Prepare the 1.2.0 advisory release

[pengfei-threemoonslab]

Prepares 1.2.0 on the advisory channel, per docs/release-runbook.md § Cutting the release (steps 1–3) and § The advisory release channel (step 1). It stops at "ready for the owner to review". No tag was pushed, no GitHub Release was created, the pypi environment was not approved, and nothing was uploaded anywhere.

Why this is a MINOR, not a patch

STABILITY.md § Versioning defines a minor as new features adhering to the contract. This cycle:

• adds a command, diff --application (feat: compare application agent wiring without prior setup #871, Application diff: derive the comparison scope from the change when --scope is omitted #875, Application diff reports 'compared, no changes' when an SDK agent built by 'return Agent(...)' gains a tool #876, Google ADK: resolve repository-local imported functions and module-qualified tool bindings #864, Google ADK: trace bounded FunctionTool factories into agent tool bindings #865, Application diff rows should name what a bound tool reaches: endpoint, action, credential, model-supplied arguments #872);
• adds host readers: hook script bytes (A hook command that runs an in-repo script: editing the script changes what runs, with no row #702), workflow agent launches (Read agent launches in CI workflows: known agent action inputs, literal agent CLI run steps and head-ref checkouts #823), and hook and MCP launch detail (Publish hook matcher, command summary and timeout, and MCP launch arguments, as bounded engine fields (#795 slice 2) #819);
• moves five schema versions:
  • runtime contract 40 → 41;
  • verifier 0.20 → 0.21;
  • capability diff 0.3 → 0.4;
  • host-grants inventory, baseline and drift 0.6 → 0.7;
  • application_comparison_schema_version 0.2, its first published version.

Channel: advisory, per docs/release-evidence-policy-decision.md § Amendment 5, as for 1.0.0 and 1.1.0. A tagged version's channel is frozen forever, so declaring 1.2.0 advisory means 1.2.0 can never carry a qualified claim.

What this changes

The release note (CHANGELOG.md). ## Unreleased is promoted to ## 1.2.0 - 2026-09-29. It is one line per change (33), with a framing paragraph, a ### Highlights block and the two pointers 1.1.0 carries. The subsection ### Application comparison without prior setup keeps its name, so the anchor README.md and docs/application-comparison.md link to still resolves. scripts/release_notes.py --tag v1.2.0 extracts it at 7,936 characters.

Claim discipline. The framing cites only measurements the entries themselves record:

• 296 rows from a derived scope, as from the root, on 48 application PRs (Application diff: derive the comparison scope from the change when --scope is omitted #875);
• rows, statuses and exit codes unchanged on 134 SDK/ADK PRs, with 9 of 96 sides naming an outbound call (Application diff rows should name what a bound tool reaches: endpoint, action, credential, model-supplied arguments #872);
• 412 s → 46 s on TencentCloud/CubeSandbox (shipgate diff repacks the repository's entire reachable history to read one base tree: 27 s per call on a 26 MB repository, 1 s on a toy one #686).

It states the limits plainly:

• diff --application reads OpenAI Agents SDK and Google ADK source wiring only;
• it compares what it observes, not what runs;
• what it cannot follow is a named gap;
• reach is static, and read is bounded.

There is no "catches issues" framing.

The record. The full reviewed prose for all 33 entries moves verbatim to docs/changelog/1.2.0.md (new), linked from docs/INDEX.md.

Version-status reconciliation. In the record:

• the application comparison entries now say the result ships as "0.2" (feat: compare application agent wiring without prior setup #871 introduced "0.1", which never shipped);
• "the unreleased runtime contract 41" and a link labelled Migration Note: Unreleased are restated.

docs/application-comparison.md and README.md said "while that entry is under Unreleased, use a source build". They now say "new in 1.2.0; until 1.2.0 is published, use a source build", which stays true after publication. docs/engineering/permission-review-guidance.md's "are unreleased" becomes "had not shipped before 1.2.0".

STABILITY.md.

• Title stamped · 1.2.0.
• 18 Migration Note: Unreleased headings become Migration Note: 1.2.0; every heading has an explicit <a id> above it, so no inbound link breaks.
• Six preamble paragraphs that opened "Unreleased" / "Also unreleased" now open "New in 1.2.0".
• Note bodies that said verifier 0.21, capability diff 0.4 and contract v41 "are unreleased" now say they "had not shipped before 1.2.0".
• "The unreleased tree" becomes "1.2.0", and one note that compared against an unshipped intermediate partial now compares against 1.1.0 and says the intermediate never shipped.

