layers
• Shared JSONL infrastructure: stable streaming scanner and metadata reducers #10 — shared stable JSONL scanner/reducer
• Snapshot exact reads: remove whole-parent 2,000-sibling failure mode #16 — exact reads independent of unrelated sibling count
• ReadBudget invariants: reject all caller-raised global ceilings #17 — no caller-raised global bounds
• source safety/correctness backlog under Cross-adapter large-session and discovery parity after v0.3.3 #18
• destination/installer safety under Destination CLI integration and installer hardening umbrella #30, especially [P0 security] recover must not delete paths supplied by an untrusted complete journal #20/[P0 security] pin installer support metadata and make lock, journal, and manifest writes no-follow and atomic #21/Installer manifest and journal need closed bounded schemas with strict validation #28/[P0 security] installer payload commits must be descriptor-relative and resist parent-directory swap escapes #31/Install must rebuild and authorize the plan under lock from the exact current manifest generation #35 before broad new direct installs are released

No new profile should increase the blast radius of known installer P0s. Source-only fixtures/adapters can proceed independently when they do not use the mutating installer.

Tier 1 — full source + destination candidates

[ ] #37 OpenClaw

Scope:

• current per-agent SQLite sessions/transcript events
• multi-agent composite identity
• branch/compaction/reset/internal-session rules
• OpenClaw workspace/global Agent Skills and local-directory install evidence

Priority: P1, high user value and strong current storage/Skill contracts.

[ ] #38 Pi

Scope:

• documented v1/v2/v3 tree JSONL
• active leaf, branch and compaction summaries
• native .pi/skills/~/.pi/agent/skills destination

Priority: P1, likely the cleanest first new source adapter after #36/#10.

[ ] #39 goose

Scope:

• current SQLite schema 15 session/message store
• user versus scheduled/subagent/hidden/gateway/ACP filtering
• legacy JSONL later
• direct Agent Skills destination

Priority: P1, implement current SQLite first.

[ ] #40 Crush

Scope:

• current project/data-root SQLite and migrations
• message parts JSON, parent/summary semantics
• CRUSH_GLOBAL_DATA and Skill roots

Priority: P1, schema qualification PR before adapter.

[ ] #41 Cline

Scope:

• authoritative per-session JSON plus SQLite index/fallback
• root versus child/team/observer privacy
• direct .cline/skills destination

Priority: P1, active product but schema/privacy qualification is substantial.

[ ] #42 OpenHands CLI

Scope:

• current per-event JSON conversation directories
• stable multi-file snapshot
• action/observation and condensation semantics
• local direct Agent Skills only; Cloud/SDK/plugins excluded

Priority: P1, exact current CLI product/version required.

[ ] #43 Hermes Agent

Scope:

• WAL SQLite state/messages
• profile/native ID, child/compaction lineage and messaging privacy
• global $HERMES_HOME/skills destination

Priority: P1, direct global Skills first; no hub/gateway/skill_manage.

Tier 1 phased — destination first, source gated

[ ] #44 GitHub Copilot CLI

Phase A:

• official project/personal Agent Skills roots
• direct and local CLI-managed Skill install/activation evidence

Phase B:

• authoritative qualification of session-state/<id>/events.jsonl and session-store.db
• source profile remains planned until rewind/compaction/subagent/fleet/cloud-sync semantics are fixture-backed

Priority: Destination P1, source qualification P1 research.

Tier 2 compatibility / fork qualification

[ ] #45 Gemini CLI

• preserve local/enterprise/API-backed histories and Gemini-specific Skill roots
• consumer Login-with-Google access ended June 18, 2026; Antigravity remains separate
• destination may land before current JSON source schema qualification
• no auth/quota/settings migration

Priority: P2 compatibility/migration.

[ ] #46 Kilo Code CLI

• destination and source profiles independent of OpenCode
• active Kilo database/storage refactor must be pinned
• reuse only generic helpers proven through Kilo/OpenCode divergence tests
• CLI direct Skills first; IDE/Cloud/Marketplace later

Priority: P1/P2 staged, source remains research until current schema fixtures exist.

Tier 3 research backlog

[ ] #47 Qualification registry

Priority candidates:

• Aider
• Amp
• Factory Droid
• Codebuff
• Kiro CLI
• Zed Agent
• Roo Code

Additional destination/source qualification:

• Warp
• Windsurf
• Junie
• Replit/Universal
• Qoder
• Trae
• CodeBuddy
• Command Code
• Bob
• Augment
• Firebender
• Cortex
• Mistral Vibe
• other current gh skill hosts

Archival/deferred review:

• Continue
• pre-SQLite Goose
• old Hermes JSONL
• old OpenHands single-file formats
• archived/renamed Kilo repositories

Research entries never enter runtime/release matrices until promoted by a dedicated implementation issue.

