handlers: audit + sharpen every agent_action string per uniform contract (U3)

[mastermanas805]

Summary

Establishes a single source of truth for every agent_action string the API returns (internal/handlers/agent_action.go) and enforces a four-requirement contract via TestAgentActionContract.

Today's strings worked but were inconsistent — some imperative, some passive; some named the tier, some said "their plan"; some included the full URL, some only /pricing. This PR audits all 25 strings across handlers/ + middleware/, sharpens every one to a uniform contract, and extracts each to a named constant or builder so reviewers audit prose in one place.

The U3 contract — every agent_action MUST:

1. Open with "Tell the user" — the LLM agent re-articulates verbatim to the human.
2. Name the specific reason rejected (tier, limit, policy, resource) — no "their plan does not allow X".
3. Name the exact next action (Upgrade, Claim, Provision a twin, Contact support) — never "try again later" without context.
4. Include a full https://instanode.dev/ URL — no relative /pricing.
5. (Soft) Under 280 chars so LLMs reproduce verbatim instead of paraphrasing.

The full contract block is committed at the top of internal/handlers/agent_action.go.

Audit counts

	Count
Strings audited (before)	25
Strings meeting all 4 contract requirements (before)	~11
Strings meeting all 4 contract requirements (after)	25

Example before/after

deployment_limit_reached (deploy.go):

• Before: "Tell the user they've hit the %s tier deployment cap (%d apps). Upgrade them to Pro for 10 medium deploys: https://instanode.dev/start?t=..." (broken placeholder URL)
• After: "Tell the user they've hit the %s tier deployment cap (%d apps). Upgrade to Pro for 10 deployments at https://instanode.dev/pricing — takes 30 seconds, no card for upgrade preview."

upgrade_required (helpers.go registry):

• Before: "Tell the user this feature requires a higher plan. Have them upgrade at https://instanode.dev/pricing." (vague — "higher plan" gives the LLM nothing concrete)
• After: "Tell the user this feature requires the Pro plan or higher. Upgrade at https://instanode.dev/pricing — takes 30 seconds."

missing_image_ref (stack.go promote):

• Before: "This stack predates the image-ref persistence migration. Redeploy the source stack first so its image is cached, then promote." (no "Tell the user" prefix, no URL)
• After: "Tell the user this stack predates the image-ref persistence migration, so promote has nothing to redeploy. Redeploy the source stack first at https://instanode.dev/app/stacks, then retry the promote."

Pushback / scope notes

• The middleware/env_policy.go string lives in a different package so it's kept inline (with a same-shape rewrite + an explicit "must stay in sync with agent_action.go" comment). No package-graph reshuffling to avoid blast radius.
• internal/plans/TestAll_ReturnsAllPlans fails on master and remains failing — pre-existing, unrelated. Filed as known-issue rather than fixed here.
• 503 / transient infra responses (provision_failed, list_failed, stream_failed, etc.) deliberately omit agent_action per codeToAgentAction curation comments — there's no action the user can take. The one exception (resource_binding_lookup_failed) keeps an agent_action because the actionable advice is concrete: "retry in ~10 seconds, check status page".

Coordination with P1, U1, U2

P1, U1, U2 add new agent_action strings — those tracks should land their strings as named constants/builders in internal/handlers/agent_action.go so the audit + the contract-test stay one-stop. The contract block at the top of that file is the canonical reference.

Test plan

• go build ./... + go vet ./... clean
• TestAgentActionContract (new) — every static const + every builder + every registry entry passes all 4 contract checks
• TestAgentActionContract_RegistryCoverage — all expected error codes registered
• TestRespondError_* — existing tests updated to match sharpened registry strings, all green
• internal/handlers/ full suite green (25.9s)
• internal/middleware/ full suite green (after one flaky-Redis retry — same flake exists on master)
• All non-e2e packages green except pre-existing internal/plans failure (unrelated)
• No deploy required for U3 — pure prose change with test gates

🤖 Generated with Claude Code