v2.5.0 — Auto-context into harness phases
Minor — first non-doc-only paired pair in the recent run (after v2.4.0/v2.4.1/v2.4.2/v2.4.3 all doc-only on this side). Harness ships real new phase behavior: every phase command (/setup, /plan, /work, /review, /release, /bugfix) now auto-invokes MemoryVault at predictable boundaries without the agent or operator having to remember to call /memory search or /memory save.
Paired with agent-toolkit v0.11.1 which ships the toolkit-side companion documentation (Cross-Repo-Memory-Protocol.md).
What changes for operators
With MEMORY_VAULT_PATH set + agent-toolkit/skills/memory/ sibling-cloned:
- Every phase auto-loads operator conventions + project-specific decisions + open-questions/known-issues (per phase) at its start.
- Phases that surface durable items offer to save them at the end (self-modulating ask — high-confidence saves silently with stderr notice; low-confidence prompts).
Without those prerequisites: zero behavior change. Every phase graceful-skips silently. Harness runs unchanged on systems where MemoryVault isn't adopted.
What's new
scripts/harness_memory.py(~520 lines, stdlib-only) — dispatcher with 4 sub-commands:recall/offer-save(self-modulating by--confidence) /plan-done-promotion(cursor-tracked) /available. 3-tier toolkit discovery (env override > sibling-clone >~/Antigravity/agent-toolkit/).scripts/vault_project.py(~200 lines, stdlib-only) —vault_projectslug resolution with 3-tier fallback (explicit field >github.repobasename > git origin).- 61 unit tests (33 + 28) wired into all 3 OS CI workflows.
- 6 phase/pipeline specs amended via sub-letter convention (§1b/§4c/§7b/§7c/§5b/§5c) preserving integer §-numbers.
- ADR 0007 — full design rationale + 5 locked design calls Q1–Q5 + 4 load-bearing assumptions.
- How-to — per-phase boundary table + 5-env-var matrix + 3 worked scenarios + 8 troubleshooting rows.
Env vars
| Variable | Default | Effect |
|---|---|---|
MEMORY_VAULT_PATH |
unset | Required for auto-context (graceful-skip when unset). |
HARNESS_AUTO_SAVE_MODE |
ask |
ask (confidence-modulated) / silent / off. |
HARNESS_AUTO_SAVE_CONFIDENCE_THRESHOLD |
0.8 |
Agent-supplied confidence ≥ threshold → silent save. |
HARNESS_RECALL_BUDGET_<PHASE> |
4k–6k | Per-phase recall token cap. |
HARNESS_MEMORY_TOOLKIT_PATH |
auto-detect | Toolkit-scripts dir override. |
Coordinated-release ordering
Paired with toolkit v0.11.1 (tagged first; this release notes URL-link back to it). 5th consecutive paired-release-as-documentation pair (v0.9.0/v0.9.2/v0.10.0/v0.11.0/v0.11.1) but first non-doc-only when measured from this side (harness MINOR bump justified by real new phase behavior).
Full changelog
v2.4.3..v2.5.0. See CHANGELOG.md.