Recommended execution order

Phase 0 — unblock architecture and safety

1. Next-wave support architecture: decouple source adapters, destination hosts, and native package surfaces #36 registry/matrix PR 1
2. ReadBudget invariants: reject all caller-raised global ceilings #17 bounds invariant
3. Shared JSONL infrastructure: stable streaming scanner and metadata reducers #10 scanner API/tests
4. installer P0 patch release work [P0 security] recover must not delete paths supplied by an untrusted complete journal #20/[P0 security] installer payload commits must be descriptor-relative and resist parent-directory swap escapes #31 and locked-plan/control-store prerequisites as required for destination installs

Phase 1 — simplest high-confidence sources

5. Pi PR 1 fixtures/decision record
6. Pi source adapter
7. OpenClaw PR 1 fixtures/schema
8. goose PR 1 schema/fixtures

Pi is recommended first because the versioned file format and Skill roots are publicly documented. OpenClaw/goose follow through the shared SQLite safety path.

Phase 2 — first destinations

9. Next-wave support architecture: decouple source adapters, destination hosts, and native package surfaces #36 dynamic destination/source registry and matrix PRs complete
10. Pi/OpenClaw/goose direct destination profiles after the installer security release gate
11. Copilot destination profile
12. exact-version isolated activation evidence per host, never inferred from the Python runner matrix

Phase 3 — remaining first wave

13. Crush
14. Cline
15. OpenHands
16. Hermes

Each starts with a decision-record/fixture PR; no large adapter/host monolith.

Phase 4 — compatibility/forks

17. Gemini destination/lifecycle decision, then source only if qualified
18. Kilo destination/storage decision, then source only if qualified

Phase 5 — qualification backlog

19. Qualify second-wave agents: Aider, Amp, Droid, Codebuff, Kiro, Zed, Roo, Warp, Windsurf, Junie, and others #47 research registry
20. open dedicated implementation issues only for qualified source/destination axes

Standard PR sequence per new tool

PR A — upstream contract and synthetic fixtures

Required:

• exact upstream repository/product/version/commit/date
• source root and override policy
• schema/version/authority ordering
• cwd/project identity
• branch/rewind/compaction/child rules
• privacy and cloud/local boundary
• destination roots/precedence/activation
• generated synthetic fixture family and manifest
• explicit blockers and go/no-go

This PR may contain no runtime support key yet.

PR B — source adapter

• profile registration
• safe probe
• metadata-only list
• exact ID/path
• selected public show/reducer
• large-resource, corruption, race, path and privacy tests
• immutable source proof

PR C — destination direct Skill

• destination registration/root provenance
• host-neutral payload profile
• direct install/verify/uninstall lifecycle
• dynamic installed-runner cells
• shadow detection/docs/localizations

Requires destination installer security gates applicable to the selected roots.

PR D — exact-version host evidence

• isolated host home/config/data
• exact package/payload digest
• native/local install or direct discovery
• explicit activation
• loaded owned-runner provenance
• sanitized evidence tied to host version and date

PR E — optional compatibility/native marketplace

Legacy providers, IDE products, cloud/team/gateway surfaces or native marketplaces each require their own contract/evidence and are not bundled into basic CLI support.

Source acceptance contract

A new source cannot be marked supported until:

• authoritative local source exists
• closed format/schema profile and synthetic fixture exist
• exact ID/path and cwd/project isolation work
• probe/list do not parse full transcripts unnecessarily
• show uses the correct source/record/transcript/output budgets
• active branch/compaction/rewind semantics are proven
• child/internal sessions cannot crowd root sessions before LIMIT
• system/reasoning/credentials/private routing/binary data is filtered
• source mutation produces retry/busy rather than mixed output
• no source process/network/API is invoked
• source tree/DB remains unchanged

Destination acceptance contract

A new destination cannot be marked supported until:

• native/default project/global root and overrides are authoritative
• precedence/alternate roots are machine-readable
• direct payload parser/frontmatter/resource constraints pass
• selected root provenance is reported
• owned runner and typed request path are deterministic
• direct install/verify/uninstall is transactional and safe
• duplicate/shadow copies are handled under Detect duplicate and shadowed Portable Resume Skills across each host's discovery roots #34
• installed-runner matrix cells pass for every enabled source
• native activation is recorded separately on an exact host version

A destination-only profile is valid and does not require a source adapter.