The general policy sentences about "unreleased main" and the historical migration-note-unreleased-* anchors are left alone.

Version sites.

File	Change
pyproject.toml	version = "1.2.0"
src/agents_shipgate/__init__.py	__version__ = "1.2.0"
.well-known/agents-shipgate.json	version only
docs/agent-contract-current.md	In-tree runtime only
.claude-plugin/marketplace.json, plugins/claude-code/.claude-plugin/plugin.json, plugins/agents-shipgate/.codex-plugin/plugin.json	plugin/manifest versions
llms.txt, docs/ai-search-summary.md	source-tree runtime, keeping the "unreleased" qualifier
README.md, docs/quickstart.md	"The source tree still reports version 1.2.0"
docs/design-partner-pilot-results.md	Source tree: 1.2.0, with a clause saying only the label moved and the cells were not re-measured
generated schemas, docs/determinism-boundary.*, llms-full.txt	regenerated (generate_schemas.py, build-llms-full.py)
docs/examples/capability-lock.v0.8.example.json	cli_version
tests/golden/codex_boundary_result/*.json (8)	tool.version
tests/test_v07_metadata_roundtrip.py	the in-tree-runtime literal

Channel declaration. .github/release-channels.json gains "1.2.0": "advisory", and tests/test_release_channel.py's committed-declaration assertion moves with it.

What this deliberately does not change

• No tag, no release, no publish.

• No published-release pins (runbook step 8). These all stay at 1.1.0 until after publication:

  • LATEST_PUBLISHED_VERSION / LATEST_PUBLISHED_CONTRACT_VERSION;
  • .well-known's release_status.latest_release and package.github_action;
  • docs/agent-contract-current.md's "Latest release";
  • @v1.1.0 / ==1.1.0 pins.

  Moving them before the tag exists reproduces init writes an adopter workflow pinned to a tag that does not exist (56 days and counting) #506.

• The two-phase labels stay. llms.txt, docs/ai-search-summary.md and the README/quickstart "not yet released" blocks keep their qualifiers until step 7.

• No deprecated ID is removed. The IDs deprecated in 1.1.0 (docs/checks.md, STABILITY.md) become eligible for hard removal in 1.2.0 under § Versioning, but nothing requires it, and they stay readable.

• No published run-of-record evidence and no GitHub issue text was edited.

Release checklist state

Runbook step	State
1. Promote the CHANGELOG heading, split the prose	done in this PR
2. Stamp STABILITY.md	done in this PR
3a. pyproject.toml at the release version	done in this PR
3b. .github/release-channels.json declares the channel	done in this PR
3c. Release Engine Smoke (unqualified)	not run: dispatch on the merged main commit
3d. Advisory Release Rehearsal	not run: after the smoke, on that same commit
4. Push the tag	owner decision
5. verify runs unattended	after the tag
6. Approve the pypi environment gate	owner only: the publication decision
7. Confirm the published release, run the fan-out checks	after publication
8. Move the published-release constants and pins	after publication

Tests

• Locally, the full suite (-n 8) passes except three environmental cases that fail the same way on main on this workstation (a global git ignore of .claude/settings.local.json, and a long worktree path truncating one hint).
• scripts/generate_schemas.py --check, scripts/verify_dependency_lock.py, ruff check . and ./shipgate doctor --config shipgate.yaml --json pass; doctor reports imported version 1.2.0 with no mismatches.
• scripts/release_cadence.py --json reports the advisory line current (7 days since v1.1.0).

Needs the owner's decision

1. The date on the heading is 2026-09-29. If the tag lands later, say so and I will re-date it before the smoke run. The release body is extracted from this commit.
2. Whether the framing is the one you want in the release body. The framing paragraphs and ### Highlights say what the entries measure and nothing more.
3. The deprecated IDs. Removing them is allowed from 1.2.0 on, but it is not done here. Keeping them costs nothing; removing them is a separate change.
4. Before tagging, re-read docs/release-runbook.md § Deployment prerequisites. The recorded platform observations are dated 2026-09-09/10. Separately, the Protect main ruleset still requires suite (1)–suite (3) only. Name what a bound tool reaches in diff --application rows (#872) #906 split the suite into four shards, so suite (4) should be added there, and to the runbook's table once it is.

🤖 Generated with Claude Code