Go 1.26.0 carries 17 standard library vulnerabilities that govulncheck reports. The security workflow takes its toolchain from go.mod, so the scan failed. Go 1.26.8 clears them.
The pi adapter writes contextWindow and maxTokens to models.json only
when the profile sets them. Pi keeps its own defaults for every other
model. Tests cover both cases.
The govulncheck job fails on 17 standard library findings. The runner
uses an outdated Go toolchain. Set check-latest so setup-go installs
the newest patch release.
Add Rule as a first-class canonical kind (rules/<name>.md, loaded like
agents) and propagate it the way each harness actually loads it:
- claude-code: native ~/.claude/rules directory symlink (preserves rule
path-scoping frontmatter); rules NOT folded into CLAUDE.md.
- opencode/kilo/zed/crush: fold rule bodies into the global AGENTS.md
always-on instructions file (frontmatter stripped). crush has no global
instructions file, so register the absolute AGENTS.md path in
options.context_paths (defaults are re-prepended at load, so nothing is
clobbered).
- goose: fold into ~/.config/goose/.goosehints.
- cagent: fold into the per-agent instruction.
Also:
- Wire markdown agents to opencode (~/.config/opencode/agents) and fix the
kilo agents symlink: plural ~/.config/kilo/agents (singular "agent" is a
silent no-op on current Kilo CLI).
- Add ManagesAgents/ManagesRules to HarnessCapabilities + ImportResult.Rules;
claude-code/opencode import rules/agents round-trip.
- Centralize instruction selection in Bundle.InstructionText /
InstructionTextWithRules (replaces per-adapter override duplication).
- init scaffolds + imports rules; README capability matrix gains
Agents/Rules columns with native-vs-folded footnotes.
Grounded each harness's sink against official docs + config source this
session. Gate: gofmt, go vet, go test -race ./..., golangci-lint (0),
govulncheck (clean).
Three robustness fixes for the apply pipeline:
1. Silent read errors on state/target files treated 'file absent' as
'file empty' and fast-forward-overwrote user edits. Add
common.ReadExistingFile that returns (nil,nil) only on
os.IsNotExist; other I/O errors propagate. Use it from apply.Run
and from every adapter's Render (claudecode, crush, kilo,
opencode, goose, zed).
2. apply.Run required callers to pre-construct *gitx.Repo; passing
nil silently skipped the commit. Run now auto-opens a repo from
Bundle.Root, gracefully no-ops when the root is not a git repo,
and uses a new gitx.HasStagedChanges() check to skip the no-op
'nothing to commit' that triggered the original silent-skip.
3. Each of 7 adapter packages re-implemented the same 25-line
Option/WithHome/New constructor. Hoist into adapter.Base +
adapter.BaseOption + NewBase. Each adapter embeds *Base and
reduces to a 9-line constructor. Public New(WithHome(h)) API
preserved; a.home becomes a.Home.
Tests:
- apply: 5 new (propagates-read-error, auto-opens-repo,
no-repo-skips-commit, skips-no-op-commit, commit-message-counts)
- gitx: 2 new (HasStagedChanges empty/after-add)
- adapter/common: 3 new (ReadExistingFile missing/content/perm)
- adapter: 4 new (NewBase default/WithHome/order/typed)
- existing adapter tests unchanged — verify a.Home still honoured.
The Zed Agent panel (new ACP-based UI) only discovers providers registered
under language_models.openai_compatible.<name>, not language_models.openai.
Each named entry appears as its own provider in the LLM Providers list.
- language_models.openai_compatible.harness-sync-gateway holds gateway URL + models
- agent.default_model.provider = 'harness-sync-gateway' (was 'openai')
- Model entries include capabilities {tools,images,parallel_tool_calls}
- Export common.GatewayProviderID so zed.go can reference it
- Fix staticcheck QF1001 (De Morgan's law) in io.go
- Fix struct field alignment in File (80→72 bytes)
- gofmt remaining files
opencode/kilo require top-level 'name' and a 'models' Record<modelID,Model>
in each provider entry for the harness UI to enumerate and display models.
Without 'name', the gateway appears unlabelled; without 'models', the provider
shows no model list in the selector.
26 bite-sized tasks across 7 phases (bootstrap, adapter framework,
render/git/merge, apply pipeline, per-harness adapters, CLI commands,
integration test + polish). TDD throughout, frequent commits, exact
code in every step. Adapter-specific schema steps include inspection
commands and reference actual harness configs on this machine.
Canonical config at ~/.config/harness-sync/ (git repo) propagates skills,
agents, MCP, LLM endpoints (via profiles), and global instructions into
each detected harness. Hybrid sync: symlinks where shapes match, rendered
files where they differ. Plugin-style adapters; conflict resolution via
git merge-file three-way merge tracked in state/.
lukaszraczylo
