[tracking202]

Backlog: 5A30 (with the check WKB1)

What changed and why

The MCP catalogue (go-cli/cmd/mcp_catalog.go, served by p202 mcp and by the hosted relay that runs this CLI as js/wasm) offered text_ad_list only, so a demo account's text ads had to be typed into Setup › Text Ads through a browser.

• Three tools: text_ad_create, text_ad_update, text_ad_delete, declared as the campaign_* / landing_page_* writes are (and as 1.9.79: update and delete tools for traffic sources, their accounts and categories in the MCP connector #280's traffic source and category writes, which now sit beside them): kinds create / update / delete (so update and delete carry destructiveHint, delete idempotentHint and the dry_run note), an id positional on update and delete (delete also takes ids), no NotStaged/NotHosted. Create's note says a tracker shows an ad through its text_ad_id (tracker_update). text_ad_list keeps 1.9.79: the hosted MCP server's list tools take no all #279's hosted omission of all.
• Hosted connector: AiConnect::WRITE_AREAS gains text-ads, so an approval with propose access now mints text-ads:stage and write access text-ads:write (TestHostedMCPToolsStayInsideTheAIKeyAreas holds every hosted tool to those lists; it fails without this). The approval page (202-account/connect-ai.php) names text ads in all three access descriptions. A key minted before the install ran 1.9.79 keeps the scope it was given and answers 403 for the three tools until the person connects again; nothing widens a stored key (the person approved a list that did not include text ads). Documented in documentation/cli/10-go-cli.md, docs/cli-agent.md and the changelog.
• The CLI commands the schemas come from (p202 text-ad create/update/delete, go-cli/cmd/crud.go): every field was already a flag of the right JSON type (string, the API validates). What was missing was Setup › Text Ads' rule: a campaign's ad (--text-ad-type 0, the default) names its campaign and an advanced landing page's ad (--text-ad-type 1) its page. Now create refuses a missing link before any request (no config needed; validation, exit 1, the hint names p202 campaign list / p202 landing-page list and the other kind of ad); update holds the ad as it will be (the flags given over the stored ad, which it GETs only when a flag sets the type or a link; an answer missing the field is an error, never read as "none"). --text-ad-type and --landing-page-type are enums 0, 1 (as the pages offer); the field help names the 100-character limits, the type each link belongs to and where ids come from; the create example includes --aff-campaign-id. Mechanism: an optional Check/CheckFields/ExampleFields on crudEntity, used only by text-ad.
• The API (32cd43c, found by review): TextAdsController refuses an advanced landing page's ad (text_ad_type 1) that names a simple landing page, with a 422 naming landing_page_id, on a create and on an update that sets the type or the page; an ad stored that way before saves when re-sent as it is. OwnedReferencesIntegrationTest has the case; docs/openapi.yaml says so.
• WKB1: landing_page_create already makes an advanced landing page: landing_page_type was a tool argument and LandingPagesController::requiredLinkMayBeNone() takes aff_campaign_id: 0 for type 1. Measured live through p202 mcp call landing_page_create (201, landing_page_type: 1, aff_campaign_id: 0; a simple page with campaign 0 is still refused 422). Nothing was missing; the enum on --landing-page-type is the only change there.
• Catalogue budget: TestMCPCatalogStaysInBudget is 88,000 characters. 1.9.79 had raised it to 85,000 for 1.9.79: update and delete tools for traffic sources, their accounts and categories in the MCP connector #280's tools; with the text ad writes (~3,400, Setup › Text Ads' fields and nothing else) the list is 87,814.
• 1.9.79: a hosted call refuses answers with more JSON values than it decodes #281's value guard: the three tools are held to it like every other (the hosted-areas and value tests loop over the catalogue). The one shape of their own is text_ad_update's read of the stored ad before its PUT: TestHostedTextAdUpdateRefusedAtItsReadSendsNoWrite holds an answer past the bound there to the refusal, one GET and no PUT, with the hint that nothing that changes data was sent.
• Agent evals (tests/fixtures/agent-eval/cases/setup-text-ads.json): setup-text-ad-001 (write an ad for EVAL Campaign B and put it on its existing tracker; checks the ad's type, campaign and copy and that the tracker carries it) and its negative twin setup-text-ad-002 (an ad for nothing is asked about, not made). Reference-agent branches for both. 001 is listed in searchKnownMisses with why (two tasks; the Text Ads page's text-ad list ranks first, text-ad create fourth).
• Tests: crud_text_ad_test.go (create refusals, what create sends, update against a stored ad: which requests go out and in what order), TestMCPToolArgv cases for the three tools, mcpWriteToolInput for the hosted-areas test, the value-guard test above. Every plant was caught by a named test: removing the create check, the update check, the missing-field refusal, reading type 1 as 0, dropping text-ads from WRITE_AREAS, dropping the type enum (each planted with a mktemp backup checked with cmp, restored and re-run green).
• Docs: docs/cli.md, docs/cli-agent.md, documentation/cli/10-go-cli.md, docs/openapi.yaml; changelogs.txt and 202-config/changelogs.json under 1.9.79, after 1.9.79's own lines.

Merged with 1.9.79 (#270, #279, #280, #281, #282)

Conflicts resolved in mcp_catalog.go (#280's update/delete tools and setupParentDeleteNote beside the text ad tools), mcp_test.go (both sets of TestMCPToolArgv cases; the budget above), docs/cli-agent.md and documentation/cli/10-go-cli.md (#281's hosted read bounds, then the text ad reconnect note) and the changelogs (1.9.79's current lines kept, the text ad lines after them).

