CI Healer: implementation backlog

Status key: Done (implemented and tested here), Done, unverified live (implemented and tested with fakes; a live check is still needed), Blocked (needs something this environment lacks), Open (not started).

Gaps from the original review

ID	Gap	Status	What was done	Remaining
G-1	State was a JSON file	Done	SQLite ledger (persistence/ledger-store.ts), schema v1 and v2, WAL, transactions	The Ruby reference still uses JSON, by design
G-2	Standalone Ruby tool, not in src/domains/	Done	TypeScript port in src/domains/ci-healer/, command nexum ci-healer	Ruby kept as reference
G-3	Setup and verify commands ran on the host	Done, unverified live	Docker sandbox with the same hardening as nexum's ShellTool; verify has no network; preflight fails closed; host mode needs an explicit acknowledgement	Real-daemon test is gated (CI_HEALER_DOCKER_TESTS=1) and was not run here: no Docker daemon. Image: the default nexum-sandbox:latest is Node-only (see B-05)
G-4	Agent inherited GH_TOKEN and other credentials	Done	Allowlisted environment for the agent and repo commands; GitHub, SSH, and GPG names are always denied	File-based credentials under HOME are not covered (see B-06)
G-5	Run discovery was not paginated and kept no cursor	Done, unverified live, replaced design	Discovery scans branch heads instead of a run window, so no cutoff can hide a current failure. A persisted cursor rotates through branches beyond max_branches_per_tick. The REST runs endpoint, filtered by branch, returns runs for one branch. per_page bounds each page	This replaces the proposal's "pagination plus run cursor". The change is a judgement call: review it before merging
G-6	Only nexum rpc was wired	Done, unverified live	Ordered runtime chain (nexum, opencode, agy) with circuit breakers and rate-limit cooldowns, stored in SQLite. Free-tier routing is ordered (runtime, model) entries	opencode and agy are not installed here. Their argument vectors come from the documentation (see B-03 and B-04)
G-7	Stale-run check on pull_request runs unverified	Done, verified against the real repo	Discovery compares run SHAs to the branch head. A mismatch logs the reported SHAs, and plan shows the result	Resolved by B-01: GitHub reports the PR head SHA

Items from the proposal's "not included" list

ID	Item	Status	Notes
N-1	CI Healer design document	Done	docs/ci-healer/DESIGN.md
N-2	Implementation backlog	Done	This file
N-3	Docker sandboxing	Done, unverified live	See G-3
N-4	SQLite migration	Done	Versioned migrations; upgrade path tested
N-5	Credential scrubbing	Done	See G-4

Live verification (remaining)

ID	Check	Blocker	Procedure
B-01	headSha semantics for pull_request runs	Resolved	plan against shubhamtaywade82/nexum (REST API) found the PR head SHA reported for pull_request runs, and no mismatch was logged. The no runs report the branch head log would mean GitHub changed this
B-02	Real-daemon sandbox behaviour	No Docker daemon	docker build -t nexum-sandbox:latest docker/nexum-sandbox/, then CI_HEALER_DOCKER_TESTS=1 npx jest tests/domains/ci-healer/docker-sandbox.gated.test.ts
B-03	opencode headless flags	Binary not installed	Run opencode run --help on the installed version. Compare with openCodeAgent in adapters/cli-agents.ts. Check whether --format json and --model provider/model still match
B-04	agy headless flags and auth	Binary not installed	Run agy --help. The documented print mode is agy -p <prompt> --output-format json. Sign in once interactively first; a headless run without cached credentials fails
B-05	Language toolchains in the sandbox image	Image is Node-only	Add toolchains to docker/nexum-sandbox/Dockerfile only when a target repo needs them, or set sandbox.image to a repo-specific image
B-06	File-based credentials under HOME	Not in scope of the env policy	For Docker mode, verify commands see no host HOME, so this is covered. For host mode, it is not covered and needs the acknowledgement. For agy, the cached sign-in is reachable by design
B-07	Ollama condenser	No Ollama reachable	Run with condenser.type = ollama against a running Ollama; confirm the fallback when it is stopped
B-08	Merge-SHA mapping	Not needed	B-01 found that GitHub reports the PR head SHA, so no mapping is required. Reopen only if GitHub starts reporting merge SHAs

Known limitations (accepted, documented)

ID	Limitation	Why accepted	Mitigation
L-1	The worktree's .git file points at the host repository, so git inside the container cannot resolve it	Mounting the host .git would expose hooks and config to repo code	Verify commands that need git history must run in host mode; most test commands do not need it
L-2	Setup has network access in Docker mode	Dependency installation needs it	Set setup_network to none for vendored dependencies
L-3	Runtime CLIs' JSON output is not parsed	Their schemas are not stable contracts	Exit status plus the verify gate decide
L-4	One failing run per head is digested	One lineage per head keeps attempts bounded	The digest is the newest failing run; other failing workflows are not merged into it
L-5	Windows and macOS paths are untested	The build environment is Linux	Docker bind mounts with commas are rejected explicitly

Open items

ID	Item	Phase
O-1	Failure classification (flaky, infrastructure, dependency, code) before repair	2
O-2	Webhook-triggered ticks in place of polling	4
O-3	Multi-repository config and per-repository execution policy	5
O-4	Repair history and runtime metrics (latency, tokens, cost)	5
O-5	Notifications and human-approval controls	5
O-6	Interactive dashboard	5

Acceptance criteria

The criteria are in DESIGN.md section 11. Each gate that protects a push is covered by a test that fails when the gate is removed.