Matrix and release-size expectations

Counts are always derived from enabled registries.

For example, after the original eight plus seven full first-wave tools:

15 supported sources × 15 supported direct destinations = 225 direct runner cells

If Copilot/Gemini/Kilo destinations land before their sources, the matrix becomes rectangular rather than symmetric. Reports must show dimensions and enabled key sets, not merely a single number.

To keep CI practical:

• shard by source/destination
• reuse one deterministic payload build per destination/profile
• generate large fixtures at runtime
• keep native proprietary host validation scheduled/manual
• do not run network/account activation in normal CI

Release milestones

Milestone N1 — architecture ready

• Next-wave support architecture: decouple source adapters, destination hosts, and native package surfaces #36 core registries/dynamic reporting
• current original eight remain green
• research/planned profiles excluded from pass totals

Milestone N2 — first new source

• Pi source probe/list/show
• no destination release claim required
• current adapters still pass their backlogs

Milestone N3 — first new source+destination pair

• Pi direct destination and exact-version activation evidence
• dynamically expanded installed-runner matrix

Milestone N4 — SQLite wave

• OpenClaw + goose current SQLite source/destination
• shared SQLite/WAL security evidence

Milestone N5 — complete first wave

• Crush, Cline, OpenHands, Hermes supported at approved axes
• Copilot destination supported/source status explicit

Milestone N6 — compatibility and second wave

• Gemini/Kilo approved phases
• Qualify second-wave agents: Aider, Amp, Droid, Codebuff, Kiro, Zed, Roo, Warp, Windsurf, Junie, and others #47 qualification promotions

Partial releases are acceptable only when STATUS/CHANGELOG name exact completed axes and keep the remaining issues open.

Verification gates

Every implementation PR reports relevant focused tests and the canonical suite:

python3 scripts/self_verify.py
python3 scripts/check_docs.py
python3 scripts/check_secrets.py
PYTHONPATH=src python3 -m unittest discover -s tests -q
PYTHONPATH=src python3 scripts/smoke_installed_matrix.py

Additional per-tool gates:

• fixture generation/manifest validation
• source-tree immutability
• no source-process static/runtime invocation
• large history boundary
• exact path/ID and cwd collision
• privacy/redaction
• destination root/shadow/runner provenance
• exact native host evidence when claimed

Completion criteria

This roadmap closes only when:

• Next-wave support architecture: decouple source adapters, destination hosts, and native package surfaces #36 is complete and all new profiles use it
• every first-wave issue Add OpenClaw source adapter and destination Skill support for the SQLite session era #37–Add Hermes Agent source adapter and destination Skill support for WAL SQLite sessions #43 is closed or explicitly reclassified/deferred with evidence
• Copilot Add GitHub Copilot CLI destination support; gate source adapter on pinned session-event schema #44 source/destination axes are separately resolved
• Gemini Preserve Gemini CLI source/destination compatibility after the Antigravity consumer transition #45 and Kilo Add Kilo Code CLI support without assuming OpenCode storage or Skill parity #46 compatibility decisions are complete
• Qualify second-wave agents: Aider, Amp, Droid, Codebuff, Kiro, Zed, Roo, Warp, Windsurf, Junie, and others #47 has machine-readable results and dedicated issues for promoted candidates
• matrices/evidence remain dynamic and layered
• source Cross-adapter large-session and discovery parity after v0.3.3 #18 and destination Destination CLI integration and installer hardening umbrella #30 safety work is not bypassed
• STATUS and CHANGELOG show provider/host/package/activation scope per tool
• no live resume/cloud/history migration claim is inferred from inert context migration

Out of scope

• supporting every name in a generic Skill installer catalog
• popularity-only profile promotion
• cloud/team/remote session retrieval
• automatically invoking candidate CLIs to inspect user stores
• combining all first-wave support in one PR
• executing recovered commands or restoring workspaces
• native live-process continuation

Security / privacy

All source fixtures, host homes and activation evidence must be generated/synthetic and isolated. Never attach real sessions, databases, WALs, event files, account/team/channel metadata, credentials, private source, tool arguments/results, worktrees, checkpoints or home paths.

[ImL1s]

Progress from PR #49 (7b5192c)

Milestone N1 / Phase 0 foundation merged. No Pi/OpenClaw/goose support claimed. Continue per-axis PR A–E behind honesty gates; see docs/STATUS.md and the next-wave plan status banner.