Live pass (local instance, MariaDB 10.11, PHP 8.3.6, before the merge)

Installed with install-instance.sh, seeded with seed.sh, CLI built from this branch, every call through p202 mcp call <tool> --input -:

• text_ad_create for EVAL Campaign B → 201 (text_ad_type 0, aff_campaign_id 2); without a campaign → validation envelope naming --aff-campaign-id; text_ad_type: "2" → refused naming the flag; a 101-character headline → 422 naming text_ad_headline.
• landing_page_create with landing_page_type 1, aff_campaign_id 0 → advanced page (WKB1); text_ad_create type 1 on it → 201.
• text_ad_update: headline only → 200 (that it reads nothing first is the unit test's); text_ad_type 1 on a campaign's ad with no page → refused; aff_campaign_id 0 → refused; type and link together → 200.
• tracker_update text_ad_id on EVAL Campaign B's tracker, read back with tracker_get (text_ad_id set).
• text_ad_list before and after; text_ad_delete with dry_run: true (preview, nothing removed per text_ad_list), then the delete, then text_ad_list (gone).
• --read-only refuses text_ad_create; --staged returns a chg_ proposal.
• Hosted path end to end: signed in, approved Read and change on connect-ai.php, exchanged the code at POST /api/v3/connect/token with its verifier → scope includes text-ads:write; with that key text_ad_create, text_ad_update, text_ad_delete (dry run, then delete) all worked. A key with 1.9.78's write scope (text-ads:read) → 403 requires 'text-ads:write'; a propose key → staged change.
• Full agent-eval suite with the reference agent on a fresh instance: 53 cases, 0 fail, 0 error, 53 needs_judge (no judge here; the deterministic halves pass).

The simple-page refusal (32cd43c) was measured live by its own commit; I did not repeat the live pass after merging 1.9.79.

Verification (after the merge)

.claude/skills/p202-verify/scripts/verify.sh --since origin/1.9.79 --changed on the merge of 1.9.79 (Go and PHP; PHP 8.3.6, full vendor/, schema tier against a scratch MariaDB 10.11 database; unit 6104 tests, 4 skipped):

Scope report
------------
  go         PASS  ran under CI's go1.25.14 via GOTOOLCHAIN (local go is 1.24)
  golangci   PASS  ran under CI's go1.25.14 via GOTOOLCHAIN (local go is 1.24)
  patterns   PASS
  phpcs      PASS
  phpstan    PASS
  schema     PASS
  syntax     PASS
  unit       PASS

Then, after merging 32cd43c (PHP and the spec only; no Go change), --since origin/1.9.79 with the PHP tiers and openapi:

Scope report
------------
  syntax     PASS
  phpstan    PASS
  phpcs      PASS
  unit       PASS
  schema     PASS
  patterns   PASS
  openapi    PASS

OwnedReferencesIntegrationTest against a scratch database (p202_ta_owned): OK, 6 tests, 240 assertions. CI status: not yet observed.

What could not run here

• CI's MySQL 8.0 (this sandbox has MariaDB 10.11); the PR's Agent Evals and Go CLI workflows will run the cases and Go suites there.
• No rubric judge: every eval case reports needs_judge here, as in CI without judge_cmd.
• The hosted relay itself (another repo): the install side of the hosted path was driven end to end, the relay's own tool offering was not.

For the reviewer to check

1. Where the text ad rule lives. The missing-link refusal is in the Go CLI (checkTextAdLink), so p202 mcp and the hosted relay get it; the API now refuses a type 1 ad on a simple page, but still stores an ad whose type's link is 0 (OwnedReferencesIntegrationTest pins a text ad with no campaign as accepted, and sync/import copy existing rows that may have none).
2. The update rule reads the stored ad (one GET) only when the update sets text_ad_type, aff_campaign_id or landing_page_id and the body alone does not decide; an edit of the copy is sent unread and never refused for what the ad already holds.
3. Budget raised to 88,000 rather than trimming other tools.
4. Existing hosted keys are not widened. An upgrade step could rewrite stored AI-connection scopes to the new lists, but that would grant text ad writes nobody approved; the person reconnects instead (documented).
5. --landing-page-type became an enum (0, 1) alongside --text-ad-type; the API still takes any TINYINT for both.

Found outside this item (one line each)

• The API (TextAdsController) still stores a text ad whose type's link is 0 and answers 201 (REST, the PHP CLI and sync are not held to the Go CLI's refusal).
• A 403 scope refusal for a hosted AI key hints "mint one: p202 user apikey create ... --scope write", which is the wrong remedy for a hosted connection (reconnect); the relay could also offer only the tools a key's scope covers (/capabilities lists principal.scopes).
• The PHP CLI's text-ad entity (cli/Application.php) requires only text_ad_name and none of the copy the API requires.

🤖 Generated with Claude Code

https://claude.ai/code/session_01VRoS2SViuxaucFaKWsFnLr

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-28T12:30:38.463845Z	6c40104	New commits

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

Code review: CLEAN

For ff2dfa4: 0 issue(s). The review is encrypted for the maintainers in the review-findings artifact of this run.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: ff2dfa4e71

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[github-actions]

Adversarial review: NON-BLOCKING

For 6c40104: 6 finding(s). The report is encrypted for the maintainers in the review-findings artifact of this run.

Rewritten on every push; the check fails on BLOCKING.

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.